sys/kern/kern_lock.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
   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(s), this list of conditions and the following disclaimer as
  12  *    the first lines of this file unmodified other than the possible
  13  *    addition of one or more copyright notices.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice(s), this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  19  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  20  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  21  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  22  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  25  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  28  * DAMAGE.
  29  */
  30
  31 #include "opt_ddb.h"
  32 #include "opt_hwpmc_hooks.h"
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include <sys/param.h>
  38 #include <sys/kdb.h>
  39 #include <sys/ktr.h>
  40 #include <sys/lock.h>
  41 #include <sys/lock_profile.h>
  42 #include <sys/lockmgr.h>
  43 #include <sys/mutex.h>
  44 #include <sys/proc.h>
  45 #include <sys/sleepqueue.h>
  46 #ifdef DEBUG_LOCKS
  47 #include <sys/stack.h>
  48 #endif
  49 #include <sys/sysctl.h>
  50 #include <sys/systm.h>
  51
  52 #include <machine/cpu.h>
  53
  54 #ifdef DDB
  55 #include <ddb/ddb.h>
  56 #endif
  57
  58 #ifdef HWPMC_HOOKS
  59 #include <sys/pmckern.h>
  60 PMC_SOFT_DECLARE( , , lock, failed);
  61 #endif
  62
  63 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
  64     (LK_ADAPTIVE | LK_NOSHARE));
  65 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
  66     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
  67
  68 #define SQ_EXCLUSIVE_QUEUE      0
  69 #define SQ_SHARED_QUEUE         1
  70
  71 #ifndef INVARIANTS
  72 #define _lockmgr_assert(lk, what, file, line)
  73 #endif
  74
  75 #define TD_SLOCKS_INC(td)       ((td)->td_lk_slocks++)
  76 #define TD_SLOCKS_DEC(td)       ((td)->td_lk_slocks--)
  77
  78 #ifndef DEBUG_LOCKS
  79 #define STACK_PRINT(lk)
  80 #define STACK_SAVE(lk)
  81 #define STACK_ZERO(lk)
  82 #else
  83 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
  84 #define STACK_SAVE(lk)  stack_save(&(lk)->lk_stack)
  85 #define STACK_ZERO(lk)  stack_zero(&(lk)->lk_stack)
  86 #endif
  87
  88 #define LOCK_LOG2(lk, string, arg1, arg2)                               \
  89         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  90                 CTR2(KTR_LOCK, (string), (arg1), (arg2))
  91 #define LOCK_LOG3(lk, string, arg1, arg2, arg3)                         \
  92         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  93                 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
  94
  95 #define GIANT_DECLARE                                                   \
  96         int _i = 0;                                                     \
  97         WITNESS_SAVE_DECL(Giant)
  98 #define GIANT_RESTORE() do {                                            \
  99         if (_i > 0) {                                                   \
 100                 while (_i--)                                            \
 101                         mtx_lock(&Giant);                               \
 102                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 103         }                                                               \
 104 } while (0)
 105 #define GIANT_SAVE() do {                                               \
 106         if (mtx_owned(&Giant)) {                                        \
 107                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
 108                 while (mtx_owned(&Giant)) {                             \
 109                         _i++;                                           \
 110                         mtx_unlock(&Giant);                             \
 111                 }                                                       \
 112         }                                                               \
 113 } while (0)
 114
 115 #define LK_CAN_SHARE(x, flags)                                          \
 116         (((x) & LK_SHARE) &&                                            \
 117         (((x) & (LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 0 || \
 118         (curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||  \
 119         (curthread->td_pflags & TDP_DEADLKTREAT)))
 120 #define LK_TRYOP(x)                                                     \
 121         ((x) & LK_NOWAIT)
 122
 123 #define LK_CAN_WITNESS(x)                                               \
 124         (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
 125 #define LK_TRYWIT(x)                                                    \
 126         (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
 127
 128 #define LK_CAN_ADAPT(lk, f)                                             \
 129         (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&             \
 130         ((f) & LK_SLEEPFAIL) == 0)
 131
 132 #define lockmgr_disowned(lk)                                            \
 133         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
 134
 135 #define lockmgr_xlocked(lk)                                             \
 136         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
 137
 138 static void     assert_lockmgr(const struct lock_object *lock, int how);
 139 #ifdef DDB
 140 static void     db_show_lockmgr(const struct lock_object *lock);
 141 #endif
 142 static void     lock_lockmgr(struct lock_object *lock, uintptr_t how);
 143 #ifdef KDTRACE_HOOKS
 144 static int      owner_lockmgr(const struct lock_object *lock,
 145                     struct thread **owner);
 146 #endif
 147 static uintptr_t unlock_lockmgr(struct lock_object *lock);
 148
 149 struct lock_class lock_class_lockmgr = {
 150         .lc_name = "lockmgr",
 151         .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
 152         .lc_assert = assert_lockmgr,
 153 #ifdef DDB
 154         .lc_ddb_show = db_show_lockmgr,
 155 #endif
 156         .lc_lock = lock_lockmgr,
 157         .lc_unlock = unlock_lockmgr,
 158 #ifdef KDTRACE_HOOKS
 159         .lc_owner = owner_lockmgr,
 160 #endif
 161 };
 162
 163 static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp,
 164     int flags);
 165 static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t x);
 166
 167 static void
 168 lockmgr_note_shared_acquire(struct lock *lk, int contested,
 169     uint64_t waittime, const char *file, int line, int flags)
 170 {
 171
 172         lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
 173             file, line);
 174         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line);
 175         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line);
 176         TD_LOCKS_INC(curthread);
 177         TD_SLOCKS_INC(curthread);
 178         STACK_SAVE(lk);
 179 }
 180
 181 static void
 182 lockmgr_note_shared_release(struct lock *lk, const char *file, int line)
 183 {
 184
 185         lock_profile_release_lock(&lk->lock_object);
 186         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 187         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 188         TD_LOCKS_DEC(curthread);
 189         TD_SLOCKS_DEC(curthread);
 190 }
 191
 192 static void
 193 lockmgr_note_exclusive_acquire(struct lock *lk, int contested,
 194     uint64_t waittime, const char *file, int line, int flags)
 195 {
 196
 197         lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
 198             file, line);
 199         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
 200         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file,
 201             line);
 202         TD_LOCKS_INC(curthread);
 203         STACK_SAVE(lk);
 204 }
 205
 206 static void
 207 lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line)
 208 {
 209
 210         lock_profile_release_lock(&lk->lock_object);
 211         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file,
 212             line);
 213         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
 214         TD_LOCKS_DEC(curthread);
 215 }
 216
 217 static void
 218 lockmgr_note_exclusive_upgrade(struct lock *lk, const char *file, int line,
 219     int flags)
 220 {
 221
 222         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 223             line);
 224         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 225             LK_TRYWIT(flags), file, line);
 226         TD_SLOCKS_DEC(curthread);
 227 }
 228
 229 static __inline struct thread *
 230 lockmgr_xholder(const struct lock *lk)
 231 {
 232         uintptr_t x;
 233
 234         x = lk->lk_lock;
 235         return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
 236 }
 237
 238 /*
 239  * It assumes sleepq_lock held and returns with this one unheld.
