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, uintptr_t how);
 146 #ifdef KDTRACE_HOOKS
 147 static int      owner_lockmgr(const struct lock_object *lock,
 148                     struct thread **owner);
 149 #endif
 150 static uintptr_t 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         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 242         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 243
 244         wakeup_swapper = 0;
 245         for (;;) {
 246                 x = lk->lk_lock;
 247
 248                 /*
 249                  * If there is more than one shared lock held, just drop one
 250                  * and return.
 251                  */
 252                 if (LK_SHARERS(x) > 1) {
 253                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
 254                             x - LK_ONE_SHARER))
 255                                 break;
 256                         continue;
 257                 }
 258
 259                 /*
 260                  * If there are not waiters on the exclusive queue, drop the
 261                  * lock quickly.
 262                  */
 263                 if ((x & LK_ALL_WAITERS) == 0) {
 264                         MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
 265                             LK_SHARERS_LOCK(1));
 266                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
 267                                 break;
 268                         continue;
 269                 }
 270
 271                 /*
 272                  * We should have a sharer with waiters, so enter the hard
 273                  * path in order to handle wakeups correctly.
 274                  */
 275                 sleepq_lock(&lk->lock_object);
 276                 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 277                 v = LK_UNLOCKED;
 278
 279                 /*
 280                  * If the lock has exclusive waiters, give them preference in
 281                  * order to avoid deadlock with shared runners up.
 282                  * If interruptible sleeps left the exclusive queue empty
 283                  * avoid a starvation for the threads sleeping on the shared
 284                  * queue by giving them precedence and cleaning up the
 285                  * exclusive waiters bit anyway.
 286                  * Please note that lk_exslpfail count may be lying about
 287                  * the real number of waiters with the LK_SLEEPFAIL flag on
 288                  * because they may be used in conjuction with interruptible
 289                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 290                  * bound, including the edge cases.
 291                  */
 292                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 293                     SQ_EXCLUSIVE_QUEUE);
 294                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 295                         if (lk->lk_exslpfail < realexslp) {
 296                                 lk->lk_exslpfail = 0;
 297                                 queue = SQ_EXCLUSIVE_QUEUE;
 298                                 v |= (x & LK_SHARED_WAITERS);
 299                         } else {
 300                                 lk->lk_exslpfail = 0;
 301                                 LOCK_LOG2(lk,
 302                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 303                                     __func__, lk);
 304                                 LOCK_LOG2(lk,
 305                             "%s: %p waking up threads on the exclusive queue",
 306                                     __func__, lk);
 307                                 wakeup_swapper =
 308                                     sleepq_broadcast(&lk->lock_object,
 309                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 310                                 queue = SQ_SHARED_QUEUE;
 311                         }
 312
 313                 } else {
 314
 315                         /*
 316                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 317                          * and using interruptible sleeps/timeout may have
 318                          * left spourious lk_exslpfail counts on, so clean
 319                          * it up anyway.
 320                          */
 321                         lk->lk_exslpfail = 0;
 322                         queue = SQ_SHARED_QUEUE;
 323                 }
 324
 325                 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
 326                     v)) {
 327                         sleepq_release(&lk->lock_object);
 328                         continue;
 329                 }
 330                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 331                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 332                     "exclusive");
 333                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 334                     0, queue);
 335                 sleepq_release(&lk->lock_object);
 336                 break;
 337         }
 338
 339         lock_profile_release_lock(&lk->lock_object);
 340         TD_LOCKS_DEC(curthread);
 341         TD_SLOCKS_DEC(curthread);
 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, uintptr_t how)
 354 {
 355
 356         panic("lockmgr locks do not support sleep interlocking");
 357 }
 358
 359 static uintptr_t
 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         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 401         lk->lk_lock = LK_UNLOCKED;
 402         lk->lk_recurse = 0;
 403         lk->lk_exslpfail = 0;
 404         lk->lk_timo = timo;
 405         lk->lk_pri = pri;
 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_TRYUPGRADE:
 501                 case LK_DOWNGRADE:
 502                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 503                             file, line);
 504                         if (flags & LK_INTERLOCK)
 505                                 class->lc_unlock(ilk);
 506                         return (0);
 507                 }
 508         }
 509
 510         wakeup_swapper = 0;
 511         switch (op) {
 512         case LK_SHARED:
 513                 if (LK_CAN_WITNESS(flags))
 514                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 515                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 516                 for (;;) {
 517                         x = lk->lk_lock;
 518
 519                         /*
 520                          * If no other thread has an exclusive lock, or
 521                          * no exclusive waiter is present, bump the count of
 522                          * sharers.  Since we have to preserve the state of
 523                          * waiters, if we fail to acquire the shared lock
 524                          * loop back and retry.
