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/sched.h> 74 #include <sys/sysctl.h> 75 #include <sys/turnstile.h> 76 77 /* 78 * Constants for the hash table of turnstile chains. TC_SHIFT is a magic 79 * number chosen because the sleep queue's use the same value for the 80 * shift. Basically, we ignore the lower 8 bits of the address. 81 * TC_TABLESIZE must be a power of two for TC_MASK to work properly. 82 */ 83 #define TC_TABLESIZE 128 /* Must be power of 2. */ 84 #define TC_MASK (TC_TABLESIZE - 1) 85 #define TC_SHIFT 8 86 #define TC_HASH(lock) (((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK) 87 #define TC_LOOKUP(lock) &turnstile_chains[TC_HASH(lock)] 88 89 /* 90 * There are three different lists of turnstiles as follows. The list 91 * connected by ts_link entries is a per-thread list of all the turnstiles 92 * attached to locks that we own. This is used to fixup our priority when 93 * a lock is released. The other two lists use the ts_hash entries. The 94 * first of these two is the turnstile chain list that a turnstile is on 95 * when it is attached to a lock. The second list to use ts_hash is the 96 * free list hung off of a turnstile that is attached to a lock. 97 * 98 * Each turnstile contains two lists of threads. The ts_blocked list is 99 * a linked list of threads blocked on the turnstile's lock. The 100 * ts_pending list is a linked list of threads previously awakened by 101 * turnstile_signal() or turnstile_wait() that are waiting to be put on 102 * the run queue. 103 * 104 * Locking key: 105 * c - turnstile chain lock 106 * q - td_contested lock 107 */ 108 struct turnstile { 109 TAILQ_HEAD(, thread) ts_blocked; /* (c + q) Blocked threads. */ 110 TAILQ_HEAD(, thread) ts_pending; /* (c) Pending threads. */ 111 LIST_ENTRY(turnstile) ts_hash; /* (c) Chain and free list. */ 112 LIST_ENTRY(turnstile) ts_link; /* (q) Contested locks. */ 113 LIST_HEAD(, turnstile) ts_free; /* (c) Free turnstiles. */ 114 struct lock_object *ts_lockobj; /* (c) Lock we reference. */ 115 struct thread *ts_owner; /* (c + q) Who owns the lock. */ 116 }; 117 118 struct turnstile_chain { 119 LIST_HEAD(, turnstile) tc_turnstiles; /* List of turnstiles. */ 120 struct mtx tc_lock; /* Spin lock for this chain. */ 121 #ifdef TURNSTILE_PROFILING 122 u_int tc_depth; /* Length of tc_queues. */ 123 u_int tc_max_depth; /* Max length of tc_queues. */ 124 #endif 125 }; 126 127 #ifdef TURNSTILE_PROFILING 128 u_int turnstile_max_depth; 129 SYSCTL_NODE(_debug, OID_AUTO, turnstile, CTLFLAG_RD, 0, "turnstile profiling"); 130 SYSCTL_NODE(_debug_turnstile, OID_AUTO, chains, CTLFLAG_RD, 0, 131 "turnstile chain stats"); 132 SYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD, 133 &turnstile_max_depth, 0, "maxmimum depth achieved of a single chain"); 134 #endif 135 static struct mtx td_contested_lock; 136 static struct turnstile_chain turnstile_chains[TC_TABLESIZE]; 137 138 static MALLOC_DEFINE(M_TURNSTILE, "turnstiles", "turnstiles"); 139 140 /* 141 * Prototypes for non-exported routines. 142 */ 143 static void init_turnstile0(void *dummy); 144 #ifdef TURNSTILE_PROFILING 145 static void init_turnstile_profiling(void *arg); 146 #endif 147 static void propagate_priority(struct thread *td); 148 static int turnstile_adjust_thread(struct turnstile *ts, 149 struct thread *td); 150 static void turnstile_setowner(struct turnstile *ts, struct thread *owner); 151 152 /* 153 * Walks the chain of turnstiles and their owners to propagate the priority 154 * of the thread being blocked to all the threads holding locks that have to 155 * release their locks before this thread can run again. 