thr_kern.c revision 117344
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 117344 2003-07-09 01:06:12Z 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#include "ksd.h" 58 59/*#define DEBUG_THREAD_KERN */ 60#ifdef DEBUG_THREAD_KERN 61#define DBG_MSG stdout_debug 62#else 63#define DBG_MSG(x...) 64#endif 65 66/* 67 * Define a high water mark for the maximum number of threads that 68 * will be cached. Once this level is reached, any extra threads 69 * will be free()'d. 70 * 71 * XXX - It doesn't make sense to worry about the maximum number of 72 * KSEs that we can cache because the system will limit us to 73 * something *much* less than the maximum number of threads 74 * that we can have. Disregarding KSEs in their own group, 75 * the maximum number of KSEs is the number of processors in 76 * the system. 77 */ 78#define MAX_CACHED_THREADS 100 79#define KSE_STACKSIZE 16384 80 81#define KSE_SET_MBOX(kse, thrd) \ 82 (kse)->k_mbx.km_curthread = &(thrd)->tmbx 83 84#define KSE_SET_EXITED(kse) (kse)->k_flags |= KF_EXITED 85 86/* 87 * Macros for manipulating the run queues. The priority queue 88 * routines use the thread's pqe link and also handle the setting 89 * and clearing of the thread's THR_FLAGS_IN_RUNQ flag. 90 */ 91#define KSE_RUNQ_INSERT_HEAD(kse, thrd) \ 92 _pq_insert_head(&(kse)->k_schedq->sq_runq, thrd) 93#define KSE_RUNQ_INSERT_TAIL(kse, thrd) \ 94 _pq_insert_tail(&(kse)->k_schedq->sq_runq, thrd) 95#define KSE_RUNQ_REMOVE(kse, thrd) \ 96 _pq_remove(&(kse)->k_schedq->sq_runq, thrd) 97#define KSE_RUNQ_FIRST(kse) _pq_first(&(kse)->k_schedq->sq_runq) 98 99#define KSE_RUNQ_THREADS(kse) ((kse)->k_schedq->sq_runq.pq_threads) 100 101/* 102 * We've got to keep track of everything that is allocated, not only 103 * to have a speedy free list, but also so they can be deallocated 104 * after a fork(). 105 */ 106static TAILQ_HEAD(, kse) active_kseq; 107static TAILQ_HEAD(, kse) free_kseq; 108static TAILQ_HEAD(, kse_group) free_kse_groupq; 109static TAILQ_HEAD(, kse_group) active_kse_groupq; 110static TAILQ_HEAD(, kse_group) gc_ksegq; 111static struct lock kse_lock; /* also used for kseg queue */ 112static int free_kse_count = 0; 113static int free_kseg_count = 0; 114static TAILQ_HEAD(, pthread) free_threadq; 115static struct lock thread_lock; 116static int free_thread_count = 0; 117static int inited = 0; 118static int active_threads = 1; 119static int active_kse_count = 0; 120static int active_kseg_count = 0; 121static u_int64_t next_uniqueid = 1; 122 123 124#ifdef DEBUG_THREAD_KERN 125static void dump_queues(struct kse *curkse); 126#endif 127static void kse_check_completed(struct kse *kse); 128static void kse_check_waitq(struct kse *kse); 129static void kse_fini(struct kse *curkse); 130static void kse_reinit(struct kse *kse); 131static void kse_sched_multi(struct kse *curkse); 132#ifdef NOT_YET 133static void kse_sched_single(struct kse *curkse); 134#endif 135static void kse_switchout_thread(struct kse *kse, struct pthread *thread); 136static void kse_wait(struct kse *kse, struct pthread *td_wait); 137static void kse_free_unlocked(struct kse *kse); 138static void kseg_free_unlocked(struct kse_group *kseg); 139static void kseg_init(struct kse_group *kseg); 140static void kseg_reinit(struct kse_group *kseg); 141static void kse_waitq_insert(struct pthread *thread); 142static void kse_wakeup_multi(struct kse *curkse); 143static void kse_wakeup_one(struct pthread *thread); 144static void thr_cleanup(struct kse *kse, struct pthread *curthread); 145static void thr_link(struct pthread *thread); 146static void thr_resume_wrapper(int sig, siginfo_t *, ucontext_t *); 147static void thr_resume_check(struct pthread *curthread, ucontext_t *ucp, 148 struct pthread_sigframe *psf); 149static int thr_timedout(struct pthread *thread, struct timespec *curtime); 150static void thr_unlink(struct pthread *thread); 151 152 153/* 154 * This is called after a fork(). 155 * No locks need to be taken here since we are guaranteed to be 156 * single threaded. 157 * 158 * XXX 159 * POSIX says for threaded process, fork() function is used 160 * only to run new programs, and the effects of calling functions 161 * that require certain resources between the call to fork() and 162 * the call to an exec function are undefined. 163 * 164 * Here it is not safe to reinitialize the library after fork(). 165 * Because memory management may be corrupted, further calling 166 * malloc()/free() may cause undefined behavior. 167 */ 168void 169_kse_single_thread(struct pthread *curthread) 170{ 171#ifdef NOTYET 172 struct kse *kse; 173 struct kse_group *kseg; 174 struct pthread *thread; 175 kse_critical_t crit; 176 int i; 177 178 179 /* 180 * Disable upcalls and clear the threaded flag. 181 * XXX - I don't think we need to disable upcalls after a fork(). 182 * but it doesn't hurt. 183 */ 184 crit = _kse_critical_enter(); 185 __isthreaded = 0; 186 active_threads = 1; 187 _thr_signal_deinit(); 188 189 /* 190 * Enter a loop to remove and free all threads other than 191 * the running thread from the active thread list: 192 */ 193 while ((thread = TAILQ_FIRST(&_thread_list)) != NULL) { 194 THR_GCLIST_REMOVE(thread); 195 /* 196 * Remove this thread from the list (the current 197 * thread will be removed but re-added by libpthread 198 * initialization. 199 */ 200 TAILQ_REMOVE(&_thread_list, thread, tle); 201 /* Make sure this isn't the running thread: */ 202 if (thread != curthread) { 203 _thr_stack_free(&thread->attr); 204 if (thread->specific != NULL) 205 free(thread->specific); 206 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 207 _lockuser_destroy(&thread->lockusers[i]); 208 } 209 _lock_destroy(&thread->lock); 210 free(thread); 211 } 212 } 213 214 TAILQ_INIT(&curthread->mutexq); /* initialize mutex queue */ 215 curthread->joiner = NULL; /* no joining threads yet */ 216 curthread->refcount = 0; 217 SIGEMPTYSET(curthread->sigpend); /* clear pending signals */ 218 if (curthread->specific != NULL) { 219 free(curthread->specific); 220 curthread->specific = NULL; 221 curthread->specific_data_count = 0; 222 } 223 224 /* Free the free KSEs: */ 225 while ((kse = TAILQ_FIRST(&free_kseq)) != NULL) { 226 TAILQ_REMOVE(&free_kseq, kse, k_qe); 227 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 228 _lockuser_destroy(&kse->k_lockusers[i]); 229 } 230 _lock_destroy(&kse->k_lock); 231 _ksd_destroy(&kse->k_ksd); 232 if (kse->k_stack.ss_sp != NULL) 233 free(kse->k_stack.ss_sp); 234 free(kse); 235 } 236 free_kse_count = 0; 237 238 /* Free the active KSEs: */ 239 while ((kse = TAILQ_FIRST(&active_kseq)) != NULL) { 240 TAILQ_REMOVE(&active_kseq, kse, k_qe); 241 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 242 _lockuser_destroy(&kse->k_lockusers[i]); 243 } 244 _lock_destroy(&kse->k_lock); 245 if (kse->k_stack.ss_sp != NULL) 246 free(kse->k_stack.ss_sp); 247 free(kse); 248 } 249 active_kse_count = 0; 250 251 /* Free the free KSEGs: */ 252 while ((kseg = TAILQ_FIRST(&free_kse_groupq)) != NULL) { 253 TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe); 254 _lock_destroy(&kseg->kg_lock); 255 _pq_free(&kseg->kg_schedq.sq_runq); 256 free(kseg); 257 } 258 free_kseg_count = 0; 259 260 /* Free the active KSEGs: */ 261 while ((kseg = TAILQ_FIRST(&active_kse_groupq)) != NULL) { 262 TAILQ_REMOVE(&active_kse_groupq, kseg, kg_qe); 263 _lock_destroy(&kseg->kg_lock); 264 _pq_free(&kseg->kg_schedq.sq_runq); 265 free(kseg); 266 } 267 active_kseg_count = 0; 268 269 /* Free the free threads. */ 270 while ((thread = TAILQ_FIRST(&free_threadq)) != NULL) { 271 TAILQ_REMOVE(&free_threadq, thread, tle); 272 if (thread->specific != NULL) 273 free(thread->specific); 274 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 275 _lockuser_destroy(&thread->lockusers[i]); 276 } 277 _lock_destroy(&thread->lock); 278 free(thread); 279 } 280 free_thread_count = 0; 281 282 /* Free the to-be-gc'd threads. */ 283 while ((thread = TAILQ_FIRST(&_thread_gc_list)) != NULL) { 284 TAILQ_REMOVE(&_thread_gc_list, thread, gcle); 285 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 286 _lockuser_destroy(&thread->lockusers[i]); 287 } 288 _lock_destroy(&thread->lock); 289 free(thread); 290 } 291 TAILQ_INIT(&gc_ksegq); 292 _gc_count = 0; 293 294 if (inited != 0) { 295 /* 296 * Destroy these locks; they'll be recreated to assure they 297 * are in the unlocked state. 