sys/kern/kern_lock.c

   1 /*-
   2  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice(s), this list of conditions and the following disclaimer as
  10  *    the first lines of this file unmodified other than the possible
  11  *    addition of one or more copyright notices.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice(s), this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  26  * DAMAGE.
  27  */
  28
  29 #include "opt_adaptive_lockmgrs.h"
  30 #include "opt_ddb.h"
  31 #include "opt_hwpmc_hooks.h"
  32 #include "opt_kdtrace.h"
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include <sys/param.h>
  38 #include <sys/ktr.h>
  39 #include <sys/lock.h>
  40 #include <sys/lock_profile.h>
  41 #include <sys/lockmgr.h>
  42 #include <sys/mutex.h>
  43 #include <sys/proc.h>
  44 #include <sys/sleepqueue.h>
  45 #ifdef DEBUG_LOCKS
  46 #include <sys/stack.h>
  47 #endif
  48 #include <sys/sysctl.h>
  49 #include <sys/systm.h>
  50
  51 #include <machine/cpu.h>
  52
  53 #ifdef DDB
  54 #include <ddb/ddb.h>
  55 #endif
  56
  57 #ifdef HWPMC_HOOKS
  58 #include <sys/pmckern.h>
  59 PMC_SOFT_DECLARE( , , lock, failed);
  60 #endif
  61
  62 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
  63     (LK_ADAPTIVE | LK_NOSHARE));
  64 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
  65     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
  66
  67 #define SQ_EXCLUSIVE_QUEUE      0
  68 #define SQ_SHARED_QUEUE         1
  69
  70 #ifdef ADAPTIVE_LOCKMGRS
  71 #define ALK_RETRIES             10
  72 #define ALK_LOOPS               10000
  73 #endif
  74
  75 #ifndef INVARIANTS
  76 #define _lockmgr_assert(lk, what, file, line)
  77 #define TD_LOCKS_INC(td)
  78 #define TD_LOCKS_DEC(td)
  79 #else
  80 #define TD_LOCKS_INC(td)        ((td)->td_locks++)
  81 #define TD_LOCKS_DEC(td)        ((td)->td_locks--)
  82 #endif
  83 #define TD_SLOCKS_INC(td)       ((td)->td_lk_slocks++)
  84 #define TD_SLOCKS_DEC(td)       ((td)->td_lk_slocks--)
  85
  86 #ifndef DEBUG_LOCKS
  87 #define STACK_PRINT(lk)
  88 #define STACK_SAVE(lk)
  89 #define STACK_ZERO(lk)
  90 #else
  91 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
  92 #define STACK_SAVE(lk)  stack_save(&(lk)->lk_stack)
  93 #define STACK_ZERO(lk)  stack_zero(&(lk)->lk_stack)
  94 #endif
  95
  96 #define LOCK_LOG2(lk, string, arg1, arg2)                               \
  97         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  98                 CTR2(KTR_LOCK, (string), (arg1), (arg2))
  99 #define LOCK_LOG3(lk, string, arg1, arg2, arg3)                         \
 100         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
 101                 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
 102
 103 #define GIANT_DECLARE                                                   \
 104         int _i = 0;                                                     \
 105         WITNESS_SAVE_DECL(Giant)
 106 #define GIANT_RESTORE() do {                                            \
 107         if (_i > 0) {                                                   \
 108                 while (_i--)                                            \
 109                         mtx_lock(&Giant);                               \
 110                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 111         }                                                               \
 112 } while (0)
 113 #define GIANT_SAVE() do {                                               \
 114         if (mtx_owned(&Giant)) {                                        \
 115                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
 116                 while (mtx_owned(&Giant)) {                             \
 117                         _i++;                                           \
 118                         mtx_unlock(&Giant);                             \
 119                 }                                                       \
 120         }                                                               \
 121 } while (0)
 122
 123 #define LK_CAN_SHARE(x)                                                 \
 124         (((x) & LK_SHARE) && (((x) & LK_EXCLUSIVE_WAITERS) == 0 ||      \
 125         ((x) & LK_EXCLUSIVE_SPINNERS) == 0 ||                           \
 126         curthread->td_lk_slocks || (curthread->td_pflags & TDP_DEADLKTREAT)))
 127 #define LK_TRYOP(x)                                                     \
 128         ((x) & LK_NOWAIT)
 129
 130 #define LK_CAN_WITNESS(x)                                               \
 131         (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
 132 #define LK_TRYWIT(x)                                                    \
 133         (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
 134
 135 #define LK_CAN_ADAPT(lk, f)                                             \
 136         (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&             \
 137         ((f) & LK_SLEEPFAIL) == 0)
 138
 139 #define lockmgr_disowned(lk)                                            \
 140         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
 141
 142 #define lockmgr_xlocked(lk)                                             \
 143         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
 144
 145 static void      assert_lockmgr(struct lock_object *lock, int how);
 146 #ifdef DDB
 147 static void      db_show_lockmgr(struct lock_object *lock);
 148 #endif
 149 static void      lock_lockmgr(struct lock_object *lock, int how);
 150 #ifdef KDTRACE_HOOKS
 151 static int       owner_lockmgr(struct lock_object *lock, struct thread **owner);
 152 #endif
 153 static int       unlock_lockmgr(struct lock_object *lock);
 154
 155 struct lock_class lock_class_lockmgr = {
 156         .lc_name = "lockmgr",
 157         .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
 158         .lc_assert = assert_lockmgr,
 159 #ifdef DDB
 160         .lc_ddb_show = db_show_lockmgr,
 161 #endif
 162         .lc_lock = lock_lockmgr,
 163         .lc_unlock = unlock_lockmgr,
 164 #ifdef KDTRACE_HOOKS
 165         .lc_owner = owner_lockmgr,
 166 #endif
 167 };
 168
 169 static __inline struct thread *
 170 lockmgr_xholder(struct lock *lk)
 171 {
 172         uintptr_t x;
 173
 174         x = lk->lk_lock;
 175         return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
 176 }
 177
 178 /*
 179  * It assumes sleepq_lock held and returns with this one unheld.
 180  * It also assumes the generic interlock is sane and previously checked.