 525                          */
 526                         if (LK_CAN_SHARE(x)) {
 527                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 528                                     x + LK_ONE_SHARER))
 529                                         break;
 530                                 continue;
 531                         }
 532 #ifdef HWPMC_HOOKS
 533                         PMC_SOFT_CALL( , , lock, failed);
 534 #endif
 535                         lock_profile_obtain_lock_failed(&lk->lock_object,
 536                             &contested, &waittime);
 537
 538                         /*
 539                          * If the lock is already held by curthread in
 540                          * exclusive way avoid a deadlock.
 541                          */
 542                         if (LK_HOLDER(x) == tid) {
 543                                 LOCK_LOG2(lk,
 544                                     "%s: %p already held in exclusive mode",
 545                                     __func__, lk);
 546                                 error = EDEADLK;
 547                                 break;
 548                         }
 549
 550                         /*
 551                          * If the lock is expected to not sleep just give up
 552                          * and return.
 553                          */
 554                         if (LK_TRYOP(flags)) {
 555                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 556                                     __func__, lk);
 557                                 error = EBUSY;
 558                                 break;
 559                         }
 560
 561 #ifdef ADAPTIVE_LOCKMGRS
 562                         /*
 563                          * If the owner is running on another CPU, spin until
 564                          * the owner stops running or the state of the lock
 565                          * changes.  We need a double-state handle here
 566                          * because for a failed acquisition the lock can be
 567                          * either held in exclusive mode or shared mode
 568                          * (for the writer starvation avoidance technique).
 569                          */
 570                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 571                             LK_HOLDER(x) != LK_KERNPROC) {
 572                                 owner = (struct thread *)LK_HOLDER(x);
 573                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 574                                         CTR3(KTR_LOCK,
 575                                             "%s: spinning on %p held by %p",
 576                                             __func__, lk, owner);
 577
 578                                 /*
 579                                  * If we are holding also an interlock drop it
 580                                  * in order to avoid a deadlock if the lockmgr
 581                                  * owner is adaptively spinning on the
 582                                  * interlock itself.
 583                                  */
 584                                 if (flags & LK_INTERLOCK) {
 585                                         class->lc_unlock(ilk);
 586                                         flags &= ~LK_INTERLOCK;
 587                                 }
 588                                 GIANT_SAVE();
 589                                 while (LK_HOLDER(lk->lk_lock) ==
 590                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 591                                         cpu_spinwait();
 592                                 GIANT_RESTORE();
 593                                 continue;
 594                         } else if (LK_CAN_ADAPT(lk, flags) &&
 595                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 596                             spintries < alk_retries) {
 597                                 if (flags & LK_INTERLOCK) {
 598                                         class->lc_unlock(ilk);
 599                                         flags &= ~LK_INTERLOCK;
 600                                 }
 601                                 GIANT_SAVE();
 602                                 spintries++;
 603                                 for (i = 0; i < alk_loops; i++) {
 604                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 605                                                 CTR4(KTR_LOCK,
 606                                     "%s: shared spinning on %p with %u and %u",
 607                                                     __func__, lk, spintries, i);
 608                                         x = lk->lk_lock;
 609                                         if ((x & LK_SHARE) == 0 ||
 610                                             LK_CAN_SHARE(x) != 0)
 611                                                 break;
 612                                         cpu_spinwait();
 613                                 }
 614                                 GIANT_RESTORE();
 615                                 if (i != alk_loops)
 616                                         continue;
 617                         }
 618 #endif
 619
 620                         /*
 621                          * Acquire the sleepqueue chain lock because we
 622                          * probabilly will need to manipulate waiters flags.
 623                          */
 624                         sleepq_lock(&lk->lock_object);
 625                         x = lk->lk_lock;
 626
 627                         /*
 628                          * if the lock can be acquired in shared mode, try
 629                          * again.
 630                          */
 631                         if (LK_CAN_SHARE(x)) {
 632                                 sleepq_release(&lk->lock_object);
 633                                 continue;
 634                         }
 635
 636 #ifdef ADAPTIVE_LOCKMGRS
 637                         /*
 638                          * The current lock owner might have started executing
 639                          * on another CPU (or the lock could have changed
 640                          * owner) while we were waiting on the turnstile
 641                          * chain lock.  If so, drop the turnstile lock and try
 642                          * again.
 643                          */
 644                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 645                             LK_HOLDER(x) != LK_KERNPROC) {
 646                                 owner = (struct thread *)LK_HOLDER(x);
 647                                 if (TD_IS_RUNNING(owner)) {
 648                                         sleepq_release(&lk->lock_object);
 649                                         continue;
 650                                 }
 651                         }
 652 #endif
 653
 654                         /*
 655                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 656                          * loop back and retry.
