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