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