 657                          */
 658                         if ((x & LK_SHARED_WAITERS) == 0) {
 659                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 660                                     x | LK_SHARED_WAITERS)) {
 661                                         sleepq_release(&lk->lock_object);
 662                                         continue;
 663                                 }
 664                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 665                                     __func__, lk);
 666                         }
 667
 668                         /*
 669                          * As far as we have been unable to acquire the
 670                          * shared lock and the shared waiters flag is set,
 671                          * we will sleep.
 672                          */
 673                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 674                             SQ_SHARED_QUEUE);
 675                         flags &= ~LK_INTERLOCK;
 676                         if (error) {
 677                                 LOCK_LOG3(lk,
 678                                     "%s: interrupted sleep for %p with %d",
 679                                     __func__, lk, error);
 680                                 break;
 681                         }
 682                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 683                             __func__, lk);
 684                 }
 685                 if (error == 0) {
 686                         lock_profile_obtain_lock_success(&lk->lock_object,
 687                             contested, waittime, file, line);
 688                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 689                             line);
 690                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 691                             line);
 692                         TD_LOCKS_INC(curthread);
 693                         TD_SLOCKS_INC(curthread);
 694                         STACK_SAVE(lk);
 695                 }
 696                 break;
 697         case LK_UPGRADE:
 698         case LK_TRYUPGRADE:
 699                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 700                 v = lk->lk_lock;
 701                 x = v & LK_ALL_WAITERS;
 702                 v &= LK_EXCLUSIVE_SPINNERS;
 703
 704                 /*
 705                  * Try to switch from one shared lock to an exclusive one.
 706                  * We need to preserve waiters flags during the operation.
 707                  */
 708                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 709                     tid | x)) {
 710                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 711                             line);
 712                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 713                             LK_TRYWIT(flags), file, line);
 714                         TD_SLOCKS_DEC(curthread);
 715                         break;
 716                 }
 717
 718                 /*
 719                  * In LK_TRYUPGRADE mode, do not drop the lock,
 720                  * returning EBUSY instead.
 721                  */
 722                 if (op == LK_TRYUPGRADE) {
 723                         LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
 724                             __func__, lk);
 725                         error = EBUSY;
 726                         break;
 727                 }
 728
 729                 /*
 730                  * We have been unable to succeed in upgrading, so just
 731                  * give up the shared lock.
 732                  */
 733                 wakeup_swapper |= wakeupshlk(lk, file, line);
 734
 735                 /* FALLTHROUGH */
 736         case LK_EXCLUSIVE:
 737                 if (LK_CAN_WITNESS(flags))
 738                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 739                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 740                             ilk : NULL);
 741
 742                 /*
 743                  * If curthread already holds the lock and this one is
 744                  * allowed to recurse, simply recurse on it.
 745                  */
 746                 if (lockmgr_xlocked(lk)) {
 747                         if ((flags & LK_CANRECURSE) == 0 &&
 748                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 749
 750                                 /*
 751                                  * If the lock is expected to not panic just
 752                                  * give up and return.
 753                                  */
 754                                 if (LK_TRYOP(flags)) {
 755                                         LOCK_LOG2(lk,
 756                                             "%s: %p fails the try operation",
 757                                             __func__, lk);
 758                                         error = EBUSY;
 759                                         break;
 760                                 }
 761                                 if (flags & LK_INTERLOCK)
 762                                         class->lc_unlock(ilk);
 763                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 764                                     __func__, iwmesg, file, line);
 765                         }
 766                         lk->lk_recurse++;
 767                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 768                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 769                             lk->lk_recurse, file, line);
 770                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 771                             LK_TRYWIT(flags), file, line);
 772                         TD_LOCKS_INC(curthread);
 773                         break;
 774                 }
 775
 776                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
 777                     tid)) {
 778 #ifdef HWPMC_HOOKS
 779                         PMC_SOFT_CALL( , , lock, failed);
 780 #endif
 781                         lock_profile_obtain_lock_failed(&lk->lock_object,
 782                             &contested, &waittime);
 783
 784                         /*
 785                          * If the lock is expected to not sleep just give up
 786                          * and return.
 787                          */
 788                         if (LK_TRYOP(flags)) {
 789                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 790                                     __func__, lk);
 791                                 error = EBUSY;
 792                                 break;
 793                         }
 794
 795 #ifdef ADAPTIVE_LOCKMGRS
 796                         /*
 797                          * If the owner is running on another CPU, spin until
 798                          * the owner stops running or the state of the lock
 799                          * changes.
 800                          */
 801                         x = lk->lk_lock;
 802                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 803                             LK_HOLDER(x) != LK_KERNPROC) {
 804                                 owner = (struct thread *)LK_HOLDER(x);
 805                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 806                                         CTR3(KTR_LOCK,
 807                                             "%s: spinning on %p held by %p",
 808                                             __func__, lk, owner);
 809
 810                                 /*
 811                                  * If we are holding also an interlock drop it
 812                                  * in order to avoid a deadlock if the lockmgr
 813                                  * owner is adaptively spinning on the
 814                                  * interlock itself.