 181  * If LK_INTERLOCK is specified the interlock is not reacquired after the
 182  * sleep.
 183  */
 184 static __inline int
 185 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
 186     const char *wmesg, int pri, int timo, int queue)
 187 {
 188         GIANT_DECLARE;
 189         struct lock_class *class;
 190         int catch, error;
 191
 192         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 193         catch = pri & PCATCH;
 194         pri &= PRIMASK;
 195         error = 0;
 196
 197         LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
 198             (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
 199
 200         if (flags & LK_INTERLOCK)
 201                 class->lc_unlock(ilk);
 202         if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
 203                 lk->lk_exslpfail++;
 204         GIANT_SAVE();
 205         sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
 206             SLEEPQ_INTERRUPTIBLE : 0), queue);
 207         if ((flags & LK_TIMELOCK) && timo)
 208                 sleepq_set_timeout(&lk->lock_object, timo);
 209
 210         /*
 211          * Decisional switch for real sleeping.
 212          */
 213         if ((flags & LK_TIMELOCK) && timo && catch)
 214                 error = sleepq_timedwait_sig(&lk->lock_object, pri);
 215         else if ((flags & LK_TIMELOCK) && timo)
 216                 error = sleepq_timedwait(&lk->lock_object, pri);
 217         else if (catch)
 218                 error = sleepq_wait_sig(&lk->lock_object, pri);
 219         else
 220                 sleepq_wait(&lk->lock_object, pri);
 221         GIANT_RESTORE();
 222         if ((flags & LK_SLEEPFAIL) && error == 0)
 223                 error = ENOLCK;
 224
 225         return (error);
 226 }
 227
 228 static __inline int
 229 wakeupshlk(struct lock *lk, const char *file, int line)
 230 {
 231         uintptr_t v, x;
 232         u_int realexslp;
 233         int queue, wakeup_swapper;
 234
 235         TD_LOCKS_DEC(curthread);
 236         TD_SLOCKS_DEC(curthread);
 237         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 238         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 239
 240         wakeup_swapper = 0;
 241         for (;;) {
 242                 x = lk->lk_lock;
 243
 244                 /*
 245                  * If there is more than one shared lock held, just drop one
 246                  * and return.
 247                  */
 248                 if (LK_SHARERS(x) > 1) {
 249                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
 250                             x - LK_ONE_SHARER))
 251                                 break;
 252                         continue;
 253                 }
 254
 255                 /*
 256                  * If there are not waiters on the exclusive queue, drop the
 257                  * lock quickly.
 258                  */
 259                 if ((x & LK_ALL_WAITERS) == 0) {
 260                         MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
 261                             LK_SHARERS_LOCK(1));
 262                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
 263                                 break;
 264                         continue;
 265                 }
 266
 267                 /*
 268                  * We should have a sharer with waiters, so enter the hard
 269                  * path in order to handle wakeups correctly.
 270                  */
 271                 sleepq_lock(&lk->lock_object);
 272                 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 273                 v = LK_UNLOCKED;
 274
 275                 /*
 276                  * If the lock has exclusive waiters, give them preference in
 277                  * order to avoid deadlock with shared runners up.
 278                  * If interruptible sleeps left the exclusive queue empty
 279                  * avoid a starvation for the threads sleeping on the shared
 280                  * queue by giving them precedence and cleaning up the
 281                  * exclusive waiters bit anyway.
 282                  * Please note that lk_exslpfail count may be lying about
 283                  * the real number of waiters with the LK_SLEEPFAIL flag on
 284                  * because they may be used in conjuction with interruptible
 285                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 286                  * bound, including the edge cases.
 287                  */
 288                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 289                     SQ_EXCLUSIVE_QUEUE);
 290                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 291                         if (lk->lk_exslpfail < realexslp) {
 292                                 lk->lk_exslpfail = 0;
 293                                 queue = SQ_EXCLUSIVE_QUEUE;
 294                                 v |= (x & LK_SHARED_WAITERS);
 295                         } else {
 296                                 lk->lk_exslpfail = 0;
 297                                 LOCK_LOG2(lk,
 298                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 299                                     __func__, lk);
 300                                 LOCK_LOG2(lk,
 301                             "%s: %p waking up threads on the exclusive queue",
 302                                     __func__, lk);
 303                                 wakeup_swapper =
 304                                     sleepq_broadcast(&lk->lock_object,
 305                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 306                                 queue = SQ_SHARED_QUEUE;
 307                         }
 308
 309                 } else {
 310
 311                         /*
 312                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 313                          * and using interruptible sleeps/timeout may have
 314                          * left spourious lk_exslpfail counts on, so clean
 315                          * it up anyway.
 316                          */
 317                         lk->lk_exslpfail = 0;
 318                         queue = SQ_SHARED_QUEUE;
 319                 }
 320
 321                 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
 322                     v)) {
 323                         sleepq_release(&lk->lock_object);
 324                         continue;
 325                 }
 326                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 327                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 328                     "exclusive");
 329                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 330                     0, queue);
 331                 sleepq_release(&lk->lock_object);
 332                 break;
 333         }
 334
 335         lock_profile_release_lock(&lk->lock_object);
 336         return (wakeup_swapper);
 337 }
 338
 339 static void
 340 assert_lockmgr(struct lock_object *lock, int what)
 341 {
 342
 343         panic("lockmgr locks do not support assertions");
 344 }
 345
 346 static void
 347 lock_lockmgr(struct lock_object *lock, int how)
 348 {
 349
 350         panic("lockmgr locks do not support sleep interlocking");
 351 }
 352
 353 static int
 354 unlock_lockmgr(struct lock_object *lock)
 355 {
 356
 357         panic("lockmgr locks do not support sleep interlocking");
 358 }
 359
 360 #ifdef KDTRACE_HOOKS
 361 static int
 362 owner_lockmgr(struct lock_object *lock, struct thread **owner)
 363 {
 364
 365         panic("lockmgr locks do not support owner inquiring");
 366 }
 367 #endif
 368
 369 void
 370 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 371 {
 372         int iflags;
 373
 374         MPASS((flags & ~LK_INIT_MASK) == 0);
 375         ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
 376             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
 377             &lk->lk_lock));
 378
 379         iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 380         if (flags & LK_CANRECURSE)
 381                 iflags |= LO_RECURSABLE;
 382         if ((flags & LK_NODUP) == 0)
 383                 iflags |= LO_DUPOK;
 384         if (flags & LK_NOPROFILE)
 385                 iflags |= LO_NOPROFILE;
 386         if ((flags & LK_NOWITNESS) == 0)
 387                 iflags |= LO_WITNESS;
 388         if (flags & LK_QUIET)
 389                 iflags |= LO_QUIET;
 390         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 391
 392         lk->lk_lock = LK_UNLOCKED;
 393         lk->lk_recurse = 0;
 394         lk->lk_exslpfail = 0;
 395         lk->lk_timo = timo;
 396         lk->lk_pri = pri;
 397         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 398         STACK_ZERO(lk);
 399 }
 400
 401 /*
 402  * XXX: Gross hacks to manipulate external lock flags after
 403  * initialization.  Used for certain vnode and buf locks.