 240  * It also assumes the generic interlock is sane and previously checked.
 241  * If LK_INTERLOCK is specified the interlock is not reacquired after the
 242  * sleep.
 243  */
 244 static __inline int
 245 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
 246     const char *wmesg, int pri, int timo, int queue)
 247 {
 248         GIANT_DECLARE;
 249         struct lock_class *class;
 250         int catch, error;
 251
 252         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 253         catch = pri & PCATCH;
 254         pri &= PRIMASK;
 255         error = 0;
 256
 257         LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
 258             (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
 259
 260         if (flags & LK_INTERLOCK)
 261                 class->lc_unlock(ilk);
 262         if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
 263                 lk->lk_exslpfail++;
 264         GIANT_SAVE();
 265         sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
 266             SLEEPQ_INTERRUPTIBLE : 0), queue);
 267         if ((flags & LK_TIMELOCK) && timo)
 268                 sleepq_set_timeout(&lk->lock_object, timo);
 269
 270         /*
 271          * Decisional switch for real sleeping.
 272          */
 273         if ((flags & LK_TIMELOCK) && timo && catch)
 274                 error = sleepq_timedwait_sig(&lk->lock_object, pri);
 275         else if ((flags & LK_TIMELOCK) && timo)
 276                 error = sleepq_timedwait(&lk->lock_object, pri);
 277         else if (catch)
 278                 error = sleepq_wait_sig(&lk->lock_object, pri);
 279         else
 280                 sleepq_wait(&lk->lock_object, pri);
 281         GIANT_RESTORE();
 282         if ((flags & LK_SLEEPFAIL) && error == 0)
 283                 error = ENOLCK;
 284
 285         return (error);
 286 }
 287
 288 static __inline int
 289 wakeupshlk(struct lock *lk, const char *file, int line)
 290 {
 291         uintptr_t v, x;
 292         u_int realexslp;
 293         int queue, wakeup_swapper;
 294
 295         wakeup_swapper = 0;
 296         for (;;) {
 297                 x = lk->lk_lock;
 298                 if (lockmgr_sunlock_try(lk, x))
 299                         break;
 300
 301                 /*
 302                  * We should have a sharer with waiters, so enter the hard
 303                  * path in order to handle wakeups correctly.
 304                  */
 305                 sleepq_lock(&lk->lock_object);
 306                 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 307                 v = LK_UNLOCKED;
 308
 309                 /*
 310                  * If the lock has exclusive waiters, give them preference in
 311                  * order to avoid deadlock with shared runners up.
 312                  * If interruptible sleeps left the exclusive queue empty
 313                  * avoid a starvation for the threads sleeping on the shared
 314                  * queue by giving them precedence and cleaning up the
 315                  * exclusive waiters bit anyway.
 316                  * Please note that lk_exslpfail count may be lying about
 317                  * the real number of waiters with the LK_SLEEPFAIL flag on
 318                  * because they may be used in conjunction with interruptible
 319                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 320                  * bound, including the edge cases.
 321                  */
 322                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 323                     SQ_EXCLUSIVE_QUEUE);
 324                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 325                         if (lk->lk_exslpfail < realexslp) {
 326                                 lk->lk_exslpfail = 0;
 327                                 queue = SQ_EXCLUSIVE_QUEUE;
 328                                 v |= (x & LK_SHARED_WAITERS);
 329                         } else {
 330                                 lk->lk_exslpfail = 0;
 331                                 LOCK_LOG2(lk,
 332                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 333                                     __func__, lk);
 334                                 LOCK_LOG2(lk,
 335                             "%s: %p waking up threads on the exclusive queue",
 336                                     __func__, lk);
 337                                 wakeup_swapper =
 338                                     sleepq_broadcast(&lk->lock_object,
 339                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 340                                 queue = SQ_SHARED_QUEUE;
 341                         }
 342
 343                 } else {
 344
 345                         /*
 346                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 347                          * and using interruptible sleeps/timeout may have
 348                          * left spourious lk_exslpfail counts on, so clean
 349                          * it up anyway.