 815                                  */
 816                                 if (flags & LK_INTERLOCK) {
 817                                         class->lc_unlock(ilk);
 818                                         flags &= ~LK_INTERLOCK;
 819                                 }
 820                                 GIANT_SAVE();
 821                                 while (LK_HOLDER(lk->lk_lock) ==
 822                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 823                                         cpu_spinwait();
 824                                 GIANT_RESTORE();
 825                                 continue;
 826                         } else if (LK_CAN_ADAPT(lk, flags) &&
 827                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 828                             spintries < alk_retries) {
 829                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 830                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 831                                     x | LK_EXCLUSIVE_SPINNERS))
 832                                         continue;
 833                                 if (flags & LK_INTERLOCK) {
 834                                         class->lc_unlock(ilk);
 835                                         flags &= ~LK_INTERLOCK;
 836                                 }
 837                                 GIANT_SAVE();
 838                                 spintries++;
 839                                 for (i = 0; i < alk_loops; i++) {
 840                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 841                                                 CTR4(KTR_LOCK,
 842                                     "%s: shared spinning on %p with %u and %u",
 843                                                     __func__, lk, spintries, i);
 844                                         if ((lk->lk_lock &
 845                                             LK_EXCLUSIVE_SPINNERS) == 0)
 846                                                 break;
 847                                         cpu_spinwait();
 848                                 }
 849                                 GIANT_RESTORE();
 850                                 if (i != alk_loops)
 851                                         continue;
 852                         }
 853 #endif
 854
 855                         /*
 856                          * Acquire the sleepqueue chain lock because we
 857                          * probabilly will need to manipulate waiters flags.
 858                          */
 859                         sleepq_lock(&lk->lock_object);
 860                         x = lk->lk_lock;
 861
 862                         /*
 863                          * if the lock has been released while we spun on
 864                          * the sleepqueue chain lock just try again.
 865                          */
 866                         if (x == LK_UNLOCKED) {
 867                                 sleepq_release(&lk->lock_object);
 868                                 continue;
 869                         }
 870
 871 #ifdef ADAPTIVE_LOCKMGRS
 872                         /*
 873                          * The current lock owner might have started executing
 874                          * on another CPU (or the lock could have changed
 875                          * owner) while we were waiting on the turnstile
 876                          * chain lock.  If so, drop the turnstile lock and try
 877                          * again.
 878                          */
 879                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 880                             LK_HOLDER(x) != LK_KERNPROC) {
 881                                 owner = (struct thread *)LK_HOLDER(x);
 882                                 if (TD_IS_RUNNING(owner)) {
 883                                         sleepq_release(&lk->lock_object);
 884                                         continue;
 885                                 }
 886                         }
 887 #endif
 888
 889                         /*
 890                          * The lock can be in the state where there is a
 891                          * pending queue of waiters, but still no owner.
 892                          * This happens when the lock is contested and an
 893                          * owner is going to claim the lock.
 894                          * If curthread is the one successfully acquiring it
 895                          * claim lock ownership and return, preserving waiters
 896                          * flags.
 897                          */
 898                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 899                         if ((x & ~v) == LK_UNLOCKED) {
 900                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 901                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 902                                     tid | v)) {
 903                                         sleepq_release(&lk->lock_object);
 904                                         LOCK_LOG2(lk,
 905                                             "%s: %p claimed by a new writer",
 906                                             __func__, lk);
 907                                         break;
 908                                 }
 909                                 sleepq_release(&lk->lock_object);
 910                                 continue;
 911                         }
 912
 913                         /*
 914                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 915                          * fail, loop back and retry.
 916                          */
 917                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 918                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 919                                     x | LK_EXCLUSIVE_WAITERS)) {
 920                                         sleepq_release(&lk->lock_object);
 921                                         continue;
 922                                 }
 923                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 924                                     __func__, lk);
 925                         }
 926
 927                         /*
 928                          * As far as we have been unable to acquire the
 929                          * exclusive lock and the exclusive waiters flag
 930                          * is set, we will sleep.
 931                          */
 932                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 933                             SQ_EXCLUSIVE_QUEUE);
 934                         flags &= ~LK_INTERLOCK;
 935                         if (error) {
 936                                 LOCK_LOG3(lk,
 937                                     "%s: interrupted sleep for %p with %d",
 938                                     __func__, lk, error);
 939                                 break;
 940                         }
 941                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 942                             __func__, lk);
 943                 }
 944                 if (error == 0) {
 945                         lock_profile_obtain_lock_success(&lk->lock_object,
 946                             contested, waittime, file, line);
 947                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 948                             lk->lk_recurse, file, line);
 949                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 950                             LK_TRYWIT(flags), file, line);
 951                         TD_LOCKS_INC(curthread);
 952                         STACK_SAVE(lk);
 953                 }
 954                 break;
 955         case LK_DOWNGRADE:
 956                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 957                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 958                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 959
 960                 /*
 961                  * Panic if the lock is recursed.
