sys/kern/kern_lock.c

   1 /*-
   2  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice(s), this list of conditions and the following disclaimer as
  10  *    the first lines of this file unmodified other than the possible
  11  *    addition of one or more copyright notices.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice(s), this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  26  * DAMAGE.
  27  */
  28
  29 #include "opt_adaptive_lockmgrs.h"
  30 #include "opt_ddb.h"
  31 #include "opt_hwpmc_hooks.h"
  32 #include "opt_kdtrace.h"
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include <sys/param.h>
  38 #include <sys/kdb.h>
  39 #include <sys/ktr.h>
  40 #include <sys/lock.h>
  41 #include <sys/lock_profile.h>
  42 #include <sys/lockmgr.h>
  43 #include <sys/mutex.h>
  44 #include <sys/proc.h>
  45 #include <sys/sleepqueue.h>
  46 #ifdef DEBUG_LOCKS
  47 #include <sys/stack.h>
  48 #endif
  49 #include <sys/sysctl.h>
  50 #include <sys/systm.h>
  51
  52 #include <machine/cpu.h>
  53
  54 #ifdef DDB
  55 #include <ddb/ddb.h>
  56 #endif
  57
  58 #ifdef HWPMC_HOOKS
  59 #include <sys/pmckern.h>
  60 PMC_SOFT_DECLARE( , , lock, failed);
  61 #endif
  62
  63 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
  64     (LK_ADAPTIVE | LK_NOSHARE));
  65 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
  66     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
  67
  68 #define SQ_EXCLUSIVE_QUEUE      0
  69 #define SQ_SHARED_QUEUE         1
  70
  71 #ifndef INVARIANTS
  72 #define _lockmgr_assert(lk, what, file, line)
  73 #define TD_LOCKS_INC(td)
  74 #define TD_LOCKS_DEC(td)
  75 #else
  76 #define TD_LOCKS_INC(td)        ((td)->td_locks++)
  77 #define TD_LOCKS_DEC(td)        ((td)->td_locks--)
  78 #endif
  79 #define TD_SLOCKS_INC(td)       ((td)->td_lk_slocks++)
  80 #define TD_SLOCKS_DEC(td)       ((td)->td_lk_slocks--)
  81
  82 #ifndef DEBUG_LOCKS
  83 #define STACK_PRINT(lk)
  84 #define STACK_SAVE(lk)
  85 #define STACK_ZERO(lk)
  86 #else
  87 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
  88 #define STACK_SAVE(lk)  stack_save(&(lk)->lk_stack)
  89 #define STACK_ZERO(lk)  stack_zero(&(lk)->lk_stack)
  90 #endif
  91
  92 #define LOCK_LOG2(lk, string, arg1, arg2)                               \
  93         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  94                 CTR2(KTR_LOCK, (string), (arg1), (arg2))
  95 #define LOCK_LOG3(lk, string, arg1, arg2, arg3)                         \
  96         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  97                 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
  98
  99 #define GIANT_DECLARE                                                   \
 100         int _i = 0;                                                     \
 101         WITNESS_SAVE_DECL(Giant)
 102 #define GIANT_RESTORE() do {                                            \
 103         if (_i > 0) {                                                   \
 104                 while (_i--)                                            \
 105                         mtx_lock(&Giant);                               \
 106                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 107         }                                                               \
 108 } while (0)
 109 #define GIANT_SAVE() do {                                               \
 110         if (mtx_owned(&Giant)) {                                        \
 111                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
 112                 while (mtx_owned(&Giant)) {                             \
 113                         _i++;                                           \
 114                         mtx_unlock(&Giant);                             \
 115                 }                                                       \
 116         }                                                               \
 117 } while (0)
 118
 119 #define LK_CAN_SHARE(x)                                                 \
 120         (((x) & LK_SHARE) && (((x) & LK_EXCLUSIVE_WAITERS) == 0 ||      \
 121         ((x) & LK_EXCLUSIVE_SPINNERS) == 0 ||                           \
 122         curthread->td_lk_slocks || (curthread->td_pflags & TDP_DEADLKTREAT)))
 123 #define LK_TRYOP(x)                                                     \
 124         ((x) & LK_NOWAIT)
 125
 126 #define LK_CAN_WITNESS(x)                                               \
 127         (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
 128 #define LK_TRYWIT(x)                                                    \
 129         (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
 130
 131 #define LK_CAN_ADAPT(lk, f)                                             \
 132         (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&             \
 133         ((f) & LK_SLEEPFAIL) == 0)
 134
 135 #define lockmgr_disowned(lk)                                            \
 136         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
 137
 138 #define lockmgr_xlocked(lk)                                             \
 139         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
 140
 141 static void     assert_lockmgr(const struct lock_object *lock, int how);
 142 #ifdef DDB
 143 static void     db_show_lockmgr(const struct lock_object *lock);
 144 #endif
 145 static void     lock_lockmgr(struct lock_object *lock, int how);
 146 #ifdef KDTRACE_HOOKS
 147 static int      owner_lockmgr(const struct lock_object *lock,
 148                     struct thread **owner);
 149 #endif
 150 static int      unlock_lockmgr(struct lock_object *lock);
 151
 152 struct lock_class lock_class_lockmgr = {
 153         .lc_name = "lockmgr",
 154         .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
 155         .lc_assert = assert_lockmgr,
 156 #ifdef DDB
 157         .lc_ddb_show = db_show_lockmgr,
 158 #endif
 159         .lc_lock = lock_lockmgr,
 160         .lc_unlock = unlock_lockmgr,
 161 #ifdef KDTRACE_HOOKS
 162         .lc_owner = owner_lockmgr,
 163 #endif
 164 };
 165
 166 #ifdef ADAPTIVE_LOCKMGRS
 167 static u_int alk_retries = 10;
 168 static u_int alk_loops = 10000;
 169 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL,
 170     "lockmgr debugging");
 171 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
 172 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
 173 #endif
 174
 175 static __inline struct thread *
 176 lockmgr_xholder(const struct lock *lk)
 177 {
 178         uintptr_t x;
 179
 180         x = lk->lk_lock;
 181         return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
 182 }
 183
 184 /*
 185  * It assumes sleepq_lock held and returns with this one unheld.