 350                          */
 351                         lk->lk_exslpfail = 0;
 352                         queue = SQ_SHARED_QUEUE;
 353                 }
 354
 355                 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
 356                     v)) {
 357                         sleepq_release(&lk->lock_object);
 358                         continue;
 359                 }
 360                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 361                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 362                     "exclusive");
 363                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 364                     0, queue);
 365                 sleepq_release(&lk->lock_object);
 366                 break;
 367         }
 368
 369         lockmgr_note_shared_release(lk, file, line);
 370         return (wakeup_swapper);
 371 }
 372
 373 static void
 374 assert_lockmgr(const struct lock_object *lock, int what)
 375 {
 376
 377         panic("lockmgr locks do not support assertions");
 378 }
 379
 380 static void
 381 lock_lockmgr(struct lock_object *lock, uintptr_t how)
 382 {
 383
 384         panic("lockmgr locks do not support sleep interlocking");
 385 }
 386
 387 static uintptr_t
 388 unlock_lockmgr(struct lock_object *lock)
 389 {
 390
 391         panic("lockmgr locks do not support sleep interlocking");
 392 }
 393
 394 #ifdef KDTRACE_HOOKS
 395 static int
 396 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
 397 {
 398
 399         panic("lockmgr locks do not support owner inquiring");
 400 }
 401 #endif
 402
 403 void
 404 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 405 {
 406         int iflags;
 407
 408         MPASS((flags & ~LK_INIT_MASK) == 0);
 409         ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
 410             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
 411             &lk->lk_lock));
 412
 413         iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 414         if (flags & LK_CANRECURSE)
 415                 iflags |= LO_RECURSABLE;
 416         if ((flags & LK_NODUP) == 0)
 417                 iflags |= LO_DUPOK;
 418         if (flags & LK_NOPROFILE)
 419                 iflags |= LO_NOPROFILE;
 420         if ((flags & LK_NOWITNESS) == 0)
 421                 iflags |= LO_WITNESS;
 422         if (flags & LK_QUIET)
 423                 iflags |= LO_QUIET;
 424         if (flags & LK_IS_VNODE)
 425                 iflags |= LO_IS_VNODE;
 426         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 427
 428         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 429         lk->lk_lock = LK_UNLOCKED;
 430         lk->lk_recurse = 0;
 431         lk->lk_exslpfail = 0;
 432         lk->lk_timo = timo;
 433         lk->lk_pri = pri;
 434         STACK_ZERO(lk);
 435 }
 436
 437 /*
 438  * XXX: Gross hacks to manipulate external lock flags after
 439  * initialization.  Used for certain vnode and buf locks.
 440  */
 441 void
 442 lockallowshare(struct lock *lk)
 443 {
 444
 445         lockmgr_assert(lk, KA_XLOCKED);
 446         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 447 }
 448
 449 void
 450 lockdisableshare(struct lock *lk)
 451 {
 452
 453         lockmgr_assert(lk, KA_XLOCKED);
 454         lk->lock_object.lo_flags |= LK_NOSHARE;
 455 }
 456
 457 void
 458 lockallowrecurse(struct lock *lk)
 459 {
 460
 461         lockmgr_assert(lk, KA_XLOCKED);
 462         lk->lock_object.lo_flags |= LO_RECURSABLE;
 463 }
 464
 465 void
 466 lockdisablerecurse(struct lock *lk)
 467 {
 468
 469         lockmgr_assert(lk, KA_XLOCKED);
 470         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 471 }
 472
 473 void
 474 lockdestroy(struct lock *lk)
 475 {
 476
 477         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 478         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 479         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 480         lock_destroy(&lk->lock_object);
 481 }
 482
 483 static bool __always_inline
 484 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags)
 485 {
 486
 487         /*
 488          * If no other thread has an exclusive lock, or
 489          * no exclusive waiter is present, bump the count of
 490          * sharers.  Since we have to preserve the state of
 491          * waiters, if we fail to acquire the shared lock
 492          * loop back and retry.
 493          */
 494         *xp = lk->lk_lock;
 495         while (LK_CAN_SHARE(*xp, flags)) {
 496                 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
 497                     *xp + LK_ONE_SHARER)) {
 498                         return (true);
 499                 }
 500         }
 501         return (false);
 502 }
 503
 504 static bool __always_inline
 505 lockmgr_sunlock_try(struct lock *lk, uintptr_t x)
 506 {
 507
 508         for (;;) {
 509                 /*
 510                  * If there is more than one shared lock held, just drop one
 511                  * and return.
 512                  */
 513                 if (LK_SHARERS(x) > 1) {
 514                         if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &x,
 515                             x - LK_ONE_SHARER))
 516                                 return (true);
 517                         continue;
 518                 }
 519
 520                 /*
 521                  * If there are not waiters on the exclusive queue, drop the
 522                  * lock quickly.
 523                  */
 524                 if ((x & LK_ALL_WAITERS) == 0) {
 525                         MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
 526                             LK_SHARERS_LOCK(1));
 527                         if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &x,
 528                             LK_UNLOCKED))
 529                                 return (true);
 530                         continue;
 531                 }
 532                 break;
 533         }
 534         return (false);
 535 }
 536
 537 int
 538 lockmgr_lock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk,
 539     const char *file, int line)
 540 {
 541         struct lock_class *class;
 542         uintptr_t x, v, tid;
 543         u_int op;
 544         bool locked;
 545
 546         op = flags & LK_TYPE_MASK;
 547         locked = false;
 548         switch (op) {
 549         case LK_SHARED:
 550                 if (LK_CAN_WITNESS(flags))
 551                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 552                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 553                 if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
 554                         break;
 555                 if (lockmgr_slock_try(lk, &x, flags)) {
 556                         lockmgr_note_shared_acquire(lk, 0, 0,
 557                             file, line, flags);
 558                         locked = true;
 559                 }
 560                 break;
 561         case LK_EXCLUSIVE:
 562                 if (LK_CAN_WITNESS(flags))
 563                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 564                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 565                             ilk : NULL);
 566                 tid = (uintptr_t)curthread;
 567                 if (lk->lk_lock == LK_UNLOCKED &&
 568                     atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
 569                         lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
 570                             flags);
 571                         locked = true;
 572                 }
 573                 break;
 574         case LK_UPGRADE:
 575         case LK_TRYUPGRADE:
 576                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 577                 tid = (uintptr_t)curthread;
 578                 v = lk->lk_lock;
 579                 x = v & LK_ALL_WAITERS;
 580                 v &= LK_EXCLUSIVE_SPINNERS;
 581                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 582                     tid | x)) {
 583                         lockmgr_note_exclusive_upgrade(lk, file, line, flags);
 584                         locked = true;
 585                 }
 586                 break;
 587         default:
 588                 break;
 589         }
 590         if (__predict_true(locked)) {
 591                 if (__predict_false(flags & LK_INTERLOCK)) {
 592                         class = LOCK_CLASS(ilk);
 593                         class->lc_unlock(ilk);
 594                 }
 595                 return (0);
 596         } else {
 597                 return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT,
 598                     LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line));
 599         }
 600 }
 601
 602 int
 603 lockmgr_unlock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk)
