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