298 */ 299 _lock_destroy(&kse_lock); 300 _lock_destroy(&thread_lock); 301 _lock_destroy(&_thread_list_lock); 302 inited = 0; 303 } 304 305 /* 306 * After a fork(), the leftover thread goes back to being 307 * scope process. 308 */ 309 curthread->attr.flags &= ~PTHREAD_SCOPE_SYSTEM; 310 curthread->attr.flags |= PTHREAD_SCOPE_PROCESS; 311 312 /* 313 * After a fork, we are still operating on the thread's original 314 * stack. Don't clear the THR_FLAGS_USER from the thread's 315 * attribute flags. 316 */ 317 318 /* Initialize the threads library. */ 319 curthread->kse = NULL; 320 curthread->kseg = NULL; 321 _kse_initial = NULL; 322 _libpthread_init(curthread); 323#else 324 _ksd_readandclear_tmbx(); 325 __isthreaded = 0; 326 active_threads = 0; 327 _thr_signal_deinit(); 328#endif 329} 330 331/* 332 * This is used to initialize housekeeping and to initialize the 333 * KSD for the KSE. 334 */ 335void 336_kse_init(void) 337{ 338 if (inited == 0) { 339 TAILQ_INIT(&active_kseq); 340 TAILQ_INIT(&active_kse_groupq); 341 TAILQ_INIT(&free_kseq); 342 TAILQ_INIT(&free_kse_groupq); 343 TAILQ_INIT(&free_threadq); 344 TAILQ_INIT(&gc_ksegq); 345 if (_lock_init(&kse_lock, LCK_ADAPTIVE, 346 _kse_lock_wait, _kse_lock_wakeup) != 0) 347 PANIC("Unable to initialize free KSE queue lock"); 348 if (_lock_init(&thread_lock, LCK_ADAPTIVE, 349 _kse_lock_wait, _kse_lock_wakeup) != 0) 350 PANIC("Unable to initialize free thread queue lock"); 351 if (_lock_init(&_thread_list_lock, LCK_ADAPTIVE, 352 _kse_lock_wait, _kse_lock_wakeup) != 0) 353 PANIC("Unable to initialize thread list lock"); 354 active_kse_count = 0; 355 active_kseg_count = 0; 356 _gc_count = 0; 357 inited = 1; 358 } 359} 360 361int 362_kse_isthreaded(void) 363{ 364 return (__isthreaded != 0); 365} 366 367/* 368 * This is called when the first thread (other than the initial 369 * thread) is created. 370 */ 371int 372_kse_setthreaded(int threaded) 373{ 374 if ((threaded != 0) && (__isthreaded == 0)) { 375 /* 376 * Locking functions in libc are required when there are 377 * threads other than the initial thread. 378 */ 379 __isthreaded = 1; 380 381 /* 382 * Tell the kernel to create a KSE for the initial thread 383 * and enable upcalls in it. 384 */ 385 _thr_signal_init(); 386 _kse_initial->k_flags |= KF_STARTED; 387 if (kse_create(&_kse_initial->k_mbx, 0) != 0) { 388 _kse_initial->k_flags &= ~KF_STARTED; 389 __isthreaded = 0; 390 /* may abort() */ 391 PANIC("kse_create() failed\n"); 392 return (-1); 393 } 394 KSE_SET_MBOX(_kse_initial, _thr_initial); 395 _thr_start_sig_daemon(); 396 _thr_setmaxconcurrency(); 397 } 398 return (0); 399} 400 401/* 402 * Lock wait and wakeup handlers for KSE locks. These are only used by 403 * KSEs, and should never be used by threads. KSE locks include the 404 * KSE group lock (used for locking the scheduling queue) and the 405 * kse_lock defined above. 406 * 407 * When a KSE lock attempt blocks, the entire KSE blocks allowing another 408 * KSE to run. For the most part, it doesn't make much sense to try and 409 * schedule another thread because you need to lock the scheduling queue 410 * in order to do that. And since the KSE lock is used to lock the scheduling 411 * queue, you would just end up blocking again. 412 */ 413void 414_kse_lock_wait(struct lock *lock, struct lockuser *lu) 415{ 416 struct kse *curkse = (struct kse *)_LCK_GET_PRIVATE(lu); 417 struct timespec ts; 418 int saved_flags; 419 420 if (curkse->k_mbx.km_curthread != NULL) 421 PANIC("kse_lock_wait does not disable upcall.\n"); 422 /* 423 * Enter a loop to wait until we get the lock. 424 */ 425 ts.tv_sec = 0; 426 ts.tv_nsec = 1000000; /* 1 sec */ 427 while (!_LCK_GRANTED(lu)) { 428 /* 429 * Yield the kse and wait to be notified when the lock 430 * is granted. 431 */ 432 saved_flags = curkse->k_mbx.km_flags; 433 curkse->k_mbx.km_flags |= KMF_NOUPCALL | KMF_NOCOMPLETED; 434 kse_release(&ts); 435 curkse->k_mbx.km_flags = saved_flags; 436 } 437} 438 439void 440_kse_lock_wakeup(struct lock *lock, struct lockuser *lu) 441{ 442 struct kse *curkse; 443 struct kse *kse; 444 struct kse_mailbox *mbx; 445 446 curkse = _get_curkse(); 447 kse = (struct kse *)_LCK_GET_PRIVATE(lu); 448 449 if (kse == curkse) 450 PANIC("KSE trying to wake itself up in lock"); 451 else { 452 mbx = &kse->k_mbx; 453 _lock_grant(lock, lu); 454 /* 455 * Notify the owning kse that it has the lock. 456 * It is safe to pass invalid address to kse_wakeup 457 * even if the mailbox is not in kernel at all, 458 * and waking up a wrong kse is also harmless. 459 */ 460 kse_wakeup(mbx); 461 } 462} 463 464/* 465 * Thread wait and wakeup handlers for thread locks. These are only used 466 * by threads, never by KSEs. Thread locks include the per-thread lock 467 * (defined in its structure), and condition variable and mutex locks. 468 */ 469void 470_thr_lock_wait(struct lock *lock, struct lockuser *lu) 471{ 472 struct pthread *curthread = (struct pthread *)lu->lu_private; 473 474 do { 475 THR_SCHED_LOCK(curthread, curthread); 476 THR_SET_STATE(curthread, PS_LOCKWAIT); 477 THR_SCHED_UNLOCK(curthread, curthread); 478 _thr_sched_switch(curthread); 479 } while (!_LCK_GRANTED(lu)); 480} 481 482void 483_thr_lock_wakeup(struct lock *lock, struct lockuser *lu) 484{ 485 struct pthread *thread; 486 struct pthread *curthread; 487 488 curthread = _get_curthread(); 489 thread = (struct pthread *)_LCK_GET_PRIVATE(lu); 490 491 THR_SCHED_LOCK(curthread, thread); 492 _lock_grant(lock, lu); 493 _thr_setrunnable_unlocked(thread); 494 THR_SCHED_UNLOCK(curthread, thread); 495} 496 497kse_critical_t 498_kse_critical_enter(void) 499{ 500 kse_critical_t crit; 501 502 crit = _ksd_readandclear_tmbx(); 503 return (crit); 504} 505 506void 507_kse_critical_leave(kse_critical_t crit) 508{ 509 struct pthread *curthread; 510 511 _ksd_set_tmbx(crit); 512 if ((crit != NULL) && ((curthread = _get_curthread()) != NULL)) 513 THR_YIELD_CHECK(curthread); 514} 515 516int 517_kse_in_critical(void) 518{ 519 return (_ksd_get_tmbx() == NULL); 520} 521 522void 523_thr_critical_enter(struct pthread *thread) 524{ 525 thread->critical_count++; 526} 527 528void 529_thr_critical_leave(struct pthread *thread) 530{ 531 thread->critical_count--; 532 THR_YIELD_CHECK(thread); 533} 534 535void 536_thr_sched_switch(struct pthread *curthread) 537{ 538 struct kse *curkse; 539 540 (void)_kse_critical_enter(); 541 curkse = _get_curkse(); 542 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 543 _thr_sched_switch_unlocked(curthread); 544} 545 546/* 547 * XXX - We may need to take the scheduling lock before calling 548 * this, or perhaps take the lock within here before 549 * doing anything else. 550 */ 551void 552_thr_sched_switch_unlocked(struct pthread *curthread) 553{ 554 struct pthread *td; 555 struct pthread_sigframe psf; 556 struct kse *curkse; 557 int ret; 558 volatile int uts_once; 559 volatile int resume_once = 0; 560 ucontext_t uc; 561 562 /* We're in the scheduler, 5 by 5: */ 563 curkse = _get_curkse(); 564 565 curthread->need_switchout = 1; /* The thread yielded on its own. */ 566 curthread->critical_yield = 0; /* No need to yield anymore. */ 567 curthread->slice_usec = -1; /* Restart the time slice. */ 568 569 /* Thread can unlock the scheduler lock. */ 570 curthread->lock_switch = 1; 571 572 /* 573 * The signal frame is allocated off the stack because 574 * a thread can be interrupted by other signals while 575 * it is running down pending signals. 576 */ 577 psf.psf_valid = 0; 578 curthread->curframe = &psf; 579 580 /* 581 * Enter the scheduler if any one of the following is true: 582 * 583 * o The current thread is dead; it's stack needs to be 584 * cleaned up and it can't be done while operating on 585 * it. 586 * o The current thread has signals pending, should 587 * let scheduler install signal trampoline for us. 588 * o There are no runnable threads. 