 604 {
 605         struct lock_class *class;
 606         uintptr_t x, tid;
 607         bool unlocked;
 608         const char *file;
 609         int line;
 610
 611         file = __FILE__;
 612         line = __LINE__;
 613
 614         _lockmgr_assert(lk, KA_LOCKED, file, line);
 615         unlocked = false;
 616         x = lk->lk_lock;
 617         if (__predict_true(x & LK_SHARE) != 0) {
 618                 if (lockmgr_sunlock_try(lk, x)) {
 619                         lockmgr_note_shared_release(lk, file, line);
 620                         unlocked = true;
 621                 }
 622         } else {
 623                 tid = (uintptr_t)curthread;
 624                 if (!lockmgr_recursed(lk) &&
 625                     atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) {
 626                         lockmgr_note_exclusive_release(lk, file, line);
 627                         unlocked = true;
 628                 }
 629         }
 630         if (__predict_true(unlocked)) {
 631                 if (__predict_false(flags & LK_INTERLOCK)) {
 632                         class = LOCK_CLASS(ilk);
 633                         class->lc_unlock(ilk);
 634                 }
 635                 return (0);
 636         } else {
 637                 return (__lockmgr_args(lk, flags | LK_RELEASE, ilk, LK_WMESG_DEFAULT,
 638                     LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, LOCK_FILE, LOCK_LINE));
 639         }
 640 }
 641
 642 int
 643 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 644     const char *wmesg, int pri, int timo, const char *file, int line)
 645 {
 646         GIANT_DECLARE;
 647         struct lock_class *class;
 648         const char *iwmesg;
 649         uintptr_t tid, v, x;
 650         u_int op, realexslp;
 651         int error, ipri, itimo, queue, wakeup_swapper;
 652 #ifdef LOCK_PROFILING
 653         uint64_t waittime = 0;
 654         int contested = 0;
 655 #endif
 656
 657         error = 0;
 658         tid = (uintptr_t)curthread;
 659         op = (flags & LK_TYPE_MASK);
 660         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 661         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 662         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 663
 664         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 665         KASSERT((op & (op - 1)) == 0,
 666             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 667         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 668             (op != LK_DOWNGRADE && op != LK_RELEASE),
 669             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 670             __func__, file, line));
 671         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 672             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 673             __func__, file, line));
 674         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 675             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
 676             lk->lock_object.lo_name, file, line));
 677
 678         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 679         if (panicstr != NULL) {
 680                 if (flags & LK_INTERLOCK)
 681                         class->lc_unlock(ilk);
 682                 return (0);
 683         }
 684
 685         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 686                 switch (op) {
 687                 case LK_SHARED:
 688                         op = LK_EXCLUSIVE;
 689                         break;
 690                 case LK_UPGRADE:
 691                 case LK_TRYUPGRADE:
 692                 case LK_DOWNGRADE:
 693                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 694                             file, line);
 695                         if (flags & LK_INTERLOCK)
 696                                 class->lc_unlock(ilk);
 697                         return (0);
 698                 }
 699         }
 700
 701         wakeup_swapper = 0;
 702         switch (op) {
 703         case LK_SHARED:
 704                 if (LK_CAN_WITNESS(flags))
 705                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 706                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 707                 for (;;) {
 708                         if (lockmgr_slock_try(lk, &x, flags))
 709                                 break;
 710 #ifdef HWPMC_HOOKS
 711                         PMC_SOFT_CALL( , , lock, failed);
 712 #endif
 713                         lock_profile_obtain_lock_failed(&lk->lock_object,
 714                             &contested, &waittime);
 715
 716                         /*
 717                          * If the lock is already held by curthread in
 718                          * exclusive way avoid a deadlock.
 719                          */
 720                         if (LK_HOLDER(x) == tid) {
 721                                 LOCK_LOG2(lk,
 722                                     "%s: %p already held in exclusive mode",
 723                                     __func__, lk);
 724                                 error = EDEADLK;
 725                                 break;
 726                         }
 727
 728                         /*
 729                          * If the lock is expected to not sleep just give up
 730                          * and return.
 731                          */
 732                         if (LK_TRYOP(flags)) {
 733                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 734                                     __func__, lk);
 735                                 error = EBUSY;
 736                                 break;
 737                         }
 738
 739                         /*
 740                          * Acquire the sleepqueue chain lock because we
 741                          * probabilly will need to manipulate waiters flags.
 742                          */
 743                         sleepq_lock(&lk->lock_object);
 744                         x = lk->lk_lock;
 745
 746                         /*
 747                          * if the lock can be acquired in shared mode, try
 748                          * again.
 749                          */
 750                         if (LK_CAN_SHARE(x, flags)) {
 751                                 sleepq_release(&lk->lock_object);
 752                                 continue;
 753                         }
 754
 755                         /*
 756                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 757                          * loop back and retry.
 758                          */
 759                         if ((x & LK_SHARED_WAITERS) == 0) {
 760                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 761                                     x | LK_SHARED_WAITERS)) {
 762                                         sleepq_release(&lk->lock_object);
 763                                         continue;
 764                                 }
 765                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 766                                     __func__, lk);
 767                         }
 768
 769                         /*
 770                          * As far as we have been unable to acquire the
 771                          * shared lock and the shared waiters flag is set,
 772                          * we will sleep.
 773                          */
 774                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 775                             SQ_SHARED_QUEUE);
 776                         flags &= ~LK_INTERLOCK;
 777                         if (error) {
 778                                 LOCK_LOG3(lk,
 779                                     "%s: interrupted sleep for %p with %d",
 780                                     __func__, lk, error);
 781                                 break;
 782                         }
 783                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 784                             __func__, lk);
 785                 }
 786                 if (error == 0) {
 787 #ifdef LOCK_PROFILING
 788                         lockmgr_note_shared_acquire(lk, contested, waittime,
 789                             file, line, flags);
 790 #else
 791                         lockmgr_note_shared_acquire(lk, 0, 0, file, line,
 792                             flags);
 793 #endif
 794                 }
 795                 break;
 796         case LK_UPGRADE:
 797         case LK_TRYUPGRADE:
 798                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 799                 v = lk->lk_lock;
 800                 x = v & LK_ALL_WAITERS;
 801                 v &= LK_EXCLUSIVE_SPINNERS;
 802
 803                 /*
 804                  * Try to switch from one shared lock to an exclusive one.
 805                  * We need to preserve waiters flags during the operation.
