1 /* 2 * Copyright (c) 2009 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 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 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 /* 35 * Implement fast persistent locks based on atomic_cmpset_int() with 36 * semantics similar to lockmgr locks but faster and taking up much less 37 * space. Taken from HAMMER's lock implementation. 38 * 39 * These are meant to complement our LWKT tokens. Tokens are only held 40 * while the thread is running. Mutexes can be held across blocking 41 * conditions. 42 * 43 * - Exclusive priority over shared to prevent SMP starvation. 44 * - locks can be aborted (async callback, if any, will be made w/ENOLCK). 45 * - locks can be asynchronous. 46 * - synchronous fast path if no blocking occurs (async callback is not 47 * made in this case). 48 * 49 * Generally speaking any caller-supplied link state must be properly 50 * initialized before use. 51 * 52 * Most of the support is in sys/mutex[2].h. We mostly provide backoff 53 * functions here. 54 */ 55 56 #include <sys/param.h> 57 #include <sys/systm.h> 58 #include <sys/kernel.h> 59 #include <sys/sysctl.h> 60 #include <sys/indefinite.h> 61 #include <sys/thread.h> 62 63 #include <machine/cpufunc.h> 64 65 #include <sys/thread2.h> 66 #include <sys/mutex2.h> 67 #include <sys/indefinite2.h> 68 69 static int mtx_chain_link_ex(mtx_t *mtx, u_int olock); 70 static int mtx_chain_link_sh(mtx_t *mtx, u_int olock); 71 static void mtx_delete_link(mtx_t *mtx, mtx_link_t *link); 72 73 /* 74 * Exclusive-lock a mutex, block until acquired unless link is async. 75 * Recursion is allowed. 76 * 77 * Returns 0 on success, the tsleep() return code on failure, EINPROGRESS 78 * if async. If immediately successful an async exclusive lock will return 0 79 * and not issue the async callback or link the link structure. The caller 80 * must handle this case (typically this is an optimal code path). 81 * 82 * A tsleep() error can only be returned if PCATCH is specified in the flags. 83 */ 84 static __inline int 85 __mtx_lock_ex(mtx_t *mtx, mtx_link_t *link, int flags, int to) 86 { 87 thread_t td; 88 u_int lock; 89 u_int nlock; 90 int error; 91 int isasync; 92 93 for (;;) { 94 lock = mtx->mtx_lock; 95 cpu_ccfence(); 96 97 if (lock == 0) { 98 nlock = MTX_EXCLUSIVE | 1; 99 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) { 100 mtx->mtx_owner = curthread; 101 cpu_sfence(); 102 link->state = MTX_LINK_ACQUIRED; 103 error = 0; 104 break; 105 } 106 continue; 107 } 108 if ((lock & MTX_EXCLUSIVE) && mtx->mtx_owner == curthread) { 109 KKASSERT((lock & MTX_MASK) != MTX_MASK); 110 nlock = lock + 1; 111 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) { 112 cpu_sfence(); 113 link->state = MTX_LINK_ACQUIRED; 114 error = 0; 115 break; 116 } 117 continue; 118 } 119 120 /* 121 * We need MTX_LINKSPIN to manipulate exlink or 122 * shlink. 123 * 124 * We must set MTX_EXWANTED with MTX_LINKSPIN to indicate 125 * pending exclusive requests. It cannot be set as a separate 126 * operation prior to acquiring MTX_LINKSPIN. 127 * 128 * To avoid unnecessary cpu cache traffic we poll 129 * for collisions. It is also possible that EXWANTED 130 * state failing the above test was spurious, so all the 131 * tests must be repeated if we cannot obtain LINKSPIN 132 * with the prior state tests intact (i.e. don't reload 133 * the (lock) variable here, for heaven's sake!). 