156 */ 157 static void 158 propagate_priority(struct thread *td) 159 { 160 struct turnstile_chain *tc; 161 struct turnstile *ts; 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 (tid %d) owns a non-sleepable lock", 191 td->td_tid)); 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 * Bump this thread's priority. 202 */ 203 sched_lend_prio(td, pri); 204 205 /* 206 * If lock holder is actually running or on the run queue 207 * then we are done. 208 */ 209 if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) { 210 MPASS(td->td_blocked == NULL); 211 return; 212 } 213 214 #ifndef SMP 215 /* 216 * For UP, we check to see if td is curthread (this shouldn't 217 * ever happen however as it would mean we are in a deadlock.) 218 */ 219 KASSERT(td != curthread, ("Deadlock detected")); 220 #endif 221 222 /* 223 * If we aren't blocked on a lock, we should be. 224 */ 225 KASSERT(TD_ON_LOCK(td), ( 226 "thread %d(%s):%d holds %s but isn't blocked on a lock\n", 227 td->td_tid, td->td_proc->p_comm, td->td_state, 228 ts->ts_lockobj->lo_name)); 229 230 /* 231 * Pick up the lock that td is blocked on. 232 */ 233 ts = td->td_blocked; 234 MPASS(ts != NULL); 235 tc = TC_LOOKUP(ts->ts_lockobj); 236 mtx_lock_spin(&tc->tc_lock); 237 238 /* Resort td on the list if needed. */ 239 if (!turnstile_adjust_thread(ts, td)) { 240 mtx_unlock_spin(&tc->tc_lock); 241 return; 242 } 243 mtx_unlock_spin(&tc->tc_lock); 244 } 245 } 246 247 /* 248 * Adjust the thread's position on a turnstile after its priority has been 249 * changed. 250 */ 251 static int 252 turnstile_adjust_thread(struct turnstile *ts, struct thread *td) 253 { 254 struct turnstile_chain *tc; 255 struct thread *td1, *td2; 256 257 mtx_assert(&sched_lock, MA_OWNED); 258 MPASS(TD_ON_LOCK(td)); 259 260 /* 261 * This thread may not be blocked on this turnstile anymore 262 * but instead might already be woken up on another CPU 263 * that is waiting on sched_lock in turnstile_unpend() to 264 * finish waking this thread up. We can detect this case 265 * by checking to see if this thread has been given a 266 * turnstile by either turnstile_signal() or 267 * turnstile_broadcast(). In this case, treat the thread as 268 * if it was already running. 269 */ 270 if (td->td_turnstile != NULL) 271 return (0); 272 273 /* 274 * Check if the thread needs to be moved on the blocked chain. 275 * It needs to be moved if either its priority is lower than 276 * the previous thread or higher than the next thread. 277 */ 278 tc = TC_LOOKUP(ts->ts_lockobj); 279 mtx_assert(&tc->tc_lock, MA_OWNED); 280 td1 = TAILQ_PREV(td, threadqueue, td_lockq); 281 td2 = TAILQ_NEXT(td, td_lockq); 282 if ((td1 != NULL && td->td_priority < td1->td_priority) || 283 (td2 != NULL && td->td_priority > td2->td_priority)) { 284 285 /* 286 * Remove thread from blocked chain and determine where 287 * it should be moved to. 288 */ 289 mtx_lock_spin(&td_contested_lock); 290 TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq); 291 TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq) { 292 MPASS(td1->td_proc->p_magic == P_MAGIC); 293 if (td1->td_priority > td->td_priority) 294 break; 295 } 296 297 if (td1 == NULL) 298 TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq); 299 else 300 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 301 mtx_unlock_spin(&td_contested_lock); 302 if (td1 == NULL) 303 CTR3(KTR_LOCK, 304 "turnstile_adjust_thread: td %d put at tail on [%p] %s", 305 td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name); 306 else 307 CTR4(KTR_LOCK, 308 "turnstile_adjust_thread: td %d moved before %d on [%p] %s", 309 td->td_tid, td1->td_tid, ts->ts_lockobj, 310 ts->ts_lockobj->lo_name); 311 } 312 return (1); 313 } 314 315 /* 316 * Early initialization of turnstiles. This is not done via a SYSINIT() 317 * since this needs to be initialized very early when mutexes are first 318 * initialized. 