 186  * It also assumes the generic interlock is sane and previously checked.
 187  * If LK_INTERLOCK is specified the interlock is not reacquired after the
 188  * sleep.
 189  */
 190 static __inline int
 191 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
 192     const char *wmesg, int pri, int timo, int queue)
 193 {
 194         GIANT_DECLARE;
 195         struct lock_class *class;
 196         int catch, error;
 197
 198         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 199         catch = pri & PCATCH;
 200         pri &= PRIMASK;
 201         error = 0;
 202
 203         LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
 204             (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
 205
 206         if (flags & LK_INTERLOCK)
 207                 class->lc_unlock(ilk);
 208         if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
 209                 lk->lk_exslpfail++;
 210         GIANT_SAVE();
 211         sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
 212             SLEEPQ_INTERRUPTIBLE : 0), queue);
 213         if ((flags & LK_TIMELOCK) && timo)
 214                 sleepq_set_timeout(&lk->lock_object, timo);
 215
 216         /*
 217          * Decisional switch for real sleeping.
 218          */
 219         if ((flags & LK_TIMELOCK) && timo && catch)
 220                 error = sleepq_timedwait_sig(&lk->lock_object, pri);
 221         else if ((flags & LK_TIMELOCK) && timo)
 222                 error = sleepq_timedwait(&lk->lock_object, pri);
 223         else if (catch)
 224                 error = sleepq_wait_sig(&lk->lock_object, pri);
 225         else
 226                 sleepq_wait(&lk->lock_object, pri);
 227         GIANT_RESTORE();
 228         if ((flags & LK_SLEEPFAIL) && error == 0)
 229                 error = ENOLCK;
 230
 231         return (error);
 232 }
 233
 234 static __inline int
 235 wakeupshlk(struct lock *lk, const char *file, int line)
 236 {
 237         uintptr_t v, x;
 238         u_int realexslp;
 239         int queue, wakeup_swapper;
 240
 241         TD_LOCKS_DEC(curthread);
 242         TD_SLOCKS_DEC(curthread);
 243         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 244         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 245
 246         wakeup_swapper = 0;
 247         for (;;) {
 248                 x = lk->lk_lock;
 249
 250                 /*
 251                  * If there is more than one shared lock held, just drop one
 252                  * and return.
 253                  */
 254                 if (LK_SHARERS(x) > 1) {
 255                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
 256                             x - LK_ONE_SHARER))
 257                                 break;
 258                         continue;
 259                 }
 260
 261                 /*
 262                  * If there are not waiters on the exclusive queue, drop the
 263                  * lock quickly.
 264                  */
 265                 if ((x & LK_ALL_WAITERS) == 0) {
 266                         MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
 267                             LK_SHARERS_LOCK(1));
 268                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
 269                                 break;
 270                         continue;
 271                 }
 272
 273                 /*
 274                  * We should have a sharer with waiters, so enter the hard
 275                  * path in order to handle wakeups correctly.
 276                  */
 277                 sleepq_lock(&lk->lock_object);
 278                 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 279                 v = LK_UNLOCKED;
 280
 281                 /*
 282                  * If the lock has exclusive waiters, give them preference in
 283                  * order to avoid deadlock with shared runners up.
 284                  * If interruptible sleeps left the exclusive queue empty
 285                  * avoid a starvation for the threads sleeping on the shared
 286                  * queue by giving them precedence and cleaning up the
 287                  * exclusive waiters bit anyway.
 288                  * Please note that lk_exslpfail count may be lying about
 289                  * the real number of waiters with the LK_SLEEPFAIL flag on
 290                  * because they may be used in conjuction with interruptible
 291                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 292                  * bound, including the edge cases.
 293                  */
 294                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 295                     SQ_EXCLUSIVE_QUEUE);
 296                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 297                         if (lk->lk_exslpfail < realexslp) {
 298                                 lk->lk_exslpfail = 0;
 299                                 queue = SQ_EXCLUSIVE_QUEUE;
 300                                 v |= (x & LK_SHARED_WAITERS);
 301                         } else {
 302                                 lk->lk_exslpfail = 0;
 303                                 LOCK_LOG2(lk,
 304                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 305                                     __func__, lk);
 306                                 LOCK_LOG2(lk,
 307                             "%s: %p waking up threads on the exclusive queue",
 308                                     __func__, lk);
 309                                 wakeup_swapper =
 310                                     sleepq_broadcast(&lk->lock_object,
 311                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 312                                 queue = SQ_SHARED_QUEUE;
 313                         }
 314
 315                 } else {
 316
 317                         /*
 318                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 319                          * and using interruptible sleeps/timeout may have
 320                          * left spourious lk_exslpfail counts on, so clean
 321                          * it up anyway.
 322                          */
 323                         lk->lk_exslpfail = 0;
 324                         queue = SQ_SHARED_QUEUE;
 325                 }
 326
 327                 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
 328                     v)) {
 329                         sleepq_release(&lk->lock_object);
 330                         continue;
 331                 }
 332                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 333                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 334                     "exclusive");
 335                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 336                     0, queue);
 337                 sleepq_release(&lk->lock_object);
 338                 break;
 339         }
 340
 341         lock_profile_release_lock(&lk->lock_object);
 342         return (wakeup_swapper);
 343 }
 344
 345 static void
 346 assert_lockmgr(const struct lock_object *lock, int what)
 347 {
 348
 349         panic("lockmgr locks do not support assertions");
 350 }
 351
 352 static void
 353 lock_lockmgr(struct lock_object *lock, int how)
 354 {
 355
 356         panic("lockmgr locks do not support sleep interlocking");
 357 }
 358
 359 static int
 360 unlock_lockmgr(struct lock_object *lock)
 361 {
 362
 363         panic("lockmgr locks do not support sleep interlocking");
 364 }
 365
 366 #ifdef KDTRACE_HOOKS
 367 static int
 368 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
 369 {
 370
 371         panic("lockmgr locks do not support owner inquiring");
 372 }
 373 #endif
 374
 375 void
 376 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 377 {
 378         int iflags;
 379
 380         MPASS((flags & ~LK_INIT_MASK) == 0);
 381         ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
 382             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
 383             &lk->lk_lock));
 384
 385         iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 386         if (flags & LK_CANRECURSE)
 387                 iflags |= LO_RECURSABLE;
 388         if ((flags & LK_NODUP) == 0)
 389                 iflags |= LO_DUPOK;
 390         if (flags & LK_NOPROFILE)
 391                 iflags |= LO_NOPROFILE;
 392         if ((flags & LK_NOWITNESS) == 0)
 393                 iflags |= LO_WITNESS;
 394         if (flags & LK_QUIET)
 395                 iflags |= LO_QUIET;
 396         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 397
 398         lk->lk_lock = LK_UNLOCKED;
 399         lk->lk_recurse = 0;
 400         lk->lk_exslpfail = 0;
 401         lk->lk_timo = timo;
 402         lk->lk_pri = pri;
 403         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 404         STACK_ZERO(lk);
 405 }
 406
 407 /*
 408  * XXX: Gross hacks to manipulate external lock flags after
 409  * initialization.  Used for certain vnode and buf locks.