589 * o The next thread to run won't unlock the scheduler 590 * lock. A side note: the current thread may be run 591 * instead of the next thread in the run queue, but 592 * we don't bother checking for that. 593 */ 594 if ((curthread->state == PS_DEAD) || 595 (((td = KSE_RUNQ_FIRST(curkse)) == NULL) && 596 (curthread->state != PS_RUNNING)) || 597 ((td != NULL) && (td->lock_switch == 0))) { 598 curkse->k_switch = 1; 599 _thread_enter_uts(&curthread->tmbx, &curkse->k_mbx); 600 } 601 else { 602 uts_once = 0; 603 THR_GETCONTEXT(&curthread->tmbx.tm_context); 604 if (uts_once == 0) { 605 uts_once = 1; 606 607 /* Switchout the current thread. */ 608 kse_switchout_thread(curkse, curthread); 609 610 /* Choose another thread to run. */ 611 td = KSE_RUNQ_FIRST(curkse); 612 KSE_RUNQ_REMOVE(curkse, td); 613 curkse->k_curthread = td; 614 615 /* 616 * Make sure the current thread's kse points to 617 * this kse. 618 */ 619 td->kse = curkse; 620 621 /* 622 * Reset accounting. 623 */ 624 td->tmbx.tm_uticks = 0; 625 td->tmbx.tm_sticks = 0; 626 627 /* 628 * Reset the time slice if this thread is running 629 * for the first time or running again after using 630 * its full time slice allocation. 631 */ 632 if (td->slice_usec == -1) 633 td->slice_usec = 0; 634 635 /* Mark the thread active. */ 636 td->active = 1; 637 638 /* Remove the frame reference. */ 639 td->curframe = NULL; 640 641 /* 642 * Continue the thread at its current frame: 643 */ 644 ret = _thread_switch(&td->tmbx, NULL); 645 /* This point should not be reached. */ 646 if (ret != 0) 647 PANIC("Bad return from _thread_switch"); 648 PANIC("Thread has returned from _thread_switch"); 649 } 650 } 651 652 if (psf.psf_valid) { 653 /* 654 * It is ugly we must increase critical count, because we 655 * have a frame saved, we must backout state in psf 656 * before we can process signals. 657 */ 658 curthread->critical_count++; 659 } 660 661 if (curthread->lock_switch != 0) { 662 /* 663 * Unlock the scheduling queue and leave the 664 * critical region. 665 */ 666 /* Don't trust this after a switch! */ 667 curkse = _get_curkse(); 668 669 curthread->lock_switch = 0; 670 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 671 _kse_critical_leave(&curthread->tmbx); 672 } 673 /* 674 * This thread is being resumed; check for cancellations. 675 */ 676 if ((psf.psf_valid || 677 (curthread->check_pending && !THR_IN_CRITICAL(curthread)))) { 678 resume_once = 0; 679 THR_GETCONTEXT(&uc); 680 if (resume_once == 0) { 681 resume_once = 1; 682 curthread->check_pending = 0; 683 thr_resume_check(curthread, &uc, &psf); 684 } 685 } 686 THR_ACTIVATE_LAST_LOCK(curthread); 687} 688 689/* 690 * This is the scheduler for a KSE which runs a scope system thread. 691 * The multi-thread KSE scheduler should also work for a single threaded 692 * KSE, but we use a separate scheduler so that it can be fine-tuned 693 * to be more efficient (and perhaps not need a separate stack for 694 * the KSE, allowing it to use the thread's stack). 695 * 696 * XXX - This probably needs some work. 697 */ 698#ifdef NOT_YET 699static void 700kse_sched_single(struct kse *curkse) 701{ 702 struct pthread *curthread = curkse->k_curthread; 703 struct pthread *td_wait; 704 struct timespec ts; 705 int level; 706 707 if (curthread->active == 0) { 708 if (curthread->state != PS_RUNNING) { 709 /* Check to see if the thread has timed out. */ 710 KSE_GET_TOD(curkse, &ts); 711 if (thr_timedout(curthread, &ts) != 0) { 712 curthread->timeout = 1; 713 curthread->state = PS_RUNNING; 714 } 715 } 716 } 717 718 /* This thread no longer needs to yield the CPU: */ 719 curthread->critical_yield = 0; 720 curthread->need_switchout = 0; 721 722 /* 723 * Lock the scheduling queue. 724 * 725 * There is no scheduling queue for single threaded KSEs, 726 * but we need a lock for protection regardless. 727 */ 728 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 729 730 /* 731 * This has to do the job of kse_switchout_thread(), only 732 * for a single threaded KSE/KSEG. 733 */ 734 735 switch (curthread->state) { 736 case PS_DEAD: 737 /* Unlock the scheduling queue and exit the KSE and thread. */ 738 thr_cleaup(curkse, curthread); 739 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 740 break; 741 742 case PS_COND_WAIT: 743 case PS_SLEEP_WAIT: 744 /* Only insert threads that can timeout: */ 745 if (curthread->wakeup_time.tv_sec != -1) { 746 /* Insert into the waiting queue: */ 747 KSE_WAITQ_INSERT(curkse, curthread); 748 } 749 break; 750 751 case PS_LOCKWAIT: 752 level = curthread->locklevel - 1; 753 if (!_LCK_GRANTED(&curthread->lockusers[level])) 754 KSE_WAITQ_INSERT(curkse, curthread); 755 else 756 THR_SET_STATE(curthread, PS_RUNNING); 757 break; 758 759 case PS_JOIN: 760 case PS_MUTEX_WAIT: 761 case PS_RUNNING: 762 case PS_SIGSUSPEND: 763 case PS_SIGWAIT: 764 case PS_SUSPENDED: 765 case PS_DEADLOCK: 766 default: 767 /* 768 * These states don't timeout and don't need 769 * to be in the waiting queue. 770 */ 771 break; 772 } 773 while (curthread->state != PS_RUNNING) { 774 curthread->active = 0; 775 td_wait = KSE_WAITQ_FIRST(curkse); 776 777 kse_wait(curkse, td_wait); 778 779 if (td_wait != NULL) { 780 KSE_GET_TOD(curkse, &ts); 781 if (thr_timedout(curthread, &ts)) { 782 /* Indicate the thread timedout: */ 783 td_wait->timeout = 1; 784 785 /* Make the thread runnable. */ 786 THR_SET_STATE(td_wait, PS_RUNNING); 787 KSE_WAITQ_REMOVE(curkse, td_wait); 788 } 789 } 790 } 791 792 /* Remove the frame reference. */ 793 curthread->curframe = NULL; 794 795 /* Unlock the scheduling queue. */ 796 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 797 798 /* 799 * Continue the thread at its current frame: 800 */ 801 DBG_MSG("Continuing bound thread %p\n", curthread); 802 _thread_switch(&curthread->tmbx, &curkse->k_mbx.km_curthread); 803 PANIC("Thread has returned from _thread_switch"); 804} 805#endif 806 807#ifdef DEBUG_THREAD_KERN 808static void 809dump_queues(struct kse *curkse) 810{ 811 struct pthread *thread; 812 813 DBG_MSG("Threads in waiting queue:\n"); 814 TAILQ_FOREACH(thread, &curkse->k_kseg->kg_schedq.sq_waitq, pqe) { 815 DBG_MSG(" thread %p, state %d, blocked %d\n", 816 thread, thread->state, thread->blocked); 817 } 818} 819#endif 820 821/* 822 * This is the scheduler for a KSE which runs multiple threads. 823 */ 824static void 825kse_sched_multi(struct kse *curkse) 826{ 827 struct pthread *curthread, *td_wait; 828 struct pthread_sigframe *curframe; 829 int ret; 830 831 THR_ASSERT(curkse->k_mbx.km_curthread == NULL, 832 "Mailbox not null in kse_sched_multi"); 833 834 /* Check for first time initialization: */ 835 if ((curkse->k_flags & KF_INITIALIZED) == 0) { 836 /* Setup this KSEs specific data. */ 837 _ksd_setprivate(&curkse->k_ksd); 838 _set_curkse(curkse); 839 840 /* Set this before grabbing the context. */ 841 curkse->k_flags |= KF_INITIALIZED; 842 } 843 844 /* This may have returned from a kse_release(). */ 845 if (KSE_WAITING(curkse)) { 846 DBG_MSG("Entered upcall when KSE is waiting."); 847 KSE_CLEAR_WAIT(curkse); 848 } 849 850 /*If this is an upcall; take the scheduler lock. */ 851 if (curkse->k_switch == 0) 852 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 853 curkse->k_switch = 0; 854 855 curthread = curkse->k_curthread; 856 857 if (KSE_IS_IDLE(curkse)) { 858 KSE_CLEAR_IDLE(curkse); 859 curkse->k_kseg->kg_idle_kses--; 860 } 861 /* 862 * If the current thread was completed in another KSE, then 863 * it will be in the run queue. Don't mark it as being blocked. 864 */ 865 if ((curthread != NULL) && 866 ((curthread->flags & THR_FLAGS_IN_RUNQ) == 0) && 867 (curthread->need_switchout == 0)) { 868 /* 869 * Assume the current thread is blocked; when the 870 * completed threads are checked and if the current 871 * thread is among the completed, the blocked flag 872 * will be cleared. 873 */ 874 curthread->blocked = 1; 875 } 876 877 /* Check for any unblocked threads in the kernel. */ 878 kse_check_completed(curkse); 879 880 /* 881 * Check for threads that have timed-out. 882 */ 883 kse_check_waitq(curkse); 884 885 /* 886 * Switchout the current thread, if necessary, as the last step 887 * so that it is inserted into the run queue (if it's runnable) 888 * _after_ any other threads that were added to it above. 889 */ 890 if (curthread == NULL) 891 ; /* Nothing to do here. */ 892 else if ((curthread->need_switchout == 0) && 893 (curthread->blocked == 0) && (THR_IN_CRITICAL(curthread))) { 894 /* 895 * Resume the thread and tell it to yield when 896 * it leaves the critical region. 