1 /*- 2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 3. Berkeley Software Design Inc's name may not be used to endorse or 13 * promote products derived from this software without specific prior 14 * written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 29 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 30 */ 31 32 /* 33 * Implementation of turnstiles used to hold queue of threads blocked on 34 * non-sleepable locks. Sleepable locks use condition variables to 35 * implement their queues. Turnstiles differ from a sleep queue in that 36 * turnstile queue's are assigned to a lock held by an owning thread. Thus, 37 * when one thread is enqueued onto a turnstile, it can lend its priority 38 * to the owning thread. 39 * 40 * We wish to avoid bloating locks with an embedded turnstile and we do not 41 * want to use back-pointers in the locks for the same reason. Thus, we 42 * use a similar approach to that of Solaris 7 as described in Solaris 43 * Internals by Jim Mauro and Richard McDougall. Turnstiles are looked up 44 * in a hash table based on the address of the lock. Each entry in the 45 * hash table is a linked-lists of turnstiles and is called a turnstile 46 * chain. Each chain contains a spin mutex that protects all of the 47 * turnstiles in the chain. 48 * 49 * Each time a thread is created, a turnstile is malloc'd and attached to 50 * that thread. When a thread blocks on a lock, if it is the first thread 51 * to block, it lends its turnstile to the lock. If the lock already has 52 * a turnstile, then it gives its turnstile to the lock's turnstile's free 53 * list. When a thread is woken up, it takes a turnstile from the free list 54 * if there are any other waiters. If it is the only thread blocked on the 55 * lock, then it reclaims the turnstile associated with the lock and removes 56 * it from the hash table. 57 */ 58 59 #include "opt_turnstile_profiling.h" 60 61 #include <sys/cdefs.h> 62 __FBSDID("$FreeBSD$"); 63 64 #include <sys/param.h> 65 #include <sys/systm.h> 66 #include <sys/kernel.h> 67 #include <sys/ktr.h> 68 #include <sys/lock.h> 69 #include <sys/malloc.h> 70 #include <sys/mutex.h> 71 #include <sys/proc.h> 72 #include <sys/queue.h> 73 #include <sys/resourcevar.h> 74 #include <sys/sched.h> 75 #include <sys/sysctl.h> 76 #include <sys/turnstile.h> 77 78 /* 79 * Constants for the hash table of turnstile chains. TC_SHIFT is a magic 80 * number chosen because the sleep queue's use the same value for the 81 * shift. Basically, we ignore the lower 8 bits of the address. 82 * TC_TABLESIZE must be a power of two for TC_MASK to work properly. 83 */ 84 #define TC_TABLESIZE 128 /* Must be power of 2. */ 85 #define TC_MASK (TC_TABLESIZE - 1) 86 #define TC_SHIFT 8 87 #define TC_HASH(lock) (((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK) 88 #define TC_LOOKUP(lock) &turnstile_chains[TC_HASH(lock)] 89 90 /* 91 * There are three different lists of turnstiles as follows. The list 92 * connected by ts_link entries is a per-thread list of all the turnstiles 93 * attached to locks that we own. This is used to fixup our priority when 94 * a lock is released. The other two lists use the ts_hash entries. The 95 * first of these two is the turnstile chain list that a turnstile is on 96 * when it is attached to a lock. The second list to use ts_hash is the 97 * free list hung off of a turnstile that is attached to a lock. 98 * 99 * Each turnstile contains two lists of threads. The ts_blocked list is 100 * a linked list of threads blocked on the turnstile's lock. The 101 * ts_pending list is a linked list of threads previously awakened by 102 * turnstile_signal() or turnstile_wait() that are waiting to be put on 103 * the run queue. 