sys/kern/kern_lock.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice(s), this list of conditions and the following disclaimer as
  12  *    the first lines of this file unmodified other than the possible
  13  *    addition of one or more copyright notices.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice(s), this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  19  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  20  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  21  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  22  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  25  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  28  * DAMAGE.
  29  */
  30
  31 #include "opt_ddb.h"
  32 #include "opt_hwpmc_hooks.h"
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include <sys/param.h>
  38 #include <sys/kdb.h>
  39 #include <sys/ktr.h>
  40 #include <sys/lock.h>
  41 #include <sys/lock_profile.h>
  42 #include <sys/lockmgr.h>
  43 #include <sys/mutex.h>
  44 #include <sys/proc.h>
  45 #include <sys/sleepqueue.h>
  46 #ifdef DEBUG_LOCKS
  47 #include <sys/stack.h>
  48 #endif
  49 #include <sys/sysctl.h>
  50 #include <sys/systm.h>
  51
  52 #include <machine/cpu.h>
  53
  54 #ifdef DDB
  55 #include <ddb/ddb.h>
  56 #endif
  57
  58 #ifdef HWPMC_HOOKS
  59 #include <sys/pmckern.h>
  60 PMC_SOFT_DECLARE( , , lock, failed);
  61 #endif
  62
  63 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
  64     (LK_ADAPTIVE | LK_NOSHARE));
  65 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
  66     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
  67
  68 #define SQ_EXCLUSIVE_QUEUE      0
  69 #define SQ_SHARED_QUEUE         1
  70
  71 #ifndef INVARIANTS
  72 #define _lockmgr_assert(lk, what, file, line)
  73 #endif
  74
  75 #define TD_SLOCKS_INC(td)       ((td)->td_lk_slocks++)
  76 #define TD_SLOCKS_DEC(td)       ((td)->td_lk_slocks--)
  77
  78 #ifndef DEBUG_LOCKS
  79 #define STACK_PRINT(lk)
  80 #define STACK_SAVE(lk)
  81 #define STACK_ZERO(lk)
  82 #else
  83 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
  84 #define STACK_SAVE(lk)  stack_save(&(lk)->lk_stack)
  85 #define STACK_ZERO(lk)  stack_zero(&(lk)->lk_stack)
  86 #endif
  87
  88 #define LOCK_LOG2(lk, string, arg1, arg2)                               \
  89         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  90                 CTR2(KTR_LOCK, (string), (arg1), (arg2))
  91 #define LOCK_LOG3(lk, string, arg1, arg2, arg3)                         \
  92         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  93                 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
  94
  95 #define GIANT_DECLARE                                                   \
  96         int _i = 0;                                                     \
  97         WITNESS_SAVE_DECL(Giant)
  98 #define GIANT_RESTORE() do {                                            \
  99         if (_i > 0) {                                                   \
 100                 while (_i--)                                            \
 101                         mtx_lock(&Giant);                               \
 102                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 103         }                                                               \
 104 } while (0)
 105 #define GIANT_SAVE() do {                                               \
 106         if (mtx_owned(&Giant)) {                                        \
 107                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
 108                 while (mtx_owned(&Giant)) {                             \
 109                         _i++;                                           \
 110                         mtx_unlock(&Giant);                             \
 111                 }                                                       \
 112         }                                                               \
 113 } while (0)
 114
 115 static bool __always_inline
 116 LK_CAN_SHARE(uintptr_t x, int flags, bool fp)
 117 {
 118
 119         if ((x & (LK_SHARE | LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) ==
 120             LK_SHARE)
 121                 return (true);
 122         if (fp || (!(x & LK_SHARE)))
 123                 return (false);
 124         if ((curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||
 125             (curthread->td_pflags & TDP_DEADLKTREAT))
 126                 return (true);
 127         return (false);
 128 }
 129
 130 #define LK_TRYOP(x)                                                     \
 131         ((x) & LK_NOWAIT)
 132
 133 #define LK_CAN_WITNESS(x)                                               \
 134         (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
 135 #define LK_TRYWIT(x)                                                    \
 136         (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
 137
 138 #define LK_CAN_ADAPT(lk, f)                                             \
 139         (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&             \
 140         ((f) & LK_SLEEPFAIL) == 0)
 141
 142 #define lockmgr_disowned(lk)                                            \
 143         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
 144
 145 #define lockmgr_xlocked_v(v)                                            \
 146         (((v) & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
 147
 148 #define lockmgr_xlocked(lk) lockmgr_xlocked_v((lk)->lk_lock)
 149
 150 static void     assert_lockmgr(const struct lock_object *lock, int how);
 151 #ifdef DDB
 152 static void     db_show_lockmgr(const struct lock_object *lock);
 153 #endif
 154 static void     lock_lockmgr(struct lock_object *lock, uintptr_t how);
 155 #ifdef KDTRACE_HOOKS
 156 static int      owner_lockmgr(const struct lock_object *lock,
 157                     struct thread **owner);
 158 #endif
 159 static uintptr_t unlock_lockmgr(struct lock_object *lock);
 160
 161 struct lock_class lock_class_lockmgr = {
 162         .lc_name = "lockmgr",
 163         .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
 164         .lc_assert = assert_lockmgr,
 165 #ifdef DDB
 166         .lc_ddb_show = db_show_lockmgr,
 167 #endif
 168         .lc_lock = lock_lockmgr,
 169         .lc_unlock = unlock_lockmgr,
 170 #ifdef KDTRACE_HOOKS
 171         .lc_owner = owner_lockmgr,
 172 #endif
 173 };
 174
 175 struct lockmgr_wait {
 176         const char *iwmesg;
 177         int ipri;
 178         int itimo;
 179 };
 180
 181 static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp,
 182     int flags, bool fp);
 183 static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp);
 184
 185 static void
 186 lockmgr_exit(u_int flags, struct lock_object *ilk, int wakeup_swapper)
 187 {
 188         struct lock_class *class;
 189
 190         if (flags & LK_INTERLOCK) {
 191                 class = LOCK_CLASS(ilk);
 192                 class->lc_unlock(ilk);
 193         }
 194
 195         if (__predict_false(wakeup_swapper))
 196                 kick_proc0();
 197 }
 198
 199 static void
 200 lockmgr_note_shared_acquire(struct lock *lk, int contested,
 201     uint64_t waittime, const char *file, int line, int flags)
 202 {
 203
 204         lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
 205             file, line);
 206         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line);
 207         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line);
 208         TD_LOCKS_INC(curthread);
 209         TD_SLOCKS_INC(curthread);
 210         STACK_SAVE(lk);
 211 }
 212
 213 static void
 214 lockmgr_note_shared_release(struct lock *lk, const char *file, int line)
 215 {
 216
 217         lock_profile_release_lock(&lk->lock_object);
 218         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 219         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 220         TD_LOCKS_DEC(curthread);
 221         TD_SLOCKS_DEC(curthread);
 222 }
 223
 224 static void
 225 lockmgr_note_exclusive_acquire(struct lock *lk, int contested,
 226     uint64_t waittime, const char *file, int line, int flags)
 227 {
 228
 229         lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime,
 230             file, line);
 231         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
 232         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file,
 233             line);
 234         TD_LOCKS_INC(curthread);
 235         STACK_SAVE(lk);
 236 }
 237
 238 static void
 239 lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line)
 240 {
 241
 242         lock_profile_release_lock(&lk->lock_object);
 243         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file,
 244             line);
 245         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
 246         TD_LOCKS_DEC(curthread);
 247 }
 248
 249 static __inline struct thread *
 250 lockmgr_xholder(const struct lock *lk)
 251 {
 252         uintptr_t x;
 253
 254         x = lk->lk_lock;
 255         return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
 256 }
 257
 258 /*
 259  * It assumes sleepq_lock held and returns with this one unheld.