134 */ 135 if (lock & MTX_LINKSPIN) { 136 cpu_pause(); 137 continue; 138 } 139 td = curthread; 140 nlock = lock | MTX_EXWANTED | MTX_LINKSPIN; 141 crit_enter_raw(td); 142 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock) == 0) { 143 crit_exit_raw(td); 144 continue; 145 } 146 147 /* 148 * Check for early abort. 149 */ 150 if (link->state == MTX_LINK_ABORTED) { 151 if (mtx->mtx_exlink == NULL) { 152 atomic_clear_int(&mtx->mtx_lock, 153 MTX_LINKSPIN | 154 MTX_EXWANTED); 155 } else { 156 atomic_clear_int(&mtx->mtx_lock, 157 MTX_LINKSPIN); 158 } 159 crit_exit_raw(td); 160 link->state = MTX_LINK_IDLE; 161 error = ENOLCK; 162 break; 163 } 164 165 /* 166 * Add our link to the exlink list and release LINKSPIN. 167 */ 168 link->owner = td; 169 link->state = MTX_LINK_LINKED_EX; 170 if (mtx->mtx_exlink) { 171 link->next = mtx->mtx_exlink; 172 link->prev = link->next->prev; 173 link->next->prev = link; 174 link->prev->next = link; 175 } else { 176 link->next = link; 177 link->prev = link; 178 mtx->mtx_exlink = link; 179 } 180 isasync = (link->callback != NULL); 181 atomic_clear_int(&mtx->mtx_lock, MTX_LINKSPIN); 182 crit_exit_raw(td); 183 184 /* 185 * If asynchronous lock request return without 186 * blocking, leave link structure linked. 187 */ 188 if (isasync) { 189 error = EINPROGRESS; 190 break; 191 } 192 193 /* 194 * Wait for lock 195 */ 196 error = mtx_wait_link(mtx, link, flags, to); 197 break; 198 } 199 return (error); 200 } 201 202 int 203 _mtx_lock_ex_link(mtx_t *mtx, mtx_link_t *link, int flags, int to) 204 { 205 return(__mtx_lock_ex(mtx, link, flags, to)); 206 } 207 208 int 209 _mtx_lock_ex(mtx_t *mtx, int flags, int to) 210 { 211 mtx_link_t link; 212 213 mtx_link_init(&link); 214 return(__mtx_lock_ex(mtx, &link, flags, to)); 215 } 216 217 int 218 _mtx_lock_ex_quick(mtx_t *mtx) 219 { 220 mtx_link_t link; 221 222 mtx_link_init(&link); 223 return(__mtx_lock_ex(mtx, &link, 0, 0)); 224 } 225 226 /* 227 * Share-lock a mutex, block until acquired. Recursion is allowed. 228 * 229 * Returns 0 on success, or the tsleep() return code on failure. 230 * An error can only be returned if PCATCH is specified in the flags. 231 * 232 * NOTE: Shared locks get a mass-wakeup so if the tsleep fails we 233 * do not have to chain the wakeup(). 234 */ 235 static __inline int 236 __mtx_lock_sh(mtx_t *mtx, mtx_link_t *link, int flags, int to) 237 { 238 thread_t td; 239 u_int lock; 240 u_int nlock; 241 int error; 242 int isasync; 243 244 for (;;) { 245 lock = mtx->mtx_lock; 246 cpu_ccfence(); 247 248 if (lock == 0) { 249 nlock = 1; 250 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) { 251 error = 0; 252 cpu_sfence(); 253 link->state = MTX_LINK_ACQUIRED; 254 break; 255 } 256 continue; 257 } 258 if ((lock & (MTX_EXCLUSIVE | MTX_EXWANTED)) == 0) { 259 KKASSERT((lock & MTX_MASK) != MTX_MASK); 260 nlock = lock + 1; 261 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) { 262 error = 0; 263 cpu_sfence(); 264 link->state = MTX_LINK_ACQUIRED; 265 break; 266 } 267 continue; 268 } 269 270 /* 271 * We need MTX_LINKSPIN to manipulate exlink or 272 * shlink. 273 * 274 * We must set MTX_SHWANTED with MTX_LINKSPIN to indicate 275 * pending shared requests. It cannot be set as a separate 276 * operation prior to acquiring MTX_LINKSPIN. 277 * 278 * To avoid unnecessary cpu cache traffic we poll 279 * for collisions. It is also possible that EXWANTED 280 * state failing the above test was spurious, so all the 281 * tests must be repeated if we cannot obtain LINKSPIN 282 * with the prior state tests intact (i.e. don't reload 283 * the (lock) variable here, for heaven's sake!). 