104 * 105 * Locking key: 106 * c - turnstile chain lock 107 * q - td_contested lock 108 */ 109 struct turnstile { 110 TAILQ_HEAD(, thread) ts_blocked; /* (c + q) Blocked threads. */ 111 TAILQ_HEAD(, thread) ts_pending; /* (c) Pending threads. */ 112 LIST_ENTRY(turnstile) ts_hash; /* (c) Chain and free list. */ 113 LIST_ENTRY(turnstile) ts_link; /* (q) Contested locks. */ 114 LIST_HEAD(, turnstile) ts_free; /* (c) Free turnstiles. */ 115 struct lock_object *ts_lockobj; /* (c) Lock we reference. */ 116 struct thread *ts_owner; /* (c + q) Who owns the lock. */ 117 }; 118 119 struct turnstile_chain { 120 LIST_HEAD(, turnstile) tc_turnstiles; /* List of turnstiles. */ 121 struct mtx tc_lock; /* Spin lock for this chain. */ 122 #ifdef TURNSTILE_PROFILING 123 u_int tc_depth; /* Length of tc_queues. */ 124 u_int tc_max_depth; /* Max length of tc_queues. */ 125 #endif 126 }; 127 128 #ifdef TURNSTILE_PROFILING 129 u_int turnstile_max_depth; 130 SYSCTL_NODE(_debug, OID_AUTO, turnstile, CTLFLAG_RD, 0, "turnstile profiling"); 131 SYSCTL_NODE(_debug_turnstile, OID_AUTO, chains, CTLFLAG_RD, 0, 132 "turnstile chain stats"); 133 SYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD, 134 &turnstile_max_depth, 0, "maxmimum depth achieved of a single chain"); 135 #endif 136 static struct mtx td_contested_lock; 137 static struct turnstile_chain turnstile_chains[TC_TABLESIZE]; 138 139 MALLOC_DEFINE(M_TURNSTILE, "turnstiles", "turnstiles"); 140 141 /* 142 * Prototypes for non-exported routines. 143 */ 144 static void init_turnstile0(void *dummy); 145 #ifdef TURNSTILE_PROFILING 146 static void init_turnstile_profiling(void *arg); 147 #endif 148 static void propagate_priority(struct thread *); 149 static void turnstile_setowner(struct turnstile *ts, struct thread *owner); 150 151 /* 152 * Walks the chain of turnstiles and their owners to propagate the priority 153 * of the thread being blocked to all the threads holding locks that have to 154 * release their locks before this thread can run again. 155 */ 156 static void 157 propagate_priority(struct thread *td) 158 { 159 struct turnstile_chain *tc; 160 struct turnstile *ts; 161 struct thread *td1; 162 int pri; 163 164 mtx_assert(&sched_lock, MA_OWNED); 165 pri = td->td_priority; 166 ts = td->td_blocked; 167 for (;;) { 168 td = ts->ts_owner; 169 170 if (td == NULL) { 171 /* 172 * This really isn't quite right. Really 173 * ought to bump priority of thread that 174 * next acquires the lock. 175 */ 176 return; 177 } 178 179 MPASS(td->td_proc != NULL); 180 MPASS(td->td_proc->p_magic == P_MAGIC); 181 182 /* 183 * XXX: The owner of a turnstile can be stale if it is the 184 * first thread to grab a slock of a sx lock. In that case 185 * it is possible for us to be at SSLEEP or some other 186 * weird state. We should probably just return if the state 187 * isn't SRUN or SLOCK. 188 */ 189 KASSERT(!TD_IS_SLEEPING(td), 190 ("sleeping thread (pid %d) owns a non-sleepable lock", 191 td->td_proc->p_pid)); 192 193 /* 194 * If this thread already has higher priority than the 195 * thread that is being blocked, we are finished. 