 410  */
 411 void
 412 lockallowshare(struct lock *lk)
 413 {
 414
 415         lockmgr_assert(lk, KA_XLOCKED);
 416         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 417 }
 418
 419 void
 420 lockallowrecurse(struct lock *lk)
 421 {
 422
 423         lockmgr_assert(lk, KA_XLOCKED);
 424         lk->lock_object.lo_flags |= LO_RECURSABLE;
 425 }
 426
 427 void
 428 lockdisablerecurse(struct lock *lk)
 429 {
 430
 431         lockmgr_assert(lk, KA_XLOCKED);
 432         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 433 }
 434
 435 void
 436 lockdestroy(struct lock *lk)
 437 {
 438
 439         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 440         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 441         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 442         lock_destroy(&lk->lock_object);
 443 }
 444
 445 int
 446 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 447     const char *wmesg, int pri, int timo, const char *file, int line)
 448 {
 449         GIANT_DECLARE;
 450         struct lock_class *class;
 451         const char *iwmesg;
 452         uintptr_t tid, v, x;
 453         u_int op, realexslp;
 454         int error, ipri, itimo, queue, wakeup_swapper;
 455 #ifdef LOCK_PROFILING
 456         uint64_t waittime = 0;
 457         int contested = 0;
 458 #endif
 459 #ifdef ADAPTIVE_LOCKMGRS
 460         volatile struct thread *owner;
 461         u_int i, spintries = 0;
 462 #endif
 463
 464         error = 0;
 465         tid = (uintptr_t)curthread;
 466         op = (flags & LK_TYPE_MASK);
 467         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 468         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 469         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 470
 471         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 472         KASSERT((op & (op - 1)) == 0,
 473             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 474         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 475             (op != LK_DOWNGRADE && op != LK_RELEASE),
 476             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 477             __func__, file, line));
 478         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 479             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 480             __func__, file, line));
 481         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 482             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
 483             lk->lock_object.lo_name, file, line));
 484
 485         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 486         if (panicstr != NULL) {
 487                 if (flags & LK_INTERLOCK)
 488                         class->lc_unlock(ilk);
 489                 return (0);
 490         }
 491
 492         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 493                 switch (op) {
 494                 case LK_SHARED:
 495                         op = LK_EXCLUSIVE;
 496                         break;
 497                 case LK_UPGRADE:
 498                 case LK_DOWNGRADE:
 499                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 500                             file, line);
 501                         if (flags & LK_INTERLOCK)
 502                                 class->lc_unlock(ilk);
 503                         return (0);
 504                 }
 505         }
 506
 507         wakeup_swapper = 0;
 508         switch (op) {
 509         case LK_SHARED:
 510                 if (LK_CAN_WITNESS(flags))
 511                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 512                             file, line, ilk);
 513                 for (;;) {
 514                         x = lk->lk_lock;
 515
 516                         /*
 517                          * If no other thread has an exclusive lock, or
 518                          * no exclusive waiter is present, bump the count of
 519                          * sharers.  Since we have to preserve the state of
 520                          * waiters, if we fail to acquire the shared lock
 521                          * loop back and retry.
 522                          */
 523                         if (LK_CAN_SHARE(x)) {
 524                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 525                                     x + LK_ONE_SHARER))
 526                                         break;
 527                                 continue;
 528                         }
 529 #ifdef HWPMC_HOOKS
 530                         PMC_SOFT_CALL( , , lock, failed);
 531 #endif
 532                         lock_profile_obtain_lock_failed(&lk->lock_object,
 533                             &contested, &waittime);
 534
 535                         /*
 536                          * If the lock is already held by curthread in
 537                          * exclusive way avoid a deadlock.
 538                          */
 539                         if (LK_HOLDER(x) == tid) {
 540                                 LOCK_LOG2(lk,
 541                                     "%s: %p already held in exclusive mode",
 542                                     __func__, lk);
 543                                 error = EDEADLK;
 544                                 break;
 545                         }
 546
 547                         /*
 548                          * If the lock is expected to not sleep just give up
 549                          * and return.
 550                          */
 551                         if (LK_TRYOP(flags)) {
 552                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 553                                     __func__, lk);
 554                                 error = EBUSY;
 555                                 break;
 556                         }
 557
 558 #ifdef ADAPTIVE_LOCKMGRS
 559                         /*
 560                          * If the owner is running on another CPU, spin until
 561                          * the owner stops running or the state of the lock
 562                          * changes.  We need a double-state handle here
 563                          * because for a failed acquisition the lock can be
 564                          * either held in exclusive mode or shared mode
 565                          * (for the writer starvation avoidance technique).
 566                          */
 567                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 568                             LK_HOLDER(x) != LK_KERNPROC) {
 569                                 owner = (struct thread *)LK_HOLDER(x);
 570                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 571                                         CTR3(KTR_LOCK,
 572                                             "%s: spinning on %p held by %p",
 573                                             __func__, lk, owner);
 574
 575                                 /*
 576                                  * If we are holding also an interlock drop it
 577                                  * in order to avoid a deadlock if the lockmgr
 578                                  * owner is adaptively spinning on the
 579                                  * interlock itself.