897 */ 898 curthread->critical_yield = 1; 899 curthread->active = 1; 900 if ((curthread->flags & THR_FLAGS_IN_RUNQ) != 0) 901 KSE_RUNQ_REMOVE(curkse, curthread); 902 curkse->k_curthread = curthread; 903 curthread->kse = curkse; 904 DBG_MSG("Continuing thread %p in critical region\n", 905 curthread); 906 kse_wakeup_multi(curkse); 907 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 908 ret = _thread_switch(&curthread->tmbx, 909 &curkse->k_mbx.km_curthread); 910 if (ret != 0) 911 PANIC("Can't resume thread in critical region\n"); 912 } 913 else if ((curthread->flags & THR_FLAGS_IN_RUNQ) == 0) 914 kse_switchout_thread(curkse, curthread); 915 curkse->k_curthread = NULL; 916 917 kse_wakeup_multi(curkse); 918 919#ifdef DEBUG_THREAD_KERN 920 dump_queues(curkse); 921#endif 922 923 /* Check if there are no threads ready to run: */ 924 while (((curthread = KSE_RUNQ_FIRST(curkse)) == NULL) && 925 (curkse->k_kseg->kg_threadcount != 0)) { 926 /* 927 * Wait for a thread to become active or until there are 928 * no more threads. 929 */ 930 td_wait = KSE_WAITQ_FIRST(curkse); 931 kse_wait(curkse, td_wait); 932 kse_check_completed(curkse); 933 kse_check_waitq(curkse); 934 } 935 936 /* Check for no more threads: */ 937 if (curkse->k_kseg->kg_threadcount == 0) { 938 /* 939 * Normally this shouldn't return, but it will if there 940 * are other KSEs running that create new threads that 941 * are assigned to this KSE[G]. For instance, if a scope 942 * system thread were to create a scope process thread 943 * and this kse[g] is the initial kse[g], then that newly 944 * created thread would be assigned to us (the initial 945 * kse[g]). 946 */ 947 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 948 kse_fini(curkse); 949 /* never returns */ 950 } 951 952 THR_ASSERT(curthread != NULL, 953 "Return from kse_wait/fini without thread."); 954 THR_ASSERT(curthread->state != PS_DEAD, 955 "Trying to resume dead thread!"); 956 KSE_RUNQ_REMOVE(curkse, curthread); 957 958 /* 959 * Make the selected thread the current thread. 960 */ 961 curkse->k_curthread = curthread; 962 963 /* 964 * Make sure the current thread's kse points to this kse. 965 */ 966 curthread->kse = curkse; 967 968 /* 969 * Reset accounting. 970 */ 971 curthread->tmbx.tm_uticks = 0; 972 curthread->tmbx.tm_sticks = 0; 973 974 /* 975 * Reset the time slice if this thread is running for the first 976 * time or running again after using its full time slice allocation. 977 */ 978 if (curthread->slice_usec == -1) 979 curthread->slice_usec = 0; 980 981 /* Mark the thread active. */ 982 curthread->active = 1; 983 984 /* Remove the frame reference. */ 985 curframe = curthread->curframe; 986 curthread->curframe = NULL; 987 988 kse_wakeup_multi(curkse); 989 990 /* 991 * The thread's current signal frame will only be NULL if it 992 * is being resumed after being blocked in the kernel. In 993 * this case, and if the thread needs to run down pending 994 * signals or needs a cancellation check, we need to add a 995 * signal frame to the thread's context. 996 */ 997#ifdef NOT_YET 998 if ((((curframe == NULL) && (curthread->check_pending != 0)) || 999 (((curthread->cancelflags & THR_AT_CANCEL_POINT) == 0) && 1000 ((curthread->cancelflags & PTHREAD_CANCEL_ASYNCHRONOUS) != 0))) && 1001 !THR_IN_CRITICAL(curthread)) 1002 signalcontext(&curthread->tmbx.tm_context, 0, 1003 (__sighandler_t *)thr_resume_wrapper); 1004#else 1005 if ((curframe == NULL) && (curthread->check_pending != 0) && 1006 !THR_IN_CRITICAL(curthread)) { 1007 curthread->check_pending = 0; 1008 signalcontext(&curthread->tmbx.tm_context, 0, 1009 (__sighandler_t *)thr_resume_wrapper); 1010 } 1011#endif 1012 /* 1013 * Continue the thread at its current frame: 1014 */ 1015 if (curthread->lock_switch != 0) { 1016 /* 1017 * This thread came from a scheduler switch; it will 1018 * unlock the scheduler lock and set the mailbox. 1019 */ 1020 ret = _thread_switch(&curthread->tmbx, NULL); 1021 } else { 1022 /* This thread won't unlock the scheduler lock. */ 1023 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1024 ret = _thread_switch(&curthread->tmbx, 1025 &curkse->k_mbx.km_curthread); 1026 } 1027 if (ret != 0) 1028 PANIC("Thread has returned from _thread_switch"); 1029 1030 /* This point should not be reached. */ 1031 PANIC("Thread has returned from _thread_switch"); 1032} 1033 1034static void 1035thr_resume_wrapper(int sig, siginfo_t *siginfo, ucontext_t *ucp) 1036{ 1037 struct pthread *curthread = _get_curthread(); 1038 struct kse *curkse; 1039 int ret, err_save = curthread->error; 1040 1041 DBG_MSG(">>> sig wrapper\n"); 1042 if (curthread->lock_switch) 1043 PANIC("thr_resume_wrapper, lock_switch != 0\n"); 1044 thr_resume_check(curthread, ucp, NULL); 1045 _kse_critical_enter(); 1046 curkse = _get_curkse(); 1047 curthread->tmbx.tm_context = *ucp; 1048 curthread->error = err_save; 1049 ret = _thread_switch(&curthread->tmbx, &curkse->k_mbx.km_curthread); 1050 if (ret != 0) 1051 PANIC("thr_resume_wrapper: thread has returned " 1052 "from _thread_switch"); 1053 /* THR_SETCONTEXT(ucp); */ /* not work, why ? */ 1054} 1055 1056static void 1057thr_resume_check(struct pthread *curthread, ucontext_t *ucp, 1058 struct pthread_sigframe *psf) 1059{ 1060 _thr_sig_rundown(curthread, ucp, psf); 1061 1062#ifdef NOT_YET 1063 if (((curthread->cancelflags & THR_AT_CANCEL_POINT) == 0) && 1064 ((curthread->cancelflags & PTHREAD_CANCEL_ASYNCHRONOUS) != 0)) 1065 pthread_testcancel(); 1066#endif 1067} 1068 1069/* 1070 * Clean up a thread. This must be called with the thread's KSE 1071 * scheduling lock held. The thread must be a thread from the 1072 * KSE's group. 1073 */ 1074static void 1075thr_cleanup(struct kse *curkse, struct pthread *thread) 1076{ 1077 struct pthread *joiner; 1078 int sys_scope; 1079 1080 if ((joiner = thread->joiner) != NULL) { 1081 /* Joinee scheduler lock held; joiner won't leave. */ 1082 if (joiner->kseg == curkse->k_kseg) { 1083 if (joiner->join_status.thread == thread) { 1084 joiner->join_status.thread = NULL; 1085 joiner->join_status.ret = thread->ret; 1086 _thr_setrunnable_unlocked(joiner); 1087 } 1088 } else { 1089 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1090 /* The joiner may have removed itself and exited. */ 1091 if (_thr_ref_add(thread, joiner, 0) == 0) { 1092 KSE_SCHED_LOCK(curkse, joiner->kseg); 1093 if (joiner->join_status.thread == thread) { 1094 joiner->join_status.thread = NULL; 1095 joiner->join_status.ret = thread->ret; 1096 _thr_setrunnable_unlocked(joiner); 1097 } 1098 KSE_SCHED_UNLOCK(curkse, joiner->kseg); 1099 _thr_ref_delete(thread, joiner); 1100 } 1101 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 1102 } 1103 thread->attr.flags |= PTHREAD_DETACHED; 1104 } 1105 1106 if (!(sys_scope = (thread->attr.flags & PTHREAD_SCOPE_SYSTEM))) { 1107 /* 1108 * Remove the thread from the KSEG's list of threads. 1109 */ 1110 KSEG_THRQ_REMOVE(thread->kseg, thread); 1111 /* 1112 * Migrate the thread to the main KSE so that this 1113 * KSE and KSEG can be cleaned when their last thread 1114 * exits. 