 260  * It also assumes the generic interlock is sane and previously checked.
 261  * If LK_INTERLOCK is specified the interlock is not reacquired after the
 262  * sleep.
 263  */
 264 static __inline int
 265 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
 266     const char *wmesg, int pri, int timo, int queue)
 267 {
 268         GIANT_DECLARE;
 269         struct lock_class *class;
 270         int catch, error;
 271
 272         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 273         catch = pri & PCATCH;
 274         pri &= PRIMASK;
 275         error = 0;
 276
 277         LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
 278             (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
 279
 280         if (flags & LK_INTERLOCK)
 281                 class->lc_unlock(ilk);
 282         if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
 283                 lk->lk_exslpfail++;
 284         GIANT_SAVE();
 285         sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
 286             SLEEPQ_INTERRUPTIBLE : 0), queue);
 287         if ((flags & LK_TIMELOCK) && timo)
 288                 sleepq_set_timeout(&lk->lock_object, timo);
 289
 290         /*
 291          * Decisional switch for real sleeping.
 292          */
 293         if ((flags & LK_TIMELOCK) && timo && catch)
 294                 error = sleepq_timedwait_sig(&lk->lock_object, pri);
 295         else if ((flags & LK_TIMELOCK) && timo)
 296                 error = sleepq_timedwait(&lk->lock_object, pri);
 297         else if (catch)
 298                 error = sleepq_wait_sig(&lk->lock_object, pri);
 299         else
 300                 sleepq_wait(&lk->lock_object, pri);
 301         GIANT_RESTORE();
 302         if ((flags & LK_SLEEPFAIL) && error == 0)
 303                 error = ENOLCK;
 304
 305         return (error);
 306 }
 307
 308 static __inline int
 309 wakeupshlk(struct lock *lk, const char *file, int line)
 310 {
 311         uintptr_t v, x, orig_x;
 312         u_int realexslp;
 313         int queue, wakeup_swapper;
 314
 315         wakeup_swapper = 0;
 316         for (;;) {
 317                 x = lk->lk_lock;
 318                 if (lockmgr_sunlock_try(lk, &x))
 319                         break;
 320
 321                 /*
 322                  * We should have a sharer with waiters, so enter the hard
 323                  * path in order to handle wakeups correctly.
 324                  */
 325                 sleepq_lock(&lk->lock_object);
 326                 orig_x = lk->lk_lock;
 327 retry_sleepq:
 328                 x = orig_x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 329                 v = LK_UNLOCKED;
 330
 331                 /*
 332                  * If the lock has exclusive waiters, give them preference in
 333                  * order to avoid deadlock with shared runners up.
 334                  * If interruptible sleeps left the exclusive queue empty
 335                  * avoid a starvation for the threads sleeping on the shared
 336                  * queue by giving them precedence and cleaning up the
 337                  * exclusive waiters bit anyway.
 338                  * Please note that lk_exslpfail count may be lying about
 339                  * the real number of waiters with the LK_SLEEPFAIL flag on
 340                  * because they may be used in conjunction with interruptible
 341                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 342                  * bound, including the edge cases.
 343                  */
 344                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 345                     SQ_EXCLUSIVE_QUEUE);
 346                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 347                         if (lk->lk_exslpfail < realexslp) {
 348                                 lk->lk_exslpfail = 0;
 349                                 queue = SQ_EXCLUSIVE_QUEUE;
 350                                 v |= (x & LK_SHARED_WAITERS);
 351                         } else {
 352                                 lk->lk_exslpfail = 0;
 353                                 LOCK_LOG2(lk,
 354                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 355                                     __func__, lk);
 356                                 LOCK_LOG2(lk,
 357                             "%s: %p waking up threads on the exclusive queue",
 358                                     __func__, lk);
 359                                 wakeup_swapper =
 360                                     sleepq_broadcast(&lk->lock_object,
 361                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 362                                 queue = SQ_SHARED_QUEUE;
 363                         }
 364
 365                 } else {
 366
 367                         /*
 368                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 369                          * and using interruptible sleeps/timeout may have
 370                          * left spourious lk_exslpfail counts on, so clean
 371                          * it up anyway.
 372                          */
 373                         lk->lk_exslpfail = 0;
 374                         queue = SQ_SHARED_QUEUE;
 375                 }
 376
 377                 if (lockmgr_sunlock_try(lk, &orig_x)) {
 378                         sleepq_release(&lk->lock_object);
 379                         break;
 380                 }
 381
 382                 x |= LK_SHARERS_LOCK(1);
 383                 if (!atomic_fcmpset_rel_ptr(&lk->lk_lock, &x, v)) {
 384                         orig_x = x;
 385                         goto retry_sleepq;
 386                 }
 387                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 388                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 389                     "exclusive");
 390                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 391                     0, queue);
 392                 sleepq_release(&lk->lock_object);
 393                 break;
 394         }
 395
 396         lockmgr_note_shared_release(lk, file, line);
 397         return (wakeup_swapper);
 398 }
 399
 400 static void
 401 assert_lockmgr(const struct lock_object *lock, int what)
 402 {
 403
 404         panic("lockmgr locks do not support assertions");
 405 }
 406
 407 static void
 408 lock_lockmgr(struct lock_object *lock, uintptr_t how)
 409 {
 410
 411         panic("lockmgr locks do not support sleep interlocking");
 412 }
 413
 414 static uintptr_t
 415 unlock_lockmgr(struct lock_object *lock)
 416 {
 417
 418         panic("lockmgr locks do not support sleep interlocking");
 419 }
 420
 421 #ifdef KDTRACE_HOOKS
 422 static int
 423 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
 424 {
 425
 426         panic("lockmgr locks do not support owner inquiring");
 427 }
 428 #endif
 429
 430 void
 431 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 432 {
 433         int iflags;
 434
 435         MPASS((flags & ~LK_INIT_MASK) == 0);
 436         ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
 437             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
 438             &lk->lk_lock));
 439
 440         iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 441         if (flags & LK_CANRECURSE)
 442                 iflags |= LO_RECURSABLE;
 443         if ((flags & LK_NODUP) == 0)
 444                 iflags |= LO_DUPOK;
 445         if (flags & LK_NOPROFILE)
 446                 iflags |= LO_NOPROFILE;
 447         if ((flags & LK_NOWITNESS) == 0)
 448                 iflags |= LO_WITNESS;
 449         if (flags & LK_QUIET)
 450                 iflags |= LO_QUIET;
 451         if (flags & LK_IS_VNODE)
 452                 iflags |= LO_IS_VNODE;
 453         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 454
 455         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 456         lk->lk_lock = LK_UNLOCKED;
 457         lk->lk_recurse = 0;
 458         lk->lk_exslpfail = 0;
 459         lk->lk_timo = timo;
 460         lk->lk_pri = pri;
 461         STACK_ZERO(lk);
 462 }
 463
 464 /*
 465  * XXX: Gross hacks to manipulate external lock flags after
 466  * initialization.  Used for certain vnode and buf locks.