284 */ 285 if (lock & MTX_LINKSPIN) { 286 cpu_pause(); 287 continue; 288 } 289 td = curthread; 290 nlock = lock | MTX_SHWANTED | MTX_LINKSPIN; 291 crit_enter_raw(td); 292 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock) == 0) { 293 crit_exit_raw(td); 294 continue; 295 } 296 297 /* 298 * Check for early abort. Other shared lock requestors 299 * could have sneaked in before we set LINKSPIN so make 300 * sure we undo the state properly. 301 */ 302 if (link->state == MTX_LINK_ABORTED) { 303 if (mtx->mtx_shlink) { 304 atomic_clear_int(&mtx->mtx_lock, 305 MTX_LINKSPIN); 306 } else { 307 atomic_clear_int(&mtx->mtx_lock, 308 MTX_LINKSPIN | 309 MTX_SHWANTED); 310 } 311 crit_exit_raw(td); 312 link->state = MTX_LINK_IDLE; 313 error = ENOLCK; 314 break; 315 } 316 317 /* 318 * Add our link to the shlink list and release LINKSPIN. 319 */ 320 link->owner = td; 321 link->state = MTX_LINK_LINKED_SH; 322 if (mtx->mtx_shlink) { 323 link->next = mtx->mtx_shlink; 324 link->prev = link->next->prev; 325 link->next->prev = link; 326 link->prev->next = link; 327 } else { 328 link->next = link; 329 link->prev = link; 330 mtx->mtx_shlink = link; 331 } 332 isasync = (link->callback != NULL); 333 atomic_clear_int(&mtx->mtx_lock, MTX_LINKSPIN); 334 crit_exit_raw(td); 335 336 /* 337 * If asynchronous lock request return without 338 * blocking, leave link structure linked. 339 */ 340 if (isasync) { 341 error = EINPROGRESS; 342 break; 343 } 344 345 /* 346 * Wait for lock 347 */ 348 error = mtx_wait_link(mtx, link, flags, to); 349 break; 350 } 351 return (error); 352 } 353 354 int 355 _mtx_lock_sh_link(mtx_t *mtx, mtx_link_t *link, int flags, int to) 356 { 357 return(__mtx_lock_sh(mtx, link, flags, to)); 358 } 359 360 int 361 _mtx_lock_sh(mtx_t *mtx, int flags, int to) 362 { 363 mtx_link_t link; 364 365 mtx_link_init(&link); 366 return(__mtx_lock_sh(mtx, &link, flags, to)); 367 } 368 369 int 370 _mtx_lock_sh_quick(mtx_t *mtx) 371 { 372 mtx_link_t link; 373 374 mtx_link_init(&link); 375 return(__mtx_lock_sh(mtx, &link, 0, 0)); 376 } 377 378 /* 379 * Get an exclusive spinlock the hard way. 380 */ 381 void 382 _mtx_spinlock(mtx_t *mtx) 383 { 384 u_int lock; 385 u_int nlock; 386 int bb = 1; 387 int bo; 388 389 for (;;) { 390 lock = mtx->mtx_lock; 391 if (lock == 0) { 392 nlock = MTX_EXCLUSIVE | 1; 393 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) { 394 mtx->mtx_owner = curthread; 395 break; 396 } 397 } else if ((lock & MTX_EXCLUSIVE) && 398 mtx->mtx_owner == curthread) { 399 KKASSERT((lock & MTX_MASK) != MTX_MASK); 400 nlock = lock + 1; 401 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 402 break; 403 } else { 404 /* MWAIT here */ 405 if (bb < 1000) 406 ++bb; 407 cpu_pause(); 408 for (bo = 0; bo < bb; ++bo) 409 ; 410 } 411 cpu_pause(); 412 } 413 } 414 415 /* 416 * Attempt to acquire a spinlock, if we fail we must undo the 417 * gd->gd_spinlocks/gd->gd_curthead->td_critcount predisposition. 418 * 419 * Returns 0 on success, EAGAIN on failure. 420 */ 421 int 422 _mtx_spinlock_try(mtx_t *mtx) 423 { 424 globaldata_t gd = mycpu; 425 u_int lock; 426 u_int nlock; 427 int res = 0; 428 429 for (;;) { 430 lock = mtx->mtx_lock; 431 if (lock == 0) { 432 nlock = MTX_EXCLUSIVE | 1; 433 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) { 434 mtx->mtx_owner = gd->gd_curthread; 435 break; 436 } 437 } else if ((lock & MTX_EXCLUSIVE) && 438 mtx->mtx_owner == gd->gd_curthread) { 439 KKASSERT((lock & MTX_MASK) != MTX_MASK); 440 nlock = lock + 1; 441 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 442 break; 443 } else { 444 --gd->gd_spinlocks; 445 cpu_ccfence(); 