 404  */
 405 void
 406 lockallowshare(struct lock *lk)
 407 {
 408
 409         lockmgr_assert(lk, KA_XLOCKED);
 410         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 411 }
 412
 413 void
 414 lockallowrecurse(struct lock *lk)
 415 {
 416
 417         lockmgr_assert(lk, KA_XLOCKED);
 418         lk->lock_object.lo_flags |= LO_RECURSABLE;
 419 }
 420
 421 void
 422 lockdisablerecurse(struct lock *lk)
 423 {
 424
 425         lockmgr_assert(lk, KA_XLOCKED);
 426         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 427 }
 428
 429 void
 430 lockdestroy(struct lock *lk)
 431 {
 432
 433         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 434         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 435         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 436         lock_destroy(&lk->lock_object);
 437 }
 438
 439 int
 440 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 441     const char *wmesg, int pri, int timo, const char *file, int line)
 442 {
 443         GIANT_DECLARE;
 444         struct lock_class *class;
 445         const char *iwmesg;
 446         uintptr_t tid, v, x;
 447         u_int op, realexslp;
 448         int error, ipri, itimo, queue, wakeup_swapper;
 449 #ifdef LOCK_PROFILING
 450         uint64_t waittime = 0;
 451         int contested = 0;
 452 #endif
 453 #ifdef ADAPTIVE_LOCKMGRS
 454         volatile struct thread *owner;
 455         u_int i, spintries = 0;
 456 #endif
 457
 458         error = 0;
 459         tid = (uintptr_t)curthread;
 460         op = (flags & LK_TYPE_MASK);
 461         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 462         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 463         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 464
 465         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 466         KASSERT((op & (op - 1)) == 0,
 467             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 468         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 469             (op != LK_DOWNGRADE && op != LK_RELEASE),
 470             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 471             __func__, file, line));
 472         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 473             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 474             __func__, file, line));
 475
 476         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 477         if (panicstr != NULL) {
 478                 if (flags & LK_INTERLOCK)
 479                         class->lc_unlock(ilk);
 480                 return (0);
 481         }
 482
 483         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 484                 switch (op) {
 485                 case LK_SHARED:
 486                         op = LK_EXCLUSIVE;
 487                         break;
 488                 case LK_UPGRADE:
 489                 case LK_DOWNGRADE:
 490                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 491                             file, line);
 492                         if (flags & LK_INTERLOCK)
 493                                 class->lc_unlock(ilk);
 494                         return (0);
 495                 }
 496         }
 497
 498         wakeup_swapper = 0;
 499         switch (op) {
 500         case LK_SHARED:
 501                 if (LK_CAN_WITNESS(flags))
 502                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 503                             file, line, ilk);
 504                 for (;;) {
 505                         x = lk->lk_lock;
 506
 507                         /*
 508                          * If no other thread has an exclusive lock, or
 509                          * no exclusive waiter is present, bump the count of
 510                          * sharers.  Since we have to preserve the state of
 511                          * waiters, if we fail to acquire the shared lock
 512                          * loop back and retry.
 513                          */
 514                         if (LK_CAN_SHARE(x)) {
 515                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 516                                     x + LK_ONE_SHARER))
 517                                         break;
 518                                 continue;
 519                         }
 520 #ifdef HWPMC_HOOKS
 521                         PMC_SOFT_CALL( , , lock, failed);
 522 #endif
 523                         lock_profile_obtain_lock_failed(&lk->lock_object,
 524                             &contested, &waittime);
 525
 526                         /*
 527                          * If the lock is already held by curthread in
 528                          * exclusive way avoid a deadlock.
 529                          */
 530                         if (LK_HOLDER(x) == tid) {
 531                                 LOCK_LOG2(lk,
 532                                     "%s: %p already held in exclusive mode",
 533                                     __func__, lk);
 534                                 error = EDEADLK;
 535                                 break;
 536                         }
 537
 538                         /*
 539                          * If the lock is expected to not sleep just give up
 540                          * and return.
 541                          */
 542                         if (LK_TRYOP(flags)) {
 543                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 544                                     __func__, lk);
 545                                 error = EBUSY;
 546                                 break;
 547                         }
 548
 549 #ifdef ADAPTIVE_LOCKMGRS
 550                         /*
 551                          * If the owner is running on another CPU, spin until
 552                          * the owner stops running or the state of the lock
 553                          * changes.  We need a double-state handle here
 554                          * because for a failed acquisition the lock can be
 555                          * either held in exclusive mode or shared mode
 556                          * (for the writer starvation avoidance technique).
 557                          */
 558                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 559                             LK_HOLDER(x) != LK_KERNPROC) {
 560                                 owner = (struct thread *)LK_HOLDER(x);
 561                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 562                                         CTR3(KTR_LOCK,
 563                                             "%s: spinning on %p held by %p",
 564                                             __func__, lk, owner);
 565
 566                                 /*
 567                                  * If we are holding also an interlock drop it
 568                                  * in order to avoid a deadlock if the lockmgr
 569                                  * owner is adaptively spinning on the
 570                                  * interlock itself.