196 */ 197 if (td->td_priority <= pri) 198 return; 199 200 /* 201 * If lock holder is actually running, just bump priority. 202 */ 203 if (TD_IS_RUNNING(td)) { 204 td->td_priority = pri; 205 return; 206 } 207 208 #ifndef SMP 209 /* 210 * For UP, we check to see if td is curthread (this shouldn't 211 * ever happen however as it would mean we are in a deadlock.) 212 */ 213 KASSERT(td != curthread, ("Deadlock detected")); 214 #endif 215 216 /* 217 * If on run queue move to new run queue, and quit. 218 * XXXKSE this gets a lot more complicated under threads 219 * but try anyhow. 220 */ 221 if (TD_ON_RUNQ(td)) { 222 MPASS(td->td_blocked == NULL); 223 sched_prio(td, pri); 224 return; 225 } 226 227 /* 228 * Bump this thread's priority. 229 */ 230 td->td_priority = pri; 231 232 /* 233 * If we aren't blocked on a lock, we should be. 234 */ 235 KASSERT(TD_ON_LOCK(td), ( 236 "process %d(%s):%d holds %s but isn't blocked on a lock\n", 237 td->td_proc->p_pid, td->td_proc->p_comm, td->td_state, 238 ts->ts_lockobj->lo_name)); 239 240 /* 241 * Pick up the lock that td is blocked on. 242 */ 243 ts = td->td_blocked; 244 MPASS(ts != NULL); 245 tc = TC_LOOKUP(ts->ts_lockobj); 246 mtx_lock_spin(&tc->tc_lock); 247 248 /* 249 * This thread may not be blocked on this turnstile anymore 250 * but instead might already be woken up on another CPU 251 * that is waiting on sched_lock in turnstile_unpend() to 252 * finish waking this thread up. We can detect this case 253 * by checking to see if this thread has been given a 254 * turnstile by either turnstile_signal() or 255 * turnstile_broadcast(). In this case, treat the thread as 256 * if it was already running. 257 */ 258 if (td->td_turnstile != NULL) { 259 mtx_unlock_spin(&tc->tc_lock); 260 return; 261 } 262 263 /* 264 * Check if the thread needs to be moved up on 265 * the blocked chain. It doesn't need to be moved 266 * if it is already at the head of the list or if 267 * the item in front of it still has a higher priority. 268 */ 269 if (td == TAILQ_FIRST(&ts->ts_blocked)) { 270 mtx_unlock_spin(&tc->tc_lock); 271 continue; 272 } 273 274 td1 = TAILQ_PREV(td, threadqueue, td_lockq); 275 if (td1->td_priority <= pri) { 276 mtx_unlock_spin(&tc->tc_lock); 277 continue; 278 } 279 280 /* 281 * Remove thread from blocked chain and determine where 282 * it should be moved up to. Since we know that td1 has 283 * a lower priority than td, we know that at least one 284 * thread in the chain has a lower priority and that 285 * td1 will thus not be NULL after the loop. 286 */ 287 mtx_lock_spin(&td_contested_lock); 288 TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq); 289 TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq) { 290 MPASS(td1->td_proc->p_magic == P_MAGIC); 291 if (td1->td_priority > pri) 292 break; 293 } 294 295 MPASS(td1 != NULL); 296 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 297 mtx_unlock_spin(&td_contested_lock); 298 CTR4(KTR_LOCK, 299 "propagate_priority: td %p moved before %p on [%p] %s", 300 td, td1, ts->ts_lockobj, ts->ts_lockobj->lo_name); 301 mtx_unlock_spin(&tc->tc_lock); 302 } 303 } 304 305 /* 306 * Early initialization of turnstiles. This is not done via a SYSINIT() 307 * since this needs to be initialized very early when mutexes are first 308 * initialized. 