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