 571                                  */
 572                                 if (flags & LK_INTERLOCK) {
 573                                         class->lc_unlock(ilk);
 574                                         flags &= ~LK_INTERLOCK;
 575                                 }
 576                                 GIANT_SAVE();
 577                                 while (LK_HOLDER(lk->lk_lock) ==
 578                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 579                                         cpu_spinwait();
 580                                 GIANT_RESTORE();
 581                                 continue;
 582                         } else if (LK_CAN_ADAPT(lk, flags) &&
 583                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 584                             spintries < ALK_RETRIES) {
 585                                 if (flags & LK_INTERLOCK) {
 586                                         class->lc_unlock(ilk);
 587                                         flags &= ~LK_INTERLOCK;
 588                                 }
 589                                 GIANT_SAVE();
 590                                 spintries++;
 591                                 for (i = 0; i < ALK_LOOPS; i++) {
 592                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 593                                                 CTR4(KTR_LOCK,
 594                                     "%s: shared spinning on %p with %u and %u",
 595                                                     __func__, lk, spintries, i);
 596                                         x = lk->lk_lock;
 597                                         if ((x & LK_SHARE) == 0 ||
 598                                             LK_CAN_SHARE(x) != 0)
 599                                                 break;
 600                                         cpu_spinwait();
 601                                 }
 602                                 GIANT_RESTORE();
 603                                 if (i != ALK_LOOPS)
 604                                         continue;
 605                         }
 606 #endif
 607
 608                         /*
 609                          * Acquire the sleepqueue chain lock because we
 610                          * probabilly will need to manipulate waiters flags.
 611                          */
 612                         sleepq_lock(&lk->lock_object);
 613                         x = lk->lk_lock;
 614
 615                         /*
 616                          * if the lock can be acquired in shared mode, try
 617                          * again.
 618                          */
 619                         if (LK_CAN_SHARE(x)) {
 620                                 sleepq_release(&lk->lock_object);
 621                                 continue;
 622                         }
 623
 624 #ifdef ADAPTIVE_LOCKMGRS
 625                         /*
 626                          * The current lock owner might have started executing
 627                          * on another CPU (or the lock could have changed
 628                          * owner) while we were waiting on the turnstile
 629                          * chain lock.  If so, drop the turnstile lock and try
 630                          * again.
 631                          */
 632                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 633                             LK_HOLDER(x) != LK_KERNPROC) {
 634                                 owner = (struct thread *)LK_HOLDER(x);
 635                                 if (TD_IS_RUNNING(owner)) {
 636                                         sleepq_release(&lk->lock_object);
 637                                         continue;
 638                                 }
 639                         }
 640 #endif
 641
 642                         /*
 643                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 644                          * loop back and retry.
 645                          */
 646                         if ((x & LK_SHARED_WAITERS) == 0) {
 647                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 648                                     x | LK_SHARED_WAITERS)) {
 649                                         sleepq_release(&lk->lock_object);
 650                                         continue;
 651                                 }
 652                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 653                                     __func__, lk);
 654                         }
 655
 656                         /*
 657                          * As far as we have been unable to acquire the
 658                          * shared lock and the shared waiters flag is set,
 659                          * we will sleep.
 660                          */
 661                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 662                             SQ_SHARED_QUEUE);
 663                         flags &= ~LK_INTERLOCK;
 664                         if (error) {
 665                                 LOCK_LOG3(lk,
 666                                     "%s: interrupted sleep for %p with %d",
 667                                     __func__, lk, error);
 668                                 break;
 669                         }
 670                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 671                             __func__, lk);
 672                 }
 673                 if (error == 0) {
 674                         lock_profile_obtain_lock_success(&lk->lock_object,
 675                             contested, waittime, file, line);
 676                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 677                             line);
 678                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 679                             line);
 680                         TD_LOCKS_INC(curthread);
 681                         TD_SLOCKS_INC(curthread);
 682                         STACK_SAVE(lk);
 683                 }
 684                 break;
 685         case LK_UPGRADE:
 686                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 687                 v = lk->lk_lock;
 688                 x = v & LK_ALL_WAITERS;
 689                 v &= LK_EXCLUSIVE_SPINNERS;
 690
 691                 /*
 692                  * Try to switch from one shared lock to an exclusive one.
 693                  * We need to preserve waiters flags during the operation.
 694                  */
 695                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 696                     tid | x)) {
 697                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 698                             line);
 699                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 700                             LK_TRYWIT(flags), file, line);
 701                         TD_SLOCKS_DEC(curthread);
 702                         break;
 703                 }
 704
 705                 /*
 706                  * We have been unable to succeed in upgrading, so just
 707                  * give up the shared lock.
 708                  */
 709                 wakeup_swapper |= wakeupshlk(lk, file, line);
 710
 711                 /* FALLTHROUGH */
 712         case LK_EXCLUSIVE:
 713                 if (LK_CAN_WITNESS(flags))
 714                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 715                             LOP_EXCLUSIVE, file, line, ilk);
 716
 717                 /*
 718                  * If curthread already holds the lock and this one is
 719                  * allowed to recurse, simply recurse on it.
 720                  */
 721                 if (lockmgr_xlocked(lk)) {
 722                         if ((flags & LK_CANRECURSE) == 0 &&
 723                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 724
 725                                 /*
 726                                  * If the lock is expected to not panic just
 727                                  * give up and return.
 728                                  */
 729                                 if (LK_TRYOP(flags)) {
 730                                         LOCK_LOG2(lk,
 731                                             "%s: %p fails the try operation",
 732                                             __func__, lk);
 733                                         error = EBUSY;
 734                                         break;
 735                                 }
 736                                 if (flags & LK_INTERLOCK)
 737                                         class->lc_unlock(ilk);
 738                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 739                                     __func__, iwmesg, file, line);
 740                         }
 741                         lk->lk_recurse++;
 742                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 743                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 744                             lk->lk_recurse, file, line);
 745                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 746                             LK_TRYWIT(flags), file, line);
 747                         TD_LOCKS_INC(curthread);
 748                         break;
 749                 }
 750
 751                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
 752                     tid)) {
 753 #ifdef HWPMC_HOOKS
 754                         PMC_SOFT_CALL( , , lock, failed);
 755 #endif
 756                         lock_profile_obtain_lock_failed(&lk->lock_object,
 757                             &contested, &waittime);
 758
 759                         /*
 760                          * If the lock is expected to not sleep just give up
 761                          * and return.