1115 */ 1116 thread->kseg = _kse_initial->k_kseg; 1117 thread->kse = _kse_initial; 1118 } 1119 thread->flags |= THR_FLAGS_GC_SAFE; 1120 1121 /* 1122 * We can't hold the thread list lock while holding the 1123 * scheduler lock. 1124 */ 1125 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1126 DBG_MSG("Adding thread %p to GC list\n", thread); 1127 KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock); 1128 THR_GCLIST_ADD(thread); 1129 /* Use thread_list_lock */ 1130 active_threads--; 1131 if (active_threads == 1) { 1132 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 1133 exit(0); 1134 } 1135 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 1136 if (sys_scope) { 1137 /* 1138 * System scope thread is single thread group, 1139 * when thread is exited, its kse and ksegrp should 1140 * be recycled as well. 1141 */ 1142 kse_exit(); 1143 PANIC("kse_exit() failed for system scope thread"); 1144 } 1145 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 1146} 1147 1148void 1149_thr_gc(struct pthread *curthread) 1150{ 1151 struct pthread *td, *td_next; 1152 kse_critical_t crit; 1153 TAILQ_HEAD(, pthread) worklist; 1154 1155 TAILQ_INIT(&worklist); 1156 crit = _kse_critical_enter(); 1157 KSE_LOCK_ACQUIRE(curthread->kse, &_thread_list_lock); 1158 1159 /* Check the threads waiting for GC. */ 1160 for (td = TAILQ_FIRST(&_thread_gc_list); td != NULL; td = td_next) { 1161 td_next = TAILQ_NEXT(td, gcle); 1162 if ((td->flags & THR_FLAGS_GC_SAFE) == 0) 1163 continue; 1164 else if (((td->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) && 1165 ((td->kse->k_mbx.km_flags & KMF_DONE) == 0)) { 1166 /* 1167 * The thread and KSE are operating on the same 1168 * stack. Wait for the KSE to exit before freeing 1169 * the thread's stack as well as everything else. 1170 */ 1171 continue; 1172 } 1173 /* 1174 * Remove the thread from the GC list. If the thread 1175 * isn't yet detached, it will get added back to the 1176 * GC list at a later time. 1177 */ 1178 THR_GCLIST_REMOVE(td); 1179 DBG_MSG("Freeing thread %p stack\n", td); 1180 /* 1181 * We can free the thread stack since it's no longer 1182 * in use. 1183 */ 1184 _thr_stack_free(&td->attr); 1185 if (((td->attr.flags & PTHREAD_DETACHED) != 0) && 1186 (td->refcount == 0)) { 1187 /* 1188 * The thread has detached and is no longer 1189 * referenced. It is safe to remove all 1190 * remnants of the thread. 1191 */ 1192 THR_LIST_REMOVE(td); 1193 TAILQ_INSERT_HEAD(&worklist, td, gcle); 1194 } 1195 } 1196 KSE_LOCK_RELEASE(curthread->kse, &_thread_list_lock); 1197 _kse_critical_leave(crit); 1198 1199 while ((td = TAILQ_FIRST(&worklist)) != NULL) { 1200 TAILQ_REMOVE(&worklist, td, gcle); 1201 1202 if ((td->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) { 1203 crit = _kse_critical_enter(); 1204 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1205 kse_free_unlocked(td->kse); 1206 kseg_free_unlocked(td->kseg); 1207 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1208 _kse_critical_leave(crit); 1209 } 1210 /* 1211 * XXX we don't free initial thread, because there might 1212 * have some code referencing initial thread. 1213 */ 1214 if (td != _thr_initial) { 1215 DBG_MSG("Freeing thread %p\n", td); 1216 _thr_free(curthread, td); 1217 } else 1218 DBG_MSG("Initial thread won't be freed\n"); 1219 } 1220 /* XXX free kse and ksegrp list should be looked as well */ 1221} 1222 1223 1224/* 1225 * Only new threads that are running or suspended may be scheduled. 1226 */ 1227int 1228_thr_schedule_add(struct pthread *curthread, struct pthread *newthread) 1229{ 1230 kse_critical_t crit; 1231 int ret; 1232 1233 /* Add the new thread. */ 1234 thr_link(newthread); 1235 1236 /* 1237 * If this is the first time creating a thread, make sure 1238 * the mailbox is set for the current thread. 1239 */ 1240 if ((newthread->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) { 1241#ifdef NOT_YET 1242 /* We use the thread's stack as the KSE's stack. */ 1243 new_thread->kse->k_mbx.km_stack.ss_sp = 1244 new_thread->attr.stackaddr_attr; 1245 new_thread->kse->k_mbx.km_stack.ss_size = 1246 new_thread->attr.stacksize_attr; 1247#endif 1248 /* 1249 * No need to lock the scheduling queue since the 1250 * KSE/KSEG pair have not yet been started. 1251 */ 1252 KSEG_THRQ_ADD(newthread->kseg, newthread); 1253 if (newthread->state == PS_RUNNING) 1254 THR_RUNQ_INSERT_TAIL(newthread); 1255 newthread->kse->k_curthread = NULL; 1256 newthread->kse->k_mbx.km_flags = 0; 1257 newthread->kse->k_mbx.km_func = (kse_func_t *)kse_sched_multi; 1258 newthread->kse->k_mbx.km_quantum = 0; 1259 1260 /* 1261 * This thread needs a new KSE and KSEG. 1262 */ 1263 newthread->kse->k_flags &= ~KF_INITIALIZED; 1264 newthread->kse->k_flags |= KF_STARTED; 1265 ret = kse_create(&newthread->kse->k_mbx, 1); 1266 if (ret != 0) 1267 ret = errno; 1268 } 1269 else { 1270 /* 1271 * Lock the KSE and add the new thread to its list of 1272 * assigned threads. If the new thread is runnable, also 1273 * add it to the KSE's run queue. 1274 */ 1275 crit = _kse_critical_enter(); 1276 KSE_SCHED_LOCK(curthread->kse, newthread->kseg); 1277 KSEG_THRQ_ADD(newthread->kseg, newthread); 1278 if (newthread->state == PS_RUNNING) 1279 THR_RUNQ_INSERT_TAIL(newthread); 1280 if ((newthread->kse->k_flags & KF_STARTED) == 0) { 1281 /* 1282 * This KSE hasn't been started yet. Start it 1283 * outside of holding the lock. 1284 */ 1285 newthread->kse->k_flags |= KF_STARTED; 1286 newthread->kse->k_mbx.km_func = 1287 (kse_func_t *)kse_sched_multi; 1288 newthread->kse->k_mbx.km_flags = 0; 1289 kse_create(&newthread->kse->k_mbx, 0); 1290 } else if ((newthread->state == PS_RUNNING) && 1291 KSE_IS_IDLE(newthread->kse)) { 1292 /* 1293 * The thread is being scheduled on another KSEG. 1294 */ 1295 kse_wakeup_one(newthread); 1296 } 1297 KSE_SCHED_UNLOCK(curthread->kse, newthread->kseg); 1298 _kse_critical_leave(crit); 1299 ret = 0; 1300 } 1301 if (ret != 0) 1302 thr_unlink(newthread); 1303 1304 return (ret); 1305} 1306 1307void 1308kse_waitq_insert(struct pthread *thread) 1309{ 1310 struct pthread *td; 1311 1312 if (thread->wakeup_time.tv_sec == -1) 1313 TAILQ_INSERT_TAIL(&thread->kse->k_schedq->sq_waitq, thread, 1314 pqe); 1315 else { 1316 td = TAILQ_FIRST(&thread->kse->k_schedq->sq_waitq); 1317 while ((td != NULL) && (td->wakeup_time.tv_sec != -1) && 1318 ((td->wakeup_time.tv_sec < thread->wakeup_time.tv_sec) || 1319 ((td->wakeup_time.tv_sec == thread->wakeup_time.tv_sec) && 1320 (td->wakeup_time.tv_nsec <= thread->wakeup_time.tv_nsec)))) 1321 td = TAILQ_NEXT(td, pqe); 1322 if (td == NULL) 1323 TAILQ_INSERT_TAIL(&thread->kse->k_schedq->sq_waitq, 1324 thread, pqe); 1325 else 1326 TAILQ_INSERT_BEFORE(td, thread, pqe); 1327 } 1328 thread->flags |= THR_FLAGS_IN_WAITQ; 1329} 1330 1331/* 1332 * This must be called with the scheduling lock held. 1333 */ 1334static void 1335kse_check_completed(struct kse *kse) 1336{ 1337 struct pthread *thread; 1338 struct kse_thr_mailbox *completed; 1339 int sig; 1340 1341 if ((completed = kse->k_mbx.km_completed) != NULL) { 1342 kse->k_mbx.km_completed = NULL; 1343 while (completed != NULL) { 1344 thread = completed->tm_udata; 1345 DBG_MSG("Found completed thread %p, name %s\n", 1346 thread, 1347 (thread->name == NULL) ? "none" : thread->name); 1348 thread->blocked = 0; 1349 if (thread != kse->k_curthread) { 1350 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1351 THR_SET_STATE(thread, PS_SUSPENDED); 1352 else 1353 KSE_RUNQ_INSERT_TAIL(kse, thread); 1354 if ((thread->kse != kse) && 1355 (thread->kse->k_curthread == thread)) { 1356 thread->kse->k_curthread = NULL; 1357 thread->active = 0; 1358 } 1359 } 1360 if ((sig = thread->tmbx.tm_syncsig.si_signo) != 0) { 1361 if (SIGISMEMBER(thread->sigmask, sig)) 1362 SIGADDSET(thread->sigpend, sig); 1363 else 1364 _thr_sig_add(thread, sig, &thread->tmbx.tm_syncsig); 1365 thread->tmbx.tm_syncsig.si_signo = 0; 1366 } 1367 completed = completed->tm_next; 1368 } 1369 } 1370} 1371 1372/* 1373 * This must be called with the scheduling lock held. 1374 */ 1375static void 1376kse_check_waitq(struct kse *kse) 1377{ 1378 struct pthread *pthread; 1379 struct timespec ts; 1380 1381 KSE_GET_TOD(kse, &ts); 1382 1383 /* 1384 * Wake up threads that have timedout. This has to be 1385 * done before adding the current thread to the run queue 1386 * so that a CPU intensive thread doesn't get preference 1387 * over waiting threads. 1388 */ 1389 while (((pthread = KSE_WAITQ_FIRST(kse)) != NULL) && 1390 thr_timedout(pthread, &ts)) { 1391 /* Remove the thread from the wait queue: */ 1392 KSE_WAITQ_REMOVE(kse, pthread); 1393 DBG_MSG("Found timedout thread %p in waitq\n", pthread); 1394 1395 /* Indicate the thread timedout: */ 1396 pthread->timeout = 1; 1397 1398 /* Add the thread to the priority queue: */ 1399 if ((pthread->flags & THR_FLAGS_SUSPENDED) != 0) 1400 THR_SET_STATE(pthread, PS_SUSPENDED); 1401 else { 1402 THR_SET_STATE(pthread, PS_RUNNING); 1403 KSE_RUNQ_INSERT_TAIL(kse, pthread); 1404 } 1405 } 1406} 1407 1408static int 1409thr_timedout(struct pthread *thread, struct timespec *curtime) 1410{ 1411 if (thread->wakeup_time.tv_sec < 0) 1412 return (0); 1413 else if (thread->wakeup_time.tv_sec > curtime->tv_sec) 1414 return (0); 1415 else if ((thread->wakeup_time.tv_sec == curtime->tv_sec) && 1416 (thread->wakeup_time.tv_nsec > curtime->tv_nsec)) 1417 return (0); 1418 else 1419 return (1); 1420} 1421 1422/* 1423 * This must be called with the scheduling lock held. 1424 * 1425 * Each thread has a time slice, a wakeup time (used when it wants 1426 * to wait for a specified amount of time), a run state, and an 1427 * active flag. 1428 * 1429 * When a thread gets run by the scheduler, the active flag is 1430 * set to non-zero (1). When a thread performs an explicit yield 1431 * or schedules a state change, it enters the scheduler and the 1432 * active flag is cleared. When the active flag is still seen 1433 * set in the scheduler, that means that the thread is blocked in 1434 * the kernel (because it is cleared before entering the scheduler 1435 * in all other instances). 