 580                                  */
 581                                 if (flags & LK_INTERLOCK) {
 582                                         class->lc_unlock(ilk);
 583                                         flags &= ~LK_INTERLOCK;
 584                                 }
 585                                 GIANT_SAVE();
 586                                 while (LK_HOLDER(lk->lk_lock) ==
 587                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 588                                         cpu_spinwait();
 589                                 GIANT_RESTORE();
 590                                 continue;
 591                         } else if (LK_CAN_ADAPT(lk, flags) &&
 592                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 593                             spintries < alk_retries) {
 594                                 if (flags & LK_INTERLOCK) {
 595                                         class->lc_unlock(ilk);
 596                                         flags &= ~LK_INTERLOCK;
 597                                 }
 598                                 GIANT_SAVE();
 599                                 spintries++;
 600                                 for (i = 0; i < alk_loops; i++) {
 601                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 602                                                 CTR4(KTR_LOCK,
 603                                     "%s: shared spinning on %p with %u and %u",
 604                                                     __func__, lk, spintries, i);
 605                                         x = lk->lk_lock;
 606                                         if ((x & LK_SHARE) == 0 ||
 607                                             LK_CAN_SHARE(x) != 0)
 608                                                 break;
 609                                         cpu_spinwait();
 610                                 }
 611                                 GIANT_RESTORE();
 612                                 if (i != alk_loops)
 613                                         continue;
 614                         }
 615 #endif
 616
 617                         /*
 618                          * Acquire the sleepqueue chain lock because we
 619                          * probabilly will need to manipulate waiters flags.
 620                          */
 621                         sleepq_lock(&lk->lock_object);
 622                         x = lk->lk_lock;
 623
 624                         /*
 625                          * if the lock can be acquired in shared mode, try
 626                          * again.
 627                          */
 628                         if (LK_CAN_SHARE(x)) {
 629                                 sleepq_release(&lk->lock_object);
 630                                 continue;
 631                         }
 632
 633 #ifdef ADAPTIVE_LOCKMGRS
 634                         /*
 635                          * The current lock owner might have started executing
 636                          * on another CPU (or the lock could have changed
 637                          * owner) while we were waiting on the turnstile
 638                          * chain lock.  If so, drop the turnstile lock and try
 639                          * again.
 640                          */
 641                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 642                             LK_HOLDER(x) != LK_KERNPROC) {
 643                                 owner = (struct thread *)LK_HOLDER(x);
 644                                 if (TD_IS_RUNNING(owner)) {
 645                                         sleepq_release(&lk->lock_object);
 646                                         continue;
 647                                 }
 648                         }
 649 #endif
 650
 651                         /*
 652                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 653                          * loop back and retry.
 654                          */
 655                         if ((x & LK_SHARED_WAITERS) == 0) {
 656                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 657                                     x | LK_SHARED_WAITERS)) {
 658                                         sleepq_release(&lk->lock_object);
 659                                         continue;
 660                                 }
 661                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 662                                     __func__, lk);
 663                         }
 664
 665                         /*
 666                          * As far as we have been unable to acquire the
 667                          * shared lock and the shared waiters flag is set,
 668                          * we will sleep.
 669                          */
 670                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 671                             SQ_SHARED_QUEUE);
 672                         flags &= ~LK_INTERLOCK;
 673                         if (error) {
 674                                 LOCK_LOG3(lk,
 675                                     "%s: interrupted sleep for %p with %d",
 676                                     __func__, lk, error);
 677                                 break;
 678                         }
 679                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 680                             __func__, lk);
 681                 }
 682                 if (error == 0) {
 683                         lock_profile_obtain_lock_success(&lk->lock_object,
 684                             contested, waittime, file, line);
 685                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 686                             line);
 687                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 688                             line);
 689                         TD_LOCKS_INC(curthread);
 690                         TD_SLOCKS_INC(curthread);
 691                         STACK_SAVE(lk);
 692                 }
 693                 break;
 694         case LK_UPGRADE:
 695                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 696                 v = lk->lk_lock;
 697                 x = v & LK_ALL_WAITERS;
 698                 v &= LK_EXCLUSIVE_SPINNERS;
 699
 700                 /*
 701                  * Try to switch from one shared lock to an exclusive one.
 702                  * We need to preserve waiters flags during the operation.
 703                  */
 704                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 705                     tid | x)) {
 706                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 707                             line);
 708                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 709                             LK_TRYWIT(flags), file, line);
 710                         TD_SLOCKS_DEC(curthread);
 711                         break;
 712                 }
 713
 714                 /*
 715                  * We have been unable to succeed in upgrading, so just
 716                  * give up the shared lock.
 717                  */
 718                 wakeup_swapper |= wakeupshlk(lk, file, line);
 719
 720                 /* FALLTHROUGH */
 721         case LK_EXCLUSIVE:
 722                 if (LK_CAN_WITNESS(flags))
 723                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 724                             LOP_EXCLUSIVE, file, line, ilk);
 725
 726                 /*
 727                  * If curthread already holds the lock and this one is
 728                  * allowed to recurse, simply recurse on it.
 729                  */
 730                 if (lockmgr_xlocked(lk)) {
 731                         if ((flags & LK_CANRECURSE) == 0 &&
 732                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 733
 734                                 /*
 735                                  * If the lock is expected to not panic just
 736                                  * give up and return.
 737                                  */
 738                                 if (LK_TRYOP(flags)) {
 739                                         LOCK_LOG2(lk,
 740                                             "%s: %p fails the try operation",
 741                                             __func__, lk);
 742                                         error = EBUSY;
 743                                         break;
 744                                 }
 745                                 if (flags & LK_INTERLOCK)
 746                                         class->lc_unlock(ilk);
 747                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 748                                     __func__, iwmesg, file, line);
 749                         }
 750                         lk->lk_recurse++;
 751                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 752                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 753                             lk->lk_recurse, file, line);
 754                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 755                             LK_TRYWIT(flags), file, line);
 756                         TD_LOCKS_INC(curthread);
 757                         break;
 758                 }
 759
 760                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
 761                     tid)) {
 762 #ifdef HWPMC_HOOKS
 763                         PMC_SOFT_CALL( , , lock, failed);
 764 #endif
 765                         lock_profile_obtain_lock_failed(&lk->lock_object,
 766                             &contested, &waittime);
 767
 768                         /*
 769                          * If the lock is expected to not sleep just give up
 770                          * and return.
