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         if (flags & LK_IS_VNODE)
 397                 iflags |= LO_IS_VNODE;
 398         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 399
 400         lk->lk_lock = LK_UNLOCKED;
 401         lk->lk_recurse = 0;
 402         lk->lk_exslpfail = 0;
 403         lk->lk_timo = timo;
 404         lk->lk_pri = pri;
 405         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 406         STACK_ZERO(lk);
 407 }
 408
 409 /*
 410  * XXX: Gross hacks to manipulate external lock flags after
 411  * initialization.  Used for certain vnode and buf locks.
 412  */
 413 void
 414 lockallowshare(struct lock *lk)
 415 {
 416
 417         lockmgr_assert(lk, KA_XLOCKED);
 418         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 419 }
 420
 421 void
 422 lockallowrecurse(struct lock *lk)
 423 {
 424
 425         lockmgr_assert(lk, KA_XLOCKED);
 426         lk->lock_object.lo_flags |= LO_RECURSABLE;
 427 }
 428
 429 void
 430 lockdisablerecurse(struct lock *lk)
 431 {
 432
 433         lockmgr_assert(lk, KA_XLOCKED);
 434         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 435 }
 436
 437 void
 438 lockdestroy(struct lock *lk)
 439 {
 440
 441         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 442         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 443         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 444         lock_destroy(&lk->lock_object);
 445 }
 446
 447 int
 448 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 449     const char *wmesg, int pri, int timo, const char *file, int line)
 450 {
 451         GIANT_DECLARE;
 452         struct lock_class *class;
 453         const char *iwmesg;
 454         uintptr_t tid, v, x;
 455         u_int op, realexslp;
 456         int error, ipri, itimo, queue, wakeup_swapper;
 457 #ifdef LOCK_PROFILING
 458         uint64_t waittime = 0;
 459         int contested = 0;
 460 #endif
 461 #ifdef ADAPTIVE_LOCKMGRS
 462         volatile struct thread *owner;
 463         u_int i, spintries = 0;
 464 #endif
 465
 466         error = 0;
 467         tid = (uintptr_t)curthread;
 468         op = (flags & LK_TYPE_MASK);
 469         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 470         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 471         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 472
 473         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 474         KASSERT((op & (op - 1)) == 0,
 475             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 476         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 477             (op != LK_DOWNGRADE && op != LK_RELEASE),
 478             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 479             __func__, file, line));
 480         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 481             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 482             __func__, file, line));
 483         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 484             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
 485             lk->lock_object.lo_name, file, line));
 486
 487         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 488         if (panicstr != NULL) {
 489                 if (flags & LK_INTERLOCK)
 490                         class->lc_unlock(ilk);
 491                 return (0);
 492         }
 493
 494         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 495                 switch (op) {
 496                 case LK_SHARED:
 497                         op = LK_EXCLUSIVE;
 498                         break;
 499                 case LK_UPGRADE:
 500                 case LK_DOWNGRADE:
 501                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 502                             file, line);
 503                         if (flags & LK_INTERLOCK)
 504                                 class->lc_unlock(ilk);
 505                         return (0);
 506                 }
 507         }
 508
 509         wakeup_swapper = 0;
 510         switch (op) {
 511         case LK_SHARED:
 512                 if (LK_CAN_WITNESS(flags))
 513                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 514                             file, line, ilk);
 515                 for (;;) {
 516                         x = lk->lk_lock;
 517
 518                         /*
 519                          * If no other thread has an exclusive lock, or
 520                          * no exclusive waiter is present, bump the count of
 521                          * sharers.  Since we have to preserve the state of
 522                          * waiters, if we fail to acquire the shared lock
 523                          * loop back and retry.
 524                          */
 525                         if (LK_CAN_SHARE(x)) {
 526                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 527                                     x + LK_ONE_SHARER))
 528                                         break;
 529                                 continue;
 530                         }
 531 #ifdef HWPMC_HOOKS
 532                         PMC_SOFT_CALL( , , lock, failed);
 533 #endif
 534                         lock_profile_obtain_lock_failed(&lk->lock_object,
 535                             &contested, &waittime);
 536
 537                         /*
 538                          * If the lock is already held by curthread in
 539                          * exclusive way avoid a deadlock.
 540                          */
 541                         if (LK_HOLDER(x) == tid) {
 542                                 LOCK_LOG2(lk,
 543                                     "%s: %p already held in exclusive mode",
 544                                     __func__, lk);
 545                                 error = EDEADLK;
 546                                 break;
 547                         }
 548
 549                         /*
 550                          * If the lock is expected to not sleep just give up
 551                          * and return.