 467  */
 468 void
 469 lockallowshare(struct lock *lk)
 470 {
 471
 472         lockmgr_assert(lk, KA_XLOCKED);
 473         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 474 }
 475
 476 void
 477 lockdisableshare(struct lock *lk)
 478 {
 479
 480         lockmgr_assert(lk, KA_XLOCKED);
 481         lk->lock_object.lo_flags |= LK_NOSHARE;
 482 }
 483
 484 void
 485 lockallowrecurse(struct lock *lk)
 486 {
 487
 488         lockmgr_assert(lk, KA_XLOCKED);
 489         lk->lock_object.lo_flags |= LO_RECURSABLE;
 490 }
 491
 492 void
 493 lockdisablerecurse(struct lock *lk)
 494 {
 495
 496         lockmgr_assert(lk, KA_XLOCKED);
 497         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 498 }
 499
 500 void
 501 lockdestroy(struct lock *lk)
 502 {
 503
 504         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 505         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 506         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 507         lock_destroy(&lk->lock_object);
 508 }
 509
 510 static bool __always_inline
 511 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp)
 512 {
 513
 514         /*
 515          * If no other thread has an exclusive lock, or
 516          * no exclusive waiter is present, bump the count of
 517          * sharers.  Since we have to preserve the state of
 518          * waiters, if we fail to acquire the shared lock
 519          * loop back and retry.
 520          */
 521         *xp = lk->lk_lock;
 522         while (LK_CAN_SHARE(*xp, flags, fp)) {
 523                 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
 524                     *xp + LK_ONE_SHARER)) {
 525                         return (true);
 526                 }
 527         }
 528         return (false);
 529 }
 530
 531 static bool __always_inline
 532 lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp)
 533 {
 534
 535         for (;;) {
 536                 if (LK_SHARERS(*xp) > 1 || !(*xp & LK_ALL_WAITERS)) {
 537                         if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp,
 538                             *xp - LK_ONE_SHARER))
 539                                 return (true);
 540                         continue;
 541                 }
 542                 break;
 543         }
 544         return (false);
 545 }
 546
 547 static __noinline int
 548 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 549     const char *file, int line, struct lockmgr_wait *lwa)
 550 {
 551         uintptr_t tid, x;
 552         int error = 0;
 553         const char *iwmesg;
 554         int ipri, itimo;
 555
 556 #ifdef LOCK_PROFILING
 557         uint64_t waittime = 0;
 558         int contested = 0;
 559 #endif
 560
 561         if (__predict_false(panicstr != NULL))
 562                 goto out;
 563
 564         tid = (uintptr_t)curthread;
 565
 566         if (LK_CAN_WITNESS(flags))
 567                 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 568                     file, line, flags & LK_INTERLOCK ? ilk : NULL);
 569         for (;;) {
 570                 if (lockmgr_slock_try(lk, &x, flags, false))
 571                         break;
 572 #ifdef HWPMC_HOOKS
 573                 PMC_SOFT_CALL( , , lock, failed);
 574 #endif
 575                 lock_profile_obtain_lock_failed(&lk->lock_object,
 576                     &contested, &waittime);
 577
 578                 /*
 579                  * If the lock is already held by curthread in
 580                  * exclusive way avoid a deadlock.
 581                  */
 582                 if (LK_HOLDER(x) == tid) {
 583                         LOCK_LOG2(lk,
 584                             "%s: %p already held in exclusive mode",
 585                             __func__, lk);
 586                         error = EDEADLK;
 587                         break;
 588                 }
 589
 590                 /*
 591                  * If the lock is expected to not sleep just give up
 592                  * and return.
 593                  */
 594                 if (LK_TRYOP(flags)) {
 595                         LOCK_LOG2(lk, "%s: %p fails the try operation",
 596                             __func__, lk);
 597                         error = EBUSY;
 598                         break;
 599                 }
 600
 601                 /*
 602                  * Acquire the sleepqueue chain lock because we
 603                  * probabilly will need to manipulate waiters flags.
 604                  */
 605                 sleepq_lock(&lk->lock_object);
 606                 x = lk->lk_lock;
 607 retry_sleepq:
 608
 609                 /*
 610                  * if the lock can be acquired in shared mode, try
 611                  * again.
 612                  */
 613                 if (LK_CAN_SHARE(x, flags, false)) {
 614                         sleepq_release(&lk->lock_object);
 615                         continue;
 616                 }
 617
 618                 /*
 619                  * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 620                  * loop back and retry.
 621                  */
 622                 if ((x & LK_SHARED_WAITERS) == 0) {
 623                         if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
 624                             x | LK_SHARED_WAITERS)) {
 625                                 goto retry_sleepq;
 626                         }
 627                         LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 628                             __func__, lk);
 629                 }
 630
 631                 if (lwa == NULL) {
 632                         iwmesg = lk->lock_object.lo_name;
 633                         ipri = lk->lk_pri;
 634                         itimo = lk->lk_timo;
 635                 } else {
 636                         iwmesg = lwa->iwmesg;
 637                         ipri = lwa->ipri;
 638                         itimo = lwa->itimo;
 639                 }
 640
 641                 /*
 642                  * As far as we have been unable to acquire the
 643                  * shared lock and the shared waiters flag is set,
 644                  * we will sleep.
