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