309 */ 310 void 311 init_turnstiles(void) 312 { 313 int i; 314 315 for (i = 0; i < TC_TABLESIZE; i++) { 316 LIST_INIT(&turnstile_chains[i].tc_turnstiles); 317 mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain", 318 NULL, MTX_SPIN); 319 } 320 mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN); 321 thread0.td_turnstile = NULL; 322 } 323 324 #ifdef TURNSTILE_PROFILING 325 static void 326 init_turnstile_profiling(void *arg) 327 { 328 struct sysctl_oid *chain_oid; 329 char chain_name[10]; 330 int i; 331 332 for (i = 0; i < TC_TABLESIZE; i++) { 333 snprintf(chain_name, sizeof(chain_name), "%d", i); 334 chain_oid = SYSCTL_ADD_NODE(NULL, 335 SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO, 336 chain_name, CTLFLAG_RD, NULL, "turnstile chain stats"); 337 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 338 "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0, 339 NULL); 340 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 341 "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth, 342 0, NULL); 343 } 344 } 345 SYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY, 346 init_turnstile_profiling, NULL); 347 #endif 348 349 static void 350 init_turnstile0(void *dummy) 351 { 352 353 thread0.td_turnstile = turnstile_alloc(); 354 } 355 SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL); 356 357 /* 358 * Set the owner of the lock this turnstile is attached to. 359 */ 360 static void 361 turnstile_setowner(struct turnstile *ts, struct thread *owner) 362 { 363 364 mtx_assert(&td_contested_lock, MA_OWNED); 365 MPASS(owner->td_proc->p_magic == P_MAGIC); 366 MPASS(ts->ts_owner == NULL); 367 ts->ts_owner = owner; 368 LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link); 369 } 370 371 /* 372 * Malloc a turnstile for a new thread, initialize it and return it. 373 */ 374 struct turnstile * 375 turnstile_alloc(void) 376 { 377 struct turnstile *ts; 378 379 ts = malloc(sizeof(struct turnstile), M_TURNSTILE, M_WAITOK | M_ZERO); 380 TAILQ_INIT(&ts->ts_blocked); 381 TAILQ_INIT(&ts->ts_pending); 382 LIST_INIT(&ts->ts_free); 383 return (ts); 384 } 385 386 /* 387 * Free a turnstile when a thread is destroyed. 388 */ 389 void 390 turnstile_free(struct turnstile *ts) 391 { 392 393 MPASS(ts != NULL); 394 MPASS(TAILQ_EMPTY(&ts->ts_blocked)); 395 MPASS(TAILQ_EMPTY(&ts->ts_pending)); 396 free(ts, M_TURNSTILE); 397 } 398 399 /* 400 * Look up the turnstile for a lock in the hash table locking the associated 401 * turnstile chain along the way. Return with the turnstile chain locked. 402 * If no turnstile is found in the hash table, NULL is returned. 403 */ 404 struct turnstile * 405 turnstile_lookup(struct lock_object *lock) 406 { 407 struct turnstile_chain *tc; 408 struct turnstile *ts; 409 410 tc = TC_LOOKUP(lock); 411 mtx_lock_spin(&tc->tc_lock); 412 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 413 if (ts->ts_lockobj == lock) 414 return (ts); 415 return (NULL); 416 } 417 418 /* 419 * Unlock the turnstile chain associated with a given lock. 420 */ 421 void 422 turnstile_release(struct lock_object *lock) 423 { 424 struct turnstile_chain *tc; 425 426 tc = TC_LOOKUP(lock); 427 mtx_unlock_spin(&tc->tc_lock); 428 } 429 430 /* 431 * Take ownership of a turnstile and adjust the priority of the new 432 * owner appropriately. 