 645                  */
 646                 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 647                     SQ_SHARED_QUEUE);
 648                 flags &= ~LK_INTERLOCK;
 649                 if (error) {
 650                         LOCK_LOG3(lk,
 651                             "%s: interrupted sleep for %p with %d",
 652                             __func__, lk, error);
 653                         break;
 654                 }
 655                 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 656                     __func__, lk);
 657         }
 658         if (error == 0) {
 659 #ifdef LOCK_PROFILING
 660                 lockmgr_note_shared_acquire(lk, contested, waittime,
 661                     file, line, flags);
 662 #else
 663                 lockmgr_note_shared_acquire(lk, 0, 0, file, line,
 664                     flags);
 665 #endif
 666         }
 667
 668 out:
 669         lockmgr_exit(flags, ilk, 0);
 670         return (error);
 671 }
 672
 673 static __noinline int
 674 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 675     const char *file, int line, struct lockmgr_wait *lwa)
 676 {
 677         struct lock_class *class;
 678         uintptr_t tid, x, v;
 679         int error = 0;
 680         const char *iwmesg;
 681         int ipri, itimo;
 682
 683 #ifdef LOCK_PROFILING
 684         uint64_t waittime = 0;
 685         int contested = 0;
 686 #endif
 687
 688         if (__predict_false(panicstr != NULL))
 689                 goto out;
 690
 691         tid = (uintptr_t)curthread;
 692
 693         if (LK_CAN_WITNESS(flags))
 694                 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 695                     LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 696                     ilk : NULL);
 697
 698         /*
 699          * If curthread already holds the lock and this one is
 700          * allowed to recurse, simply recurse on it.
 701          */
 702         if (lockmgr_xlocked(lk)) {
 703                 if ((flags & LK_CANRECURSE) == 0 &&
 704                     (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 705                         /*
 706                          * If the lock is expected to not panic just
 707                          * give up and return.
 708                          */
 709                         if (LK_TRYOP(flags)) {
 710                                 LOCK_LOG2(lk,
 711                                     "%s: %p fails the try operation",
 712                                     __func__, lk);
 713                                 error = EBUSY;
 714                                 goto out;
 715                         }
 716                         if (flags & LK_INTERLOCK) {
 717                                 class = LOCK_CLASS(ilk);
 718                                 class->lc_unlock(ilk);
 719                         }
 720                         panic("%s: recursing on non recursive lockmgr %p "
 721                             "@ %s:%d\n", __func__, lk, file, line);
 722                 }
 723                 lk->lk_recurse++;
 724                 LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 725                 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 726                     lk->lk_recurse, file, line);
 727                 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 728                     LK_TRYWIT(flags), file, line);
 729                 TD_LOCKS_INC(curthread);
 730                 goto out;
 731         }
 732
 733         for (;;) {
 734                 if (lk->lk_lock == LK_UNLOCKED &&
 735                     atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
 736                         break;
 737 #ifdef HWPMC_HOOKS
 738                 PMC_SOFT_CALL( , , lock, failed);
 739 #endif
 740                 lock_profile_obtain_lock_failed(&lk->lock_object,
 741                     &contested, &waittime);
 742
 743                 /*
 744                  * If the lock is expected to not sleep just give up
 745                  * and return.
 746                  */
 747                 if (LK_TRYOP(flags)) {
 748                         LOCK_LOG2(lk, "%s: %p fails the try operation",
 749                             __func__, lk);
 750                         error = EBUSY;
 751                         break;
 752                 }
 753
 754                 /*
 755                  * Acquire the sleepqueue chain lock because we
 756                  * probabilly will need to manipulate waiters flags.
 757                  */
 758                 sleepq_lock(&lk->lock_object);
 759                 x = lk->lk_lock;
 760 retry_sleepq:
 761
 762                 /*
 763                  * if the lock has been released while we spun on
 764                  * the sleepqueue chain lock just try again.
 765                  */
 766                 if (x == LK_UNLOCKED) {
 767                         sleepq_release(&lk->lock_object);
 768                         continue;
 769                 }
 770
 771                 /*
 772                  * The lock can be in the state where there is a
 773                  * pending queue of waiters, but still no owner.
 774                  * This happens when the lock is contested and an
 775                  * owner is going to claim the lock.
 776                  * If curthread is the one successfully acquiring it
 777                  * claim lock ownership and return, preserving waiters
 778                  * flags.
 779                  */
 780                 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 781                 if ((x & ~v) == LK_UNLOCKED) {
 782                         v &= ~LK_EXCLUSIVE_SPINNERS;
 783                         if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
 784                             tid | v)) {
 785                                 sleepq_release(&lk->lock_object);
 786                                 LOCK_LOG2(lk,
 787                                     "%s: %p claimed by a new writer",
 788                                     __func__, lk);
 789                                 break;
 790                         }
 791                         goto retry_sleepq;
 792                 }
 793
 794                 /*
 795                  * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 796                  * fail, loop back and retry.
 797                  */
 798                 if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 799                         if (!atomic_fcmpset_ptr(&lk->lk_lock, &x,
 800                             x | LK_EXCLUSIVE_WAITERS)) {
 801                                 goto retry_sleepq;
 802                         }
 803                         LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 804                             __func__, lk);
 805                 }
 806
 807                 if (lwa == NULL) {
 808                         iwmesg = lk->lock_object.lo_name;
 809                         ipri = lk->lk_pri;
 810                         itimo = lk->lk_timo;
 811                 } else {
 812                         iwmesg = lwa->iwmesg;
 813                         ipri = lwa->ipri;
 814                         itimo = lwa->itimo;
 815                 }
 816
 817                 /*
 818                  * As far as we have been unable to acquire the
 819                  * exclusive lock and the exclusive waiters flag
 820                  * is set, we will sleep.
 821                  */
 822                 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 823                     SQ_EXCLUSIVE_QUEUE);
 824                 flags &= ~LK_INTERLOCK;
 825                 if (error) {
 826                         LOCK_LOG3(lk,
 827                             "%s: interrupted sleep for %p with %d",
 828                             __func__, lk, error);
 829                         break;
 830                 }
 831                 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 832                     __func__, lk);
 833         }
 834         if (error == 0) {
 835 #ifdef LOCK_PROFILING
 836                 lockmgr_note_exclusive_acquire(lk, contested, waittime,
 837                     file, line, flags);
 838 #else
 839                 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
 840                     flags);
 841 #endif
 842         }
 843
 844 out:
 845         lockmgr_exit(flags, ilk, 0);
 846         return (error);
 847 }
 848
 849 static __noinline int
 850 lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk,
 851     const char *file, int line, struct lockmgr_wait *lwa)
 852 {
 853         uintptr_t tid, x, v;
 854         int error = 0;
 855         int wakeup_swapper = 0;
 856         int op;
 857
 858         if (__predict_false(panicstr != NULL))
 859                 goto out;
 860
 861         tid = (uintptr_t)curthread;
 862
 863         _lockmgr_assert(lk, KA_SLOCKED, file, line);
 864         v = lk->lk_lock;
 865         x = v & LK_ALL_WAITERS;
 866         v &= LK_EXCLUSIVE_SPINNERS;
 867
 868         /*
 869          * Try to switch from one shared lock to an exclusive one.