 552                          */
 553                         if (LK_TRYOP(flags)) {
 554                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 555                                     __func__, lk);
 556                                 error = EBUSY;
 557                                 break;
 558                         }
 559
 560 #ifdef ADAPTIVE_LOCKMGRS
 561                         /*
 562                          * If the owner is running on another CPU, spin until
 563                          * the owner stops running or the state of the lock
 564                          * changes.  We need a double-state handle here
 565                          * because for a failed acquisition the lock can be
 566                          * either held in exclusive mode or shared mode
 567                          * (for the writer starvation avoidance technique).
 568                          */
 569                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 570                             LK_HOLDER(x) != LK_KERNPROC) {
 571                                 owner = (struct thread *)LK_HOLDER(x);
 572                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 573                                         CTR3(KTR_LOCK,
 574                                             "%s: spinning on %p held by %p",
 575                                             __func__, lk, owner);
 576
 577                                 /*
 578                                  * If we are holding also an interlock drop it
 579                                  * in order to avoid a deadlock if the lockmgr
 580                                  * owner is adaptively spinning on the
 581                                  * interlock itself.
 582                                  */
 583                                 if (flags & LK_INTERLOCK) {
 584                                         class->lc_unlock(ilk);
 585                                         flags &= ~LK_INTERLOCK;
 586                                 }
 587                                 GIANT_SAVE();
 588                                 while (LK_HOLDER(lk->lk_lock) ==
 589                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 590                                         cpu_spinwait();
 591                                 GIANT_RESTORE();
 592                                 continue;
 593                         } else if (LK_CAN_ADAPT(lk, flags) &&
 594                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 595                             spintries < alk_retries) {
 596                                 if (flags & LK_INTERLOCK) {
 597                                         class->lc_unlock(ilk);
 598                                         flags &= ~LK_INTERLOCK;
 599                                 }
 600                                 GIANT_SAVE();
 601                                 spintries++;
 602                                 for (i = 0; i < alk_loops; i++) {
 603                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 604                                                 CTR4(KTR_LOCK,
 605                                     "%s: shared spinning on %p with %u and %u",
 606                                                     __func__, lk, spintries, i);
 607                                         x = lk->lk_lock;
 608                                         if ((x & LK_SHARE) == 0 ||
 609                                             LK_CAN_SHARE(x) != 0)
 610                                                 break;
 611                                         cpu_spinwait();
 612                                 }
 613                                 GIANT_RESTORE();
 614                                 if (i != alk_loops)
 615                                         continue;
 616                         }
 617 #endif
 618
 619                         /*
 620                          * Acquire the sleepqueue chain lock because we
 621                          * probabilly will need to manipulate waiters flags.
 622                          */
 623                         sleepq_lock(&lk->lock_object);
 624                         x = lk->lk_lock;
 625
 626                         /*
 627                          * if the lock can be acquired in shared mode, try
 628                          * again.
 629                          */
 630                         if (LK_CAN_SHARE(x)) {
 631                                 sleepq_release(&lk->lock_object);
 632                                 continue;
 633                         }
 634
 635 #ifdef ADAPTIVE_LOCKMGRS
 636                         /*
 637                          * The current lock owner might have started executing
 638                          * on another CPU (or the lock could have changed
 639                          * owner) while we were waiting on the turnstile
 640                          * chain lock.  If so, drop the turnstile lock and try
 641                          * again.
 642                          */
 643                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 644                             LK_HOLDER(x) != LK_KERNPROC) {
 645                                 owner = (struct thread *)LK_HOLDER(x);
 646                                 if (TD_IS_RUNNING(owner)) {
 647                                         sleepq_release(&lk->lock_object);
 648                                         continue;
 649                                 }
 650                         }
 651 #endif
 652
 653                         /*
 654                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 655                          * loop back and retry.
 656                          */
 657                         if ((x & LK_SHARED_WAITERS) == 0) {
 658                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 659                                     x | LK_SHARED_WAITERS)) {
 660                                         sleepq_release(&lk->lock_object);
 661                                         continue;
 662                                 }
 663                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 664                                     __func__, lk);
 665                         }
 666
 667                         /*
 668                          * As far as we have been unable to acquire the
 669                          * shared lock and the shared waiters flag is set,
 670                          * we will sleep.
