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