 762                          */
 763                         if (LK_TRYOP(flags)) {
 764                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 765                                     __func__, lk);
 766                                 error = EBUSY;
 767                                 break;
 768                         }
 769
 770 #ifdef ADAPTIVE_LOCKMGRS
 771                         /*
 772                          * If the owner is running on another CPU, spin until
 773                          * the owner stops running or the state of the lock
 774                          * changes.
 775                          */
 776                         x = lk->lk_lock;
 777                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 778                             LK_HOLDER(x) != LK_KERNPROC) {
 779                                 owner = (struct thread *)LK_HOLDER(x);
 780                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 781                                         CTR3(KTR_LOCK,
 782                                             "%s: spinning on %p held by %p",
 783                                             __func__, lk, owner);
 784
 785                                 /*
 786                                  * If we are holding also an interlock drop it
 787                                  * in order to avoid a deadlock if the lockmgr
 788                                  * owner is adaptively spinning on the
 789                                  * interlock itself.
 790                                  */
 791                                 if (flags & LK_INTERLOCK) {
 792                                         class->lc_unlock(ilk);
 793                                         flags &= ~LK_INTERLOCK;
 794                                 }
 795                                 GIANT_SAVE();
 796                                 while (LK_HOLDER(lk->lk_lock) ==
 797                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 798                                         cpu_spinwait();
 799                                 GIANT_RESTORE();
 800                                 continue;
 801                         } else if (LK_CAN_ADAPT(lk, flags) &&
 802                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 803                             spintries < ALK_RETRIES) {
 804                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 805                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 806                                     x | LK_EXCLUSIVE_SPINNERS))
 807                                         continue;
 808                                 if (flags & LK_INTERLOCK) {
 809                                         class->lc_unlock(ilk);
 810                                         flags &= ~LK_INTERLOCK;
 811                                 }
 812                                 GIANT_SAVE();
 813                                 spintries++;
 814                                 for (i = 0; i < ALK_LOOPS; i++) {
 815                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 816                                                 CTR4(KTR_LOCK,
 817                                     "%s: shared spinning on %p with %u and %u",
 818                                                     __func__, lk, spintries, i);
 819                                         if ((lk->lk_lock &
 820                                             LK_EXCLUSIVE_SPINNERS) == 0)
 821                                                 break;
 822                                         cpu_spinwait();
 823                                 }
 824                                 GIANT_RESTORE();
 825                                 if (i != ALK_LOOPS)
 826                                         continue;
 827                         }
 828 #endif
 829
 830                         /*
 831                          * Acquire the sleepqueue chain lock because we
 832                          * probabilly will need to manipulate waiters flags.
 833                          */
 834                         sleepq_lock(&lk->lock_object);
 835                         x = lk->lk_lock;
 836
 837                         /*
 838                          * if the lock has been released while we spun on
 839                          * the sleepqueue chain lock just try again.
 840                          */
 841                         if (x == LK_UNLOCKED) {
 842                                 sleepq_release(&lk->lock_object);
 843                                 continue;
 844                         }
 845
 846 #ifdef ADAPTIVE_LOCKMGRS
 847                         /*
 848                          * The current lock owner might have started executing
 849                          * on another CPU (or the lock could have changed
 850                          * owner) while we were waiting on the turnstile
 851                          * chain lock.  If so, drop the turnstile lock and try
 852                          * again.
 853                          */
 854                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 855                             LK_HOLDER(x) != LK_KERNPROC) {
 856                                 owner = (struct thread *)LK_HOLDER(x);
 857                                 if (TD_IS_RUNNING(owner)) {
 858                                         sleepq_release(&lk->lock_object);
 859                                         continue;
 860                                 }
 861                         }
 862 #endif
 863
 864                         /*
 865                          * The lock can be in the state where there is a
 866                          * pending queue of waiters, but still no owner.
 867                          * This happens when the lock is contested and an
 868                          * owner is going to claim the lock.
 869                          * If curthread is the one successfully acquiring it
 870                          * claim lock ownership and return, preserving waiters
 871                          * flags.
 872                          */
 873                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 874                         if ((x & ~v) == LK_UNLOCKED) {
 875                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 876                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 877                                     tid | v)) {
 878                                         sleepq_release(&lk->lock_object);
 879                                         LOCK_LOG2(lk,
 880                                             "%s: %p claimed by a new writer",
 881                                             __func__, lk);
 882                                         break;
 883                                 }
 884                                 sleepq_release(&lk->lock_object);
 885                                 continue;
 886                         }
 887
 888                         /*
 889                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 890                          * fail, loop back and retry.
 891                          */
 892                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 893                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 894                                     x | LK_EXCLUSIVE_WAITERS)) {
 895                                         sleepq_release(&lk->lock_object);
 896                                         continue;
 897                                 }
 898                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 899                                     __func__, lk);
 900                         }
 901
 902                         /*
 903                          * As far as we have been unable to acquire the
 904                          * exclusive lock and the exclusive waiters flag
 905                          * is set, we will sleep.
 906                          */
 907                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 908                             SQ_EXCLUSIVE_QUEUE);
 909                         flags &= ~LK_INTERLOCK;
 910                         if (error) {
 911                                 LOCK_LOG3(lk,
 912                                     "%s: interrupted sleep for %p with %d",
 913                                     __func__, lk, error);
 914                                 break;
 915                         }
 916                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 917                             __func__, lk);
 918                 }
 919                 if (error == 0) {
 920                         lock_profile_obtain_lock_success(&lk->lock_object,
 921                             contested, waittime, file, line);
 922                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 923                             lk->lk_recurse, file, line);
 924                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 925                             LK_TRYWIT(flags), file, line);
 926                         TD_LOCKS_INC(curthread);
 927                         STACK_SAVE(lk);
 928                 }
 929                 break;
 930         case LK_DOWNGRADE:
 931                 _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line);
 932                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 933                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 934                 TD_SLOCKS_INC(curthread);
 935
 936                 /*
 937                  * In order to preserve waiters flags, just spin.