 962                  */
 963                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 964                         if (flags & LK_INTERLOCK)
 965                                 class->lc_unlock(ilk);
 966                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 967                             __func__, iwmesg, file, line);
 968                 }
 969                 TD_SLOCKS_INC(curthread);
 970
 971                 /*
 972                  * In order to preserve waiters flags, just spin.
 973                  */
 974                 for (;;) {
 975                         x = lk->lk_lock;
 976                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
 977                         x &= LK_ALL_WAITERS;
 978                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
 979                             LK_SHARERS_LOCK(1) | x))
 980                                 break;
 981                         cpu_spinwait();
 982                 }
 983                 break;
 984         case LK_RELEASE:
 985                 _lockmgr_assert(lk, KA_LOCKED, file, line);
 986                 x = lk->lk_lock;
 987
 988                 if ((x & LK_SHARE) == 0) {
 989
 990                         /*
 991                          * As first option, treact the lock as if it has not
 992                          * any waiter.
 993                          * Fix-up the tid var if the lock has been disowned.
 994                          */
 995                         if (LK_HOLDER(x) == LK_KERNPROC)
 996                                 tid = LK_KERNPROC;
 997                         else {
 998                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
 999                                     file, line);
1000                                 TD_LOCKS_DEC(curthread);
1001                         }
1002                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
1003                             lk->lk_recurse, file, line);
1004
1005                         /*
1006                          * The lock is held in exclusive mode.
1007                          * If the lock is recursed also, then unrecurse it.
1008                          */
1009                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1010                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1011                                     lk);
1012                                 lk->lk_recurse--;
1013                                 break;
1014                         }
1015                         if (tid != LK_KERNPROC)
1016                                 lock_profile_release_lock(&lk->lock_object);
1017
1018                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1019                             LK_UNLOCKED))
1020                                 break;
1021
1022                         sleepq_lock(&lk->lock_object);
1023                         x = lk->lk_lock;
1024                         v = LK_UNLOCKED;
1025
1026                         /*
1027                          * If the lock has exclusive waiters, give them
1028                          * preference in order to avoid deadlock with
1029                          * shared runners up.
1030                          * If interruptible sleeps left the exclusive queue
1031                          * empty avoid a starvation for the threads sleeping
1032                          * on the shared queue by giving them precedence
1033                          * and cleaning up the exclusive waiters bit anyway.
1034                          * Please note that lk_exslpfail count may be lying
1035                          * about the real number of waiters with the
1036                          * LK_SLEEPFAIL flag on because they may be used in
1037                          * conjuction with interruptible sleeps so
1038                          * lk_exslpfail might be considered an 'upper limit'
1039                          * bound, including the edge cases.
1040                          */
1041                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1042                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1043                             SQ_EXCLUSIVE_QUEUE);
1044                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1045                                 if (lk->lk_exslpfail < realexslp) {
1046                                         lk->lk_exslpfail = 0;
1047                                         queue = SQ_EXCLUSIVE_QUEUE;
1048                                         v |= (x & LK_SHARED_WAITERS);
1049                                 } else {
1050                                         lk->lk_exslpfail = 0;
1051                                         LOCK_LOG2(lk,
1052                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1053                                             __func__, lk);
1054                                         LOCK_LOG2(lk,
1055                         "%s: %p waking up threads on the exclusive queue",
1056                                             __func__, lk);
1057                                         wakeup_swapper =
1058                                             sleepq_broadcast(&lk->lock_object,
1059                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1060                                         queue = SQ_SHARED_QUEUE;
1061                                 }
1062                         } else {
1063
1064                                 /*
1065                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1066                                  * on and using interruptible sleeps/timeout
1067                                  * may have left spourious lk_exslpfail counts
1068                                  * on, so clean it up anyway.
1069                                  */
1070                                 lk->lk_exslpfail = 0;
1071                                 queue = SQ_SHARED_QUEUE;
1072                         }
1073
1074                         LOCK_LOG3(lk,
1075                             "%s: %p waking up threads on the %s queue",
1076                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1077                             "exclusive");
1078                         atomic_store_rel_ptr(&lk->lk_lock, v);
1079                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1080                             SLEEPQ_LK, 0, queue);
1081                         sleepq_release(&lk->lock_object);
1082                         break;
1083                 } else
1084                         wakeup_swapper = wakeupshlk(lk, file, line);
1085                 break;
1086         case LK_DRAIN:
1087                 if (LK_CAN_WITNESS(flags))
1088                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1089                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1090                             ilk : NULL);
1091
1092                 /*
1093                  * Trying to drain a lock we already own will result in a
1094                  * deadlock.
