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 (__predict_false(_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 (__predict_false(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         if (flags & LK_NEW)
 454                 iflags |= LO_NEW;
 455         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 456
 457         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 458         lk->lk_lock = LK_UNLOCKED;
 459         lk->lk_recurse = 0;
 460         lk->lk_exslpfail = 0;
 461         lk->lk_timo = timo;
 462         lk->lk_pri = pri;
 463         STACK_ZERO(lk);
 464 }
 465
 466 /*
 467  * XXX: Gross hacks to manipulate external lock flags after
 468  * initialization.  Used for certain vnode and buf locks.
 469  */
 470 void
 471 lockallowshare(struct lock *lk)
 472 {
 473
 474         lockmgr_assert(lk, KA_XLOCKED);
 475         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 476 }
 477
 478 void
 479 lockdisableshare(struct lock *lk)
 480 {
 481
 482         lockmgr_assert(lk, KA_XLOCKED);
 483         lk->lock_object.lo_flags |= LK_NOSHARE;
 484 }
 485
 486 void
 487 lockallowrecurse(struct lock *lk)
 488 {
 489
 490         lockmgr_assert(lk, KA_XLOCKED);
 491         lk->lock_object.lo_flags |= LO_RECURSABLE;
 492 }
 493
 494 void
 495 lockdisablerecurse(struct lock *lk)
 496 {
 497
 498         lockmgr_assert(lk, KA_XLOCKED);
 499         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 500 }
 501
 502 void
 503 lockdestroy(struct lock *lk)
 504 {
 505
 506         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 507         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 508         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 509         lock_destroy(&lk->lock_object);
 510 }
 511
 512 static bool __always_inline
 513 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp)
 514 {
 515
 516         /*
 517          * If no other thread has an exclusive lock, or
 518          * no exclusive waiter is present, bump the count of
 519          * sharers.  Since we have to preserve the state of
 520          * waiters, if we fail to acquire the shared lock
 521          * loop back and retry.
 522          */
 523         *xp = lk->lk_lock;
 524         while (LK_CAN_SHARE(*xp, flags, fp)) {
 525                 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
 526                     *xp + LK_ONE_SHARER)) {
 527                         return (true);
 528                 }
 529         }
 530         return (false);
 531 }
 532
 533 static bool __always_inline
 534 lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp)
 535 {
 536
 537         for (;;) {
 538                 if (LK_SHARERS(*xp) > 1 || !(*xp & LK_ALL_WAITERS)) {
 539                         if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp,
 540                             *xp - LK_ONE_SHARER))
 541                                 return (true);
 542                         continue;
 543                 }
 544                 break;
 545         }
 546         return (false);
 547 }
 548
 549 static __noinline int
 550 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 551     const char *file, int line, struct lockmgr_wait *lwa)
 552 {
 553         uintptr_t tid, x;
 554         int error = 0;
 555         const char *iwmesg;
 556         int ipri, itimo;
 557
 558 #ifdef LOCK_PROFILING
 559         uint64_t waittime = 0;
 560         int contested = 0;
 561 #endif
 562
 563         if (__predict_false(panicstr != NULL))
 564                 goto out;
 565
 566         tid = (uintptr_t)curthread;
 567
 568         if (LK_CAN_WITNESS(flags))
 569                 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 570                     file, line, flags & LK_INTERLOCK ? ilk : NULL);
 571         for (;;) {
 572                 if (lockmgr_slock_try(lk, &x, flags, false))
 573                         break;
 574 #ifdef HWPMC_HOOKS
 575                 PMC_SOFT_CALL( , , lock, failed);
 576 #endif
 577                 lock_profile_obtain_lock_failed(&lk->lock_object,
 578                     &contested, &waittime);
 579
 580                 /*
 581                  * If the lock is already held by curthread in
 582                  * exclusive way avoid a deadlock.
 583                  */
 584                 if (LK_HOLDER(x) == tid) {
 585                         LOCK_LOG2(lk,
 586                             "%s: %p already held in exclusive mode",
 587                             __func__, lk);
 588                         error = EDEADLK;
 589                         break;
 590                 }
 591
 592                 /*
 593                  * If the lock is expected to not sleep just give up
 594                  * and return.