 938                  */
 939                 for (;;) {
 940                         x = lk->lk_lock;
 941                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
 942                         x &= LK_ALL_WAITERS;
 943                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
 944                             LK_SHARERS_LOCK(1) | x))
 945                                 break;
 946                         cpu_spinwait();
 947                 }
 948                 break;
 949         case LK_RELEASE:
 950                 _lockmgr_assert(lk, KA_LOCKED, file, line);
 951                 x = lk->lk_lock;
 952
 953                 if ((x & LK_SHARE) == 0) {
 954
 955                         /*
 956                          * As first option, treact the lock as if it has not
 957                          * any waiter.
 958                          * Fix-up the tid var if the lock has been disowned.
 959                          */
 960                         if (LK_HOLDER(x) == LK_KERNPROC)
 961                                 tid = LK_KERNPROC;
 962                         else {
 963                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
 964                                     file, line);
 965                                 TD_LOCKS_DEC(curthread);
 966                         }
 967                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
 968                             lk->lk_recurse, file, line);
 969
 970                         /*
 971                          * The lock is held in exclusive mode.
 972                          * If the lock is recursed also, then unrecurse it.
 973                          */
 974                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 975                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
 976                                     lk);
 977                                 lk->lk_recurse--;
 978                                 break;
 979                         }
 980                         if (tid != LK_KERNPROC)
 981                                 lock_profile_release_lock(&lk->lock_object);
 982
 983                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
 984                             LK_UNLOCKED))
 985                                 break;
 986
 987                         sleepq_lock(&lk->lock_object);
 988                         x = lk->lk_lock;
 989                         v = LK_UNLOCKED;
 990
 991                         /*
 992                          * If the lock has exclusive waiters, give them
 993                          * preference in order to avoid deadlock with
 994                          * shared runners up.
 995                          * If interruptible sleeps left the exclusive queue
 996                          * empty avoid a starvation for the threads sleeping
 997                          * on the shared queue by giving them precedence
 998                          * and cleaning up the exclusive waiters bit anyway.
 999                          * Please note that lk_exslpfail count may be lying
1000                          * about the real number of waiters with the
1001                          * LK_SLEEPFAIL flag on because they may be used in
1002                          * conjuction with interruptible sleeps so
1003                          * lk_exslpfail might be considered an 'upper limit'
1004                          * bound, including the edge cases.
1005                          */
1006                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1007                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1008                             SQ_EXCLUSIVE_QUEUE);
1009                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1010                                 if (lk->lk_exslpfail < realexslp) {
1011                                         lk->lk_exslpfail = 0;
1012                                         queue = SQ_EXCLUSIVE_QUEUE;
1013                                         v |= (x & LK_SHARED_WAITERS);
1014                                 } else {
1015                                         lk->lk_exslpfail = 0;
1016                                         LOCK_LOG2(lk,
1017                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1018                                             __func__, lk);
1019                                         LOCK_LOG2(lk,
1020                         "%s: %p waking up threads on the exclusive queue",
1021                                             __func__, lk);
1022                                         wakeup_swapper =
1023                                             sleepq_broadcast(&lk->lock_object,
1024                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1025                                         queue = SQ_SHARED_QUEUE;
1026                                 }
1027                         } else {
1028
1029                                 /*
1030                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1031                                  * on and using interruptible sleeps/timeout
1032                                  * may have left spourious lk_exslpfail counts
1033                                  * on, so clean it up anyway.
1034                                  */
1035                                 lk->lk_exslpfail = 0;
1036                                 queue = SQ_SHARED_QUEUE;
1037                         }
1038
1039                         LOCK_LOG3(lk,
1040                             "%s: %p waking up threads on the %s queue",
1041                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1042                             "exclusive");
1043                         atomic_store_rel_ptr(&lk->lk_lock, v);
1044                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1045                             SLEEPQ_LK, 0, queue);
1046                         sleepq_release(&lk->lock_object);
1047                         break;
1048                 } else
1049                         wakeup_swapper = wakeupshlk(lk, file, line);
1050                 break;
1051         case LK_DRAIN:
1052                 if (LK_CAN_WITNESS(flags))
1053                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1054                             LOP_EXCLUSIVE, file, line, ilk);
1055
1056                 /*
1057                  * Trying to drain a lock we already own will result in a
1058                  * deadlock.
1059                  */
1060                 if (lockmgr_xlocked(lk)) {
1061                         if (flags & LK_INTERLOCK)
1062                                 class->lc_unlock(ilk);
1063                         panic("%s: draining %s with the lock held @ %s:%d\n",
1064                             __func__, iwmesg, file, line);
1065                 }
1066
1067                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1068 #ifdef HWPMC_HOOKS
1069                         PMC_SOFT_CALL( , , lock, failed);
1070 #endif
1071                         lock_profile_obtain_lock_failed(&lk->lock_object,
1072                             &contested, &waittime);
1073
1074                         /*
1075                          * If the lock is expected to not sleep just give up
1076                          * and return.
1077                          */
1078                         if (LK_TRYOP(flags)) {
1079                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1080                                     __func__, lk);
1081                                 error = EBUSY;
1082                                 break;
1083                         }
1084
1085                         /*
1086                          * Acquire the sleepqueue chain lock because we
1087                          * probabilly will need to manipulate waiters flags.
1088                          */
1089                         sleepq_lock(&lk->lock_object);
1090                         x = lk->lk_lock;
1091
1092                         /*
1093                          * if the lock has been released while we spun on
1094                          * the sleepqueue chain lock just try again.
1095                          */
1096                         if (x == LK_UNLOCKED) {
1097                                 sleepq_release(&lk->lock_object);
1098                                 continue;
1099                         }
1100
1101                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1102                         if ((x & ~v) == LK_UNLOCKED) {
1103                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1104
1105                                 /*
1106                                  * If interruptible sleeps left the exclusive
1107                                  * queue empty avoid a starvation for the
1108                                  * threads sleeping on the shared queue by
1109                                  * giving them precedence and cleaning up the
1110                                  * exclusive waiters bit anyway.
1111                                  * Please note that lk_exslpfail count may be
1112                                  * lying about the real number of waiters with
1113                                  * the LK_SLEEPFAIL flag on because they may
1114                                  * be used in conjuction with interruptible
1115                                  * sleeps so lk_exslpfail might be considered
1116                                  * an 'upper limit' bound, including the edge
1117                                  * cases.