 806                  */
 807                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 808                     tid | x)) {
 809                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 810                             line);
 811                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 812                             LK_TRYWIT(flags), file, line);
 813                         TD_SLOCKS_DEC(curthread);
 814                         break;
 815                 }
 816
 817                 /*
 818                  * In LK_TRYUPGRADE mode, do not drop the lock,
 819                  * returning EBUSY instead.
 820                  */
 821                 if (op == LK_TRYUPGRADE) {
 822                         LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
 823                             __func__, lk);
 824                         error = EBUSY;
 825                         break;
 826                 }
 827
 828                 /*
 829                  * We have been unable to succeed in upgrading, so just
 830                  * give up the shared lock.
 831                  */
 832                 wakeup_swapper |= wakeupshlk(lk, file, line);
 833
 834                 /* FALLTHROUGH */
 835         case LK_EXCLUSIVE:
 836                 if (LK_CAN_WITNESS(flags))
 837                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 838                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 839                             ilk : NULL);
 840
 841                 /*
 842                  * If curthread already holds the lock and this one is
 843                  * allowed to recurse, simply recurse on it.
 844                  */
 845                 if (lockmgr_xlocked(lk)) {
 846                         if ((flags & LK_CANRECURSE) == 0 &&
 847                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 848
 849                                 /*
 850                                  * If the lock is expected to not panic just
 851                                  * give up and return.
 852                                  */
 853                                 if (LK_TRYOP(flags)) {
 854                                         LOCK_LOG2(lk,
 855                                             "%s: %p fails the try operation",
 856                                             __func__, lk);
 857                                         error = EBUSY;
 858                                         break;
 859                                 }
 860                                 if (flags & LK_INTERLOCK)
 861                                         class->lc_unlock(ilk);
 862                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 863                                     __func__, iwmesg, file, line);
 864                         }
 865                         lk->lk_recurse++;
 866                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 867                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 868                             lk->lk_recurse, file, line);
 869                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 870                             LK_TRYWIT(flags), file, line);
 871                         TD_LOCKS_INC(curthread);
 872                         break;
 873                 }
 874
 875                 for (;;) {
 876                         if (lk->lk_lock == LK_UNLOCKED &&
 877                             atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
 878                                 break;
 879 #ifdef HWPMC_HOOKS
 880                         PMC_SOFT_CALL( , , lock, failed);
 881 #endif
 882                         lock_profile_obtain_lock_failed(&lk->lock_object,
 883                             &contested, &waittime);
 884
 885                         /*
 886                          * If the lock is expected to not sleep just give up
 887                          * and return.
 888                          */
 889                         if (LK_TRYOP(flags)) {
 890                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 891                                     __func__, lk);
 892                                 error = EBUSY;
 893                                 break;
 894                         }
 895
 896                         /*
 897                          * Acquire the sleepqueue chain lock because we
 898                          * probabilly will need to manipulate waiters flags.
 899                          */
 900                         sleepq_lock(&lk->lock_object);
 901                         x = lk->lk_lock;
 902
 903                         /*
 904                          * if the lock has been released while we spun on
 905                          * the sleepqueue chain lock just try again.
 906                          */
 907                         if (x == LK_UNLOCKED) {
 908                                 sleepq_release(&lk->lock_object);
 909                                 continue;
 910                         }
 911
 912                         /*
 913                          * The lock can be in the state where there is a
 914                          * pending queue of waiters, but still no owner.
 915                          * This happens when the lock is contested and an
 916                          * owner is going to claim the lock.
 917                          * If curthread is the one successfully acquiring it
 918                          * claim lock ownership and return, preserving waiters
 919                          * flags.
 920                          */
 921                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 922                         if ((x & ~v) == LK_UNLOCKED) {
 923                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 924                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 925                                     tid | v)) {
 926                                         sleepq_release(&lk->lock_object);
 927                                         LOCK_LOG2(lk,
 928                                             "%s: %p claimed by a new writer",
 929                                             __func__, lk);
 930                                         break;
 931                                 }
 932                                 sleepq_release(&lk->lock_object);
 933                                 continue;
 934                         }
 935
 936                         /*
 937                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 938                          * fail, loop back and retry.
 939                          */
 940                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 941                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 942                                     x | LK_EXCLUSIVE_WAITERS)) {
 943                                         sleepq_release(&lk->lock_object);
 944                                         continue;
 945                                 }
 946                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 947                                     __func__, lk);
 948                         }
 949
 950                         /*
 951                          * As far as we have been unable to acquire the
 952                          * exclusive lock and the exclusive waiters flag
 953                          * is set, we will sleep.
 954                          */
 955                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 956                             SQ_EXCLUSIVE_QUEUE);
 957                         flags &= ~LK_INTERLOCK;
 958                         if (error) {
 959                                 LOCK_LOG3(lk,
 960                                     "%s: interrupted sleep for %p with %d",
 961                                     __func__, lk, error);
 962                                 break;
 963                         }
 964                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 965                             __func__, lk);
 966                 }
 967                 if (error == 0) {
 968 #ifdef LOCK_PROFILING
 969                         lockmgr_note_exclusive_acquire(lk, contested, waittime,
 970                             file, line, flags);
 971 #else
 972                         lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
 973                             flags);
 974 #endif
 975                 }
 976                 break;
 977         case LK_DOWNGRADE:
 978                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 979                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 980                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 981
 982                 /*
 983                  * Panic if the lock is recursed.
 984                  */
 985                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 986                         if (flags & LK_INTERLOCK)
 987                                 class->lc_unlock(ilk);
 988                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 989                             __func__, iwmesg, file, line);
 990                 }
 991                 TD_SLOCKS_INC(curthread);
 992
 993                 /*
 994                  * In order to preserve waiters flags, just spin.
 995                  */
 996                 for (;;) {
 997                         x = lk->lk_lock;
 998                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
 999                         x &= LK_ALL_WAITERS;
1000                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1001                             LK_SHARERS_LOCK(1) | x))
1002                                 break;
1003                         cpu_spinwait();
1004                 }
1005                 break;
1006         case LK_RELEASE:
1007                 _lockmgr_assert(lk, KA_LOCKED, file, line);
1008                 x = lk->lk_lock;
1009
1010                 if ((x & LK_SHARE) == 0) {
1011
1012                         /*
1013                          * As first option, treact the lock as if it has not
1014                          * any waiter.