433 */ 434 void 435 turnstile_claim(struct turnstile *ts) 436 { 437 struct turnstile_chain *tc; 438 struct thread *td, *owner; 439 440 tc = TC_LOOKUP(ts->ts_lockobj); 441 mtx_assert(&tc->tc_lock, MA_OWNED); 442 443 owner = curthread; 444 mtx_lock_spin(&td_contested_lock); 445 turnstile_setowner(ts, owner); 446 mtx_unlock_spin(&td_contested_lock); 447 448 td = TAILQ_FIRST(&ts->ts_blocked); 449 MPASS(td != NULL); 450 MPASS(td->td_proc->p_magic == P_MAGIC); 451 mtx_unlock_spin(&tc->tc_lock); 452 453 /* 454 * Update the priority of the new owner if needed. 455 */ 456 mtx_lock_spin(&sched_lock); 457 if (td->td_priority < owner->td_priority) 458 owner->td_priority = td->td_priority; 459 mtx_unlock_spin(&sched_lock); 460 } 461 462 /* 463 * Block the current thread on the turnstile ts. This function will context 464 * switch and not return until this thread has been woken back up. This 465 * function must be called with the appropriate turnstile chain locked and 466 * will return with it unlocked. 467 */ 468 void 469 turnstile_wait(struct turnstile *ts, struct lock_object *lock, 470 struct thread *owner) 471 { 472 struct turnstile_chain *tc; 473 struct thread *td, *td1; 474 475 td = curthread; 476 tc = TC_LOOKUP(lock); 477 mtx_assert(&tc->tc_lock, MA_OWNED); 478 MPASS(td->td_turnstile != NULL); 479 MPASS(owner != NULL); 480 MPASS(owner->td_proc->p_magic == P_MAGIC); 481 482 /* If the passed in turnstile is NULL, use this thread's turnstile. */ 483 if (ts == NULL) { 484 #ifdef TURNSTILE_PROFILING 485 tc->tc_depth++; 486 if (tc->tc_depth > tc->tc_max_depth) { 487 tc->tc_max_depth = tc->tc_depth; 488 if (tc->tc_max_depth > turnstile_max_depth) 489 turnstile_max_depth = tc->tc_max_depth; 490 } 491 #endif 492 ts = td->td_turnstile; 493 LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash); 494 KASSERT(TAILQ_EMPTY(&ts->ts_pending), 495 ("thread's turnstile has pending threads")); 496 KASSERT(TAILQ_EMPTY(&ts->ts_blocked), 497 ("thread's turnstile has a non-empty queue")); 498 KASSERT(LIST_EMPTY(&ts->ts_free), 499 ("thread's turnstile has a non-empty free list")); 500 KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer")); 501 ts->ts_lockobj = lock; 502 mtx_lock_spin(&td_contested_lock); 503 TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq); 504 turnstile_setowner(ts, owner); 505 mtx_unlock_spin(&td_contested_lock); 506 } else { 507 TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq) 508 if (td1->td_priority > td->td_priority) 509 break; 510 mtx_lock_spin(&td_contested_lock); 511 if (td1 != NULL) 512 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 513 else 514 TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq); 515 mtx_unlock_spin(&td_contested_lock); 516 MPASS(td->td_turnstile != NULL); 517 LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash); 518 MPASS(owner == ts->ts_owner); 519 } 520 td->td_turnstile = NULL; 521 mtx_unlock_spin(&tc->tc_lock); 522 523 mtx_lock_spin(&sched_lock); 524 /* 525 * Handle race condition where a thread on another CPU that owns 526 * lock 'lock' could have woken us in between us dropping the 527 * turnstile chain lock and acquiring the sched_lock. 528 */ 529 if (td->td_flags & TDF_TSNOBLOCK) { 530 td->td_flags &= ~TDF_TSNOBLOCK; 531 mtx_unlock_spin(&sched_lock); 532 return; 533 } 534 535 #ifdef notyet 536 /* 537 * If we're borrowing an interrupted thread's VM context, we 538 * must clean up before going to sleep. 