1118                                  */
1119                                 if (v & LK_EXCLUSIVE_WAITERS) {
1120                                         queue = SQ_EXCLUSIVE_QUEUE;
1121                                         v &= ~LK_EXCLUSIVE_WAITERS;
1122                                 } else {
1123
1124                                         /*
1125                                          * Exclusive waiters sleeping with
1126                                          * LK_SLEEPFAIL on and using
1127                                          * interruptible sleeps/timeout may
1128                                          * have left spourious lk_exslpfail
1129                                          * counts on, so clean it up anyway.
1130                                          */
1131                                         MPASS(v & LK_SHARED_WAITERS);
1132                                         lk->lk_exslpfail = 0;
1133                                         queue = SQ_SHARED_QUEUE;
1134                                         v &= ~LK_SHARED_WAITERS;
1135                                 }
1136                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1137                                         realexslp =
1138                                             sleepq_sleepcnt(&lk->lock_object,
1139                                             SQ_EXCLUSIVE_QUEUE);
1140                                         if (lk->lk_exslpfail >= realexslp) {
1141                                                 lk->lk_exslpfail = 0;
1142                                                 queue = SQ_SHARED_QUEUE;
1143                                                 v &= ~LK_SHARED_WAITERS;
1144                                                 if (realexslp != 0) {
1145                                                         LOCK_LOG2(lk,
1146                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1147                                                             __func__, lk);
1148                                                         LOCK_LOG2(lk,
1149                         "%s: %p waking up threads on the exclusive queue",
1150                                                             __func__, lk);
1151                                                         wakeup_swapper =
1152                                                             sleepq_broadcast(
1153                                                             &lk->lock_object,
1154                                                             SLEEPQ_LK, 0,
1155                                                             SQ_EXCLUSIVE_QUEUE);
1156                                                 }
1157                                         } else
1158                                                 lk->lk_exslpfail = 0;
1159                                 }
1160                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1161                                         sleepq_release(&lk->lock_object);
1162                                         continue;
1163                                 }
1164                                 LOCK_LOG3(lk,
1165                                 "%s: %p waking up all threads on the %s queue",
1166                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1167                                     "shared" : "exclusive");
1168                                 wakeup_swapper |= sleepq_broadcast(
1169                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1170
1171                                 /*
1172                                  * If shared waiters have been woken up we need
1173                                  * to wait for one of them to acquire the lock
1174                                  * before to set the exclusive waiters in
1175                                  * order to avoid a deadlock.
1176                                  */
1177                                 if (queue == SQ_SHARED_QUEUE) {
1178                                         for (v = lk->lk_lock;
1179                                             (v & LK_SHARE) && !LK_SHARERS(v);
1180                                             v = lk->lk_lock)
1181                                                 cpu_spinwait();
1182                                 }
1183                         }
1184
1185                         /*
1186                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1187                          * fail, loop back and retry.
1188                          */
1189                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1190                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1191                                     x | LK_EXCLUSIVE_WAITERS)) {
1192                                         sleepq_release(&lk->lock_object);
1193                                         continue;
1194                                 }
1195                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1196                                     __func__, lk);
1197                         }
1198
1199                         /*
1200                          * As far as we have been unable to acquire the
1201                          * exclusive lock and the exclusive waiters flag
1202                          * is set, we will sleep.
1203                          */
1204                         if (flags & LK_INTERLOCK) {
1205                                 class->lc_unlock(ilk);
1206                                 flags &= ~LK_INTERLOCK;
1207                         }
1208                         GIANT_SAVE();
1209                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1210                             SQ_EXCLUSIVE_QUEUE);
1211                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1212                         GIANT_RESTORE();
1213                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1214                             __func__, lk);
1215                 }
1216
1217                 if (error == 0) {
1218                         lock_profile_obtain_lock_success(&lk->lock_object,
1219                             contested, waittime, file, line);
1220                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1221                             lk->lk_recurse, file, line);
1222                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1223                             LK_TRYWIT(flags), file, line);
1224                         TD_LOCKS_INC(curthread);
1225                         STACK_SAVE(lk);
1226                 }
1227                 break;
1228         default:
1229                 if (flags & LK_INTERLOCK)
1230                         class->lc_unlock(ilk);
1231                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1232         }
1233
1234         if (flags & LK_INTERLOCK)
1235                 class->lc_unlock(ilk);
1236         if (wakeup_swapper)
1237                 kick_proc0();
1238
1239         return (error);
1240 }
1241
1242 void
1243 _lockmgr_disown(struct lock *lk, const char *file, int line)
1244 {
1245         uintptr_t tid, x;
1246
1247         if (SCHEDULER_STOPPED())
1248                 return;
1249
1250         tid = (uintptr_t)curthread;
1251         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line);
1252
1253         /*
1254          * If the owner is already LK_KERNPROC just skip the whole operation.
1255          */
1256         if (LK_HOLDER(lk->lk_lock) != tid)
1257                 return;
1258         lock_profile_release_lock(&lk->lock_object);
1259         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1260         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1261         TD_LOCKS_DEC(curthread);
1262         STACK_SAVE(lk);
1263
1264         /*
1265          * In order to preserve waiters flags, just spin.