1095                  */
1096                 if (lockmgr_xlocked(lk)) {
1097                         if (flags & LK_INTERLOCK)
1098                                 class->lc_unlock(ilk);
1099                         panic("%s: draining %s with the lock held @ %s:%d\n",
1100                             __func__, iwmesg, file, line);
1101                 }
1102
1103                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1104 #ifdef HWPMC_HOOKS
1105                         PMC_SOFT_CALL( , , lock, failed);
1106 #endif
1107                         lock_profile_obtain_lock_failed(&lk->lock_object,
1108                             &contested, &waittime);
1109
1110                         /*
1111                          * If the lock is expected to not sleep just give up
1112                          * and return.
1113                          */
1114                         if (LK_TRYOP(flags)) {
1115                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1116                                     __func__, lk);
1117                                 error = EBUSY;
1118                                 break;
1119                         }
1120
1121                         /*
1122                          * Acquire the sleepqueue chain lock because we
1123                          * probabilly will need to manipulate waiters flags.
1124                          */
1125                         sleepq_lock(&lk->lock_object);
1126                         x = lk->lk_lock;
1127
1128                         /*
1129                          * if the lock has been released while we spun on
1130                          * the sleepqueue chain lock just try again.
1131                          */
1132                         if (x == LK_UNLOCKED) {
1133                                 sleepq_release(&lk->lock_object);
1134                                 continue;
1135                         }
1136
1137                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1138                         if ((x & ~v) == LK_UNLOCKED) {
1139                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1140
1141                                 /*
1142                                  * If interruptible sleeps left the exclusive
1143                                  * queue empty avoid a starvation for the
1144                                  * threads sleeping on the shared queue by
1145                                  * giving them precedence and cleaning up the
1146                                  * exclusive waiters bit anyway.
1147                                  * Please note that lk_exslpfail count may be
1148                                  * lying about the real number of waiters with
1149                                  * the LK_SLEEPFAIL flag on because they may
1150                                  * be used in conjuction with interruptible
1151                                  * sleeps so lk_exslpfail might be considered
1152                                  * an 'upper limit' bound, including the edge
1153                                  * cases.
1154                                  */
1155                                 if (v & LK_EXCLUSIVE_WAITERS) {
1156                                         queue = SQ_EXCLUSIVE_QUEUE;
1157                                         v &= ~LK_EXCLUSIVE_WAITERS;
1158                                 } else {
1159
1160                                         /*
1161                                          * Exclusive waiters sleeping with
1162                                          * LK_SLEEPFAIL on and using
1163                                          * interruptible sleeps/timeout may
1164                                          * have left spourious lk_exslpfail
1165                                          * counts on, so clean it up anyway.
1166                                          */
1167                                         MPASS(v & LK_SHARED_WAITERS);
1168                                         lk->lk_exslpfail = 0;
1169                                         queue = SQ_SHARED_QUEUE;
1170                                         v &= ~LK_SHARED_WAITERS;
1171                                 }
1172                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1173                                         realexslp =
1174                                             sleepq_sleepcnt(&lk->lock_object,
1175                                             SQ_EXCLUSIVE_QUEUE);
1176                                         if (lk->lk_exslpfail >= realexslp) {
1177                                                 lk->lk_exslpfail = 0;
1178                                                 queue = SQ_SHARED_QUEUE;
1179                                                 v &= ~LK_SHARED_WAITERS;
1180                                                 if (realexslp != 0) {
1181                                                         LOCK_LOG2(lk,
1182                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1183                                                             __func__, lk);
1184                                                         LOCK_LOG2(lk,
1185                         "%s: %p waking up threads on the exclusive queue",
1186                                                             __func__, lk);
1187                                                         wakeup_swapper =
1188                                                             sleepq_broadcast(
1189                                                             &lk->lock_object,
1190                                                             SLEEPQ_LK, 0,
1191                                                             SQ_EXCLUSIVE_QUEUE);
1192                                                 }
1193                                         } else
1194                                                 lk->lk_exslpfail = 0;
1195                                 }
1196                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1197                                         sleepq_release(&lk->lock_object);
1198                                         continue;
1199                                 }
1200                                 LOCK_LOG3(lk,
1201                                 "%s: %p waking up all threads on the %s queue",
1202                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1203                                     "shared" : "exclusive");
1204                                 wakeup_swapper |= sleepq_broadcast(
1205                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1206
1207                                 /*
1208                                  * If shared waiters have been woken up we need
1209                                  * to wait for one of them to acquire the lock
1210                                  * before to set the exclusive waiters in
1211                                  * order to avoid a deadlock.
1212                                  */
1213                                 if (queue == SQ_SHARED_QUEUE) {
1214                                         for (v = lk->lk_lock;
1215                                             (v & LK_SHARE) && !LK_SHARERS(v);
1216                                             v = lk->lk_lock)
1217                                                 cpu_spinwait();
1218                                 }
1219                         }
1220
1221                         /*
1222                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1223                          * fail, loop back and retry.