 671                          */
 672                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 673                             SQ_SHARED_QUEUE);
 674                         flags &= ~LK_INTERLOCK;
 675                         if (error) {
 676                                 LOCK_LOG3(lk,
 677                                     "%s: interrupted sleep for %p with %d",
 678                                     __func__, lk, error);
 679                                 break;
 680                         }
 681                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 682                             __func__, lk);
 683                 }
 684                 if (error == 0) {
 685                         lock_profile_obtain_lock_success(&lk->lock_object,
 686                             contested, waittime, file, line);
 687                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 688                             line);
 689                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 690                             line);
 691                         TD_LOCKS_INC(curthread);
 692                         TD_SLOCKS_INC(curthread);
 693                         STACK_SAVE(lk);
 694                 }
 695                 break;
 696         case LK_UPGRADE:
 697                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 698                 v = lk->lk_lock;
 699                 x = v & LK_ALL_WAITERS;
 700                 v &= LK_EXCLUSIVE_SPINNERS;
 701
 702                 /*
 703                  * Try to switch from one shared lock to an exclusive one.
 704                  * We need to preserve waiters flags during the operation.
 705                  */
 706                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 707                     tid | x)) {
 708                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 709                             line);
 710                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 711                             LK_TRYWIT(flags), file, line);
 712                         TD_SLOCKS_DEC(curthread);
 713                         break;
 714                 }
 715
 716                 /*
 717                  * We have been unable to succeed in upgrading, so just
 718                  * give up the shared lock.
 719                  */
 720                 wakeup_swapper |= wakeupshlk(lk, file, line);
 721
 722                 /* FALLTHROUGH */
 723         case LK_EXCLUSIVE:
 724                 if (LK_CAN_WITNESS(flags))
 725                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 726                             LOP_EXCLUSIVE, file, line, ilk);
 727
 728                 /*
 729                  * If curthread already holds the lock and this one is
 730                  * allowed to recurse, simply recurse on it.
 731                  */
 732                 if (lockmgr_xlocked(lk)) {
 733                         if ((flags & LK_CANRECURSE) == 0 &&
 734                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 735
 736                                 /*
 737                                  * If the lock is expected to not panic just
 738                                  * give up and return.
 739                                  */
 740                                 if (LK_TRYOP(flags)) {
 741                                         LOCK_LOG2(lk,
 742                                             "%s: %p fails the try operation",
 743                                             __func__, lk);
 744                                         error = EBUSY;
 745                                         break;
 746                                 }
 747                                 if (flags & LK_INTERLOCK)
 748                                         class->lc_unlock(ilk);
 749                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 750                                     __func__, iwmesg, file, line);
 751                         }
 752                         lk->lk_recurse++;
 753                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 754                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 755                             lk->lk_recurse, file, line);
 756                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 757                             LK_TRYWIT(flags), file, line);
 758                         TD_LOCKS_INC(curthread);
 759                         break;
 760                 }
 761
 762                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
 763                     tid)) {
 764 #ifdef HWPMC_HOOKS
 765                         PMC_SOFT_CALL( , , lock, failed);
 766 #endif
 767                         lock_profile_obtain_lock_failed(&lk->lock_object,
 768                             &contested, &waittime);
 769
 770                         /*
 771                          * If the lock is expected to not sleep just give up
 772                          * and return.
 773                          */
 774                         if (LK_TRYOP(flags)) {
 775                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 776                                     __func__, lk);
 777                                 error = EBUSY;
 778                                 break;
 779                         }
 780
 781 #ifdef ADAPTIVE_LOCKMGRS
 782                         /*
 783                          * If the owner is running on another CPU, spin until
 784                          * the owner stops running or the state of the lock
 785                          * changes.
 786                          */
 787                         x = lk->lk_lock;
 788                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 789                             LK_HOLDER(x) != LK_KERNPROC) {
 790                                 owner = (struct thread *)LK_HOLDER(x);
 791                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 792                                         CTR3(KTR_LOCK,
 793                                             "%s: spinning on %p held by %p",
 794                                             __func__, lk, owner);
 795
 796                                 /*
 797                                  * If we are holding also an interlock drop it
 798                                  * in order to avoid a deadlock if the lockmgr
 799                                  * owner is adaptively spinning on the
 800                                  * interlock itself.
 801                                  */
 802                                 if (flags & LK_INTERLOCK) {
 803                                         class->lc_unlock(ilk);
 804                                         flags &= ~LK_INTERLOCK;
 805                                 }
 806                                 GIANT_SAVE();
 807                                 while (LK_HOLDER(lk->lk_lock) ==
 808                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 809                                         cpu_spinwait();
 810                                 GIANT_RESTORE();
 811                                 continue;
 812                         } else if (LK_CAN_ADAPT(lk, flags) &&
 813                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 814                             spintries < alk_retries) {
 815                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 816                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 817                                     x | LK_EXCLUSIVE_SPINNERS))
 818                                         continue;
 819                                 if (flags & LK_INTERLOCK) {
 820                                         class->lc_unlock(ilk);
 821                                         flags &= ~LK_INTERLOCK;
 822                                 }
 823                                 GIANT_SAVE();
 824                                 spintries++;
 825                                 for (i = 0; i < alk_loops; i++) {
 826                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 827                                                 CTR4(KTR_LOCK,
 828                                     "%s: shared spinning on %p with %u and %u",
 829                                                     __func__, lk, spintries, i);
 830                                         if ((lk->lk_lock &
 831                                             LK_EXCLUSIVE_SPINNERS) == 0)
 832                                                 break;
 833                                         cpu_spinwait();
 834                                 }
 835                                 GIANT_RESTORE();
 836                                 if (i != alk_loops)
 837                                         continue;
 838                         }
 839 #endif
 840
 841                         /*
 842                          * Acquire the sleepqueue chain lock because we
 843                          * probabilly will need to manipulate waiters flags.
