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