319 */ 320 void 321 init_turnstiles(void) 322 { 323 int i; 324 325 for (i = 0; i < TC_TABLESIZE; i++) { 326 LIST_INIT(&turnstile_chains[i].tc_turnstiles); 327 mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain", 328 NULL, MTX_SPIN); 329 } 330 mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN); 331 thread0.td_turnstile = NULL; 332 } 333 334 #ifdef TURNSTILE_PROFILING 335 static void 336 init_turnstile_profiling(void *arg) 337 { 338 struct sysctl_oid *chain_oid; 339 char chain_name[10]; 340 int i; 341 342 for (i = 0; i < TC_TABLESIZE; i++) { 343 snprintf(chain_name, sizeof(chain_name), "%d", i); 344 chain_oid = SYSCTL_ADD_NODE(NULL, 345 SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO, 346 chain_name, CTLFLAG_RD, NULL, "turnstile chain stats"); 347 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 348 "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0, 349 NULL); 350 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 351 "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth, 352 0, NULL); 353 } 354 } 355 SYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY, 356 init_turnstile_profiling, NULL); 357 #endif 358 359 static void 360 init_turnstile0(void *dummy) 361 { 362 363 thread0.td_turnstile = turnstile_alloc(); 364 } 365 SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL); 366 367 /* 368 * Update a thread on the turnstile list after it's priority has been changed. 369 * The old priority is passed in as an argument. 370 */ 371 void 372 turnstile_adjust(struct thread *td, u_char oldpri) 373 { 374 struct turnstile_chain *tc; 375 struct turnstile *ts; 376 377 mtx_assert(&sched_lock, MA_OWNED); 378 MPASS(TD_ON_LOCK(td)); 379 380 /* 381 * Pick up the lock that td is blocked on. 382 */ 383 ts = td->td_blocked; 384 MPASS(ts != NULL); 385 tc = TC_LOOKUP(ts->ts_lockobj); 386 mtx_lock_spin(&tc->tc_lock); 387 388 /* Resort the turnstile on the list. */ 389 if (!turnstile_adjust_thread(ts, td)) { 390 mtx_unlock_spin(&tc->tc_lock); 391 return; 392 } 393 394 /* 395 * If our priority was lowered and we are at the head of the 396 * turnstile, then propagate our new priority up the chain. 397 * Note that we currently don't try to revoke lent priorities 398 * when our priority goes up. 399 */ 400 if (td == TAILQ_FIRST(&ts->ts_blocked) && td->td_priority < oldpri) { 401 mtx_unlock_spin(&tc->tc_lock); 402 propagate_priority(td); 403 } else 404 mtx_unlock_spin(&tc->tc_lock); 405 } 406 407 /* 408 * Set the owner of the lock this turnstile is attached to. 409 */ 410 static void 411 turnstile_setowner(struct turnstile *ts, struct thread *owner) 412 { 413 414 mtx_assert(&td_contested_lock, MA_OWNED); 415 MPASS(owner->td_proc->p_magic == P_MAGIC); 416 MPASS(ts->ts_owner == NULL); 417 ts->ts_owner = owner; 418 LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link); 419 } 420 421 /* 422 * Malloc a turnstile for a new thread, initialize it and return it. 423 */ 424 struct turnstile * 425 turnstile_alloc(void) 426 { 427 struct turnstile *ts; 428 429 ts = malloc(sizeof(struct turnstile), M_TURNSTILE, M_WAITOK | M_ZERO); 430 TAILQ_INIT(&ts->ts_blocked); 431 TAILQ_INIT(&ts->ts_pending); 432 LIST_INIT(&ts->ts_free); 433 return (ts); 434 } 435 436 /* 437 * Free a turnstile when a thread is destroyed. 438 */ 439 void 440 turnstile_free(struct turnstile *ts) 441 { 442 443 MPASS(ts != NULL); 444 MPASS(TAILQ_EMPTY(&ts->ts_blocked)); 445 MPASS(TAILQ_EMPTY(&ts->ts_pending)); 446 free(ts, M_TURNSTILE); 447 } 448 449 /* 450 * Lock the turnstile chain associated with the specified lock. 451 */ 452 void 453 turnstile_lock(struct lock_object *lock) 454 { 455 struct turnstile_chain *tc; 456 457 tc = TC_LOOKUP(lock); 458 mtx_lock_spin(&tc->tc_lock); 459 } 460 461 /* 462 * Look up the turnstile for a lock in the hash table locking the associated 463 * turnstile chain along the way. If no turnstile is found in the hash 464 * table, NULL is returned. 