1015                          * Fix-up the tid var if the lock has been disowned.
1016                          */
1017                         if (LK_HOLDER(x) == LK_KERNPROC)
1018                                 tid = LK_KERNPROC;
1019                         else {
1020                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
1021                                     file, line);
1022                                 TD_LOCKS_DEC(curthread);
1023                         }
1024                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
1025                             lk->lk_recurse, file, line);
1026
1027                         /*
1028                          * The lock is held in exclusive mode.
1029                          * If the lock is recursed also, then unrecurse it.
1030                          */
1031                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1032                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1033                                     lk);
1034                                 lk->lk_recurse--;
1035                                 break;
1036                         }
1037                         if (tid != LK_KERNPROC)
1038                                 lock_profile_release_lock(&lk->lock_object);
1039
1040                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1041                             LK_UNLOCKED))
1042                                 break;
1043
1044                         sleepq_lock(&lk->lock_object);
1045                         x = lk->lk_lock;
1046                         v = LK_UNLOCKED;
1047
1048                         /*
1049                          * If the lock has exclusive waiters, give them
1050                          * preference in order to avoid deadlock with
1051                          * shared runners up.
1052                          * If interruptible sleeps left the exclusive queue
1053                          * empty avoid a starvation for the threads sleeping
1054                          * on the shared queue by giving them precedence
1055                          * and cleaning up the exclusive waiters bit anyway.
1056                          * Please note that lk_exslpfail count may be lying
1057                          * about the real number of waiters with the
1058                          * LK_SLEEPFAIL flag on because they may be used in
1059                          * conjunction with interruptible sleeps so
1060                          * lk_exslpfail might be considered an 'upper limit'
1061                          * bound, including the edge cases.
1062                          */
1063                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1064                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1065                             SQ_EXCLUSIVE_QUEUE);
1066                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1067                                 if (lk->lk_exslpfail < realexslp) {
1068                                         lk->lk_exslpfail = 0;
1069                                         queue = SQ_EXCLUSIVE_QUEUE;
1070                                         v |= (x & LK_SHARED_WAITERS);
1071                                 } else {
1072                                         lk->lk_exslpfail = 0;
1073                                         LOCK_LOG2(lk,
1074                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1075                                             __func__, lk);
1076                                         LOCK_LOG2(lk,
1077                         "%s: %p waking up threads on the exclusive queue",
1078                                             __func__, lk);
1079                                         wakeup_swapper =
1080                                             sleepq_broadcast(&lk->lock_object,
1081                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1082                                         queue = SQ_SHARED_QUEUE;
1083                                 }
1084                         } else {
1085
1086                                 /*
1087                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1088                                  * on and using interruptible sleeps/timeout
1089                                  * may have left spourious lk_exslpfail counts
1090                                  * on, so clean it up anyway.
1091                                  */
1092                                 lk->lk_exslpfail = 0;
1093                                 queue = SQ_SHARED_QUEUE;
1094                         }
1095
1096                         LOCK_LOG3(lk,
1097                             "%s: %p waking up threads on the %s queue",
1098                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1099                             "exclusive");
1100                         atomic_store_rel_ptr(&lk->lk_lock, v);
1101                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1102                             SLEEPQ_LK, 0, queue);
1103                         sleepq_release(&lk->lock_object);
1104                         break;
1105                 } else
1106                         wakeup_swapper = wakeupshlk(lk, file, line);
1107                 break;
1108         case LK_DRAIN:
1109                 if (LK_CAN_WITNESS(flags))
1110                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1111                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1112                             ilk : NULL);
1113
1114                 /*
1115                  * Trying to drain a lock we already own will result in a
1116                  * deadlock.
1117                  */
1118                 if (lockmgr_xlocked(lk)) {
1119                         if (flags & LK_INTERLOCK)
1120                                 class->lc_unlock(ilk);
1121                         panic("%s: draining %s with the lock held @ %s:%d\n",
1122                             __func__, iwmesg, file, line);
1123                 }
1124
1125                 for (;;) {
1126                         if (lk->lk_lock == LK_UNLOCKED &&
1127                             atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1128                                 break;
1129
1130 #ifdef HWPMC_HOOKS
1131                         PMC_SOFT_CALL( , , lock, failed);
1132 #endif
1133                         lock_profile_obtain_lock_failed(&lk->lock_object,
1134                             &contested, &waittime);
1135
1136                         /*
1137                          * If the lock is expected to not sleep just give up
1138                          * and return.
1139                          */
1140                         if (LK_TRYOP(flags)) {
1141                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1142                                     __func__, lk);
1143                                 error = EBUSY;
1144                                 break;
1145                         }
1146
1147                         /*
1148                          * Acquire the sleepqueue chain lock because we
1149                          * probabilly will need to manipulate waiters flags.
1150                          */
1151                         sleepq_lock(&lk->lock_object);
1152                         x = lk->lk_lock;
1153
1154                         /*
1155                          * if the lock has been released while we spun on
1156                          * the sleepqueue chain lock just try again.
1157                          */
1158                         if (x == LK_UNLOCKED) {
1159                                 sleepq_release(&lk->lock_object);
1160                                 continue;
1161                         }
1162
1163                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1164                         if ((x & ~v) == LK_UNLOCKED) {
1165                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1166
1167                                 /*
1168                                  * If interruptible sleeps left the exclusive
1169                                  * queue empty avoid a starvation for the
1170                                  * threads sleeping on the shared queue by
1171                                  * giving them precedence and cleaning up the
1172                                  * exclusive waiters bit anyway.
1173                                  * Please note that lk_exslpfail count may be
1174                                  * lying about the real number of waiters with
1175                                  * the LK_SLEEPFAIL flag on because they may
1176                                  * be used in conjunction with interruptible
1177                                  * sleeps so lk_exslpfail might be considered
1178                                  * an 'upper limit' bound, including the edge
1179                                  * cases.