539 */ 540 if (td->td_ithd != NULL) { 541 struct ithd *it = td->td_ithd; 542 543 if (it->it_interrupted) { 544 if (LOCK_LOG_TEST(lock, 0)) 545 CTR3(KTR_LOCK, "%s: %p interrupted %p", 546 __func__, it, it->it_interrupted); 547 intr_thd_fixup(it); 548 } 549 } 550 #endif 551 552 /* Save who we are blocked on and switch. */ 553 td->td_blocked = ts; 554 td->td_lockname = lock->lo_name; 555 TD_SET_LOCK(td); 556 propagate_priority(td); 557 558 if (LOCK_LOG_TEST(lock, 0)) 559 CTR4(KTR_LOCK, "%s: td %p blocked on [%p] %s", __func__, td, 560 lock, lock->lo_name); 561 562 mi_switch(SW_VOL, NULL); 563 564 if (LOCK_LOG_TEST(lock, 0)) 565 CTR4(KTR_LOCK, "%s: td %p free from blocked on [%p] %s", 566 __func__, td, lock, lock->lo_name); 567 568 mtx_unlock_spin(&sched_lock); 569 } 570 571 /* 572 * Pick the highest priority thread on this turnstile and put it on the 573 * pending list. This must be called with the turnstile chain locked. 574 */ 575 int 576 turnstile_signal(struct turnstile *ts) 577 { 578 struct turnstile_chain *tc; 579 struct thread *td; 580 int empty; 581 582 MPASS(ts != NULL); 583 MPASS(curthread->td_proc->p_magic == P_MAGIC); 584 MPASS(ts->ts_owner == curthread); 585 tc = TC_LOOKUP(ts->ts_lockobj); 586 mtx_assert(&tc->tc_lock, MA_OWNED); 587 588 /* 589 * Pick the highest priority thread blocked on this lock and 590 * move it to the pending list. 591 */ 592 td = TAILQ_FIRST(&ts->ts_blocked); 593 MPASS(td->td_proc->p_magic == P_MAGIC); 594 mtx_lock_spin(&td_contested_lock); 595 TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq); 596 mtx_unlock_spin(&td_contested_lock); 597 TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq); 598 599 /* 600 * If the turnstile is now empty, remove it from its chain and 601 * give it to the about-to-be-woken thread. Otherwise take a 602 * turnstile from the free list and give it to the thread. 603 */ 604 empty = TAILQ_EMPTY(&ts->ts_blocked); 605 if (empty) { 606 MPASS(LIST_EMPTY(&ts->ts_free)); 607 #ifdef TURNSTILE_PROFILING 608 tc->tc_depth--; 609 #endif 610 } else 611 ts = LIST_FIRST(&ts->ts_free); 612 MPASS(ts != NULL); 613 LIST_REMOVE(ts, ts_hash); 614 td->td_turnstile = ts; 615 616 return (empty); 617 } 618 619 /* 620 * Put all blocked threads on the pending list. This must be called with 621 * the turnstile chain locked. 622 */ 623 void 624 turnstile_broadcast(struct turnstile *ts) 625 { 626 struct turnstile_chain *tc; 627 struct turnstile *ts1; 628 struct thread *td; 629 630 MPASS(ts != NULL); 631 MPASS(curthread->td_proc->p_magic == P_MAGIC); 632 MPASS(ts->ts_owner == curthread); 633 tc = TC_LOOKUP(ts->ts_lockobj); 634 mtx_assert(&tc->tc_lock, MA_OWNED); 635 636 /* 637 * Transfer the blocked list to the pending list. 638 */ 639 mtx_lock_spin(&td_contested_lock); 640 TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked, td_lockq); 641 mtx_unlock_spin(&td_contested_lock); 642 643 /* 644 * Give a turnstile to each thread. The last thread gets 645 * this turnstile. 646 */ 647 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) { 648 if (LIST_EMPTY(&ts->ts_free)) { 649 MPASS(TAILQ_NEXT(td, td_lockq) == NULL); 650 ts1 = ts; 651 #ifdef TURNSTILE_PROFILING 652 tc->tc_depth--; 653 #endif 654 } else 655 ts1 = LIST_FIRST(&ts->ts_free); 656 MPASS(ts1 != NULL); 657 LIST_REMOVE(ts1, ts_hash); 658 td->td_turnstile = ts1; 659 } 660 } 661 662 /* 663 * Wakeup all threads on the pending list and adjust the priority of the 664 * current thread appropriately. This must be called with the turnstile 665 * chain locked. 