 844                          */
 845                         sleepq_lock(&lk->lock_object);
 846                         x = lk->lk_lock;
 847
 848                         /*
 849                          * if the lock has been released while we spun on
 850                          * the sleepqueue chain lock just try again.
 851                          */
 852                         if (x == LK_UNLOCKED) {
 853                                 sleepq_release(&lk->lock_object);
 854                                 continue;
 855                         }
 856
 857 #ifdef ADAPTIVE_LOCKMGRS
 858                         /*
 859                          * The current lock owner might have started executing
 860                          * on another CPU (or the lock could have changed
 861                          * owner) while we were waiting on the turnstile
 862                          * chain lock.  If so, drop the turnstile lock and try
 863                          * again.
 864                          */
 865                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 866                             LK_HOLDER(x) != LK_KERNPROC) {
 867                                 owner = (struct thread *)LK_HOLDER(x);
 868                                 if (TD_IS_RUNNING(owner)) {
 869                                         sleepq_release(&lk->lock_object);
 870                                         continue;
 871                                 }
 872                         }
 873 #endif
 874
 875                         /*
 876                          * The lock can be in the state where there is a
 877                          * pending queue of waiters, but still no owner.
 878                          * This happens when the lock is contested and an
 879                          * owner is going to claim the lock.
 880                          * If curthread is the one successfully acquiring it
 881                          * claim lock ownership and return, preserving waiters
 882                          * flags.
 883                          */
 884                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 885                         if ((x & ~v) == LK_UNLOCKED) {
 886                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 887                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 888                                     tid | v)) {
 889                                         sleepq_release(&lk->lock_object);
 890                                         LOCK_LOG2(lk,
 891                                             "%s: %p claimed by a new writer",
 892                                             __func__, lk);
 893                                         break;
 894                                 }
 895                                 sleepq_release(&lk->lock_object);
 896                                 continue;
 897                         }
 898
 899                         /*
 900                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 901                          * fail, loop back and retry.
 902                          */
 903                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 904                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 905                                     x | LK_EXCLUSIVE_WAITERS)) {
 906                                         sleepq_release(&lk->lock_object);
 907                                         continue;
 908                                 }
 909                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 910                                     __func__, lk);
 911                         }
 912
 913                         /*
 914                          * As far as we have been unable to acquire the
 915                          * exclusive lock and the exclusive waiters flag
 916                          * is set, we will sleep.
 917                          */
 918                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 919                             SQ_EXCLUSIVE_QUEUE);
 920                         flags &= ~LK_INTERLOCK;
 921                         if (error) {
 922                                 LOCK_LOG3(lk,
 923                                     "%s: interrupted sleep for %p with %d",
 924                                     __func__, lk, error);
 925                                 break;
 926                         }
 927                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 928                             __func__, lk);
 929                 }
 930                 if (error == 0) {
 931                         lock_profile_obtain_lock_success(&lk->lock_object,
 932                             contested, waittime, file, line);
 933                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 934                             lk->lk_recurse, file, line);
 935                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 936                             LK_TRYWIT(flags), file, line);
 937                         TD_LOCKS_INC(curthread);
 938                         STACK_SAVE(lk);
 939                 }
 940                 break;
 941         case LK_DOWNGRADE:
 942                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 943                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 944                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 945
 946                 /*
 947                  * Panic if the lock is recursed.
 948                  */
 949                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 950                         if (flags & LK_INTERLOCK)
 951                                 class->lc_unlock(ilk);
 952                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 953                             __func__, iwmesg, file, line);
 954                 }
 955                 TD_SLOCKS_INC(curthread);
 956
 957                 /*
 958                  * In order to preserve waiters flags, just spin.