446 crit_exit_raw(gd->gd_curthread); 447 res = EAGAIN; 448 break; 449 } 450 cpu_pause(); 451 } 452 return res; 453 } 454 455 #if 0 456 457 void 458 _mtx_spinlock_sh(mtx_t *mtx) 459 { 460 u_int lock; 461 u_int nlock; 462 int bb = 1; 463 int bo; 464 465 for (;;) { 466 lock = mtx->mtx_lock; 467 if ((lock & MTX_EXCLUSIVE) == 0) { 468 KKASSERT((lock & MTX_MASK) != MTX_MASK); 469 nlock = lock + 1; 470 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 471 break; 472 } else { 473 /* MWAIT here */ 474 if (bb < 1000) 475 ++bb; 476 cpu_pause(); 477 for (bo = 0; bo < bb; ++bo) 478 ; 479 } 480 cpu_pause(); 481 } 482 } 483 484 #endif 485 486 int 487 _mtx_lock_ex_try(mtx_t *mtx) 488 { 489 u_int lock; 490 u_int nlock; 491 int error; 492 493 for (;;) { 494 lock = mtx->mtx_lock; 495 if (lock == 0) { 496 nlock = MTX_EXCLUSIVE | 1; 497 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) { 498 mtx->mtx_owner = curthread; 499 error = 0; 500 break; 501 } 502 } else if ((lock & MTX_EXCLUSIVE) && 503 mtx->mtx_owner == curthread) { 504 KKASSERT((lock & MTX_MASK) != MTX_MASK); 505 nlock = lock + 1; 506 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) { 507 error = 0; 508 break; 509 } 510 } else { 511 error = EAGAIN; 512 break; 513 } 514 cpu_pause(); 515 } 516 return (error); 517 } 518 519 int 520 _mtx_lock_sh_try(mtx_t *mtx) 521 { 522 u_int lock; 523 u_int nlock; 524 int error = 0; 525 526 for (;;) { 527 lock = mtx->mtx_lock; 528 if ((lock & MTX_EXCLUSIVE) == 0) { 529 KKASSERT((lock & MTX_MASK) != MTX_MASK); 530 nlock = lock + 1; 531 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 532 break; 533 } else { 534 error = EAGAIN; 535 break; 536 } 537 cpu_pause(); 538 } 539 return (error); 540 } 541 542 /* 543 * If the lock is held exclusively it must be owned by the caller. If the 544 * lock is already a shared lock this operation is a NOP. A panic will 545 * occur if the lock is not held either shared or exclusive. 546 * 547 * The exclusive count is converted to a shared count. 548 */ 549 void 550 _mtx_downgrade(mtx_t *mtx) 551 { 552 u_int lock; 553 u_int nlock; 554 555 for (;;) { 556 lock = mtx->mtx_lock; 557 cpu_ccfence(); 558 559 /* 560 * NOP if already shared. 561 */ 562 if ((lock & MTX_EXCLUSIVE) == 0) { 563 KKASSERT((lock & MTX_MASK) > 0); 564 break; 565 } 566 567 /* 568 * Transfer count to shared. Any additional pending shared 569 * waiters must be woken up. 570 */ 571 if (lock & MTX_SHWANTED) { 572 if (mtx_chain_link_sh(mtx, lock)) 573 break; 574 /* retry */ 575 } else { 576 nlock = lock & ~MTX_EXCLUSIVE; 577 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 578 break; 579 /* retry */ 580 } 581 cpu_pause(); 582 } 583 } 584 585 /* 586 * Upgrade a shared lock to an exclusive lock. The upgrade will fail if 587 * the shared lock has a count other then 1. Optimize the most likely case 588 * but note that a single cmpset can fail due to WANTED races. 589 * 590 * If the lock is held exclusively it must be owned by the caller and 591 * this function will simply return without doing anything. A panic will 592 * occur if the lock is held exclusively by someone other then the caller. 593 * 594 * Returns 0 on success, EDEADLK on failure. 595 */ 596 int 597 _mtx_upgrade_try(mtx_t *mtx) 598 { 599 u_int lock; 600 u_int nlock; 601 int error = 0; 602 603 for (;;) { 604 lock = mtx->mtx_lock; 605 cpu_ccfence(); 606 607 if ((lock & ~MTX_EXWANTED) == 1) { 608 nlock = lock | MTX_EXCLUSIVE; 609 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) { 610 mtx->mtx_owner = curthread; 611 break; 612 } 613 } else if (lock & MTX_EXCLUSIVE) { 614 KKASSERT(mtx->mtx_owner == curthread); 615 break; 616 } else { 617 error = EDEADLK; 618 break; 619 } 620 cpu_pause(); 621 } 622 return (error); 623 } 624 625 /* 626 * Unlock a lock. The caller must hold the lock either shared or exclusive. 