 771                          */
 772                         if (LK_TRYOP(flags)) {
 773                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 774                                     __func__, lk);
 775                                 error = EBUSY;
 776                                 break;
 777                         }
 778
 779 #ifdef ADAPTIVE_LOCKMGRS
 780                         /*
 781                          * If the owner is running on another CPU, spin until
 782                          * the owner stops running or the state of the lock
 783                          * changes.
 784                          */
 785                         x = lk->lk_lock;
 786                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 787                             LK_HOLDER(x) != LK_KERNPROC) {
 788                                 owner = (struct thread *)LK_HOLDER(x);
 789                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 790                                         CTR3(KTR_LOCK,
 791                                             "%s: spinning on %p held by %p",
 792                                             __func__, lk, owner);
 793
 794                                 /*
 795                                  * If we are holding also an interlock drop it
 796                                  * in order to avoid a deadlock if the lockmgr
 797                                  * owner is adaptively spinning on the
 798                                  * interlock itself.
 799                                  */
 800                                 if (flags & LK_INTERLOCK) {
 801                                         class->lc_unlock(ilk);
 802                                         flags &= ~LK_INTERLOCK;
 803                                 }
 804                                 GIANT_SAVE();
 805                                 while (LK_HOLDER(lk->lk_lock) ==
 806                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 807                                         cpu_spinwait();
 808                                 GIANT_RESTORE();
 809                                 continue;
 810                         } else if (LK_CAN_ADAPT(lk, flags) &&
 811                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 812                             spintries < alk_retries) {
 813                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 814                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 815                                     x | LK_EXCLUSIVE_SPINNERS))
 816                                         continue;
 817                                 if (flags & LK_INTERLOCK) {
 818                                         class->lc_unlock(ilk);
 819                                         flags &= ~LK_INTERLOCK;
 820                                 }
 821                                 GIANT_SAVE();
 822                                 spintries++;
 823                                 for (i = 0; i < alk_loops; i++) {
 824                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 825                                                 CTR4(KTR_LOCK,
 826                                     "%s: shared spinning on %p with %u and %u",
 827                                                     __func__, lk, spintries, i);
 828                                         if ((lk->lk_lock &
 829                                             LK_EXCLUSIVE_SPINNERS) == 0)
 830                                                 break;
 831                                         cpu_spinwait();
 832                                 }
 833                                 GIANT_RESTORE();
 834                                 if (i != alk_loops)
 835                                         continue;
 836                         }
 837 #endif
 838
 839                         /*
 840                          * Acquire the sleepqueue chain lock because we
 841                          * probabilly will need to manipulate waiters flags.
 842                          */
 843                         sleepq_lock(&lk->lock_object);
 844                         x = lk->lk_lock;
 845
 846                         /*
 847                          * if the lock has been released while we spun on
 848                          * the sleepqueue chain lock just try again.
 849                          */
 850                         if (x == LK_UNLOCKED) {
 851                                 sleepq_release(&lk->lock_object);
 852                                 continue;
 853                         }
 854
 855 #ifdef ADAPTIVE_LOCKMGRS
 856                         /*
 857                          * The current lock owner might have started executing
 858                          * on another CPU (or the lock could have changed
 859                          * owner) while we were waiting on the turnstile
 860                          * chain lock.  If so, drop the turnstile lock and try
 861                          * again.
 862                          */
 863                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 864                             LK_HOLDER(x) != LK_KERNPROC) {
 865                                 owner = (struct thread *)LK_HOLDER(x);
 866                                 if (TD_IS_RUNNING(owner)) {
 867                                         sleepq_release(&lk->lock_object);
 868                                         continue;
 869                                 }
 870                         }
 871 #endif
 872
 873                         /*
 874                          * The lock can be in the state where there is a
 875                          * pending queue of waiters, but still no owner.
 876                          * This happens when the lock is contested and an
 877                          * owner is going to claim the lock.
 878                          * If curthread is the one successfully acquiring it
 879                          * claim lock ownership and return, preserving waiters
 880                          * flags.
 881                          */
 882                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 883                         if ((x & ~v) == LK_UNLOCKED) {
 884                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 885                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 886                                     tid | v)) {
 887                                         sleepq_release(&lk->lock_object);
 888                                         LOCK_LOG2(lk,
 889                                             "%s: %p claimed by a new writer",
 890                                             __func__, lk);
 891                                         break;
 892                                 }
 893                                 sleepq_release(&lk->lock_object);
 894                                 continue;
 895                         }
 896
 897                         /*
 898                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 899                          * fail, loop back and retry.
 900                          */
 901                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 902                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 903                                     x | LK_EXCLUSIVE_WAITERS)) {
 904                                         sleepq_release(&lk->lock_object);
 905                                         continue;
 906                                 }
 907                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 908                                     __func__, lk);
 909                         }
 910
 911                         /*
 912                          * As far as we have been unable to acquire the
 913                          * exclusive lock and the exclusive waiters flag
 914                          * is set, we will sleep.
 915                          */
 916                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 917                             SQ_EXCLUSIVE_QUEUE);
 918                         flags &= ~LK_INTERLOCK;
 919                         if (error) {
 920                                 LOCK_LOG3(lk,
 921                                     "%s: interrupted sleep for %p with %d",
 922                                     __func__, lk, error);
 923                                 break;
 924                         }
 925                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 926                             __func__, lk);
 927                 }
 928                 if (error == 0) {
 929                         lock_profile_obtain_lock_success(&lk->lock_object,
 930                             contested, waittime, file, line);
 931                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 932                             lk->lk_recurse, file, line);
 933                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 934                             LK_TRYWIT(flags), file, line);
 935                         TD_LOCKS_INC(curthread);
 936                         STACK_SAVE(lk);
 937                 }
 938                 break;
 939         case LK_DOWNGRADE:
 940                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 941                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 942                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 943
 944                 /*
 945                  * Panic if the lock is recursed.
 946                  */
 947                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 948                         if (flags & LK_INTERLOCK)
 949                                 class->lc_unlock(ilk);
 950                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 951                             __func__, iwmesg, file, line);
 952                 }
 953                 TD_SLOCKS_INC(curthread);
 954
 955                 /*
 956                  * In order to preserve waiters flags, just spin.