465 */ 466 struct turnstile * 467 turnstile_lookup(struct lock_object *lock) 468 { 469 struct turnstile_chain *tc; 470 struct turnstile *ts; 471 472 tc = TC_LOOKUP(lock); 473 mtx_assert(&tc->tc_lock, MA_OWNED); 474 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 475 if (ts->ts_lockobj == lock) 476 return (ts); 477 return (NULL); 478 } 479 480 /* 481 * Unlock the turnstile chain associated with a given lock. 482 */ 483 void 484 turnstile_release(struct lock_object *lock) 485 { 486 struct turnstile_chain *tc; 487 488 tc = TC_LOOKUP(lock); 489 mtx_unlock_spin(&tc->tc_lock); 490 } 491 492 /* 493 * Take ownership of a turnstile and adjust the priority of the new 494 * owner appropriately. 495 */ 496 void 497 turnstile_claim(struct lock_object *lock) 498 { 499 struct turnstile_chain *tc; 500 struct turnstile *ts; 501 struct thread *td, *owner; 502 503 tc = TC_LOOKUP(lock); 504 mtx_assert(&tc->tc_lock, MA_OWNED); 505 ts = turnstile_lookup(lock); 506 MPASS(ts != NULL); 507 508 owner = curthread; 509 mtx_lock_spin(&td_contested_lock); 510 turnstile_setowner(ts, owner); 511 mtx_unlock_spin(&td_contested_lock); 512 513 td = TAILQ_FIRST(&ts->ts_blocked); 514 MPASS(td != NULL); 515 MPASS(td->td_proc->p_magic == P_MAGIC); 516 mtx_unlock_spin(&tc->tc_lock); 517 518 /* 519 * Update the priority of the new owner if needed. 520 */ 521 mtx_lock_spin(&sched_lock); 522 if (td->td_priority < owner->td_priority) 523 sched_lend_prio(owner, td->td_priority); 524 mtx_unlock_spin(&sched_lock); 525 } 526 527 /* 528 * Block the current thread on the turnstile assicated with 'lock'. This 529 * function will context switch and not return until this thread has been 530 * woken back up. This function must be called with the appropriate 531 * turnstile chain locked and will return with it unlocked. 532 */ 533 void 534 turnstile_wait(struct lock_object *lock, struct thread *owner) 535 { 536 struct turnstile_chain *tc; 537 struct turnstile *ts; 538 struct thread *td, *td1; 539 540 td = curthread; 541 tc = TC_LOOKUP(lock); 542 mtx_assert(&tc->tc_lock, MA_OWNED); 543 MPASS(td->td_turnstile != NULL); 544 MPASS(owner != NULL); 545 MPASS(owner->td_proc->p_magic == P_MAGIC); 546 547 /* Look up the turnstile associated with the lock 'lock'. */ 548 ts = turnstile_lookup(lock); 549 550 /* 551 * If the lock does not already have a turnstile, use this thread's 552 * turnstile. Otherwise insert the current thread into the 553 * turnstile already in use by this lock. 554 */ 555 if (ts == NULL) { 556 #ifdef TURNSTILE_PROFILING 557 tc->tc_depth++; 558 if (tc->tc_depth > tc->tc_max_depth) { 559 tc->tc_max_depth = tc->tc_depth; 560 if (tc->tc_max_depth > turnstile_max_depth) 561 turnstile_max_depth = tc->tc_max_depth; 562 } 563 #endif 564 ts = td->td_turnstile; 565 LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash); 566 KASSERT(TAILQ_EMPTY(&ts->ts_pending), 567 ("thread's turnstile has pending threads")); 568 KASSERT(TAILQ_EMPTY(&ts->ts_blocked), 569 ("thread's turnstile has a non-empty queue")); 570 KASSERT(LIST_EMPTY(&ts->ts_free), 571 ("thread's turnstile has a non-empty free list")); 572 KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer")); 573 ts->ts_lockobj = lock; 574 mtx_lock_spin(&td_contested_lock); 575 TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq); 576 turnstile_setowner(ts, owner); 577 mtx_unlock_spin(&td_contested_lock); 578 } else { 579 TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq) 580 if (td1->td_priority > td->td_priority) 581 break; 582 mtx_lock_spin(&td_contested_lock); 583 if (td1 != NULL) 584 