 959                  */
 960                 for (;;) {
 961                         x = lk->lk_lock;
 962                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
 963                         x &= LK_ALL_WAITERS;
 964                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
 965                             LK_SHARERS_LOCK(1) | x))
 966                                 break;
 967                         cpu_spinwait();
 968                 }
 969                 break;
 970         case LK_RELEASE:
 971                 _lockmgr_assert(lk, KA_LOCKED, file, line);
 972                 x = lk->lk_lock;
 973
 974                 if ((x & LK_SHARE) == 0) {
 975
 976                         /*
 977                          * As first option, treact the lock as if it has not
 978                          * any waiter.
 979                          * Fix-up the tid var if the lock has been disowned.
 980                          */
 981                         if (LK_HOLDER(x) == LK_KERNPROC)
 982                                 tid = LK_KERNPROC;
 983                         else {
 984                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
 985                                     file, line);
 986                                 TD_LOCKS_DEC(curthread);
 987                         }
 988                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
 989                             lk->lk_recurse, file, line);
 990
 991                         /*
 992                          * The lock is held in exclusive mode.
 993                          * If the lock is recursed also, then unrecurse it.
 994                          */
 995                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 996                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
 997                                     lk);
 998                                 lk->lk_recurse--;
 999                                 break;
1000                         }
1001                         if (tid != LK_KERNPROC)
1002                                 lock_profile_release_lock(&lk->lock_object);
1003
1004                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1005                             LK_UNLOCKED))
1006                                 break;
1007
1008                         sleepq_lock(&lk->lock_object);
1009                         x = lk->lk_lock;
1010                         v = LK_UNLOCKED;
1011
1012                         /*
1013                          * If the lock has exclusive waiters, give them
1014                          * preference in order to avoid deadlock with
1015                          * shared runners up.
1016                          * If interruptible sleeps left the exclusive queue
1017                          * empty avoid a starvation for the threads sleeping
1018                          * on the shared queue by giving them precedence
1019                          * and cleaning up the exclusive waiters bit anyway.
1020                          * Please note that lk_exslpfail count may be lying
1021                          * about the real number of waiters with the
1022                          * LK_SLEEPFAIL flag on because they may be used in
1023                          * conjuction with interruptible sleeps so
1024                          * lk_exslpfail might be considered an 'upper limit'
1025                          * bound, including the edge cases.
1026                          */
1027                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1028                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1029                             SQ_EXCLUSIVE_QUEUE);
1030                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1031                                 if (lk->lk_exslpfail < realexslp) {
1032                                         lk->lk_exslpfail = 0;
1033                                         queue = SQ_EXCLUSIVE_QUEUE;
1034                                         v |= (x & LK_SHARED_WAITERS);
1035                                 } else {
1036                                         lk->lk_exslpfail = 0;
1037                                         LOCK_LOG2(lk,
1038                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1039                                             __func__, lk);
1040                                         LOCK_LOG2(lk,
1041                         "%s: %p waking up threads on the exclusive queue",
1042                                             __func__, lk);
1043                                         wakeup_swapper =
1044                                             sleepq_broadcast(&lk->lock_object,
1045                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1046                                         queue = SQ_SHARED_QUEUE;
1047                                 }
1048                         } else {
1049
1050                                 /*
1051                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1052                                  * on and using interruptible sleeps/timeout
1053                                  * may have left spourious lk_exslpfail counts
1054                                  * on, so clean it up anyway.
1055                                  */
1056                                 lk->lk_exslpfail = 0;
1057                                 queue = SQ_SHARED_QUEUE;
1058                         }
1059
1060                         LOCK_LOG3(lk,
1061                             "%s: %p waking up threads on the %s queue",
1062                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1063                             "exclusive");
1064                         atomic_store_rel_ptr(&lk->lk_lock, v);
1065                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1066                             SLEEPQ_LK, 0, queue);
1067                         sleepq_release(&lk->lock_object);
1068                         break;
1069                 } else
1070                         wakeup_swapper = wakeupshlk(lk, file, line);
1071                 break;
1072         case LK_DRAIN:
1073                 if (LK_CAN_WITNESS(flags))
1074                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1075                             LOP_EXCLUSIVE, file, line, ilk);
1076
1077                 /*
1078                  * Trying to drain a lock we already own will result in a
1079                  * deadlock.
1080                  */
1081                 if (lockmgr_xlocked(lk)) {
1082                         if (flags & LK_INTERLOCK)
1083                                 class->lc_unlock(ilk);
1084                         panic("%s: draining %s with the lock held @ %s:%d\n",
1085                             __func__, iwmesg, file, line);
1086                 }
1087
1088                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1089 #ifdef HWPMC_HOOKS
1090                         PMC_SOFT_CALL( , , lock, failed);
1091 #endif
1092                         lock_profile_obtain_lock_failed(&lk->lock_object,
1093                             &contested, &waittime);
1094
1095                         /*
1096                          * If the lock is expected to not sleep just give up
1097                          * and return.