666 */ 667 void 668 turnstile_unpend(struct turnstile *ts) 669 { 670 TAILQ_HEAD( ,thread) pending_threads; 671 struct turnstile_chain *tc; 672 struct thread *td; 673 int cp, pri; 674 675 MPASS(ts != NULL); 676 MPASS(ts->ts_owner == curthread); 677 tc = TC_LOOKUP(ts->ts_lockobj); 678 mtx_assert(&tc->tc_lock, MA_OWNED); 679 MPASS(!TAILQ_EMPTY(&ts->ts_pending)); 680 681 /* 682 * Move the list of pending threads out of the turnstile and 683 * into a local variable. 684 */ 685 TAILQ_INIT(&pending_threads); 686 TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq); 687 #ifdef INVARIANTS 688 if (TAILQ_EMPTY(&ts->ts_blocked)) 689 ts->ts_lockobj = NULL; 690 #endif 691 692 /* 693 * Remove the turnstile from this thread's list of contested locks 694 * since this thread doesn't own it anymore. New threads will 695 * not be blocking on the turnstile until it is claimed by a new 696 * owner. 697 */ 698 mtx_lock_spin(&td_contested_lock); 699 ts->ts_owner = NULL; 700 LIST_REMOVE(ts, ts_link); 701 mtx_unlock_spin(&td_contested_lock); 702 mtx_unlock_spin(&tc->tc_lock); 703 704 /* 705 * Adjust the priority of curthread based on other contested 706 * locks it owns. Don't lower the priority below the base 707 * priority however. 708 */ 709 td = curthread; 710 pri = PRI_MAX; 711 mtx_lock_spin(&sched_lock); 712 mtx_lock_spin(&td_contested_lock); 713 LIST_FOREACH(ts, &td->td_contested, ts_link) { 714 cp = TAILQ_FIRST(&ts->ts_blocked)->td_priority; 715 if (cp < pri) 716 pri = cp; 717 } 718 mtx_unlock_spin(&td_contested_lock); 719 if (pri > td->td_base_pri) 720 pri = td->td_base_pri; 721 td->td_priority = pri; 722 723 /* 724 * Wake up all the pending threads. If a thread is not blocked 725 * on a lock, then it is currently executing on another CPU in 726 * turnstile_wait() or sitting on a run queue waiting to resume 727 * in turnstile_wait(). Set a flag to force it to try to acquire 728 * the lock again instead of blocking. 729 */ 730 while (!TAILQ_EMPTY(&pending_threads)) { 731 td = TAILQ_FIRST(&pending_threads); 732 TAILQ_REMOVE(&pending_threads, td, td_lockq); 733 MPASS(td->td_proc->p_magic == P_MAGIC); 734 if (TD_ON_LOCK(td)) { 735 td->td_blocked = NULL; 736 td->td_lockname = NULL; 737 TD_CLR_LOCK(td); 738 MPASS(TD_CAN_RUN(td)); 739 setrunqueue(td); 740 } else { 741 td->td_flags |= TDF_TSNOBLOCK; 742 MPASS(TD_IS_RUNNING(td) || TD_ON_RUNQ(td)); 743 } 744 } 745 mtx_unlock_spin(&sched_lock); 746 } 747 748 /* 749 * Return the first thread in a turnstile. 750 */ 751 struct thread * 752 turnstile_head(struct turnstile *ts) 753 { 754 #ifdef INVARIANTS 755 struct turnstile_chain *tc; 756 757 MPASS(ts != NULL); 758 tc = TC_LOOKUP(ts->ts_lockobj); 759 mtx_assert(&tc->tc_lock, MA_OWNED); 760 #endif 761 return (TAILQ_FIRST(&ts->ts_blocked)); 762 } 763 764 /* 765 * Returns true if a turnstile is empty. 766 */ 767 int 768 turnstile_empty(struct turnstile *ts) 769 { 770 #ifdef INVARIANTS 771 struct turnstile_chain *tc; 772 773 MPASS(ts != NULL); 774 tc = TC_LOOKUP(ts->ts_lockobj); 775 mtx_assert(&tc->tc_lock, MA_OWNED); 776 #endif 777 return (TAILQ_EMPTY(&ts->ts_blocked)); 778 } 779