TAILQ_INSERT_BEFORE(td1, td, td_lockq); 585 else 586 TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq); 587 mtx_unlock_spin(&td_contested_lock); 588 MPASS(td->td_turnstile != NULL); 589 LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash); 590 MPASS(owner == ts->ts_owner); 591 } 592 td->td_turnstile = NULL; 593 mtx_unlock_spin(&tc->tc_lock); 594 595 mtx_lock_spin(&sched_lock); 596 /* 597 * Handle race condition where a thread on another CPU that owns 598 * lock 'lock' could have woken us in between us dropping the 599 * turnstile chain lock and acquiring the sched_lock. 600 */ 601 if (td->td_flags & TDF_TSNOBLOCK) { 602 td->td_flags &= ~TDF_TSNOBLOCK; 603 mtx_unlock_spin(&sched_lock); 604 return; 605 } 606 607 #ifdef notyet 608 /* 609 * If we're borrowing an interrupted thread's VM context, we 610 * must clean up before going to sleep. 611 */ 612 if (td->td_ithd != NULL) { 613 struct ithd *it = td->td_ithd; 614 615 if (it->it_interrupted) { 616 if (LOCK_LOG_TEST(lock, 0)) 617 CTR3(KTR_LOCK, "%s: %p interrupted %p", 618 __func__, it, it->it_interrupted); 619 intr_thd_fixup(it); 620 } 621 } 622 #endif 623 624 /* Save who we are blocked on and switch. */ 625 td->td_blocked = ts; 626 td->td_lockname = lock->lo_name; 627 TD_SET_LOCK(td); 628 propagate_priority(td); 629 630 if (LOCK_LOG_TEST(lock, 0)) 631 CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__, 632 td->td_tid, lock, lock->lo_name); 633 634 mi_switch(SW_VOL, NULL); 635 636 if (LOCK_LOG_TEST(lock, 0)) 637 CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s", 638 __func__, td->td_tid, lock, lock->lo_name); 639 640 mtx_unlock_spin(&sched_lock); 641 } 642 643 /* 644 * Pick the highest priority thread on this turnstile and put it on the 645 * pending list. This must be called with the turnstile chain locked. 646 */ 647 int 648 turnstile_signal(struct turnstile *ts) 649 { 650 struct turnstile_chain *tc; 651 struct thread *td; 652 int empty; 653 654 MPASS(ts != NULL); 655 MPASS(curthread->td_proc->p_magic == P_MAGIC); 656 MPASS(ts->ts_owner == curthread); 657 tc = TC_LOOKUP(ts->ts_lockobj); 658 mtx_assert(&tc->tc_lock, MA_OWNED); 659 660 /* 661 * Pick the highest priority thread blocked on this lock and 662 * move it to the pending list. 663 */ 664 td = TAILQ_FIRST(&ts->ts_blocked); 665 MPASS(td->td_proc->p_magic == P_MAGIC); 666 mtx_lock_spin(&td_contested_lock); 667 TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq); 668 mtx_unlock_spin(&td_contested_lock); 669 TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq); 670 671 /* 672 * If the turnstile is now empty, remove it from its chain and 673 * give it to the about-to-be-woken thread. Otherwise take a 674 * turnstile from the free list and give it to the thread. 675 */ 676 empty = TAILQ_EMPTY(&ts->ts_blocked); 677 if (empty) { 678 MPASS(LIST_EMPTY(&ts->ts_free)); 679 #ifdef TURNSTILE_PROFILING 680 tc->tc_depth--; 681 #endif 682 } else 683 ts = LIST_FIRST(&ts->ts_free); 684 MPASS(ts != NULL); 685 LIST_REMOVE(ts, ts_hash); 686 td->td_turnstile = ts; 687 688 return (empty); 689 } 690 691 /* 692 * Put all blocked threads on the pending list. This must be called with 693 * the turnstile chain locked. 694 */ 695 void 696 turnstile_broadcast(struct turnstile *ts) 697 { 698 struct turnstile_chain *tc; 699 struct turnstile *ts1; 700 struct thread *td; 701 702 MPASS(ts != NULL); 703 MPASS(curthread->td_proc->p_magic == P_MAGIC); 704 MPASS(ts->ts_owner == curthread); 705 tc = TC_LOOKUP(ts->ts_lockobj); 706 mtx_assert(&tc->tc_lock, MA_OWNED); 707 708 /* 709 * Transfer the blocked list to the pending list. 710 */ 711 mtx_lock_spin(&td_contested_lock); 712 TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked, td_lockq); 713 mtx_unlock_spin(&td_contested_lock); 714 715 /* 716 * Give a turnstile to each thread. The last thread gets 717 * this turnstile. 