627 * 628 * On the last release we handle any pending chains. 629 */ 630 void 631 _mtx_unlock(mtx_t *mtx) 632 { 633 u_int lock; 634 u_int nlock; 635 636 for (;;) { 637 lock = mtx->mtx_lock; 638 cpu_ccfence(); 639 640 switch(lock) { 641 case MTX_EXCLUSIVE | 1: 642 /* 643 * Last release, exclusive lock. 644 * No exclusive or shared requests pending. 645 */ 646 mtx->mtx_owner = NULL; 647 nlock = 0; 648 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 649 goto done; 650 break; 651 case MTX_EXCLUSIVE | MTX_EXWANTED | 1: 652 case MTX_EXCLUSIVE | MTX_EXWANTED | MTX_SHWANTED | 1: 653 /* 654 * Last release, exclusive lock. 655 * Exclusive requests pending. 656 * Exclusive requests have priority over shared reqs. 657 */ 658 if (mtx_chain_link_ex(mtx, lock)) 659 goto done; 660 break; 661 case MTX_EXCLUSIVE | MTX_SHWANTED | 1: 662 /* 663 * Last release, exclusive lock. 664 * 665 * Shared requests are pending. Transfer our count (1) 666 * to the first shared request, wakeup all shared reqs. 667 */ 668 if (mtx_chain_link_sh(mtx, lock)) 669 goto done; 670 break; 671 case 1: 672 /* 673 * Last release, shared lock. 674 * No exclusive or shared requests pending. 675 */ 676 nlock = 0; 677 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 678 goto done; 679 break; 680 case MTX_EXWANTED | 1: 681 case MTX_EXWANTED | MTX_SHWANTED | 1: 682 /* 683 * Last release, shared lock. 684 * 685 * Exclusive requests are pending. Upgrade this 686 * final shared lock to exclusive and transfer our 687 * count (1) to the next exclusive request. 688 * 689 * Exclusive requests have priority over shared reqs. 690 */ 691 if (mtx_chain_link_ex(mtx, lock)) 692 goto done; 693 break; 694 case MTX_SHWANTED | 1: 695 /* 696 * Last release, shared lock. 697 * Shared requests pending. 698 */ 699 if (mtx_chain_link_sh(mtx, lock)) 700 goto done; 701 break; 702 default: 703 /* 704 * We have to loop if this is the last release but 705 * someone is fiddling with LINKSPIN. 706 */ 707 if ((lock & MTX_MASK) == 1) { 708 KKASSERT(lock & MTX_LINKSPIN); 709 break; 710 } 711 712 /* 713 * Not the last release (shared or exclusive) 714 */ 715 nlock = lock - 1; 716 KKASSERT((nlock & MTX_MASK) != MTX_MASK); 717 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 718 goto done; 719 break; 720 } 721 /* loop try again */ 722 cpu_pause(); 723 } 724 done: 725 ; 726 } 727 728 /* 729 * Chain pending links. Called on the last release of an exclusive or 730 * shared lock when the appropriate WANTED bit is set. mtx_lock old state 731 * is passed in with the count left at 1, which we can inherit, and other 732 * bits which we must adjust in a single atomic operation. 733 * 734 * Return non-zero on success, 0 if caller needs to retry. 735 * 736 * NOTE: It's ok if MTX_EXWANTED is in an indeterminant state while we are 737 * acquiring LINKSPIN as all other cases will also need to acquire 738 * LINKSPIN when handling the EXWANTED case. 739 */ 740 static int 741 mtx_chain_link_ex(mtx_t *mtx, u_int olock) 742 { 743 thread_t td = curthread; 744 mtx_link_t *link; 745 u_int nlock; 746 747 olock &= ~MTX_LINKSPIN; 748 nlock = olock | MTX_LINKSPIN | MTX_EXCLUSIVE; /* upgrade if necc */ 749 crit_enter_raw(td); 750 if (atomic_cmpset_int(&mtx->mtx_lock, olock, nlock)) { 751 link = mtx->mtx_exlink; 