1098                          */
1099                         if (LK_TRYOP(flags)) {
1100                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1101                                     __func__, lk);
1102                                 error = EBUSY;
1103                                 break;
1104                         }
1105
1106                         /*
1107                          * Acquire the sleepqueue chain lock because we
1108                          * probabilly will need to manipulate waiters flags.
1109                          */
1110                         sleepq_lock(&lk->lock_object);
1111                         x = lk->lk_lock;
1112
1113                         /*
1114                          * if the lock has been released while we spun on
1115                          * the sleepqueue chain lock just try again.
1116                          */
1117                         if (x == LK_UNLOCKED) {
1118                                 sleepq_release(&lk->lock_object);
1119                                 continue;
1120                         }
1121
1122                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1123                         if ((x & ~v) == LK_UNLOCKED) {
1124                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1125
1126                                 /*
1127                                  * If interruptible sleeps left the exclusive
1128                                  * queue empty avoid a starvation for the
1129                                  * threads sleeping on the shared queue by
1130                                  * giving them precedence and cleaning up the
1131                                  * exclusive waiters bit anyway.
1132                                  * Please note that lk_exslpfail count may be
1133                                  * lying about the real number of waiters with
1134                                  * the LK_SLEEPFAIL flag on because they may
1135                                  * be used in conjuction with interruptible
1136                                  * sleeps so lk_exslpfail might be considered
1137                                  * an 'upper limit' bound, including the edge
1138                                  * cases.
1139                                  */
1140                                 if (v & LK_EXCLUSIVE_WAITERS) {
1141                                         queue = SQ_EXCLUSIVE_QUEUE;
1142                                         v &= ~LK_EXCLUSIVE_WAITERS;
1143                                 } else {
1144
1145                                         /*
1146                                          * Exclusive waiters sleeping with
1147                                          * LK_SLEEPFAIL on and using
1148                                          * interruptible sleeps/timeout may
1149                                          * have left spourious lk_exslpfail
1150                                          * counts on, so clean it up anyway.
1151                                          */
1152                                         MPASS(v & LK_SHARED_WAITERS);
1153                                         lk->lk_exslpfail = 0;
1154                                         queue = SQ_SHARED_QUEUE;
1155                                         v &= ~LK_SHARED_WAITERS;
1156                                 }
1157                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1158                                         realexslp =
1159                                             sleepq_sleepcnt(&lk->lock_object,
1160                                             SQ_EXCLUSIVE_QUEUE);
1161                                         if (lk->lk_exslpfail >= realexslp) {
1162                                                 lk->lk_exslpfail = 0;
1163                                                 queue = SQ_SHARED_QUEUE;
1164                                                 v &= ~LK_SHARED_WAITERS;
1165                                                 if (realexslp != 0) {
1166                                                         LOCK_LOG2(lk,
1167                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1168                                                             __func__, lk);
1169                                                         LOCK_LOG2(lk,
1170                         "%s: %p waking up threads on the exclusive queue",
1171                                                             __func__, lk);
1172                                                         wakeup_swapper =
1173                                                             sleepq_broadcast(
1174                                                             &lk->lock_object,
1175                                                             SLEEPQ_LK, 0,
1176                                                             SQ_EXCLUSIVE_QUEUE);
1177                                                 }
1178                                         } else
1179                                                 lk->lk_exslpfail = 0;
1180                                 }
1181                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1182                                         sleepq_release(&lk->lock_object);
1183                                         continue;
1184                                 }
1185                                 LOCK_LOG3(lk,
1186                                 "%s: %p waking up all threads on the %s queue",
1187                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1188                                     "shared" : "exclusive");
1189                                 wakeup_swapper |= sleepq_broadcast(
1190                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1191
1192                                 /*
1193                                  * If shared waiters have been woken up we need
1194                                  * to wait for one of them to acquire the lock
1195                                  * before to set the exclusive waiters in
1196                                  * order to avoid a deadlock.
1197                                  */
1198                                 if (queue == SQ_SHARED_QUEUE) {
1199                                         for (v = lk->lk_lock;
1200                                             (v & LK_SHARE) && !LK_SHARERS(v);
1201                                             v = lk->lk_lock)
1202                                                 cpu_spinwait();
1203                                 }
1204                         }
1205
1206                         /*
1207                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1208                          * fail, loop back and retry.
1209                          */
1210                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1211                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1212                                     x | LK_EXCLUSIVE_WAITERS)) {
1213                                         sleepq_release(&lk->lock_object);
1214                                         continue;
1215                                 }
1216                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1217                                     __func__, lk);
1218                         }
1219
1220                         /*
1221                          * As far as we have been unable to acquire the
1222                          * exclusive lock and the exclusive waiters flag
1223                          * is set, we will sleep.