1180                                  */
1181                                 if (v & LK_EXCLUSIVE_WAITERS) {
1182                                         queue = SQ_EXCLUSIVE_QUEUE;
1183                                         v &= ~LK_EXCLUSIVE_WAITERS;
1184                                 } else {
1185
1186                                         /*
1187                                          * Exclusive waiters sleeping with
1188                                          * LK_SLEEPFAIL on and using
1189                                          * interruptible sleeps/timeout may
1190                                          * have left spourious lk_exslpfail
1191                                          * counts on, so clean it up anyway.
1192                                          */
1193                                         MPASS(v & LK_SHARED_WAITERS);
1194                                         lk->lk_exslpfail = 0;
1195                                         queue = SQ_SHARED_QUEUE;
1196                                         v &= ~LK_SHARED_WAITERS;
1197                                 }
1198                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1199                                         realexslp =
1200                                             sleepq_sleepcnt(&lk->lock_object,
1201                                             SQ_EXCLUSIVE_QUEUE);
1202                                         if (lk->lk_exslpfail >= realexslp) {
1203                                                 lk->lk_exslpfail = 0;
1204                                                 queue = SQ_SHARED_QUEUE;
1205                                                 v &= ~LK_SHARED_WAITERS;
1206                                                 if (realexslp != 0) {
1207                                                         LOCK_LOG2(lk,
1208                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1209                                                             __func__, lk);
1210                                                         LOCK_LOG2(lk,
1211                         "%s: %p waking up threads on the exclusive queue",
1212                                                             __func__, lk);
1213                                                         wakeup_swapper =
1214                                                             sleepq_broadcast(
1215                                                             &lk->lock_object,
1216                                                             SLEEPQ_LK, 0,
1217                                                             SQ_EXCLUSIVE_QUEUE);
1218                                                 }
1219                                         } else
1220                                                 lk->lk_exslpfail = 0;
1221                                 }
1222                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1223                                         sleepq_release(&lk->lock_object);
1224                                         continue;
1225                                 }
1226                                 LOCK_LOG3(lk,
1227                                 "%s: %p waking up all threads on the %s queue",
1228                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1229                                     "shared" : "exclusive");
1230                                 wakeup_swapper |= sleepq_broadcast(
1231                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1232
1233                                 /*
1234                                  * If shared waiters have been woken up we need
1235                                  * to wait for one of them to acquire the lock
1236                                  * before to set the exclusive waiters in
1237                                  * order to avoid a deadlock.
1238                                  */
1239                                 if (queue == SQ_SHARED_QUEUE) {
1240                                         for (v = lk->lk_lock;
1241                                             (v & LK_SHARE) && !LK_SHARERS(v);
1242                                             v = lk->lk_lock)
1243                                                 cpu_spinwait();
1244                                 }
1245                         }
1246
1247                         /*
1248                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1249                          * fail, loop back and retry.
1250                          */
1251                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1252                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1253                                     x | LK_EXCLUSIVE_WAITERS)) {
1254                                         sleepq_release(&lk->lock_object);
1255                                         continue;
1256                                 }
1257                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1258                                     __func__, lk);
1259                         }
1260
1261                         /*
1262                          * As far as we have been unable to acquire the
1263                          * exclusive lock and the exclusive waiters flag
1264                          * is set, we will sleep.
1265                          */
1266                         if (flags & LK_INTERLOCK) {
1267                                 class->lc_unlock(ilk);
1268                                 flags &= ~LK_INTERLOCK;
1269                         }
1270                         GIANT_SAVE();
1271                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1272                             SQ_EXCLUSIVE_QUEUE);
1273                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1274                         GIANT_RESTORE();
1275                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1276                             __func__, lk);
1277                 }
1278
1279                 if (error == 0) {
1280                         lock_profile_obtain_lock_success(&lk->lock_object,
1281                             contested, waittime, file, line);
1282                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1283                             lk->lk_recurse, file, line);
1284                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1285                             LK_TRYWIT(flags), file, line);
1286                         TD_LOCKS_INC(curthread);
1287                         STACK_SAVE(lk);
1288                 }
1289                 break;
1290         default:
1291                 if (flags & LK_INTERLOCK)
1292                         class->lc_unlock(ilk);
1293                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1294         }
1295
1296         if (flags & LK_INTERLOCK)
1297                 class->lc_unlock(ilk);
1298         if (wakeup_swapper)
1299                 kick_proc0();
1300
1301         return (error);
1302 }
1303
1304 void
1305 _lockmgr_disown(struct lock *lk, const char *file, int line)
1306 {
1307         uintptr_t tid, x;
1308
1309         if (SCHEDULER_STOPPED())
1310                 return;
1311
1312         tid = (uintptr_t)curthread;
1313         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1314
1315         /*
1316          * Panic if the lock is recursed.
1317          */
1318         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1319                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1320                     __func__,  file, line);
1321
1322         /*
1323          * If the owner is already LK_KERNPROC just skip the whole operation.
1324          */
1325         if (LK_HOLDER(lk->lk_lock) != tid)
1326                 return;
1327         lock_profile_release_lock(&lk->lock_object);
1328         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1329         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1330         TD_LOCKS_DEC(curthread);
1331         STACK_SAVE(lk);
1332
1333         /*
1334          * In order to preserve waiters flags, just spin.