 870          * We need to preserve waiters flags during the operation.
 871          */
 872         if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 873             tid | x)) {
 874                 LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 875                     line);
 876                 WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 877                     LK_TRYWIT(flags), file, line);
 878                 TD_SLOCKS_DEC(curthread);
 879                 goto out;
 880         }
 881
 882         op = flags & LK_TYPE_MASK;
 883
 884         /*
 885          * In LK_TRYUPGRADE mode, do not drop the lock,
 886          * returning EBUSY instead.
 887          */
 888         if (op == LK_TRYUPGRADE) {
 889                 LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
 890                     __func__, lk);
 891                 error = EBUSY;
 892                 goto out;
 893         }
 894
 895         /*
 896          * We have been unable to succeed in upgrading, so just
 897          * give up the shared lock.
 898          */
 899         wakeup_swapper |= wakeupshlk(lk, file, line);
 900         error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa);
 901         flags &= ~LK_INTERLOCK;
 902 out:
 903         lockmgr_exit(flags, ilk, wakeup_swapper);
 904         return (error);
 905 }
 906
 907 int
 908 lockmgr_lock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk,
 909     const char *file, int line)
 910 {
 911         struct lock_class *class;
 912         uintptr_t x, tid;
 913         u_int op;
 914         bool locked;
 915
 916         if (__predict_false(panicstr != NULL))
 917                 return (0);
 918
 919         op = flags & LK_TYPE_MASK;
 920         locked = false;
 921         switch (op) {
 922         case LK_SHARED:
 923                 if (LK_CAN_WITNESS(flags))
 924                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 925                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 926                 if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
 927                         break;
 928                 if (lockmgr_slock_try(lk, &x, flags, true)) {
 929                         lockmgr_note_shared_acquire(lk, 0, 0,
 930                             file, line, flags);
 931                         locked = true;
 932                 } else {
 933                         return (lockmgr_slock_hard(lk, flags, ilk, file, line,
 934                             NULL));
 935                 }
 936                 break;
 937         case LK_EXCLUSIVE:
 938                 if (LK_CAN_WITNESS(flags))
 939                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 940                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 941                             ilk : NULL);
 942                 tid = (uintptr_t)curthread;
 943                 if (lk->lk_lock == LK_UNLOCKED &&
 944                     atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
 945                         lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
 946                             flags);
 947                         locked = true;
 948                 } else {
 949                         return (lockmgr_xlock_hard(lk, flags, ilk, file, line,
 950                             NULL));
 951                 }
 952                 break;
 953         case LK_UPGRADE:
 954         case LK_TRYUPGRADE:
 955                 return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL));
 956         default:
 957                 break;
 958         }
 959         if (__predict_true(locked)) {
 960                 if (__predict_false(flags & LK_INTERLOCK)) {
 961                         class = LOCK_CLASS(ilk);
 962                         class->lc_unlock(ilk);
 963                 }
 964                 return (0);
 965         } else {
 966                 return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT,
 967                     LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line));
 968         }
 969 }
 970
 971 static __noinline int
 972 lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
 973     const char *file, int line)
 974
 975 {
 976         int wakeup_swapper = 0;
 977
 978         if (__predict_false(panicstr != NULL))
 979                 goto out;
 980
 981         wakeup_swapper = wakeupshlk(lk, file, line);
 982
 983 out:
 984         lockmgr_exit(flags, ilk, wakeup_swapper);
 985         return (0);
 986 }
 987
 988 static __noinline int
 989 lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
 990     const char *file, int line)
 991 {
 992         uintptr_t tid, v;
 993         int wakeup_swapper = 0;
 994         u_int realexslp;
 995         int queue;
 996
 997         if (__predict_false(panicstr != NULL))
 998                 goto out;
 999
1000         tid = (uintptr_t)curthread;
1001
1002         /*
1003          * As first option, treact the lock as if it has not
1004          * any waiter.
1005          * Fix-up the tid var if the lock has been disowned.
1006          */
1007         if (LK_HOLDER(x) == LK_KERNPROC)
1008                 tid = LK_KERNPROC;
1009         else {
1010                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1011                 TD_LOCKS_DEC(curthread);
1012         }
1013         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
1014
1015         /*
1016          * The lock is held in exclusive mode.
1017          * If the lock is recursed also, then unrecurse it.
1018          */
1019         if (lockmgr_xlocked_v(x) && lockmgr_recursed(lk)) {
1020                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk);
1021                 lk->lk_recurse--;
1022                 goto out;
1023         }
1024         if (tid != LK_KERNPROC)
1025                 lock_profile_release_lock(&lk->lock_object);
1026
1027         if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED))
1028                 goto out;
1029
1030         sleepq_lock(&lk->lock_object);
1031         x = lk->lk_lock;
1032         v = LK_UNLOCKED;
1033
1034         /*
1035          * If the lock has exclusive waiters, give them
1036          * preference in order to avoid deadlock with
1037          * shared runners up.
1038          * If interruptible sleeps left the exclusive queue
1039          * empty avoid a starvation for the threads sleeping
1040          * on the shared queue by giving them precedence
1041          * and cleaning up the exclusive waiters bit anyway.
1042          * Please note that lk_exslpfail count may be lying
1043          * about the real number of waiters with the
1044          * LK_SLEEPFAIL flag on because they may be used in
1045          * conjunction with interruptible sleeps so
1046          * lk_exslpfail might be considered an 'upper limit'
1047          * bound, including the edge cases.
1048          */
1049         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1050         realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE);
1051         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1052                 if (lk->lk_exslpfail < realexslp) {
1053                         lk->lk_exslpfail = 0;
1054                         queue = SQ_EXCLUSIVE_QUEUE;
1055                         v |= (x & LK_SHARED_WAITERS);
1056                 } else {
1057                         lk->lk_exslpfail = 0;
1058                         LOCK_LOG2(lk,
1059                             "%s: %p has only LK_SLEEPFAIL sleepers",
1060                             __func__, lk);
1061                         LOCK_LOG2(lk,
1062                             "%s: %p waking up threads on the exclusive queue",
1063                             __func__, lk);
1064                         wakeup_swapper = sleepq_broadcast(&lk->lock_object,
1065                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1066                         queue = SQ_SHARED_QUEUE;
1067                 }
1068         } else {
1069
1070                 /*
1071                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1072                  * on and using interruptible sleeps/timeout
1073                  * may have left spourious lk_exslpfail counts
1074                  * on, so clean it up anyway.