 595                  */
 596                 if (LK_TRYOP(flags)) {
 597                         LOCK_LOG2(lk, "%s: %p fails the try operation",
 598                             __func__, lk);
 599                         error = EBUSY;
 600                         break;
 601                 }
 602
 603                 /*
 604                  * Acquire the sleepqueue chain lock because we
 605                  * probabilly will need to manipulate waiters flags.
 606                  */
 607                 sleepq_lock(&lk->lock_object);
 608                 x = lk->lk_lock;
 609 retry_sleepq:
 610
 611                 /*
 612                  * if the lock can be acquired in shared mode, try
 613                  * again.
 614                  */
 615                 if (LK_CAN_SHARE(x, flags, false)) {
 616                         sleepq_release(&lk->lock_object);
 617                         continue;
 618                 }
 619
 620                 /*
 621                  * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 622                  * loop back and retry.
 623                  */
 624                 if ((x & LK_SHARED_WAITERS) == 0) {
 625                         if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
 626                             x | LK_SHARED_WAITERS)) {
 627                                 goto retry_sleepq;
 628                         }
 629                         LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 630                             __func__, lk);
 631                 }
 632
 633                 if (lwa == NULL) {
 634                         iwmesg = lk->lock_object.lo_name;
 635                         ipri = lk->lk_pri;
 636                         itimo = lk->lk_timo;
 637                 } else {
 638                         iwmesg = lwa->iwmesg;
 639                         ipri = lwa->ipri;
 640                         itimo = lwa->itimo;
 641                 }
 642
 643                 /*
 644                  * As far as we have been unable to acquire the
 645                  * shared lock and the shared waiters flag is set,
 646                  * we will sleep.
 647                  */
 648                 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 649                     SQ_SHARED_QUEUE);
 650                 flags &= ~LK_INTERLOCK;
 651                 if (error) {
 652                         LOCK_LOG3(lk,
 653                             "%s: interrupted sleep for %p with %d",
 654                             __func__, lk, error);
 655                         break;
 656                 }
 657                 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 658                     __func__, lk);
 659         }
 660         if (error == 0) {
 661 #ifdef LOCK_PROFILING
 662                 lockmgr_note_shared_acquire(lk, contested, waittime,
 663                     file, line, flags);
 664 #else
 665                 lockmgr_note_shared_acquire(lk, 0, 0, file, line,
 666                     flags);
 667 #endif
 668         }
 669
 670 out:
 671         lockmgr_exit(flags, ilk, 0);
 672         return (error);
 673 }
 674
 675 static __noinline int
 676 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 677     const char *file, int line, struct lockmgr_wait *lwa)
 678 {
 679         struct lock_class *class;
 680         uintptr_t tid, x, v;
 681         int error = 0;
 682         const char *iwmesg;
 683         int ipri, itimo;
 684
 685 #ifdef LOCK_PROFILING
 686         uint64_t waittime = 0;
 687         int contested = 0;
 688 #endif
 689
 690         if (__predict_false(panicstr != NULL))
 691                 goto out;
 692
 693         tid = (uintptr_t)curthread;
 694
 695         if (LK_CAN_WITNESS(flags))
 696                 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 697                     LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 698                     ilk : NULL);
 699
 700         /*
 701          * If curthread already holds the lock and this one is
 702          * allowed to recurse, simply recurse on it.
 703          */
 704         if (lockmgr_xlocked(lk)) {
 705                 if ((flags & LK_CANRECURSE) == 0 &&
 706                     (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 707                         /*
 708                          * If the lock is expected to not panic just
 709                          * give up and return.
 710                          */
 711                         if (LK_TRYOP(flags)) {
 712                                 LOCK_LOG2(lk,
 713                                     "%s: %p fails the try operation",
 714                                     __func__, lk);
 715                                 error = EBUSY;
 716                                 goto out;
 717                         }
 718                         if (flags & LK_INTERLOCK) {
 719                                 class = LOCK_CLASS(ilk);
 720                                 class->lc_unlock(ilk);
 721                         }
 722                         panic("%s: recursing on non recursive lockmgr %p "
 723                             "@ %s:%d\n", __func__, lk, file, line);
 724                 }
 725                 lk->lk_recurse++;
 726                 LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 727                 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 728                     lk->lk_recurse, file, line);
 729                 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 730                     LK_TRYWIT(flags), file, line);
 731                 TD_LOCKS_INC(curthread);
 732                 goto out;
 733         }
 734
 735         for (;;) {
 736                 if (lk->lk_lock == LK_UNLOCKED &&
 737                     atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
 738                         break;
 739 #ifdef HWPMC_HOOKS
 740                 PMC_SOFT_CALL( , , lock, failed);
 741 #endif
 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                 /*
 757                  * Acquire the sleepqueue chain lock because we
 758                  * probabilly will need to manipulate waiters flags.