718 */ 719 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) { 720 if (LIST_EMPTY(&ts->ts_free)) { 721 MPASS(TAILQ_NEXT(td, td_lockq) == NULL); 722 ts1 = ts; 723 #ifdef TURNSTILE_PROFILING 724 tc->tc_depth--; 725 #endif 726 } else 727 ts1 = LIST_FIRST(&ts->ts_free); 728 MPASS(ts1 != NULL); 729 LIST_REMOVE(ts1, ts_hash); 730 td->td_turnstile = ts1; 731 } 732 } 733 734 /* 735 * Wakeup all threads on the pending list and adjust the priority of the 736 * current thread appropriately. This must be called with the turnstile 737 * chain locked. 738 */ 739 void 740 turnstile_unpend(struct turnstile *ts) 741 { 742 TAILQ_HEAD( ,thread) pending_threads; 743 struct turnstile_chain *tc; 744 struct thread *td; 745 u_char cp, pri; 746 747 MPASS(ts != NULL); 748 MPASS(ts->ts_owner == curthread); 749 tc = TC_LOOKUP(ts->ts_lockobj); 750 mtx_assert(&tc->tc_lock, MA_OWNED); 751 MPASS(!TAILQ_EMPTY(&ts->ts_pending)); 752 753 /* 754 * Move the list of pending threads out of the turnstile and 755 * into a local variable. 756 */ 757 TAILQ_INIT(&pending_threads); 758 TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq); 759 #ifdef INVARIANTS 760 if (TAILQ_EMPTY(&ts->ts_blocked)) 761 ts->ts_lockobj = NULL; 762 #endif 763 764 /* 765 * Remove the turnstile from this thread's list of contested locks 766 * since this thread doesn't own it anymore. New threads will 767 * not be blocking on the turnstile until it is claimed by a new 768 * owner. 769 */ 770 mtx_lock_spin(&td_contested_lock); 771 ts->ts_owner = NULL; 772 LIST_REMOVE(ts, ts_link); 773 mtx_unlock_spin(&td_contested_lock); 774 critical_enter(); 775 mtx_unlock_spin(&tc->tc_lock); 776 777 /* 778 * Adjust the priority of curthread based on other contested 779 * locks it owns. Don't lower the priority below the base 780 * priority however. 781 */ 782 td = curthread; 783 pri = PRI_MAX; 784 mtx_lock_spin(&sched_lock); 785 mtx_lock_spin(&td_contested_lock); 786 LIST_FOREACH(ts, &td->td_contested, ts_link) { 787 cp = TAILQ_FIRST(&ts->ts_blocked)->td_priority; 788 if (cp < pri) 789 pri = cp; 790 } 791 mtx_unlock_spin(&td_contested_lock); 792 sched_unlend_prio(td, pri); 793 794 /* 795 * Wake up all the pending threads. If a thread is not blocked 796 * on a lock, then it is currently executing on another CPU in 797 * turnstile_wait() or sitting on a run queue waiting to resume 798 * in turnstile_wait(). Set a flag to force it to try to acquire 799 * the lock again instead of blocking. 800 */ 801 while (!TAILQ_EMPTY(&pending_threads)) { 802 td = TAILQ_FIRST(&pending_threads); 803 TAILQ_REMOVE(&pending_threads, td, td_lockq); 804 MPASS(td->td_proc->p_magic == P_MAGIC); 805 if (TD_ON_LOCK(td)) { 806 td->td_blocked = NULL; 807 td->td_lockname = NULL; 808 TD_CLR_LOCK(td); 809 MPASS(TD_CAN_RUN(td)); 810 setrunqueue(td, SRQ_BORING); 811 } else { 812 td->td_flags |= TDF_TSNOBLOCK; 813 MPASS(TD_IS_RUNNING(td) || TD_ON_RUNQ(td)); 814 } 815 } 816 critical_exit(); 817 mtx_unlock_spin(&sched_lock); 818 } 819 820 /* 821 * Return the first thread in a turnstile. 822 */ 823 struct thread * 824 turnstile_head(struct turnstile *ts) 825 { 826 #ifdef INVARIANTS 827 struct turnstile_chain *tc; 828 829 MPASS(ts != NULL); 830 tc = TC_LOOKUP(ts->ts_lockobj); 831 mtx_assert(&tc->tc_lock, MA_OWNED); 832 #endif 833 return (TAILQ_FIRST(&ts->ts_blocked)); 834 } 835 836 /* 837 * Returns true if a turnstile is empty. 838 */ 839 int 840 turnstile_empty(struct turnstile *ts) 841 { 842 #ifdef INVARIANTS 843 struct turnstile_chain *tc; 844 845 MPASS(ts != NULL); 846 tc = TC_LOOKUP(ts->ts_lockobj); 847 mtx_assert(&tc->tc_lock, MA_OWNED); 848 #endif 849 return (TAILQ_EMPTY(&ts->ts_blocked)); 850 } 851