1436 * 1437 * The wakeup time is only set for those states that can timeout. 1438 * It is set to (-1, -1) for all other instances. 1439 * 1440 * The thread's run state, aside from being useful when debugging, 1441 * is used to place the thread in an appropriate queue. There 1442 * are 2 basic queues: 1443 * 1444 * o run queue - queue ordered by priority for all threads 1445 * that are runnable 1446 * o waiting queue - queue sorted by wakeup time for all threads 1447 * that are not otherwise runnable (not blocked 1448 * in kernel, not waiting for locks) 1449 * 1450 * The thread's time slice is used for round-robin scheduling 1451 * (the default scheduling policy). While a SCHED_RR thread 1452 * is runnable it's time slice accumulates. When it reaches 1453 * the time slice interval, it gets reset and added to the end 1454 * of the queue of threads at its priority. When a thread no 1455 * longer becomes runnable (blocks in kernel, waits, etc), its 1456 * time slice is reset. 1457 * 1458 * The job of kse_switchout_thread() is to handle all of the above. 1459 */ 1460static void 1461kse_switchout_thread(struct kse *kse, struct pthread *thread) 1462{ 1463 int level; 1464 int i; 1465 int restart; 1466 siginfo_t siginfo; 1467 1468 /* 1469 * Place the currently running thread into the 1470 * appropriate queue(s). 1471 */ 1472 DBG_MSG("Switching out thread %p, state %d\n", thread, thread->state); 1473 1474 THR_DEACTIVATE_LAST_LOCK(thread); 1475 if (thread->blocked != 0) { 1476 thread->active = 0; 1477 thread->need_switchout = 0; 1478 /* This thread must have blocked in the kernel. */ 1479 /* thread->slice_usec = -1;*/ /* restart timeslice */ 1480 if ((thread->slice_usec != -1) && 1481 (thread->attr.sched_policy != SCHED_FIFO)) 1482 thread->slice_usec += (thread->tmbx.tm_uticks 1483 + thread->tmbx.tm_sticks) * _clock_res_usec; 1484 /* 1485 * Check for pending signals for this thread to 1486 * see if we need to interrupt it in the kernel. 1487 */ 1488 if (thread->check_pending != 0) { 1489 for (i = 1; i <= _SIG_MAXSIG; ++i) { 1490 if (SIGISMEMBER(thread->sigpend, i) && 1491 !SIGISMEMBER(thread->sigmask, i)) { 1492 restart = _thread_sigact[1 - 1].sa_flags & SA_RESTART; 1493 kse_thr_interrupt(&thread->tmbx, 1494 restart ? -2 : -1); 1495 break; 1496 } 1497 } 1498 } 1499 } 1500 else { 1501 switch (thread->state) { 1502 case PS_DEAD: 1503 /* 1504 * The scheduler is operating on a different 1505 * stack. It is safe to do garbage collecting 1506 * here. 1507 */ 1508 thread->active = 0; 1509 thread->need_switchout = 0; 1510 thr_cleanup(kse, thread); 1511 return; 1512 break; 1513 1514 case PS_RUNNING: 1515 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1516 THR_SET_STATE(thread, PS_SUSPENDED); 1517 break; 1518 1519 case PS_COND_WAIT: 1520 case PS_SLEEP_WAIT: 1521 /* Insert into the waiting queue: */ 1522 KSE_WAITQ_INSERT(kse, thread); 1523 break; 1524 1525 case PS_LOCKWAIT: 1526 /* 1527 * This state doesn't timeout. 1528 */ 1529 thread->wakeup_time.tv_sec = -1; 1530 thread->wakeup_time.tv_nsec = -1; 1531 level = thread->locklevel - 1; 1532 if (!_LCK_GRANTED(&thread->lockusers[level])) 1533 KSE_WAITQ_INSERT(kse, thread); 1534 else 1535 THR_SET_STATE(thread, PS_RUNNING); 1536 break; 1537 1538 case PS_SIGWAIT: 1539 KSE_WAITQ_INSERT(kse, thread); 1540 break; 1541 case PS_JOIN: 1542 case PS_MUTEX_WAIT: 1543 case PS_SIGSUSPEND: 1544 case PS_SUSPENDED: 1545 case PS_DEADLOCK: 1546 default: 1547 /* 1548 * These states don't timeout. 1549 */ 1550 thread->wakeup_time.tv_sec = -1; 1551 thread->wakeup_time.tv_nsec = -1; 1552 1553 /* Insert into the waiting queue: */ 1554 KSE_WAITQ_INSERT(kse, thread); 1555 break; 1556 } 1557 if (thread->state != PS_RUNNING) { 1558 /* Restart the time slice: */ 1559 thread->slice_usec = -1; 1560 } else { 1561 if (thread->need_switchout != 0) 1562 /* 1563 * The thread yielded on its own; 1564 * restart the timeslice. 1565 */ 1566 thread->slice_usec = -1; 1567 else if ((thread->slice_usec != -1) && 1568 (thread->attr.sched_policy != SCHED_FIFO)) { 1569 thread->slice_usec += (thread->tmbx.tm_uticks 1570 + thread->tmbx.tm_sticks) * _clock_res_usec; 1571 /* Check for time quantum exceeded: */ 1572 if (thread->slice_usec > TIMESLICE_USEC) 1573 thread->slice_usec = -1; 1574 } 1575 if (thread->slice_usec == -1) { 1576 /* 1577 * The thread exceeded its time quantum or 1578 * it yielded the CPU; place it at the tail 1579 * of the queue for its priority. 1580 */ 1581 KSE_RUNQ_INSERT_TAIL(kse, thread); 1582 } else { 1583 /* 1584 * The thread hasn't exceeded its interval 1585 * Place it at the head of the queue for its 1586 * priority. 1587 */ 1588 KSE_RUNQ_INSERT_HEAD(kse, thread); 1589 } 1590 } 1591 } 1592 thread->active = 0; 1593 thread->need_switchout = 0; 1594 if (thread->check_pending != 0) { 1595 /* Install pending signals into the frame. */ 1596 thread->check_pending = 0; 1597 KSE_LOCK_ACQUIRE(kse, &_thread_signal_lock); 1598 for (i = 1; i <= _SIG_MAXSIG; i++) { 1599 if (SIGISMEMBER(thread->sigmask, i)) 1600 continue; 1601 if (SIGISMEMBER(thread->sigpend, i)) 1602 _thr_sig_add(thread, i, &thread->siginfo[i-1]); 1603 else if (SIGISMEMBER(_thr_proc_sigpending, i) && 1604 _thr_getprocsig_unlocked(i, &siginfo)) { 1605 _thr_sig_add(thread, i, &siginfo); 1606 } 1607 } 1608 KSE_LOCK_RELEASE(kse, &_thread_signal_lock); 1609 } 1610} 1611 1612/* 1613 * This function waits for the smallest timeout value of any waiting 1614 * thread, or until it receives a message from another KSE. 1615 * 1616 * This must be called with the scheduling lock held. 1617 */ 1618static void 1619kse_wait(struct kse *kse, struct pthread *td_wait) 1620{ 1621 struct timespec ts, ts_sleep; 1622 int saved_flags; 1623 1624 KSE_GET_TOD(kse, &ts); 1625 1626 if ((td_wait == NULL) || (td_wait->wakeup_time.tv_sec < 0)) { 1627 /* Limit sleep to no more than 1 minute. */ 1628 ts_sleep.tv_sec = 60; 1629 ts_sleep.tv_nsec = 0; 1630 } else { 1631 TIMESPEC_SUB(&ts_sleep, &td_wait->wakeup_time, &ts); 1632 if (ts_sleep.tv_sec > 60) { 1633 ts_sleep.tv_sec = 60; 1634 ts_sleep.tv_nsec = 0; 1635 } 1636 } 1637 /* Don't sleep for negative times. */ 1638 if ((ts_sleep.tv_sec >= 0) && (ts_sleep.tv_nsec >= 0)) { 1639 KSE_SET_IDLE(kse); 1640 kse->k_kseg->kg_idle_kses++; 1641 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1642 saved_flags = kse->k_mbx.km_flags; 1643 kse->k_mbx.km_flags |= KMF_NOUPCALL; 1644 kse_release(&ts_sleep); 1645 kse->k_mbx.km_flags = saved_flags; 1646 KSE_SCHED_LOCK(kse, kse->k_kseg); 1647 if (KSE_IS_IDLE(kse)) { 1648 KSE_CLEAR_IDLE(kse); 1649 kse->k_kseg->kg_idle_kses--; 1650 } 1651 } 1652} 1653 1654/* 1655 * Avoid calling this kse_exit() so as not to confuse it with the 1656 * system call of the same name. 1657 */ 1658static void 1659kse_fini(struct kse *kse) 1660{ 1661 /* struct kse_group *free_kseg = NULL; */ 1662 struct timespec ts; 1663 1664 /* 1665 * Check to see if this is one of the main kses. 1666 */ 1667 if (kse->k_kseg != _kse_initial->k_kseg) { 1668 PANIC("shouldn't get here"); 1669 /* This is for supporting thread groups. */ 1670#ifdef NOT_YET 1671 /* Remove this KSE from the KSEG's list of KSEs. */ 1672 KSE_SCHED_LOCK(kse, kse->k_kseg); 1673 TAILQ_REMOVE(&kse->k_kseg->kg_kseq, kse, k_kgqe); 1674 kse->k_kseg->kg_ksecount--; 1675 if (TAILQ_EMPTY(&kse->k_kseg->kg_kseq)) 1676 free_kseg = kse->k_kseg; 1677 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1678 1679 /* 1680 * Add this KSE to the list of free KSEs along with 1681 * the KSEG if is now orphaned. 