 957                  */
 958                 for (;;) {
 959                         x = lk->lk_lock;
 960                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
 961                         x &= LK_ALL_WAITERS;
 962                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
 963                             LK_SHARERS_LOCK(1) | x))
 964                                 break;
 965                         cpu_spinwait();
 966                 }
 967                 break;
 968         case LK_RELEASE:
 969                 _lockmgr_assert(lk, KA_LOCKED, file, line);
 970                 x = lk->lk_lock;
 971
 972                 if ((x & LK_SHARE) == 0) {
 973
 974                         /*
 975                          * As first option, treact the lock as if it has not
 976                          * any waiter.
 977                          * Fix-up the tid var if the lock has been disowned.
 978                          */
 979                         if (LK_HOLDER(x) == LK_KERNPROC)
 980                                 tid = LK_KERNPROC;
 981                         else {
 982                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
 983                                     file, line);
 984                                 TD_LOCKS_DEC(curthread);
 985                         }
 986                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
 987                             lk->lk_recurse, file, line);
 988
 989                         /*
 990                          * The lock is held in exclusive mode.
 991                          * If the lock is recursed also, then unrecurse it.
 992                          */
 993                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 994                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
 995                                     lk);
 996                                 lk->lk_recurse--;
 997                                 break;
 998                         }
 999                         if (tid != LK_KERNPROC)
1000                                 lock_profile_release_lock(&lk->lock_object);
1001
1002                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1003                             LK_UNLOCKED))
1004                                 break;
1005
1006                         sleepq_lock(&lk->lock_object);
1007                         x = lk->lk_lock;
1008                         v = LK_UNLOCKED;
1009
1010                         /*
1011                          * If the lock has exclusive waiters, give them
1012                          * preference in order to avoid deadlock with
1013                          * shared runners up.
1014                          * If interruptible sleeps left the exclusive queue
1015                          * empty avoid a starvation for the threads sleeping
1016                          * on the shared queue by giving them precedence
1017                          * and cleaning up the exclusive waiters bit anyway.
1018                          * Please note that lk_exslpfail count may be lying
1019                          * about the real number of waiters with the
1020                          * LK_SLEEPFAIL flag on because they may be used in
1021                          * conjuction with interruptible sleeps so
1022                          * lk_exslpfail might be considered an 'upper limit'
1023                          * bound, including the edge cases.
1024                          */
1025                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1026                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1027                             SQ_EXCLUSIVE_QUEUE);
1028                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1029                                 if (lk->lk_exslpfail < realexslp) {
1030                                         lk->lk_exslpfail = 0;
1031                                         queue = SQ_EXCLUSIVE_QUEUE;
1032                                         v |= (x & LK_SHARED_WAITERS);
1033                                 } else {
1034                                         lk->lk_exslpfail = 0;
1035                                         LOCK_LOG2(lk,
1036                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1037                                             __func__, lk);
1038                                         LOCK_LOG2(lk,
1039                         "%s: %p waking up threads on the exclusive queue",
1040                                             __func__, lk);
1041                                         wakeup_swapper =
1042                                             sleepq_broadcast(&lk->lock_object,
1043                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1044                                         queue = SQ_SHARED_QUEUE;
1045                                 }
1046                         } else {
1047
1048                                 /*
1049                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1050                                  * on and using interruptible sleeps/timeout
1051                                  * may have left spourious lk_exslpfail counts
1052                                  * on, so clean it up anyway.
1053                                  */
1054                                 lk->lk_exslpfail = 0;
1055                                 queue = SQ_SHARED_QUEUE;
1056                         }
1057
1058                         LOCK_LOG3(lk,
1059                             "%s: %p waking up threads on the %s queue",
1060                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1061                             "exclusive");
1062                         atomic_store_rel_ptr(&lk->lk_lock, v);
1063                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1064                             SLEEPQ_LK, 0, queue);
1065                         sleepq_release(&lk->lock_object);
1066                         break;
1067                 } else
1068                         wakeup_swapper = wakeupshlk(lk, file, line);
1069                 break;
1070         case LK_DRAIN:
1071                 if (LK_CAN_WITNESS(flags))
1072                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1073                             LOP_EXCLUSIVE, file, line, ilk);
1074
1075                 /*
1076                  * Trying to drain a lock we already own will result in a
1077                  * deadlock.
1078                  */
1079                 if (lockmgr_xlocked(lk)) {
1080                         if (flags & LK_INTERLOCK)
1081                                 class->lc_unlock(ilk);
1082                         panic("%s: draining %s with the lock held @ %s:%d\n",
1083                             __func__, iwmesg, file, line);
1084                 }
1085
1086                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1087 #ifdef HWPMC_HOOKS
1088                         PMC_SOFT_CALL( , , lock, failed);
1089 #endif
1090                         lock_profile_obtain_lock_failed(&lk->lock_object,
1091                             &contested, &waittime);
1092
1093                         /*
1094                          * If the lock is expected to not sleep just give up
1095                          * and return.
1096                          */
1097                         if (LK_TRYOP(flags)) {
1098                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1099                                     __func__, lk);
1100                                 error = EBUSY;
1101                                 break;
1102                         }
1103
1104                         /*
1105                          * Acquire the sleepqueue chain lock because we
1106                          * probabilly will need to manipulate waiters flags.
1107                          */
1108                         sleepq_lock(&lk->lock_object);
1109                         x = lk->lk_lock;
1110
1111                         /*
1112                          * if the lock has been released while we spun on
1113                          * the sleepqueue chain lock just try again.
1114                          */
1115                         if (x == LK_UNLOCKED) {
1116                                 sleepq_release(&lk->lock_object);
1117                                 continue;
1118                         }
1119
1120                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1121                         if ((x & ~v) == LK_UNLOCKED) {
1122                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1123
1124                                 /*
1125                                  * If interruptible sleeps left the exclusive
1126                                  * queue empty avoid a starvation for the
1127                                  * threads sleeping on the shared queue by
1128                                  * giving them precedence and cleaning up the
1129                                  * exclusive waiters bit anyway.
1130                                  * Please note that lk_exslpfail count may be
1131                                  * lying about the real number of waiters with
1132                                  * the LK_SLEEPFAIL flag on because they may
1133                                  * be used in conjuction with interruptible
1134                                  * sleeps so lk_exslpfail might be considered
1135                                  * an 'upper limit' bound, including the edge
1136                                  * cases.