1224                          */
1225                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1226                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1227                                     x | LK_EXCLUSIVE_WAITERS)) {
1228                                         sleepq_release(&lk->lock_object);
1229                                         continue;
1230                                 }
1231                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1232                                     __func__, lk);
1233                         }
1234
1235                         /*
1236                          * As far as we have been unable to acquire the
1237                          * exclusive lock and the exclusive waiters flag
1238                          * is set, we will sleep.
1239                          */
1240                         if (flags & LK_INTERLOCK) {
1241                                 class->lc_unlock(ilk);
1242                                 flags &= ~LK_INTERLOCK;
1243                         }
1244                         GIANT_SAVE();
1245                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1246                             SQ_EXCLUSIVE_QUEUE);
1247                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1248                         GIANT_RESTORE();
1249                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1250                             __func__, lk);
1251                 }
1252
1253                 if (error == 0) {
1254                         lock_profile_obtain_lock_success(&lk->lock_object,
1255                             contested, waittime, file, line);
1256                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1257                             lk->lk_recurse, file, line);
1258                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1259                             LK_TRYWIT(flags), file, line);
1260                         TD_LOCKS_INC(curthread);
1261                         STACK_SAVE(lk);
1262                 }
1263                 break;
1264         default:
1265                 if (flags & LK_INTERLOCK)
1266                         class->lc_unlock(ilk);
1267                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1268         }
1269
1270         if (flags & LK_INTERLOCK)
1271                 class->lc_unlock(ilk);
1272         if (wakeup_swapper)
1273                 kick_proc0();
1274
1275         return (error);
1276 }
1277
1278 void
1279 _lockmgr_disown(struct lock *lk, const char *file, int line)
1280 {
1281         uintptr_t tid, x;
1282
1283         if (SCHEDULER_STOPPED())
1284                 return;
1285
1286         tid = (uintptr_t)curthread;
1287         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1288
1289         /*
1290          * Panic if the lock is recursed.
1291          */
1292         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1293                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1294                     __func__,  file, line);
1295
1296         /*
1297          * If the owner is already LK_KERNPROC just skip the whole operation.
1298          */
1299         if (LK_HOLDER(lk->lk_lock) != tid)
1300                 return;
1301         lock_profile_release_lock(&lk->lock_object);
1302         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1303         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1304         TD_LOCKS_DEC(curthread);
1305         STACK_SAVE(lk);
1306
1307         /*
1308          * In order to preserve waiters flags, just spin.
1309          */
1310         for (;;) {
1311                 x = lk->lk_lock;
1312                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1313                 x &= LK_ALL_WAITERS;
1314                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1315                     LK_KERNPROC | x))
1316                         return;
1317                 cpu_spinwait();
1318         }
1319 }
1320
1321 void
1322 lockmgr_printinfo(const struct lock *lk)
1323 {
1324         struct thread *td;
1325         uintptr_t x;
1326
1327         if (lk->lk_lock == LK_UNLOCKED)
1328                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1329         else if (lk->lk_lock & LK_SHARE)
1330                 printf("lock type %s: SHARED (count %ju)\n",
1331                     lk->lock_object.lo_name,
1332                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1333         else {
1334                 td = lockmgr_xholder(lk);
1335                 printf("lock type %s: EXCL by thread %p "
1336                     "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, td,
1337                     td->td_proc->p_pid, td->td_proc->p_comm, td->td_tid);
1338         }
1339
1340         x = lk->lk_lock;
1341         if (x & LK_EXCLUSIVE_WAITERS)
1342                 printf(" with exclusive waiters pending\n");
1343         if (x & LK_SHARED_WAITERS)
1344                 printf(" with shared waiters pending\n");
1345         if (x & LK_EXCLUSIVE_SPINNERS)
1346                 printf(" with exclusive spinners pending\n");
1347
1348         STACK_PRINT(lk);
1349 }
1350
1351 int
1352 lockstatus(const struct lock *lk)
1353 {
1354         uintptr_t v, x;
1355         int ret;
1356
1357         ret = LK_SHARED;
1358         x = lk->lk_lock;
1359         v = LK_HOLDER(x);
1360
1361         if ((x & LK_SHARE) == 0) {
1362                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1363                         ret = LK_EXCLUSIVE;
1364                 else
1365                         ret = LK_EXCLOTHER;
1366         } else if (x == LK_UNLOCKED)
1367                 ret = 0;
1368
1369         return (ret);
1370 }
1371
1372 #ifdef INVARIANT_SUPPORT
1373
1374 FEATURE(invariant_support,
1375     "Support for modules compiled with INVARIANTS option");
1376
1377 #ifndef INVARIANTS
1378 #undef  _lockmgr_assert
1379 #endif
1380
1381 void
1382 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1383 {
1384         int slocked = 0;
1385
1386         if (panicstr != NULL)
1387                 return;
1388         switch (what) {
1389         case KA_SLOCKED:
1390         case KA_SLOCKED | KA_NOTRECURSED:
1391         case KA_SLOCKED | KA_RECURSED:
1392                 slocked = 1;
1393         case KA_LOCKED:
1394         case KA_LOCKED | KA_NOTRECURSED:
1395         case KA_LOCKED | KA_RECURSED:
1396 #ifdef WITNESS
1397
1398                 /*
1399                  * We cannot trust WITNESS if the lock is held in exclusive
1400                  * mode and a call to lockmgr_disown() happened.