1075                  */
1076                 lk->lk_exslpfail = 0;
1077                 queue = SQ_SHARED_QUEUE;
1078         }
1079
1080         LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
1081             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1082             "exclusive");
1083         atomic_store_rel_ptr(&lk->lk_lock, v);
1084         wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue);
1085         sleepq_release(&lk->lock_object);
1086
1087 out:
1088         lockmgr_exit(flags, ilk, wakeup_swapper);
1089         return (0);
1090 }
1091
1092 int
1093 lockmgr_unlock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk)
1094 {
1095         struct lock_class *class;
1096         uintptr_t x, tid;
1097         const char *file;
1098         int line;
1099
1100         if (__predict_false(panicstr != NULL))
1101                 return (0);
1102
1103         file = __FILE__;
1104         line = __LINE__;
1105
1106         _lockmgr_assert(lk, KA_LOCKED, file, line);
1107         x = lk->lk_lock;
1108         if (__predict_true(x & LK_SHARE) != 0) {
1109                 if (lockmgr_sunlock_try(lk, &x)) {
1110                         lockmgr_note_shared_release(lk, file, line);
1111                 } else {
1112                         return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1113                 }
1114         } else {
1115                 tid = (uintptr_t)curthread;
1116                 if (!lockmgr_recursed(lk) &&
1117                     atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) {
1118                         lockmgr_note_exclusive_release(lk, file, line);
1119                 } else {
1120                         return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1121                 }
1122         }
1123         if (__predict_false(flags & LK_INTERLOCK)) {
1124                 class = LOCK_CLASS(ilk);
1125                 class->lc_unlock(ilk);
1126         }
1127         return (0);
1128 }
1129
1130 int
1131 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
1132     const char *wmesg, int pri, int timo, const char *file, int line)
1133 {
1134         GIANT_DECLARE;
1135         struct lockmgr_wait lwa;
1136         struct lock_class *class;
1137         const char *iwmesg;
1138         uintptr_t tid, v, x;
1139         u_int op, realexslp;
1140         int error, ipri, itimo, queue, wakeup_swapper;
1141 #ifdef LOCK_PROFILING
1142         uint64_t waittime = 0;
1143         int contested = 0;
1144 #endif
1145
1146         if (panicstr != NULL)
1147                 return (0);
1148
1149         error = 0;
1150         tid = (uintptr_t)curthread;
1151         op = (flags & LK_TYPE_MASK);
1152         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
1153         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
1154         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
1155
1156         lwa.iwmesg = iwmesg;
1157         lwa.ipri = ipri;
1158         lwa.itimo = itimo;
1159
1160         MPASS((flags & ~LK_TOTAL_MASK) == 0);
1161         KASSERT((op & (op - 1)) == 0,
1162             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
1163         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
1164             (op != LK_DOWNGRADE && op != LK_RELEASE),
1165             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
1166             __func__, file, line));
1167         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
1168             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
1169             __func__, file, line));
1170         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
1171             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
1172             lk->lock_object.lo_name, file, line));
1173
1174         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
1175
1176         if (lk->lock_object.lo_flags & LK_NOSHARE) {
1177                 switch (op) {
1178                 case LK_SHARED:
1179                         op = LK_EXCLUSIVE;
1180                         break;
1181                 case LK_UPGRADE:
1182                 case LK_TRYUPGRADE:
1183                 case LK_DOWNGRADE:
1184                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
1185                             file, line);
1186                         if (flags & LK_INTERLOCK)
1187                                 class->lc_unlock(ilk);
1188                         return (0);
1189                 }
1190         }
1191
1192         wakeup_swapper = 0;
1193         switch (op) {
1194         case LK_SHARED:
1195                 return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa));
1196                 break;
1197         case LK_UPGRADE:
1198         case LK_TRYUPGRADE:
1199                 return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa));
1200                 break;
1201         case LK_EXCLUSIVE:
1202                 return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa));
1203                 break;
1204         case LK_DOWNGRADE:
1205                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
1206                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
1207                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
1208
1209                 /*
1210                  * Panic if the lock is recursed.
1211                  */
1212                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1213                         if (flags & LK_INTERLOCK)
1214                                 class->lc_unlock(ilk);
1215                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
1216                             __func__, iwmesg, file, line);
1217                 }
1218                 TD_SLOCKS_INC(curthread);
1219
1220                 /*
1221                  * In order to preserve waiters flags, just spin.
1222                  */
1223                 for (;;) {
1224                         x = lk->lk_lock;
1225                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1226                         x &= LK_ALL_WAITERS;
1227                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1228                             LK_SHARERS_LOCK(1) | x))
1229                                 break;
1230                         cpu_spinwait();
1231                 }
1232                 break;
1233         case LK_RELEASE:
1234                 _lockmgr_assert(lk, KA_LOCKED, file, line);
1235                 x = lk->lk_lock;
1236
1237                 if (__predict_true(x & LK_SHARE) != 0) {
1238                         return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1239                 } else {
1240                         return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1241                 }
1242                 break;
1243         case LK_DRAIN:
1244                 if (LK_CAN_WITNESS(flags))
1245                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1246                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1247                             ilk : NULL);
1248
1249                 /*
1250                  * Trying to drain a lock we already own will result in a
1251                  * deadlock.
1252                  */
1253                 if (lockmgr_xlocked(lk)) {
1254                         if (flags & LK_INTERLOCK)
1255                                 class->lc_unlock(ilk);
1256                         panic("%s: draining %s with the lock held @ %s:%d\n",
1257                             __func__, iwmesg, file, line);
1258                 }
1259
1260                 for (;;) {
1261                         if (lk->lk_lock == LK_UNLOCKED &&
1262                             atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1263                                 break;
1264
1265 #ifdef HWPMC_HOOKS
1266                         PMC_SOFT_CALL( , , lock, failed);
1267 #endif
1268                         lock_profile_obtain_lock_failed(&lk->lock_object,
1269                             &contested, &waittime);
1270
1271                         /*
1272                          * If the lock is expected to not sleep just give up
1273                          * and return.
1274                          */
1275                         if (LK_TRYOP(flags)) {
1276                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1277                                     __func__, lk);
1278                                 error = EBUSY;
1279                                 break;
1280                         }
1281
1282                         /*
1283                          * Acquire the sleepqueue chain lock because we
1284                          * probabilly will need to manipulate waiters flags.
1285                          */
1286                         sleepq_lock(&lk->lock_object);
1287                         x = lk->lk_lock;
1288
1289                         /*
1290                          * if the lock has been released while we spun on
1291                          * the sleepqueue chain lock just try again.