 759                  */
 760                 sleepq_lock(&lk->lock_object);
 761                 x = lk->lk_lock;
 762 retry_sleepq:
 763
 764                 /*
 765                  * if the lock has been released while we spun on
 766                  * the sleepqueue chain lock just try again.
 767                  */
 768                 if (x == LK_UNLOCKED) {
 769                         sleepq_release(&lk->lock_object);
 770                         continue;
 771                 }
 772
 773                 /*
 774                  * The lock can be in the state where there is a
 775                  * pending queue of waiters, but still no owner.
 776                  * This happens when the lock is contested and an
 777                  * owner is going to claim the lock.
 778                  * If curthread is the one successfully acquiring it
 779                  * claim lock ownership and return, preserving waiters
 780                  * flags.
 781                  */
 782                 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 783                 if ((x & ~v) == LK_UNLOCKED) {
 784                         v &= ~LK_EXCLUSIVE_SPINNERS;
 785                         if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x,
 786                             tid | v)) {
 787                                 sleepq_release(&lk->lock_object);
 788                                 LOCK_LOG2(lk,
 789                                     "%s: %p claimed by a new writer",
 790                                     __func__, lk);
 791                                 break;
 792                         }
 793                         goto retry_sleepq;
 794                 }
 795
 796                 /*
 797                  * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 798                  * fail, loop back and retry.
 799                  */
 800                 if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 801                         if (!atomic_fcmpset_ptr(&lk->lk_lock, &x,
 802                             x | LK_EXCLUSIVE_WAITERS)) {
 803                                 goto retry_sleepq;
 804                         }
 805                         LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 806                             __func__, lk);
 807                 }
 808
 809                 if (lwa == NULL) {
 810                         iwmesg = lk->lock_object.lo_name;
 811                         ipri = lk->lk_pri;
 812                         itimo = lk->lk_timo;
 813                 } else {
 814                         iwmesg = lwa->iwmesg;
 815                         ipri = lwa->ipri;
 816                         itimo = lwa->itimo;
 817                 }
 818
 819                 /*
 820                  * As far as we have been unable to acquire the
 821                  * exclusive lock and the exclusive waiters flag
 822                  * is set, we will sleep.
 823                  */
 824                 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 825                     SQ_EXCLUSIVE_QUEUE);
 826                 flags &= ~LK_INTERLOCK;
 827                 if (error) {
 828                         LOCK_LOG3(lk,
 829                             "%s: interrupted sleep for %p with %d",
 830                             __func__, lk, error);
 831                         break;
 832                 }
 833                 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 834                     __func__, lk);
 835         }
 836         if (error == 0) {
 837 #ifdef LOCK_PROFILING
 838                 lockmgr_note_exclusive_acquire(lk, contested, waittime,
 839                     file, line, flags);
 840 #else
 841                 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
 842                     flags);
 843 #endif
 844         }
 845
 846 out:
 847         lockmgr_exit(flags, ilk, 0);
 848         return (error);
 849 }
 850
 851 static __noinline int
 852 lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk,
 853     const char *file, int line, struct lockmgr_wait *lwa)
 854 {
 855         uintptr_t tid, x, v;
 856         int error = 0;
 857         int wakeup_swapper = 0;
 858         int op;
 859
 860         if (__predict_false(panicstr != NULL))
 861                 goto out;
 862
 863         tid = (uintptr_t)curthread;
 864
 865         _lockmgr_assert(lk, KA_SLOCKED, file, line);
 866         v = lk->lk_lock;
 867         x = v & LK_ALL_WAITERS;
 868         v &= LK_EXCLUSIVE_SPINNERS;
 869
 870         /*
 871          * Try to switch from one shared lock to an exclusive one.