1682 */ 1683 KSE_LOCK_ACQUIRE(kse, &kse_lock); 1684 if (free_kseg != NULL) 1685 kseg_free_unlocked(free_kseg); 1686 kse_free_unlocked(kse); 1687 KSE_LOCK_RELEASE(kse, &kse_lock); 1688 kse_exit(); 1689 /* Never returns. */ 1690 PANIC("kse_exit()"); 1691#endif 1692 } else { 1693#ifdef NOT_YET 1694 /* 1695 * In future, we might allow program to kill 1696 * kse in initial group. 1697 */ 1698 if (kse != _kse_initial) { 1699 KSE_SCHED_LOCK(kse, kse->k_kseg); 1700 TAILQ_REMOVE(&kse->k_kseg->kg_kseq, kse, k_kgqe); 1701 kse->k_kseg->kg_ksecount--; 1702 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1703 KSE_LOCK_ACQUIRE(kse, &kse_lock); 1704 kse_free_unlocked(kse); 1705 KSE_LOCK_RELEASE(kse, &kse_lock); 1706 kse_exit(); 1707 /* Never returns. */ 1708 PANIC("kse_exit() failed for initial kseg"); 1709 } 1710#endif 1711 KSE_SCHED_LOCK(kse, kse->k_kseg); 1712 KSE_SET_IDLE(kse); 1713 kse->k_kseg->kg_idle_kses++; 1714 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1715 ts.tv_sec = 120; 1716 ts.tv_nsec = 0; 1717 kse->k_mbx.km_flags = 0; 1718 kse_release(&ts); 1719 /* Never reach */ 1720 } 1721} 1722 1723void 1724_thr_set_timeout(const struct timespec *timeout) 1725{ 1726 struct pthread *curthread = _get_curthread(); 1727 struct timespec ts; 1728 1729 /* Reset the timeout flag for the running thread: */ 1730 curthread->timeout = 0; 1731 1732 /* Check if the thread is to wait forever: */ 1733 if (timeout == NULL) { 1734 /* 1735 * Set the wakeup time to something that can be recognised as 1736 * different to an actual time of day: 1737 */ 1738 curthread->wakeup_time.tv_sec = -1; 1739 curthread->wakeup_time.tv_nsec = -1; 1740 } 1741 /* Check if no waiting is required: */ 1742 else if ((timeout->tv_sec == 0) && (timeout->tv_nsec == 0)) { 1743 /* Set the wake up time to 'immediately': */ 1744 curthread->wakeup_time.tv_sec = 0; 1745 curthread->wakeup_time.tv_nsec = 0; 1746 } else { 1747 /* Calculate the time for the current thread to wakeup: */ 1748 KSE_GET_TOD(curthread->kse, &ts); 1749 TIMESPEC_ADD(&curthread->wakeup_time, &ts, timeout); 1750 } 1751} 1752 1753void 1754_thr_panic_exit(char *file, int line, char *msg) 1755{ 1756 char buf[256]; 1757 1758 snprintf(buf, sizeof(buf), "(%s:%d) %s\n", file, line, msg); 1759 __sys_write(2, buf, strlen(buf)); 1760 abort(); 1761} 1762 1763void 1764_thr_setrunnable(struct pthread *curthread, struct pthread *thread) 1765{ 1766 kse_critical_t crit; 1767 1768 crit = _kse_critical_enter(); 1769 KSE_SCHED_LOCK(curthread->kse, thread->kseg); 1770 _thr_setrunnable_unlocked(thread); 1771 KSE_SCHED_UNLOCK(curthread->kse, thread->kseg); 1772 _kse_critical_leave(crit); 1773} 1774 1775void 1776_thr_setrunnable_unlocked(struct pthread *thread) 1777{ 1778 if ((thread->kseg->kg_flags & KGF_SINGLE_THREAD) != 0) { 1779 /* No silly queues for these threads. */ 1780 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1781 THR_SET_STATE(thread, PS_SUSPENDED); 1782 else 1783 THR_SET_STATE(thread, PS_RUNNING); 1784 } else if (thread->state != PS_RUNNING) { 1785 if ((thread->flags & THR_FLAGS_IN_WAITQ) != 0) 1786 KSE_WAITQ_REMOVE(thread->kse, thread); 1787 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1788 THR_SET_STATE(thread, PS_SUSPENDED); 1789 else { 1790 THR_SET_STATE(thread, PS_RUNNING); 1791 if ((thread->blocked == 0) && (thread->active == 0) && 1792 (thread->flags & THR_FLAGS_IN_RUNQ) == 0) 1793 THR_RUNQ_INSERT_TAIL(thread); 1794 } 1795 } 1796 /* 1797 * XXX - Threads are not yet assigned to specific KSEs; they are 1798 * assigned to the KSEG. So the fact that a thread's KSE is 1799 * waiting doesn't necessarily mean that it will be the KSE 1800 * that runs the thread after the lock is granted. But we 1801 * don't know if the other KSEs within the same KSEG are 1802 * also in a waiting state or not so we err on the side of 1803 * caution and wakeup the thread's last known KSE. We 1804 * ensure that the threads KSE doesn't change while it's 1805 * scheduling lock is held so it is safe to reference it 1806 * (the KSE). If the KSE wakes up and doesn't find any more 1807 * work it will again go back to waiting so no harm is done. 1808 */ 1809 kse_wakeup_one(thread); 1810} 1811 1812static void 1813kse_wakeup_one(struct pthread *thread) 1814{ 1815 struct kse *ke; 1816 1817 if (KSE_IS_IDLE(thread->kse)) { 1818 KSE_CLEAR_IDLE(thread->kse); 1819 thread->kseg->kg_idle_kses--; 1820 KSE_WAKEUP(thread->kse); 1821 } else { 1822 TAILQ_FOREACH(ke, &thread->kseg->kg_kseq, k_kgqe) { 1823 if (KSE_IS_IDLE(ke)) { 1824 KSE_CLEAR_IDLE(ke); 1825 ke->k_kseg->kg_idle_kses--; 1826 KSE_WAKEUP(ke); 1827 return; 1828 } 1829 } 1830 } 1831} 1832 1833static void 1834kse_wakeup_multi(struct kse *curkse) 1835{ 1836 struct kse *ke; 1837 int tmp; 1838 1839 if ((tmp = KSE_RUNQ_THREADS(curkse)) && curkse->k_kseg->kg_idle_kses) { 1840 TAILQ_FOREACH(ke, &curkse->k_kseg->kg_kseq, k_kgqe) { 1841 if (KSE_IS_IDLE(ke)) { 1842 KSE_CLEAR_IDLE(ke); 1843 ke->k_kseg->kg_idle_kses--; 1844 KSE_WAKEUP(ke); 1845 if (--tmp == 0) 1846 break; 1847 } 1848 } 1849 } 1850} 1851 1852struct pthread * 1853_get_curthread(void) 1854{ 1855 return (_ksd_curthread()); 1856} 1857 1858/* This assumes the caller has disabled upcalls. */ 1859struct kse * 1860_get_curkse(void) 1861{ 1862 return (_ksd_curkse()); 1863} 1864 1865void 1866_set_curkse(struct kse *kse) 1867{ 1868 _ksd_setprivate(&kse->k_ksd); 1869} 1870 1871/* 1872 * Allocate a new KSEG. 1873 * 1874 * We allow the current thread to be NULL in the case that this 1875 * is the first time a KSEG is being created (library initialization). 1876 * In this case, we don't need to (and can't) take any locks. 1877 */ 1878struct kse_group * 1879_kseg_alloc(struct pthread *curthread) 1880{ 1881 struct kse_group *kseg = NULL; 1882 kse_critical_t crit; 1883 1884 if ((curthread != NULL) && (free_kseg_count > 0)) { 1885 /* Use the kse lock for the kseg queue. */ 1886 crit = _kse_critical_enter(); 1887 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1888 if ((kseg = TAILQ_FIRST(&free_kse_groupq)) != NULL) { 1889 TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe); 1890 free_kseg_count--; 1891 active_kseg_count++; 1892 TAILQ_INSERT_TAIL(&active_kse_groupq, kseg, kg_qe); 1893 } 1894 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1895 _kse_critical_leave(crit); 1896 if (kseg) 1897 kseg_reinit(kseg); 1898 } 1899 1900 /* 1901 * If requested, attempt to allocate a new KSE group only if the 1902 * KSE allocation was successful and a KSE group wasn't found in 1903 * the free list. 1904 */ 1905 if ((kseg == NULL) && 1906 ((kseg = (struct kse_group *)malloc(sizeof(*kseg))) != NULL)) { 1907 if (_pq_alloc(&kseg->kg_schedq.sq_runq, 1908 THR_MIN_PRIORITY, THR_LAST_PRIORITY) != 0) { 1909 free(kseg); 1910 kseg = NULL; 1911 } else { 1912 kseg_init(kseg); 1913 /* Add the KSEG to the list of active KSEGs. */ 1914 if (curthread != NULL) { 1915 crit = _kse_critical_enter(); 1916 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1917 active_kseg_count++; 1918 TAILQ_INSERT_TAIL(&active_kse_groupq, 1919 kseg, kg_qe); 1920 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1921 _kse_critical_leave(crit); 1922 } else { 1923 active_kseg_count++; 1924 TAILQ_INSERT_TAIL(&active_kse_groupq, 1925 kseg, kg_qe); 1926 } 1927 } 1928 } 1929 return (kseg); 1930} 1931 1932/* 1933 * This must be called with the kse lock held and when there are 1934 * no more threads that reference it. 1935 */ 1936static void 1937kseg_free_unlocked(struct kse_group *kseg) 1938{ 1939 TAILQ_REMOVE(&active_kse_groupq, kseg, kg_qe); 1940 TAILQ_INSERT_HEAD(&free_kse_groupq, kseg, kg_qe); 1941 free_kseg_count++; 1942 active_kseg_count--; 1943} 1944 1945void 1946_kseg_free(struct kse_group *kseg) 1947{ 1948 struct kse *curkse; 1949 kse_critical_t crit; 1950 1951 crit = _kse_critical_enter(); 1952 curkse = _get_curkse(); 1953 KSE_LOCK_ACQUIRE(curkse, &kse_lock); 1954 kseg_free_unlocked(kseg); 1955 KSE_LOCK_RELEASE(curkse, &kse_lock); 1956 _kse_critical_leave(crit); 1957} 1958 1959/* 1960 * Allocate a new KSE. 1961 * 1962 * We allow the current thread to be NULL in the case that this 1963 * is the first time a KSE is being created (library initialization). 