1292                          */
1293                         if (x == LK_UNLOCKED) {
1294                                 sleepq_release(&lk->lock_object);
1295                                 continue;
1296                         }
1297
1298                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1299                         if ((x & ~v) == LK_UNLOCKED) {
1300                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1301
1302                                 /*
1303                                  * If interruptible sleeps left the exclusive
1304                                  * queue empty avoid a starvation for the
1305                                  * threads sleeping on the shared queue by
1306                                  * giving them precedence and cleaning up the
1307                                  * exclusive waiters bit anyway.
1308                                  * Please note that lk_exslpfail count may be
1309                                  * lying about the real number of waiters with
1310                                  * the LK_SLEEPFAIL flag on because they may
1311                                  * be used in conjunction with interruptible
1312                                  * sleeps so lk_exslpfail might be considered
1313                                  * an 'upper limit' bound, including the edge
1314                                  * cases.
1315                                  */
1316                                 if (v & LK_EXCLUSIVE_WAITERS) {
1317                                         queue = SQ_EXCLUSIVE_QUEUE;
1318                                         v &= ~LK_EXCLUSIVE_WAITERS;
1319                                 } else {
1320
1321                                         /*
1322                                          * Exclusive waiters sleeping with
1323                                          * LK_SLEEPFAIL on and using
1324                                          * interruptible sleeps/timeout may
1325                                          * have left spourious lk_exslpfail
1326                                          * counts on, so clean it up anyway.
1327                                          */
1328                                         MPASS(v & LK_SHARED_WAITERS);
1329                                         lk->lk_exslpfail = 0;
1330                                         queue = SQ_SHARED_QUEUE;
1331                                         v &= ~LK_SHARED_WAITERS;
1332                                 }
1333                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1334                                         realexslp =
1335                                             sleepq_sleepcnt(&lk->lock_object,
1336                                             SQ_EXCLUSIVE_QUEUE);
1337                                         if (lk->lk_exslpfail >= realexslp) {
1338                                                 lk->lk_exslpfail = 0;
1339                                                 queue = SQ_SHARED_QUEUE;
1340                                                 v &= ~LK_SHARED_WAITERS;
1341                                                 if (realexslp != 0) {
1342                                                         LOCK_LOG2(lk,
1343                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1344                                                             __func__, lk);
1345                                                         LOCK_LOG2(lk,
1346                         "%s: %p waking up threads on the exclusive queue",
1347                                                             __func__, lk);
1348                                                         wakeup_swapper =
1349                                                             sleepq_broadcast(
1350                                                             &lk->lock_object,
1351                                                             SLEEPQ_LK, 0,
1352                                                             SQ_EXCLUSIVE_QUEUE);
1353                                                 }
1354                                         } else
1355                                                 lk->lk_exslpfail = 0;
1356                                 }
1357                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1358                                         sleepq_release(&lk->lock_object);
1359                                         continue;
1360                                 }
1361                                 LOCK_LOG3(lk,
1362                                 "%s: %p waking up all threads on the %s queue",
1363                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1364                                     "shared" : "exclusive");
1365                                 wakeup_swapper |= sleepq_broadcast(
1366                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1367
1368                                 /*
1369                                  * If shared waiters have been woken up we need
1370                                  * to wait for one of them to acquire the lock
1371                                  * before to set the exclusive waiters in
1372                                  * order to avoid a deadlock.
1373                                  */
1374                                 if (queue == SQ_SHARED_QUEUE) {
1375                                         for (v = lk->lk_lock;
1376                                             (v & LK_SHARE) && !LK_SHARERS(v);
1377                                             v = lk->lk_lock)
1378                                                 cpu_spinwait();
1379                                 }
1380                         }
1381
1382                         /*
1383                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1384                          * fail, loop back and retry.
1385                          */
1386                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1387                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1388                                     x | LK_EXCLUSIVE_WAITERS)) {
1389                                         sleepq_release(&lk->lock_object);
1390                                         continue;
1391                                 }
1392                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1393                                     __func__, lk);
1394                         }
1395
1396                         /*
1397                          * As far as we have been unable to acquire the
1398                          * exclusive lock and the exclusive waiters flag
1399                          * is set, we will sleep.
1400                          */
1401                         if (flags & LK_INTERLOCK) {
1402                                 class->lc_unlock(ilk);
1403                                 flags &= ~LK_INTERLOCK;
1404                         }
1405                         GIANT_SAVE();
1406                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1407                             SQ_EXCLUSIVE_QUEUE);
1408                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1409                         GIANT_RESTORE();
1410                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1411                             __func__, lk);
1412                 }
1413
1414                 if (error == 0) {
1415                         lock_profile_obtain_lock_success(&lk->lock_object,
1416                             contested, waittime, file, line);
1417                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1418                             lk->lk_recurse, file, line);
1419                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1420                             LK_TRYWIT(flags), file, line);
1421                         TD_LOCKS_INC(curthread);
1422                         STACK_SAVE(lk);
1423                 }
1424                 break;
1425         default:
1426                 if (flags & LK_INTERLOCK)
1427                         class->lc_unlock(ilk);
1428                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1429         }
1430
1431         if (flags & LK_INTERLOCK)
1432                 class->lc_unlock(ilk);
1433         if (wakeup_swapper)
1434                 kick_proc0();
1435
1436         return (error);
1437 }
1438
1439 void
1440 _lockmgr_disown(struct lock *lk, const char *file, int line)
1441 {
1442         uintptr_t tid, x;
1443
1444         if (SCHEDULER_STOPPED())
1445                 return;
1446
1447         tid = (uintptr_t)curthread;
1448         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1449
1450         /*
1451          * Panic if the lock is recursed.
1452          */
1453         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1454                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1455                     __func__,  file, line);
1456
1457         /*
1458          * If the owner is already LK_KERNPROC just skip the whole operation.
1459          */
1460         if (LK_HOLDER(lk->lk_lock) != tid)
1461                 return;
1462         lock_profile_release_lock(&lk->lock_object);
1463         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1464         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1465         TD_LOCKS_DEC(curthread);
1466         STACK_SAVE(lk);
1467
1468         /*
1469          * In order to preserve waiters flags, just spin.