1401                  * Workaround this skipping the check if the lock is held in
1402                  * exclusive mode even for the KA_LOCKED case.
1403                  */
1404                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1405                         witness_assert(&lk->lock_object, what, file, line);
1406                         break;
1407                 }
1408 #endif
1409                 if (lk->lk_lock == LK_UNLOCKED ||
1410                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1411                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1412                         panic("Lock %s not %slocked @ %s:%d\n",
1413                             lk->lock_object.lo_name, slocked ? "share" : "",
1414                             file, line);
1415
1416                 if ((lk->lk_lock & LK_SHARE) == 0) {
1417                         if (lockmgr_recursed(lk)) {
1418                                 if (what & KA_NOTRECURSED)
1419                                         panic("Lock %s recursed @ %s:%d\n",
1420                                             lk->lock_object.lo_name, file,
1421                                             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                 }
1426                 break;
1427         case KA_XLOCKED:
1428         case KA_XLOCKED | KA_NOTRECURSED:
1429         case KA_XLOCKED | KA_RECURSED:
1430                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1431                         panic("Lock %s not exclusively locked @ %s:%d\n",
1432                             lk->lock_object.lo_name, file, line);
1433                 if (lockmgr_recursed(lk)) {
1434                         if (what & KA_NOTRECURSED)
1435                                 panic("Lock %s recursed @ %s:%d\n",
1436                                     lk->lock_object.lo_name, file, line);
1437                 } else if (what & KA_RECURSED)
1438                         panic("Lock %s not recursed @ %s:%d\n",
1439                             lk->lock_object.lo_name, file, line);
1440                 break;
1441         case KA_UNLOCKED:
1442                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1443                         panic("Lock %s exclusively locked @ %s:%d\n",
1444                             lk->lock_object.lo_name, file, line);
1445                 break;
1446         default:
1447                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1448                     line);
1449         }
1450 }
1451 #endif
1452
1453 #ifdef DDB
1454 int
1455 lockmgr_chain(struct thread *td, struct thread **ownerp)
1456 {
1457         struct lock *lk;
1458
1459         lk = td->td_wchan;
1460
1461         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1462                 return (0);
1463         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1464         if (lk->lk_lock & LK_SHARE)
1465                 db_printf("SHARED (count %ju)\n",
1466                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1467         else
1468                 db_printf("EXCL\n");
1469         *ownerp = lockmgr_xholder(lk);
1470
1471         return (1);
1472 }
1473
1474 static void
1475 db_show_lockmgr(const struct lock_object *lock)
1476 {
1477         struct thread *td;
1478         const struct lock *lk;
1479
1480         lk = (const struct lock *)lock;
1481
1482         db_printf(" state: ");
1483         if (lk->lk_lock == LK_UNLOCKED)
1484                 db_printf("UNLOCKED\n");
1485         else if (lk->lk_lock & LK_SHARE)
1486                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1487         else {
1488                 td = lockmgr_xholder(lk);
1489                 if (td == (struct thread *)LK_KERNPROC)
1490                         db_printf("XLOCK: LK_KERNPROC\n");
1491                 else
1492                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1493                             td->td_tid, td->td_proc->p_pid,
1494                             td->td_proc->p_comm);
1495                 if (lockmgr_recursed(lk))
1496                         db_printf(" recursed: %d\n", lk->lk_recurse);
1497         }
1498         db_printf(" waiters: ");
1499         switch (lk->lk_lock & LK_ALL_WAITERS) {
1500         case LK_SHARED_WAITERS:
1501                 db_printf("shared\n");
1502                 break;
1503         case LK_EXCLUSIVE_WAITERS:
1504                 db_printf("exclusive\n");
1505                 break;
1506         case LK_ALL_WAITERS:
1507                 db_printf("shared and exclusive\n");
1508                 break;
1509         default:
1510                 db_printf("none\n");
1511         }
1512         db_printf(" spinners: ");
1513         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1514                 db_printf("exclusive\n");
1515         else
1516                 db_printf("none\n");
1517 }
1518 #endif