1964 * In this case, we don't need to (and can't) take any locks. 1965 */ 1966struct kse * 1967_kse_alloc(struct pthread *curthread) 1968{ 1969 struct kse *kse = NULL; 1970 kse_critical_t crit; 1971 int need_ksd = 0; 1972 int i; 1973 1974 if ((curthread != NULL) && (free_kse_count > 0)) { 1975 crit = _kse_critical_enter(); 1976 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1977 /* Search for a finished KSE. */ 1978 kse = TAILQ_FIRST(&free_kseq); 1979 while ((kse != NULL) && 1980 ((kse->k_mbx.km_flags & KMF_DONE) == 0)) { 1981 kse = TAILQ_NEXT(kse, k_qe); 1982 } 1983 if (kse != NULL) { 1984 DBG_MSG("found an unused kse.\n"); 1985 TAILQ_REMOVE(&free_kseq, kse, k_qe); 1986 free_kse_count--; 1987 TAILQ_INSERT_TAIL(&active_kseq, kse, k_qe); 1988 active_kse_count++; 1989 } 1990 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1991 _kse_critical_leave(crit); 1992 if (kse != NULL) 1993 kse_reinit(kse); 1994 } 1995 if ((kse == NULL) && 1996 ((kse = (struct kse *)malloc(sizeof(*kse))) != NULL)) { 1997 bzero(kse, sizeof(*kse)); 1998 1999 /* Initialize the lockusers. */ 2000 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 2001 _lockuser_init(&kse->k_lockusers[i], (void *)kse); 2002 _LCK_SET_PRIVATE2(&kse->k_lockusers[i], NULL); 2003 } 2004 /* _lock_init(kse->k_lock, ...) */ 2005 2006 /* We had to malloc a kse; mark it as needing a new ID.*/ 2007 need_ksd = 1; 2008 2009 /* 2010 * Create the KSE context. 2011 * 2012 * XXX - For now this is done here in the allocation. 2013 * In the future, we may want to have it done 2014 * outside the allocation so that scope system 2015 * threads (one thread per KSE) are not required 2016 * to have a stack for an unneeded kse upcall. 2017 */ 2018 kse->k_mbx.km_func = (kse_func_t *)kse_sched_multi; 2019 kse->k_mbx.km_stack.ss_sp = (char *)malloc(KSE_STACKSIZE); 2020 kse->k_mbx.km_stack.ss_size = KSE_STACKSIZE; 2021 kse->k_mbx.km_udata = (void *)kse; 2022 kse->k_mbx.km_quantum = 20000; 2023 /* 2024 * We need to keep a copy of the stack in case it 2025 * doesn't get used; a KSE running a scope system 2026 * thread will use that thread's stack. 2027 */ 2028 kse->k_stack.ss_sp = kse->k_mbx.km_stack.ss_sp; 2029 kse->k_stack.ss_size = kse->k_mbx.km_stack.ss_size; 2030 if (kse->k_mbx.km_stack.ss_sp == NULL) { 2031 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 2032 _lockuser_destroy(&kse->k_lockusers[i]); 2033 } 2034 /* _lock_destroy(&kse->k_lock); */ 2035 free(kse); 2036 kse = NULL; 2037 } 2038 } 2039 if ((kse != NULL) && (need_ksd != 0)) { 2040 /* This KSE needs initialization. */ 2041 if (curthread != NULL) { 2042 crit = _kse_critical_enter(); 2043 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2044 } 2045 /* Initialize KSD inside of the lock. */ 2046 if (_ksd_create(&kse->k_ksd, (void *)kse, sizeof(*kse)) != 0) { 2047 if (curthread != NULL) { 2048 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2049 _kse_critical_leave(crit); 2050 } 2051 free(kse->k_mbx.km_stack.ss_sp); 2052 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 2053 _lockuser_destroy(&kse->k_lockusers[i]); 2054 } 2055 free(kse); 2056 return (NULL); 2057 } 2058 kse->k_flags = 0; 2059 TAILQ_INSERT_TAIL(&active_kseq, kse, k_qe); 2060 active_kse_count++; 2061 if (curthread != NULL) { 2062 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2063 _kse_critical_leave(crit); 2064 } 2065 } 2066 return (kse); 2067} 2068 2069static void 2070kse_reinit(struct kse *kse) 2071{ 2072 /* 2073 * XXX - For now every kse has its stack. 2074 * In the future, we may want to have it done 2075 * outside the allocation so that scope system 2076 * threads (one thread per KSE) are not required 2077 * to have a stack for an unneeded kse upcall. 2078 */ 2079 kse->k_mbx.km_flags = 0; 2080 kse->k_curthread = 0; 2081 kse->k_kseg = 0; 2082 kse->k_schedq = 0; 2083 kse->k_locklevel = 0; 2084 SIGEMPTYSET(kse->k_sigmask); 2085 bzero(&kse->k_sigq, sizeof(kse->k_sigq)); 2086 kse->k_check_sigq = 0; 2087 kse->k_flags = 0; 2088 kse->k_waiting = 0; 2089 kse->k_idle = 0; 2090 kse->k_error = 0; 2091 kse->k_cpu = 0; 2092 kse->k_done = 0; 2093 kse->k_switch = 0; 2094} 2095 2096void 2097kse_free_unlocked(struct kse *kse) 2098{ 2099 TAILQ_REMOVE(&active_kseq, kse, k_qe); 2100 active_kse_count--; 2101 kse->k_kseg = NULL; 2102 kse->k_mbx.km_quantum = 20000; 2103 kse->k_flags = 0; 2104 TAILQ_INSERT_HEAD(&free_kseq, kse, k_qe); 2105 free_kse_count++; 2106} 2107 2108void 2109_kse_free(struct pthread *curthread, struct kse *kse) 2110{ 2111 kse_critical_t crit; 2112 2113 if (curthread == NULL) 2114 kse_free_unlocked(kse); 2115 else { 2116 crit = _kse_critical_enter(); 2117 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2118 kse_free_unlocked(kse); 2119 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2120 _kse_critical_leave(crit); 2121 } 2122} 2123 2124static void 2125kseg_init(struct kse_group *kseg) 2126{ 2127 kseg_reinit(kseg); 2128 _lock_init(&kseg->kg_lock, LCK_ADAPTIVE, _kse_lock_wait, 2129 _kse_lock_wakeup); 2130} 2131 2132static void 2133kseg_reinit(struct kse_group *kseg) 2134{ 2135 TAILQ_INIT(&kseg->kg_kseq); 2136 TAILQ_INIT(&kseg->kg_threadq); 2137 TAILQ_INIT(&kseg->kg_schedq.sq_waitq); 2138 kseg->kg_threadcount = 0; 2139 kseg->kg_ksecount = 0; 2140 kseg->kg_idle_kses = 0; 2141 kseg->kg_flags = 0; 2142} 2143 2144struct pthread * 2145_thr_alloc(struct pthread *curthread) 2146{ 2147 kse_critical_t crit; 2148 void *p; 2149 struct pthread *thread = NULL; 2150 2151 if (curthread != NULL) { 2152 if (GC_NEEDED()) 2153 _thr_gc(curthread); 2154 if (free_thread_count > 0) { 2155 crit = _kse_critical_enter(); 2156 KSE_LOCK_ACQUIRE(curthread->kse, &thread_lock); 2157 if ((thread = TAILQ_FIRST(&free_threadq)) != NULL) { 2158 TAILQ_REMOVE(&free_threadq, thread, tle); 2159 free_thread_count--; 2160 } 2161 KSE_LOCK_RELEASE(curthread->kse, &thread_lock); 2162 _kse_critical_leave(crit); 2163 } 2164 } 2165 if (thread == NULL) { 2166 p = malloc(sizeof(struct pthread) + THR_ALIGNBYTES); 2167 if (p != NULL) { 2168 thread = (struct pthread *)THR_ALIGN(p); 2169 thread->alloc_addr = p; 2170 } 2171 } 2172 return (thread); 2173} 2174 2175void 2176_thr_free(struct pthread *curthread, struct pthread *thread) 2177{ 2178 kse_critical_t crit; 2179 int i; 2180 2181 DBG_MSG("Freeing thread %p\n", thread); 2182 if ((curthread == NULL) || (free_thread_count >= MAX_CACHED_THREADS)) { 2183 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 2184 _lockuser_destroy(&thread->lockusers[i]); 2185 } 2186 _lock_destroy(&thread->lock); 2187 free(thread->alloc_addr); 2188 } 2189 else { 2190 crit = _kse_critical_enter(); 2191 KSE_LOCK_ACQUIRE(curthread->kse, &thread_lock); 2192 TAILQ_INSERT_TAIL(&free_threadq, thread, tle); 2193 free_thread_count++; 2194 KSE_LOCK_RELEASE(curthread->kse, &thread_lock); 2195 _kse_critical_leave(crit); 2196 } 2197} 2198 2199/* 2200 * Add an active thread: 2201 * 2202 * o Assign the thread a unique id (which GDB uses to track 2203 * threads. 2204 * o Add the thread to the list of all threads and increment 2205 * number of active threads. 2206 */ 2207static void 2208thr_link(struct pthread *thread) 2209{ 2210 kse_critical_t crit; 2211 struct kse *curkse; 2212 struct pthread *curthread; 2213 2214 crit = _kse_critical_enter(); 2215 curkse = _get_curkse(); 2216 curthread = _get_curthread(); 2217 thread->sigmask = curthread->sigmask; 2218 KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock); 2219 /* 2220 * Initialize the unique id (which GDB uses to track 2221 * threads), add the thread to the list of all threads, 2222 * and 2223 */ 2224 thread->uniqueid = next_uniqueid++; 2225 THR_LIST_ADD(thread); 2226 active_threads++; 2227 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 2228 2229 _kse_critical_leave(crit); 2230} 2231 2232/* 2233 * Remove an active thread. 2234 */ 2235static void 2236thr_unlink(struct pthread *thread) 2237{ 2238 kse_critical_t crit; 2239 struct kse *curkse; 2240 2241 crit = _kse_critical_enter(); 2242 curkse = _get_curkse(); 2243 2244 KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock); 2245 THR_LIST_REMOVE(thread); 2246 active_threads--; 2247 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 2248 2249 _kse_critical_leave(crit); 2250} 2251