 872          * We need to preserve waiters flags during the operation.
 873          */
 874         if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 875             tid | x)) {
 876                 LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 877                     line);
 878                 WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 879                     LK_TRYWIT(flags), file, line);
 880                 TD_SLOCKS_DEC(curthread);
 881                 goto out;
 882         }
 883
 884         op = flags & LK_TYPE_MASK;
 885
 886         /*
 887          * In LK_TRYUPGRADE mode, do not drop the lock,
 888          * returning EBUSY instead.
 889          */
 890         if (op == LK_TRYUPGRADE) {
 891                 LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
 892                     __func__, lk);
 893                 error = EBUSY;
 894                 goto out;
 895         }
 896
 897         /*
 898          * We have been unable to succeed in upgrading, so just
 899          * give up the shared lock.
 900          */
 901         wakeup_swapper |= wakeupshlk(lk, file, line);
 902         error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa);
 903         flags &= ~LK_INTERLOCK;
 904 out:
 905         lockmgr_exit(flags, ilk, wakeup_swapper);
 906         return (error);
 907 }
 908
 909 int
 910 lockmgr_lock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk,
 911     const char *file, int line)
 912 {
 913         struct lock_class *class;
 914         uintptr_t x, tid;
 915         u_int op;
 916         bool locked;
 917
 918         if (__predict_false(panicstr != NULL))
 919                 return (0);
 920
 921         op = flags & LK_TYPE_MASK;
 922         locked = false;
 923         switch (op) {
 924         case LK_SHARED:
 925                 if (LK_CAN_WITNESS(flags))
 926                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 927                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 928                 if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
 929                         break;
 930                 if (lockmgr_slock_try(lk, &x, flags, true)) {
 931                         lockmgr_note_shared_acquire(lk, 0, 0,
 932                             file, line, flags);
 933                         locked = true;
 934                 } else {
 935                         return (lockmgr_slock_hard(lk, flags, ilk, file, line,
 936                             NULL));
 937                 }
 938                 break;
 939         case LK_EXCLUSIVE:
 940                 if (LK_CAN_WITNESS(flags))
 941                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 942                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 943                             ilk : NULL);
 944                 tid = (uintptr_t)curthread;
 945                 if (lk->lk_lock == LK_UNLOCKED &&
 946                     atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
 947                         lockmgr_note_exclusive_acquire(lk, 0, 0, file, line,
 948                             flags);
 949                         locked = true;
 950                 } else {
 951                         return (lockmgr_xlock_hard(lk, flags, ilk, file, line,
 952                             NULL));
 953                 }
 954                 break;
 955         case LK_UPGRADE:
 956         case LK_TRYUPGRADE:
 957                 return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL));
 958         default:
 959                 break;
 960         }
 961         if (__predict_true(locked)) {
 962                 if (__predict_false(flags & LK_INTERLOCK)) {
 963                         class = LOCK_CLASS(ilk);
 964                         class->lc_unlock(ilk);
 965                 }
 966                 return (0);
 967         } else {
 968                 return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT,
 969                     LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line));
 970         }
 971 }
 972
 973 static __noinline int
 974 lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
 975     const char *file, int line)
 976
 977 {
 978         int wakeup_swapper = 0;
 979
 980         if (__predict_false(panicstr != NULL))
 981                 goto out;
 982
 983         wakeup_swapper = wakeupshlk(lk, file, line);
 984
 985 out:
 986         lockmgr_exit(flags, ilk, wakeup_swapper);
 987         return (0);
 988 }
 989
 990 static __noinline int
 991 lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk,
 992     const char *file, int line)
 993 {
 994         uintptr_t tid, v;
 995         int wakeup_swapper = 0;
 996         u_int realexslp;
 997         int queue;
 998
 999         if (__predict_false(panicstr != NULL))
1000                 goto out;
1001
1002         tid = (uintptr_t)curthread;
1003
1004         /*
1005          * As first option, treact the lock as if it has not
1006          * any waiter.
1007          * Fix-up the tid var if the lock has been disowned.
1008          */
1009         if (LK_HOLDER(x) == LK_KERNPROC)
1010                 tid = LK_KERNPROC;
1011         else {
1012                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1013                 TD_LOCKS_DEC(curthread);
1014         }
1015         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line);
1016
1017         /*
1018          * The lock is held in exclusive mode.