1224                          */
1225                         if (flags & LK_INTERLOCK) {
1226                                 class->lc_unlock(ilk);
1227                                 flags &= ~LK_INTERLOCK;
1228                         }
1229                         GIANT_SAVE();
1230                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1231                             SQ_EXCLUSIVE_QUEUE);
1232                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1233                         GIANT_RESTORE();
1234                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1235                             __func__, lk);
1236                 }
1237
1238                 if (error == 0) {
1239                         lock_profile_obtain_lock_success(&lk->lock_object,
1240                             contested, waittime, file, line);
1241                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1242                             lk->lk_recurse, file, line);
1243                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1244                             LK_TRYWIT(flags), file, line);
1245                         TD_LOCKS_INC(curthread);
1246                         STACK_SAVE(lk);
1247                 }
1248                 break;
1249         default:
1250                 if (flags & LK_INTERLOCK)
1251                         class->lc_unlock(ilk);
1252                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1253         }
1254
1255         if (flags & LK_INTERLOCK)
1256                 class->lc_unlock(ilk);
1257         if (wakeup_swapper)
1258                 kick_proc0();
1259
1260         return (error);
1261 }
1262
1263 void
1264 _lockmgr_disown(struct lock *lk, const char *file, int line)
1265 {
1266         uintptr_t tid, x;
1267
1268         if (SCHEDULER_STOPPED())
1269                 return;
1270
1271         tid = (uintptr_t)curthread;
1272         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1273
1274         /*
1275          * Panic if the lock is recursed.
1276          */
1277         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1278                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1279                     __func__,  file, line);
1280
1281         /*
1282          * If the owner is already LK_KERNPROC just skip the whole operation.
1283          */
1284         if (LK_HOLDER(lk->lk_lock) != tid)
1285                 return;
1286         lock_profile_release_lock(&lk->lock_object);
1287         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1288         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1289         TD_LOCKS_DEC(curthread);
1290         STACK_SAVE(lk);
1291
1292         /*
1293          * In order to preserve waiters flags, just spin.
1294          */
1295         for (;;) {
1296                 x = lk->lk_lock;
1297                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1298                 x &= LK_ALL_WAITERS;
1299                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1300                     LK_KERNPROC | x))
1301                         return;
1302                 cpu_spinwait();
1303         }
1304 }
1305
1306 void
1307 lockmgr_printinfo(const struct lock *lk)
1308 {
1309         struct thread *td;
1310         uintptr_t x;
1311
1312         if (lk->lk_lock == LK_UNLOCKED)
1313                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1314         else if (lk->lk_lock & LK_SHARE)
1315                 printf("lock type %s: SHARED (count %ju)\n",
1316                     lk->lock_object.lo_name,
1317                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1318         else {
1319                 td = lockmgr_xholder(lk);
1320                 printf("lock type %s: EXCL by thread %p "
1321                     "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, td,
1322                     td->td_proc->p_pid, td->td_proc->p_comm, td->td_tid);
1323         }
1324
1325         x = lk->lk_lock;
1326         if (x & LK_EXCLUSIVE_WAITERS)
1327                 printf(" with exclusive waiters pending\n");
1328         if (x & LK_SHARED_WAITERS)
1329                 printf(" with shared waiters pending\n");
1330         if (x & LK_EXCLUSIVE_SPINNERS)
1331                 printf(" with exclusive spinners pending\n");
1332
1333         STACK_PRINT(lk);
1334 }
1335
1336 int
1337 lockstatus(const struct lock *lk)
1338 {
1339         uintptr_t v, x;
1340         int ret;
1341
1342         ret = LK_SHARED;
1343         x = lk->lk_lock;
1344         v = LK_HOLDER(x);
1345
1346         if ((x & LK_SHARE) == 0) {
1347                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1348                         ret = LK_EXCLUSIVE;
1349                 else
1350                         ret = LK_EXCLOTHER;
1351         } else if (x == LK_UNLOCKED)
1352                 ret = 0;
1353
1354         return (ret);
1355 }
1356
1357 #ifdef INVARIANT_SUPPORT
1358
1359 FEATURE(invariant_support,
1360     "Support for modules compiled with INVARIANTS option");
1361
1362 #ifndef INVARIANTS
1363 #undef  _lockmgr_assert
1364 #endif
1365
1366 void
1367 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1368 {
1369         int slocked = 0;
1370
1371         if (panicstr != NULL)
1372                 return;
1373         switch (what) {
1374         case KA_SLOCKED:
1375         case KA_SLOCKED | KA_NOTRECURSED:
1376         case KA_SLOCKED | KA_RECURSED:
1377                 slocked = 1;
1378         case KA_LOCKED:
1379         case KA_LOCKED | KA_NOTRECURSED:
1380         case KA_LOCKED | KA_RECURSED:
1381 #ifdef WITNESS
1382
1383                 /*
1384                  * We cannot trust WITNESS if the lock is held in exclusive
1385                  * mode and a call to lockmgr_disown() happened.