752 KKASSERT(link != NULL); 753 if (link->next == link) { 754 mtx->mtx_exlink = NULL; 755 nlock = MTX_LINKSPIN | MTX_EXWANTED; /* to clear */ 756 } else { 757 mtx->mtx_exlink = link->next; 758 link->next->prev = link->prev; 759 link->prev->next = link->next; 760 nlock = MTX_LINKSPIN; /* to clear */ 761 } 762 KKASSERT(link->state == MTX_LINK_LINKED_EX); 763 mtx->mtx_owner = link->owner; 764 cpu_sfence(); 765 766 /* 767 * WARNING! The callback can only be safely 768 * made with LINKSPIN still held 769 * and in a critical section. 770 * 771 * WARNING! The link can go away after the 772 * state is set, or after the 773 * callback. 774 */ 775 if (link->callback) { 776 link->state = MTX_LINK_CALLEDBACK; 777 link->callback(link, link->arg, 0); 778 } else { 779 link->state = MTX_LINK_ACQUIRED; 780 wakeup(link); 781 } 782 atomic_clear_int(&mtx->mtx_lock, nlock); 783 crit_exit_raw(td); 784 return 1; 785 } 786 /* retry */ 787 crit_exit_raw(td); 788 789 return 0; 790 } 791 792 /* 793 * Flush waiting shared locks. The lock's prior state is passed in and must 794 * be adjusted atomically only if it matches and LINKSPIN is not set. 795 * 796 * IMPORTANT! The caller has left one active count on the lock for us to 797 * consume. We will apply this to the first link, but must add 798 * additional counts for any other links. 799 */ 800 static int 801 mtx_chain_link_sh(mtx_t *mtx, u_int olock) 802 { 803 thread_t td = curthread; 804 mtx_link_t *link; 805 u_int addcount; 806 u_int nlock; 807 808 olock &= ~MTX_LINKSPIN; 809 nlock = olock | MTX_LINKSPIN; 810 nlock &= ~MTX_EXCLUSIVE; 811 crit_enter_raw(td); 812 if (atomic_cmpset_int(&mtx->mtx_lock, olock, nlock)) { 813 /* 814 * It should not be possible for SHWANTED to be set without 815 * any links pending. 816 */ 817 KKASSERT(mtx->mtx_shlink != NULL); 818 819 /* 820 * We have to process the count for all shared locks before 821 * we process any of the links. Count the additional shared 822 * locks beyond the first link (which is already accounted 823 * for) and associate the full count with the lock 824 * immediately. 825 */ 826 addcount = 0; 827 for (link = mtx->mtx_shlink->next; link != mtx->mtx_shlink; 828 link = link->next) { 829 ++addcount; 830 } 831 if (addcount > 0) 832 atomic_add_int(&mtx->mtx_lock, addcount); 833 834 /* 835 * We can wakeup all waiting shared locks. 836 */ 837 while ((link = mtx->mtx_shlink) != NULL) { 838 KKASSERT(link->state == MTX_LINK_LINKED_SH); 839 if (link->next == link) { 840 mtx->mtx_shlink = NULL; 841 } else { 842 mtx->mtx_shlink = link->next; 843 link->next->prev = link->prev; 844 link->prev->next = link->next; 845 } 846 link->next = NULL; 847 link->prev = NULL; 848 cpu_sfence(); 849 if (link->callback) { 850 link->state = MTX_LINK_CALLEDBACK; 851 link->callback(link, link->arg, 0); 852 } else { 853 cpu_sfence(); 854 link->state = MTX_LINK_ACQUIRED; 855 wakeup(link); 856 } 857 } 858 atomic_clear_int(&mtx->mtx_lock, MTX_LINKSPIN | 859 MTX_SHWANTED); 860 crit_exit_raw(td); 861 return 1; 862 } 863 /* retry */ 864 crit_exit_raw(td); 865 866 return 0; 867 } 868 869 /* 870 * Delete a link structure after tsleep has failed. This code is not 871 * in the critical path as most exclusive waits are chained. 872 */ 873 static 874 void 875 mtx_delete_link(mtx_t *mtx, mtx_link_t *link) 876 { 877 thread_t td = curthread; 878 u_int lock; 879 u_int nlock; 880 881 /* 882 * Acquire MTX_LINKSPIN. 883 * 884 * Do not use cmpxchg to wait for LINKSPIN to clear as this might 885 * result in too much cpu cache traffic. 