1137                                  */
1138                                 if (v & LK_EXCLUSIVE_WAITERS) {
1139                                         queue = SQ_EXCLUSIVE_QUEUE;
1140                                         v &= ~LK_EXCLUSIVE_WAITERS;
1141                                 } else {
1142
1143                                         /*
1144                                          * Exclusive waiters sleeping with
1145                                          * LK_SLEEPFAIL on and using
1146                                          * interruptible sleeps/timeout may
1147                                          * have left spourious lk_exslpfail
1148                                          * counts on, so clean it up anyway.
1149                                          */
1150                                         MPASS(v & LK_SHARED_WAITERS);
1151                                         lk->lk_exslpfail = 0;
1152                                         queue = SQ_SHARED_QUEUE;
1153                                         v &= ~LK_SHARED_WAITERS;
1154                                 }
1155                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1156                                         realexslp =
1157                                             sleepq_sleepcnt(&lk->lock_object,
1158                                             SQ_EXCLUSIVE_QUEUE);
1159                                         if (lk->lk_exslpfail >= realexslp) {
1160                                                 lk->lk_exslpfail = 0;
1161                                                 queue = SQ_SHARED_QUEUE;
1162                                                 v &= ~LK_SHARED_WAITERS;
1163                                                 if (realexslp != 0) {
1164                                                         LOCK_LOG2(lk,
1165                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1166                                                             __func__, lk);
1167                                                         LOCK_LOG2(lk,
1168                         "%s: %p waking up threads on the exclusive queue",
1169                                                             __func__, lk);
1170                                                         wakeup_swapper =
1171                                                             sleepq_broadcast(
1172                                                             &lk->lock_object,
1173                                                             SLEEPQ_LK, 0,
1174                                                             SQ_EXCLUSIVE_QUEUE);
1175                                                 }
1176                                         } else
1177                                                 lk->lk_exslpfail = 0;
1178                                 }
1179                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1180                                         sleepq_release(&lk->lock_object);
1181                                         continue;
1182                                 }
1183                                 LOCK_LOG3(lk,
1184                                 "%s: %p waking up all threads on the %s queue",
1185                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1186                                     "shared" : "exclusive");
1187                                 wakeup_swapper |= sleepq_broadcast(
1188                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1189
1190                                 /*
1191                                  * If shared waiters have been woken up we need
1192                                  * to wait for one of them to acquire the lock
1193                                  * before to set the exclusive waiters in
1194                                  * order to avoid a deadlock.
1195                                  */
1196                                 if (queue == SQ_SHARED_QUEUE) {
1197                                         for (v = lk->lk_lock;
1198                                             (v & LK_SHARE) && !LK_SHARERS(v);
1199                                             v = lk->lk_lock)
1200                                                 cpu_spinwait();
1201                                 }
1202                         }
1203
1204                         /*
1205                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1206                          * fail, loop back and retry.
1207                          */
1208                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1209                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1210                                     x | LK_EXCLUSIVE_WAITERS)) {
1211                                         sleepq_release(&lk->lock_object);
1212                                         continue;
1213                                 }
1214                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1215                                     __func__, lk);
1216                         }
1217
1218                         /*
1219                          * As far as we have been unable to acquire the
1220                          * exclusive lock and the exclusive waiters flag
1221                          * is set, we will sleep.
1222                          */
1223                         if (flags & LK_INTERLOCK) {
1224                                 class->lc_unlock(ilk);
1225                                 flags &= ~LK_INTERLOCK;
1226                         }
1227                         GIANT_SAVE();
1228                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1229                             SQ_EXCLUSIVE_QUEUE);
1230                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1231                         GIANT_RESTORE();
1232                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1233                             __func__, lk);
1234                 }
1235
1236                 if (error == 0) {
1237                         lock_profile_obtain_lock_success(&lk->lock_object,
1238                             contested, waittime, file, line);
1239                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1240                             lk->lk_recurse, file, line);
1241                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1242                             LK_TRYWIT(flags), file, line);
1243                         TD_LOCKS_INC(curthread);
1244                         STACK_SAVE(lk);
1245                 }
1246                 break;
1247         default:
1248                 if (flags & LK_INTERLOCK)
1249                         class->lc_unlock(ilk);
1250                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1251         }
1252
1253         if (flags & LK_INTERLOCK)
1254                 class->lc_unlock(ilk);
1255         if (wakeup_swapper)
1256                 kick_proc0();
1257
1258         return (error);
1259 }
1260
1261 void
1262 _lockmgr_disown(struct lock *lk, const char *file, int line)
1263 {
1264         uintptr_t tid, x;
1265
1266         if (SCHEDULER_STOPPED())
1267                 return;
1268
1269         tid = (uintptr_t)curthread;
1270         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1271
1272         /*
1273          * Panic if the lock is recursed.
1274          */
1275         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1276                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1277                     __func__,  file, line);
1278
1279         /*
1280          * If the owner is already LK_KERNPROC just skip the whole operation.
1281          */
1282         if (LK_HOLDER(lk->lk_lock) != tid)
1283                 return;
1284         lock_profile_release_lock(&lk->lock_object);
1285         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1286         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1287         TD_LOCKS_DEC(curthread);
1288         STACK_SAVE(lk);
1289
1290         /*
1291          * In order to preserve waiters flags, just spin.