1386                  * Workaround this skipping the check if the lock is held in
1387                  * exclusive mode even for the KA_LOCKED case.
1388                  */
1389                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1390                         witness_assert(&lk->lock_object, what, file, line);
1391                         break;
1392                 }
1393 #endif
1394                 if (lk->lk_lock == LK_UNLOCKED ||
1395                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1396                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1397                         panic("Lock %s not %slocked @ %s:%d\n",
1398                             lk->lock_object.lo_name, slocked ? "share" : "",
1399                             file, line);
1400
1401                 if ((lk->lk_lock & LK_SHARE) == 0) {
1402                         if (lockmgr_recursed(lk)) {
1403                                 if (what & KA_NOTRECURSED)
1404                                         panic("Lock %s recursed @ %s:%d\n",
1405                                             lk->lock_object.lo_name, file,
1406                                             line);
1407                         } else if (what & KA_RECURSED)
1408                                 panic("Lock %s not recursed @ %s:%d\n",
1409                                     lk->lock_object.lo_name, file, line);
1410                 }
1411                 break;
1412         case KA_XLOCKED:
1413         case KA_XLOCKED | KA_NOTRECURSED:
1414         case KA_XLOCKED | KA_RECURSED:
1415                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1416                         panic("Lock %s not exclusively locked @ %s:%d\n",
1417                             lk->lock_object.lo_name, file, line);
1418                 if (lockmgr_recursed(lk)) {
1419                         if (what & KA_NOTRECURSED)
1420                                 panic("Lock %s recursed @ %s:%d\n",
1421                                     lk->lock_object.lo_name, file, line);
1422                 } else if (what & KA_RECURSED)
1423                         panic("Lock %s not recursed @ %s:%d\n",
1424                             lk->lock_object.lo_name, file, line);
1425                 break;
1426         case KA_UNLOCKED:
1427                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1428                         panic("Lock %s exclusively locked @ %s:%d\n",
1429                             lk->lock_object.lo_name, file, line);
1430                 break;
1431         default:
1432                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1433                     line);
1434         }
1435 }
1436 #endif
1437
1438 #ifdef DDB
1439 int
1440 lockmgr_chain(struct thread *td, struct thread **ownerp)
1441 {
1442         struct lock *lk;
1443
1444         lk = td->td_wchan;
1445
1446         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1447                 return (0);
1448         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1449         if (lk->lk_lock & LK_SHARE)
1450                 db_printf("SHARED (count %ju)\n",
1451                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1452         else
1453                 db_printf("EXCL\n");
1454         *ownerp = lockmgr_xholder(lk);
1455
1456         return (1);
1457 }
1458
1459 static void
1460 db_show_lockmgr(const struct lock_object *lock)
1461 {
1462         struct thread *td;
1463         const struct lock *lk;
1464
1465         lk = (const struct lock *)lock;
1466
1467         db_printf(" state: ");
1468         if (lk->lk_lock == LK_UNLOCKED)
1469                 db_printf("UNLOCKED\n");
1470         else if (lk->lk_lock & LK_SHARE)
1471                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1472         else {
1473                 td = lockmgr_xholder(lk);
1474                 if (td == (struct thread *)LK_KERNPROC)
1475                         db_printf("XLOCK: LK_KERNPROC\n");
1476                 else
1477                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1478                             td->td_tid, td->td_proc->p_pid,
1479                             td->td_proc->p_comm);
1480                 if (lockmgr_recursed(lk))
1481                         db_printf(" recursed: %d\n", lk->lk_recurse);
1482         }
1483         db_printf(" waiters: ");
1484         switch (lk->lk_lock & LK_ALL_WAITERS) {
1485         case LK_SHARED_WAITERS:
1486                 db_printf("shared\n");
1487                 break;
1488         case LK_EXCLUSIVE_WAITERS:
1489                 db_printf("exclusive\n");
1490                 break;
1491         case LK_ALL_WAITERS:
1492                 db_printf("shared and exclusive\n");
1493                 break;
1494         default:
1495                 db_printf("none\n");
1496         }
1497         db_printf(" spinners: ");
1498         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1499                 db_printf("exclusive\n");
1500         else
1501                 db_printf("none\n");
1502 }
1503 #endif