886 */ 887 crit_enter_raw(td); 888 for (;;) { 889 lock = mtx->mtx_lock; 890 if (lock & MTX_LINKSPIN) { 891 cpu_pause(); 892 continue; 893 } 894 nlock = lock | MTX_LINKSPIN; 895 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 896 break; 897 cpu_pause(); 898 } 899 900 /* 901 * Delete the link and release LINKSPIN. 902 */ 903 nlock = MTX_LINKSPIN; /* to clear */ 904 905 switch(link->state) { 906 case MTX_LINK_LINKED_EX: 907 if (link->next == link) { 908 mtx->mtx_exlink = NULL; 909 nlock |= MTX_EXWANTED; /* to clear */ 910 } else { 911 mtx->mtx_exlink = link->next; 912 link->next->prev = link->prev; 913 link->prev->next = link->next; 914 } 915 break; 916 case MTX_LINK_LINKED_SH: 917 if (link->next == link) { 918 mtx->mtx_shlink = NULL; 919 nlock |= MTX_SHWANTED; /* to clear */ 920 } else { 921 mtx->mtx_shlink = link->next; 922 link->next->prev = link->prev; 923 link->prev->next = link->next; 924 } 925 break; 926 default: 927 /* no change */ 928 break; 929 } 930 atomic_clear_int(&mtx->mtx_lock, nlock); 931 crit_exit_raw(td); 932 } 933 934 /* 935 * Wait for async lock completion or abort. Returns ENOLCK if an abort 936 * occurred. 937 */ 938 int 939 mtx_wait_link(mtx_t *mtx, mtx_link_t *link, int flags, int to) 940 { 941 indefinite_info_t info; 942 int error; 943 944 indefinite_init(&info, mtx->mtx_ident, 1, 945 ((link->state & MTX_LINK_LINKED_SH) ? 'm' : 'M')); 946 947 /* 948 * Sleep. Handle false wakeups, interruptions, etc. 949 * The link may also have been aborted. The LINKED 950 * bit was set by this cpu so we can test it without 951 * fences. 952 */ 953 error = 0; 954 while (link->state & MTX_LINK_LINKED) { 955 tsleep_interlock(link, 0); 956 cpu_lfence(); 957 if (link->state & MTX_LINK_LINKED) { 958 error = tsleep(link, flags | PINTERLOCKED, 959 mtx->mtx_ident, to); 960 if (error) 961 break; 962 } 963 if ((mtx->mtx_flags & MTXF_NOCOLLSTATS) == 0) 964 indefinite_check(&info); 965 } 966 967 /* 968 * We need at least a lfence (load fence) to ensure our cpu does not 969 * reorder loads (of data outside the lock structure) prior to the 970 * remote cpu's release, since the above test may have run without 971 * any atomic interactions. 972 * 973 * If we do not do this then state updated by the other cpu before 974 * releasing its lock may not be read cleanly by our cpu when this 975 * function returns. Even though the other cpu ordered its stores, 976 * our loads can still be out of order. 977 */ 978 cpu_mfence(); 979 980 /* 981 * We are done, make sure the link structure is unlinked. 982 * It may still be on the list due to e.g. EINTR or 983 * EWOULDBLOCK. 984 * 985 * It is possible for the tsleep to race an ABORT and cause 986 * error to be 0. 987 * 988 * The tsleep() can be woken up for numerous reasons and error 989 * might be zero in situations where we intend to return an error. 990 * 991 * (This is the synchronous case so state cannot be CALLEDBACK) 992 */ 993 switch(link->state) { 994 case MTX_LINK_ACQUIRED: 995 case MTX_LINK_CALLEDBACK: 996 error = 0; 997 break; 998 case MTX_LINK_ABORTED: 999 error = ENOLCK; 1000 break; 1001 case MTX_LINK_LINKED_EX: 1002 case MTX_LINK_LINKED_SH: 1003 mtx_delete_link(mtx, link); 1004 /* fall through */ 1005 default: 1006 if (error == 0) 1007 error = EWOULDBLOCK; 1008 break; 1009 } 1010 1011 /* 1012 * Clear state on status returned. 