1470          */
1471         for (;;) {
1472                 x = lk->lk_lock;
1473                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1474                 x &= LK_ALL_WAITERS;
1475                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1476                     LK_KERNPROC | x))
1477                         return;
1478                 cpu_spinwait();
1479         }
1480 }
1481
1482 void
1483 lockmgr_printinfo(const struct lock *lk)
1484 {
1485         struct thread *td;
1486         uintptr_t x;
1487
1488         if (lk->lk_lock == LK_UNLOCKED)
1489                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1490         else if (lk->lk_lock & LK_SHARE)
1491                 printf("lock type %s: SHARED (count %ju)\n",
1492                     lk->lock_object.lo_name,
1493                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1494         else {
1495                 td = lockmgr_xholder(lk);
1496                 if (td == (struct thread *)LK_KERNPROC)
1497                         printf("lock type %s: EXCL by KERNPROC\n",
1498                             lk->lock_object.lo_name);
1499                 else
1500                         printf("lock type %s: EXCL by thread %p "
1501                             "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1502                             td, td->td_proc->p_pid, td->td_proc->p_comm,
1503                             td->td_tid);
1504         }
1505
1506         x = lk->lk_lock;
1507         if (x & LK_EXCLUSIVE_WAITERS)
1508                 printf(" with exclusive waiters pending\n");
1509         if (x & LK_SHARED_WAITERS)
1510                 printf(" with shared waiters pending\n");
1511         if (x & LK_EXCLUSIVE_SPINNERS)
1512                 printf(" with exclusive spinners pending\n");
1513
1514         STACK_PRINT(lk);
1515 }
1516
1517 int
1518 lockstatus(const struct lock *lk)
1519 {
1520         uintptr_t v, x;
1521         int ret;
1522
1523         ret = LK_SHARED;
1524         x = lk->lk_lock;
1525         v = LK_HOLDER(x);
1526
1527         if ((x & LK_SHARE) == 0) {
1528                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1529                         ret = LK_EXCLUSIVE;
1530                 else
1531                         ret = LK_EXCLOTHER;
1532         } else if (x == LK_UNLOCKED)
1533                 ret = 0;
1534
1535         return (ret);
1536 }
1537
1538 #ifdef INVARIANT_SUPPORT
1539
1540 FEATURE(invariant_support,
1541     "Support for modules compiled with INVARIANTS option");
1542
1543 #ifndef INVARIANTS
1544 #undef  _lockmgr_assert
1545 #endif
1546
1547 void
1548 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1549 {
1550         int slocked = 0;
1551
1552         if (panicstr != NULL)
1553                 return;
1554         switch (what) {
1555         case KA_SLOCKED:
1556         case KA_SLOCKED | KA_NOTRECURSED:
1557         case KA_SLOCKED | KA_RECURSED:
1558                 slocked = 1;
1559         case KA_LOCKED:
1560         case KA_LOCKED | KA_NOTRECURSED:
1561         case KA_LOCKED | KA_RECURSED:
1562 #ifdef WITNESS
1563
1564                 /*
1565                  * We cannot trust WITNESS if the lock is held in exclusive
1566                  * mode and a call to lockmgr_disown() happened.
1567                  * Workaround this skipping the check if the lock is held in
1568                  * exclusive mode even for the KA_LOCKED case.
1569                  */
1570                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1571                         witness_assert(&lk->lock_object, what, file, line);
1572                         break;
1573                 }
1574 #endif
1575                 if (lk->lk_lock == LK_UNLOCKED ||
1576                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1577                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1578                         panic("Lock %s not %slocked @ %s:%d\n",
1579                             lk->lock_object.lo_name, slocked ? "share" : "",
1580                             file, line);
1581
1582                 if ((lk->lk_lock & LK_SHARE) == 0) {
1583                         if (lockmgr_recursed(lk)) {
1584                                 if (what & KA_NOTRECURSED)
1585                                         panic("Lock %s recursed @ %s:%d\n",
1586                                             lk->lock_object.lo_name, file,
1587                                             line);
1588                         } else if (what & KA_RECURSED)
1589                                 panic("Lock %s not recursed @ %s:%d\n",
1590                                     lk->lock_object.lo_name, file, line);
1591                 }
1592                 break;
1593         case KA_XLOCKED:
1594         case KA_XLOCKED | KA_NOTRECURSED:
1595         case KA_XLOCKED | KA_RECURSED:
1596                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1597                         panic("Lock %s not exclusively locked @ %s:%d\n",
1598                             lk->lock_object.lo_name, file, line);
1599                 if (lockmgr_recursed(lk)) {
1600                         if (what & KA_NOTRECURSED)
1601                                 panic("Lock %s recursed @ %s:%d\n",
1602                                     lk->lock_object.lo_name, file, line);
1603                 } else if (what & KA_RECURSED)
1604                         panic("Lock %s not recursed @ %s:%d\n",
1605                             lk->lock_object.lo_name, file, line);
1606                 break;
1607         case KA_UNLOCKED:
1608                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1609                         panic("Lock %s exclusively locked @ %s:%d\n",
1610                             lk->lock_object.lo_name, file, line);
1611                 break;
1612         default:
1613                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1614                     line);
1615         }
1616 }
1617 #endif
1618
1619 #ifdef DDB
1620 int
1621 lockmgr_chain(struct thread *td, struct thread **ownerp)
1622 {
1623         struct lock *lk;
1624
1625         lk = td->td_wchan;
1626
1627         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1628                 return (0);
1629         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1630         if (lk->lk_lock & LK_SHARE)
1631                 db_printf("SHARED (count %ju)\n",
1632                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1633         else
1634                 db_printf("EXCL\n");
1635         *ownerp = lockmgr_xholder(lk);
1636
1637         return (1);
1638 }
1639
1640 static void
1641 db_show_lockmgr(const struct lock_object *lock)
1642 {
1643         struct thread *td;
1644         const struct lock *lk;
1645
1646         lk = (const struct lock *)lock;
1647
1648         db_printf(" state: ");
1649         if (lk->lk_lock == LK_UNLOCKED)
1650                 db_printf("UNLOCKED\n");
1651         else if (lk->lk_lock & LK_SHARE)
1652                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1653         else {
1654                 td = lockmgr_xholder(lk);
1655                 if (td == (struct thread *)LK_KERNPROC)
1656                         db_printf("XLOCK: LK_KERNPROC\n");
1657                 else
1658                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1659                             td->td_tid, td->td_proc->p_pid,
1660                             td->td_proc->p_comm);
1661                 if (lockmgr_recursed(lk))
1662                         db_printf(" recursed: %d\n", lk->lk_recurse);
1663         }
1664         db_printf(" waiters: ");
1665         switch (lk->lk_lock & LK_ALL_WAITERS) {
1666         case LK_SHARED_WAITERS:
1667                 db_printf("shared\n");
1668                 break;
1669         case LK_EXCLUSIVE_WAITERS:
1670                 db_printf("exclusive\n");
1671                 break;
1672         case LK_ALL_WAITERS:
1673                 db_printf("shared and exclusive\n");
1674                 break;
1675         default:
1676                 db_printf("none\n");
1677         }
1678         db_printf(" spinners: ");
1679         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1680                 db_printf("exclusive\n");
1681         else
1682                 db_printf("none\n");
1683 }
1684 #endif