1292          */
1293         for (;;) {
1294                 x = lk->lk_lock;
1295                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1296                 x &= LK_ALL_WAITERS;
1297                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1298                     LK_KERNPROC | x))
1299                         return;
1300                 cpu_spinwait();
1301         }
1302 }
1303
1304 void
1305 lockmgr_printinfo(const struct lock *lk)
1306 {
1307         struct thread *td;
1308         uintptr_t x;
1309
1310         if (lk->lk_lock == LK_UNLOCKED)
1311                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1312         else if (lk->lk_lock & LK_SHARE)
1313                 printf("lock type %s: SHARED (count %ju)\n",
1314                     lk->lock_object.lo_name,
1315                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1316         else {
1317                 td = lockmgr_xholder(lk);
1318                 printf("lock type %s: EXCL by thread %p "
1319                     "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, td,
1320                     td->td_proc->p_pid, td->td_proc->p_comm, td->td_tid);
1321         }
1322
1323         x = lk->lk_lock;
1324         if (x & LK_EXCLUSIVE_WAITERS)
1325                 printf(" with exclusive waiters pending\n");
1326         if (x & LK_SHARED_WAITERS)
1327                 printf(" with shared waiters pending\n");
1328         if (x & LK_EXCLUSIVE_SPINNERS)
1329                 printf(" with exclusive spinners pending\n");
1330
1331         STACK_PRINT(lk);
1332 }
1333
1334 int
1335 lockstatus(const struct lock *lk)
1336 {
1337         uintptr_t v, x;
1338         int ret;
1339
1340         ret = LK_SHARED;
1341         x = lk->lk_lock;
1342         v = LK_HOLDER(x);
1343
1344         if ((x & LK_SHARE) == 0) {
1345                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1346                         ret = LK_EXCLUSIVE;
1347                 else
1348                         ret = LK_EXCLOTHER;
1349         } else if (x == LK_UNLOCKED)
1350                 ret = 0;
1351
1352         return (ret);
1353 }
1354
1355 #ifdef INVARIANT_SUPPORT
1356
1357 FEATURE(invariant_support,
1358     "Support for modules compiled with INVARIANTS option");
1359
1360 #ifndef INVARIANTS
1361 #undef  _lockmgr_assert
1362 #endif
1363
1364 void
1365 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1366 {
1367         int slocked = 0;
1368
1369         if (panicstr != NULL)
1370                 return;
1371         switch (what) {
1372         case KA_SLOCKED:
1373         case KA_SLOCKED | KA_NOTRECURSED:
1374         case KA_SLOCKED | KA_RECURSED:
1375                 slocked = 1;
1376         case KA_LOCKED:
1377         case KA_LOCKED | KA_NOTRECURSED:
1378         case KA_LOCKED | KA_RECURSED:
1379 #ifdef WITNESS
1380
1381                 /*
1382                  * We cannot trust WITNESS if the lock is held in exclusive
1383                  * mode and a call to lockmgr_disown() happened.
1384                  * Workaround this skipping the check if the lock is held in
1385                  * exclusive mode even for the KA_LOCKED case.
1386                  */
1387                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1388                         witness_assert(&lk->lock_object, what, file, line);
1389                         break;
1390                 }
1391 #endif
1392                 if (lk->lk_lock == LK_UNLOCKED ||
1393                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1394                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1395                         panic("Lock %s not %slocked @ %s:%d\n",
1396                             lk->lock_object.lo_name, slocked ? "share" : "",
1397                             file, line);
1398
1399                 if ((lk->lk_lock & LK_SHARE) == 0) {
1400                         if (lockmgr_recursed(lk)) {
1401                                 if (what & KA_NOTRECURSED)
1402                                         panic("Lock %s recursed @ %s:%d\n",
1403                                             lk->lock_object.lo_name, file,
1404                                             line);
1405                         } else if (what & KA_RECURSED)
1406                                 panic("Lock %s not recursed @ %s:%d\n",
1407                                     lk->lock_object.lo_name, file, line);
1408                 }
1409                 break;
1410         case KA_XLOCKED:
1411         case KA_XLOCKED | KA_NOTRECURSED:
1412         case KA_XLOCKED | KA_RECURSED:
1413                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1414                         panic("Lock %s not exclusively locked @ %s:%d\n",
1415                             lk->lock_object.lo_name, file, line);
1416                 if (lockmgr_recursed(lk)) {
1417                         if (what & KA_NOTRECURSED)
1418                                 panic("Lock %s recursed @ %s:%d\n",
1419                                     lk->lock_object.lo_name, file, line);
1420                 } else if (what & KA_RECURSED)
1421                         panic("Lock %s not recursed @ %s:%d\n",
1422                             lk->lock_object.lo_name, file, line);
1423                 break;
1424         case KA_UNLOCKED:
1425                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1426                         panic("Lock %s exclusively locked @ %s:%d\n",
1427                             lk->lock_object.lo_name, file, line);
1428                 break;
1429         default:
1430                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1431                     line);
1432         }
1433 }
1434 #endif
1435
1436 #ifdef DDB
1437 int
1438 lockmgr_chain(struct thread *td, struct thread **ownerp)
1439 {
1440         struct lock *lk;
1441
1442         lk = td->td_wchan;
1443
1444         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1445                 return (0);
1446         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1447         if (lk->lk_lock & LK_SHARE)
1448                 db_printf("SHARED (count %ju)\n",
1449                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1450         else
1451                 db_printf("EXCL\n");
1452         *ownerp = lockmgr_xholder(lk);
1453
1454         return (1);
1455 }
1456
1457 static void
1458 db_show_lockmgr(const struct lock_object *lock)
1459 {
1460         struct thread *td;
1461         const struct lock *lk;
1462
1463         lk = (const struct lock *)lock;
1464
1465         db_printf(" state: ");
1466         if (lk->lk_lock == LK_UNLOCKED)
1467                 db_printf("UNLOCKED\n");
1468         else if (lk->lk_lock & LK_SHARE)
1469                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1470         else {
1471                 td = lockmgr_xholder(lk);
1472                 if (td == (struct thread *)LK_KERNPROC)
1473                         db_printf("XLOCK: LK_KERNPROC\n");
1474                 else
1475                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1476                             td->td_tid, td->td_proc->p_pid,
1477                             td->td_proc->p_comm);
1478                 if (lockmgr_recursed(lk))
1479                         db_printf(" recursed: %d\n", lk->lk_recurse);
1480         }
1481         db_printf(" waiters: ");
1482         switch (lk->lk_lock & LK_ALL_WAITERS) {
1483         case LK_SHARED_WAITERS:
1484                 db_printf("shared\n");
1485                 break;
1486         case LK_EXCLUSIVE_WAITERS:
1487                 db_printf("exclusive\n");
1488                 break;
1489         case LK_ALL_WAITERS:
1490                 db_printf("shared and exclusive\n");
1491                 break;
1492         default:
1493                 db_printf("none\n");
1494         }
1495         db_printf(" spinners: ");
1496         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1497                 db_printf("exclusive\n");
1498         else
1499                 db_printf("none\n");
1500 }
1501 #endif