1013 */ 1014 link->state = MTX_LINK_IDLE; 1015 1016 if ((mtx->mtx_flags & MTXF_NOCOLLSTATS) == 0) 1017 indefinite_done(&info); 1018 1019 return error; 1020 } 1021 1022 /* 1023 * Abort a mutex locking operation, causing mtx_lock_ex_link() to 1024 * return ENOLCK. This may be called at any time after the mtx_link 1025 * is initialized or the status from a previous lock has been 1026 * returned. If called prior to the next (non-try) lock attempt, the 1027 * next lock attempt using this link structure will abort instantly. 1028 * 1029 * Caller must still wait for the operation to complete, either from a 1030 * blocking call that is still in progress or by calling mtx_wait_link(). 1031 * 1032 * If an asynchronous lock request is possibly in-progress, the caller 1033 * should call mtx_wait_link() synchronously. Note that the asynchronous 1034 * lock callback will NOT be called if a successful abort occurred. XXX 1035 */ 1036 void 1037 mtx_abort_link(mtx_t *mtx, mtx_link_t *link) 1038 { 1039 thread_t td = curthread; 1040 u_int lock; 1041 u_int nlock; 1042 1043 /* 1044 * Acquire MTX_LINKSPIN 1045 */ 1046 crit_enter_raw(td); 1047 for (;;) { 1048 lock = mtx->mtx_lock; 1049 if (lock & MTX_LINKSPIN) { 1050 cpu_pause(); 1051 continue; 1052 } 1053 nlock = lock | MTX_LINKSPIN; 1054 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) 1055 break; 1056 cpu_pause(); 1057 } 1058 1059 /* 1060 * Do the abort. 1061 * 1062 * WARNING! Link structure can disappear once link->state is set. 1063 */ 1064 nlock = MTX_LINKSPIN; /* to clear */ 1065 1066 switch(link->state) { 1067 case MTX_LINK_IDLE: 1068 /* 1069 * Link not started yet 1070 */ 1071 link->state = MTX_LINK_ABORTED; 1072 break; 1073 case MTX_LINK_LINKED_EX: 1074 /* 1075 * de-link, mark aborted, and potentially wakeup the thread 1076 * or issue the callback. 1077 */ 1078 if (link->next == link) { 1079 if (mtx->mtx_exlink == link) { 1080 mtx->mtx_exlink = NULL; 1081 nlock |= MTX_EXWANTED; /* to clear */ 1082 } 1083 } else { 1084 if (mtx->mtx_exlink == link) 1085 mtx->mtx_exlink = link->next; 1086 link->next->prev = link->prev; 1087 link->prev->next = link->next; 1088 } 1089 1090 /* 1091 * When aborting the async callback is still made. We must 1092 * not set the link status to ABORTED in the callback case 1093 * since there is nothing else to clear its status if the 1094 * link is reused. 1095 */ 1096 if (link->callback) { 1097 link->state = MTX_LINK_CALLEDBACK; 1098 link->callback(link, link->arg, ENOLCK); 1099 } else { 1100 link->state = MTX_LINK_ABORTED; 1101 wakeup(link); 1102 } 1103 break; 1104 case MTX_LINK_LINKED_SH: 1105 /* 1106 * de-link, mark aborted, and potentially wakeup the thread 1107 * or issue the callback. 1108 */ 1109 if (link->next == link) { 1110 if (mtx->mtx_shlink == link) { 1111 mtx->mtx_shlink = NULL; 1112 nlock |= MTX_SHWANTED; /* to clear */ 1113 } 1114 } else { 1115 if (mtx->mtx_shlink == link) 1116 mtx->mtx_shlink = link->next; 1117 link->next->prev = link->prev; 1118 link->prev->next = link->next; 1119 } 1120 1121 /* 1122 * When aborting the async callback is still made. We must 1123 * not set the link status to ABORTED in the callback case 1124 * since there is nothing else to clear its status if the 1125 * link is reused. 1126 */ 1127 if (link->callback) { 1128 link->state = MTX_LINK_CALLEDBACK; 1129 link->callback(link, link->arg, ENOLCK); 1130 } else { 1131 link->state = MTX_LINK_ABORTED; 1132 wakeup(link); 1133 } 1134 break; 1135 case MTX_LINK_ACQUIRED: 1136 case MTX_LINK_CALLEDBACK: 1137 /* 1138 * Too late, the lock was acquired. Let it complete. 1139 */ 1140 break; 1141 default: 1142 /* 1143 * link already aborted, do nothing. 1144 */ 1145 break; 1146 } 1147 atomic_clear_int(&mtx->mtx_lock, nlock); 1148 crit_exit_raw(td); 1149 } 1150