1335          */
1336         for (;;) {
1337                 x = lk->lk_lock;
1338                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1339                 x &= LK_ALL_WAITERS;
1340                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1341                     LK_KERNPROC | x))
1342                         return;
1343                 cpu_spinwait();
1344         }
1345 }
1346
1347 void
1348 lockmgr_printinfo(const struct lock *lk)
1349 {
1350         struct thread *td;
1351         uintptr_t x;
1352
1353         if (lk->lk_lock == LK_UNLOCKED)
1354                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1355         else if (lk->lk_lock & LK_SHARE)
1356                 printf("lock type %s: SHARED (count %ju)\n",
1357                     lk->lock_object.lo_name,
1358                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1359         else {
1360                 td = lockmgr_xholder(lk);
1361                 if (td == (struct thread *)LK_KERNPROC)
1362                         printf("lock type %s: EXCL by KERNPROC\n",
1363                             lk->lock_object.lo_name);
1364                 else
1365                         printf("lock type %s: EXCL by thread %p "
1366                             "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1367                             td, td->td_proc->p_pid, td->td_proc->p_comm,
1368                             td->td_tid);
1369         }
1370
1371         x = lk->lk_lock;
1372         if (x & LK_EXCLUSIVE_WAITERS)
1373                 printf(" with exclusive waiters pending\n");
1374         if (x & LK_SHARED_WAITERS)
1375                 printf(" with shared waiters pending\n");
1376         if (x & LK_EXCLUSIVE_SPINNERS)
1377                 printf(" with exclusive spinners pending\n");
1378
1379         STACK_PRINT(lk);
1380 }
1381
1382 int
1383 lockstatus(const struct lock *lk)
1384 {
1385         uintptr_t v, x;
1386         int ret;
1387
1388         ret = LK_SHARED;
1389         x = lk->lk_lock;
1390         v = LK_HOLDER(x);
1391
1392         if ((x & LK_SHARE) == 0) {
1393                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1394                         ret = LK_EXCLUSIVE;
1395                 else
1396                         ret = LK_EXCLOTHER;
1397         } else if (x == LK_UNLOCKED)
1398                 ret = 0;
1399
1400         return (ret);
1401 }
1402
1403 #ifdef INVARIANT_SUPPORT
1404
1405 FEATURE(invariant_support,
1406     "Support for modules compiled with INVARIANTS option");
1407
1408 #ifndef INVARIANTS
1409 #undef  _lockmgr_assert
1410 #endif
1411
1412 void
1413 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1414 {
1415         int slocked = 0;
1416
1417         if (panicstr != NULL)
1418                 return;
1419         switch (what) {
1420         case KA_SLOCKED:
1421         case KA_SLOCKED | KA_NOTRECURSED:
1422         case KA_SLOCKED | KA_RECURSED:
1423                 slocked = 1;
1424         case KA_LOCKED:
1425         case KA_LOCKED | KA_NOTRECURSED:
1426         case KA_LOCKED | KA_RECURSED:
1427 #ifdef WITNESS
1428
1429                 /*
1430                  * We cannot trust WITNESS if the lock is held in exclusive
1431                  * mode and a call to lockmgr_disown() happened.
1432                  * Workaround this skipping the check if the lock is held in
1433                  * exclusive mode even for the KA_LOCKED case.
1434                  */
1435                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1436                         witness_assert(&lk->lock_object, what, file, line);
1437                         break;
1438                 }
1439 #endif
1440                 if (lk->lk_lock == LK_UNLOCKED ||
1441                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1442                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1443                         panic("Lock %s not %slocked @ %s:%d\n",
1444                             lk->lock_object.lo_name, slocked ? "share" : "",
1445                             file, line);
1446
1447                 if ((lk->lk_lock & LK_SHARE) == 0) {
1448                         if (lockmgr_recursed(lk)) {
1449                                 if (what & KA_NOTRECURSED)
1450                                         panic("Lock %s recursed @ %s:%d\n",
1451                                             lk->lock_object.lo_name, file,
1452                                             line);
1453                         } else if (what & KA_RECURSED)
1454                                 panic("Lock %s not recursed @ %s:%d\n",
1455                                     lk->lock_object.lo_name, file, line);
1456                 }
1457                 break;
1458         case KA_XLOCKED:
1459         case KA_XLOCKED | KA_NOTRECURSED:
1460         case KA_XLOCKED | KA_RECURSED:
1461                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1462                         panic("Lock %s not exclusively locked @ %s:%d\n",
1463                             lk->lock_object.lo_name, file, line);
1464                 if (lockmgr_recursed(lk)) {
1465                         if (what & KA_NOTRECURSED)
1466                                 panic("Lock %s recursed @ %s:%d\n",
1467                                     lk->lock_object.lo_name, file, line);
1468                 } else if (what & KA_RECURSED)
1469                         panic("Lock %s not recursed @ %s:%d\n",
1470                             lk->lock_object.lo_name, file, line);
1471                 break;
1472         case KA_UNLOCKED:
1473                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1474                         panic("Lock %s exclusively locked @ %s:%d\n",
1475                             lk->lock_object.lo_name, file, line);
1476                 break;
1477         default:
1478                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1479                     line);
1480         }
1481 }
1482 #endif
1483
1484 #ifdef DDB
1485 int
1486 lockmgr_chain(struct thread *td, struct thread **ownerp)
1487 {
1488         struct lock *lk;
1489
1490         lk = td->td_wchan;
1491
1492         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1493                 return (0);
1494         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1495         if (lk->lk_lock & LK_SHARE)
1496                 db_printf("SHARED (count %ju)\n",
1497                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1498         else
1499                 db_printf("EXCL\n");
1500         *ownerp = lockmgr_xholder(lk);
1501
1502         return (1);
1503 }
1504
1505 static void
1506 db_show_lockmgr(const struct lock_object *lock)
1507 {
1508         struct thread *td;
1509         const struct lock *lk;
1510
1511         lk = (const struct lock *)lock;
1512
1513         db_printf(" state: ");
1514         if (lk->lk_lock == LK_UNLOCKED)
1515                 db_printf("UNLOCKED\n");
1516         else if (lk->lk_lock & LK_SHARE)
1517                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1518         else {
1519                 td = lockmgr_xholder(lk);
1520                 if (td == (struct thread *)LK_KERNPROC)
1521                         db_printf("XLOCK: LK_KERNPROC\n");
1522                 else
1523                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1524                             td->td_tid, td->td_proc->p_pid,
1525                             td->td_proc->p_comm);
1526                 if (lockmgr_recursed(lk))
1527                         db_printf(" recursed: %d\n", lk->lk_recurse);
1528         }
1529         db_printf(" waiters: ");
1530         switch (lk->lk_lock & LK_ALL_WAITERS) {
1531         case LK_SHARED_WAITERS:
1532                 db_printf("shared\n");
1533                 break;
1534         case LK_EXCLUSIVE_WAITERS:
1535                 db_printf("exclusive\n");
1536                 break;
1537         case LK_ALL_WAITERS:
1538                 db_printf("shared and exclusive\n");
1539                 break;
1540         default:
1541                 db_printf("none\n");
1542         }
1543         db_printf(" spinners: ");
1544         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1545                 db_printf("exclusive\n");
1546         else
1547                 db_printf("none\n");
1548 }
1549 #endif