1266          */
1267         for (;;) {
1268                 x = lk->lk_lock;
1269                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1270                 x &= LK_ALL_WAITERS;
1271                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1272                     LK_KERNPROC | x))
1273                         return;
1274                 cpu_spinwait();
1275         }
1276 }
1277
1278 void
1279 lockmgr_printinfo(struct lock *lk)
1280 {
1281         struct thread *td;
1282         uintptr_t x;
1283
1284         if (lk->lk_lock == LK_UNLOCKED)
1285                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1286         else if (lk->lk_lock & LK_SHARE)
1287                 printf("lock type %s: SHARED (count %ju)\n",
1288                     lk->lock_object.lo_name,
1289                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1290         else {
1291                 td = lockmgr_xholder(lk);
1292                 printf("lock type %s: EXCL by thread %p (pid %d)\n",
1293                     lk->lock_object.lo_name, td, td->td_proc->p_pid);
1294         }
1295
1296         x = lk->lk_lock;
1297         if (x & LK_EXCLUSIVE_WAITERS)
1298                 printf(" with exclusive waiters pending\n");
1299         if (x & LK_SHARED_WAITERS)
1300                 printf(" with shared waiters pending\n");
1301         if (x & LK_EXCLUSIVE_SPINNERS)
1302                 printf(" with exclusive spinners pending\n");
1303
1304         STACK_PRINT(lk);
1305 }
1306
1307 int
1308 lockstatus(struct lock *lk)
1309 {
1310         uintptr_t v, x;
1311         int ret;
1312
1313         ret = LK_SHARED;
1314         x = lk->lk_lock;
1315         v = LK_HOLDER(x);
1316
1317         if ((x & LK_SHARE) == 0) {
1318                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1319                         ret = LK_EXCLUSIVE;
1320                 else
1321                         ret = LK_EXCLOTHER;
1322         } else if (x == LK_UNLOCKED)
1323                 ret = 0;
1324
1325         return (ret);
1326 }
1327
1328 #ifdef INVARIANT_SUPPORT
1329
1330 FEATURE(invariant_support,
1331     "Support for modules compiled with INVARIANTS option");
1332
1333 #ifndef INVARIANTS
1334 #undef  _lockmgr_assert
1335 #endif
1336
1337 void
1338 _lockmgr_assert(struct lock *lk, int what, const char *file, int line)
1339 {
1340         int slocked = 0;
1341
1342         if (panicstr != NULL)
1343                 return;
1344         switch (what) {
1345         case KA_SLOCKED:
1346         case KA_SLOCKED | KA_NOTRECURSED:
1347         case KA_SLOCKED | KA_RECURSED:
1348                 slocked = 1;
1349         case KA_LOCKED:
1350         case KA_LOCKED | KA_NOTRECURSED:
1351         case KA_LOCKED | KA_RECURSED:
1352 #ifdef WITNESS
1353
1354                 /*
1355                  * We cannot trust WITNESS if the lock is held in exclusive
1356                  * mode and a call to lockmgr_disown() happened.
1357                  * Workaround this skipping the check if the lock is held in
1358                  * exclusive mode even for the KA_LOCKED case.
1359                  */
1360                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1361                         witness_assert(&lk->lock_object, what, file, line);
1362                         break;
1363                 }
1364 #endif
1365                 if (lk->lk_lock == LK_UNLOCKED ||
1366                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1367                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1368                         panic("Lock %s not %slocked @ %s:%d\n",
1369                             lk->lock_object.lo_name, slocked ? "share" : "",
1370                             file, line);
1371
1372                 if ((lk->lk_lock & LK_SHARE) == 0) {
1373                         if (lockmgr_recursed(lk)) {
1374                                 if (what & KA_NOTRECURSED)
1375                                         panic("Lock %s recursed @ %s:%d\n",
1376                                             lk->lock_object.lo_name, file,
1377                                             line);
1378                         } else if (what & KA_RECURSED)
1379                                 panic("Lock %s not recursed @ %s:%d\n",
1380                                     lk->lock_object.lo_name, file, line);
1381                 }
1382                 break;
1383         case KA_XLOCKED:
1384         case KA_XLOCKED | KA_NOTRECURSED:
1385         case KA_XLOCKED | KA_RECURSED:
1386                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1387                         panic("Lock %s not exclusively locked @ %s:%d\n",
1388                             lk->lock_object.lo_name, file, line);
1389                 if (lockmgr_recursed(lk)) {
1390                         if (what & KA_NOTRECURSED)
1391                                 panic("Lock %s recursed @ %s:%d\n",
1392                                     lk->lock_object.lo_name, file, line);
1393                 } else if (what & KA_RECURSED)
1394                         panic("Lock %s not recursed @ %s:%d\n",
1395                             lk->lock_object.lo_name, file, line);
1396                 break;
1397         case KA_UNLOCKED:
1398                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1399                         panic("Lock %s exclusively locked @ %s:%d\n",
1400                             lk->lock_object.lo_name, file, line);
1401                 break;
1402         default:
1403                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1404                     line);
1405         }
1406 }
1407 #endif
1408
1409 #ifdef DDB
1410 int
1411 lockmgr_chain(struct thread *td, struct thread **ownerp)
1412 {
1413         struct lock *lk;
1414
1415         lk = td->td_wchan;
1416
1417         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1418                 return (0);
1419         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1420         if (lk->lk_lock & LK_SHARE)
1421                 db_printf("SHARED (count %ju)\n",
1422                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1423         else
1424                 db_printf("EXCL\n");
1425         *ownerp = lockmgr_xholder(lk);
1426
1427         return (1);
1428 }
1429
1430 static void
1431 db_show_lockmgr(struct lock_object *lock)
1432 {
1433         struct thread *td;
1434         struct lock *lk;
1435
1436         lk = (struct lock *)lock;
1437
1438         db_printf(" state: ");
1439         if (lk->lk_lock == LK_UNLOCKED)
1440                 db_printf("UNLOCKED\n");
1441         else if (lk->lk_lock & LK_SHARE)
1442                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1443         else {
1444                 td = lockmgr_xholder(lk);
1445                 if (td == (struct thread *)LK_KERNPROC)
1446                         db_printf("XLOCK: LK_KERNPROC\n");
1447                 else
1448                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1449                             td->td_tid, td->td_proc->p_pid,
1450                             td->td_proc->p_comm);
1451                 if (lockmgr_recursed(lk))
1452                         db_printf(" recursed: %d\n", lk->lk_recurse);
1453         }
1454         db_printf(" waiters: ");
1455         switch (lk->lk_lock & LK_ALL_WAITERS) {
1456         case LK_SHARED_WAITERS:
1457                 db_printf("shared\n");
1458                 break;
1459         case LK_EXCLUSIVE_WAITERS:
1460                 db_printf("exclusive\n");
1461                 break;
1462         case LK_ALL_WAITERS:
1463                 db_printf("shared and exclusive\n");
1464                 break;
1465         default:
1466                 db_printf("none\n");
1467         }
1468         db_printf(" spinners: ");
1469         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1470                 db_printf("exclusive\n");
1471         else
1472                 db_printf("none\n");
1473 }
1474 #endif