1019          * If the lock is recursed also, then unrecurse it.
1020          */
1021         if (lockmgr_xlocked_v(x) && lockmgr_recursed(lk)) {
1022                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk);
1023                 lk->lk_recurse--;
1024                 goto out;
1025         }
1026         if (tid != LK_KERNPROC)
1027                 lock_profile_release_lock(&lk->lock_object);
1028
1029         if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED))
1030                 goto out;
1031
1032         sleepq_lock(&lk->lock_object);
1033         x = lk->lk_lock;
1034         v = LK_UNLOCKED;
1035
1036         /*
1037          * If the lock has exclusive waiters, give them
1038          * preference in order to avoid deadlock with
1039          * shared runners up.
1040          * If interruptible sleeps left the exclusive queue
1041          * empty avoid a starvation for the threads sleeping
1042          * on the shared queue by giving them precedence
1043          * and cleaning up the exclusive waiters bit anyway.
1044          * Please note that lk_exslpfail count may be lying
1045          * about the real number of waiters with the
1046          * LK_SLEEPFAIL flag on because they may be used in
1047          * conjunction with interruptible sleeps so
1048          * lk_exslpfail might be considered an 'upper limit'
1049          * bound, including the edge cases.
1050          */
1051         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1052         realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE);
1053         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1054                 if (lk->lk_exslpfail < realexslp) {
1055                         lk->lk_exslpfail = 0;
1056                         queue = SQ_EXCLUSIVE_QUEUE;
1057                         v |= (x & LK_SHARED_WAITERS);
1058                 } else {
1059                         lk->lk_exslpfail = 0;
1060                         LOCK_LOG2(lk,
1061                             "%s: %p has only LK_SLEEPFAIL sleepers",
1062                             __func__, lk);
1063                         LOCK_LOG2(lk,
1064                             "%s: %p waking up threads on the exclusive queue",
1065                             __func__, lk);
1066                         wakeup_swapper = sleepq_broadcast(&lk->lock_object,
1067                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1068                         queue = SQ_SHARED_QUEUE;
1069                 }
1070         } else {
1071
1072                 /*
1073                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1074                  * on and using interruptible sleeps/timeout
1075                  * may have left spourious lk_exslpfail counts
1076                  * on, so clean it up anyway.
1077                  */
1078                 lk->lk_exslpfail = 0;
1079                 queue = SQ_SHARED_QUEUE;
1080         }
1081
1082         LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
1083             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1084             "exclusive");
1085         atomic_store_rel_ptr(&lk->lk_lock, v);
1086         wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue);
1087         sleepq_release(&lk->lock_object);
1088
1089 out:
1090         lockmgr_exit(flags, ilk, wakeup_swapper);
1091         return (0);
1092 }
1093
1094 int
1095 lockmgr_unlock_fast_path(struct lock *lk, u_int flags, struct lock_object *ilk)
1096 {
1097         struct lock_class *class;
1098         uintptr_t x, tid;
1099         const char *file;
1100         int line;
1101
1102         if (__predict_false(panicstr != NULL))
1103                 return (0);
1104
1105         file = __FILE__;
1106         line = __LINE__;
1107
1108         _lockmgr_assert(lk, KA_LOCKED, file, line);
1109         x = lk->lk_lock;
1110         if (__predict_true(x & LK_SHARE) != 0) {
1111                 if (lockmgr_sunlock_try(lk, &x)) {
1112                         lockmgr_note_shared_release(lk, file, line);
1113                 } else {
1114                         return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1115                 }
1116         } else {
1117                 tid = (uintptr_t)curthread;
1118                 if (!lockmgr_recursed(lk) &&
1119                     atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) {
1120                         lockmgr_note_exclusive_release(lk, file, line);
1121                 } else {
1122                         return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1123                 }
1124         }
1125         if (__predict_false(flags & LK_INTERLOCK)) {
1126                 class = LOCK_CLASS(ilk);
1127                 class->lc_unlock(ilk);
1128         }
1129         return (0);
1130 }
1131
1132 int
1133 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
1134     const char *wmesg, int pri, int timo, const char *file, int line)
1135 {
1136         GIANT_DECLARE;
1137         struct lockmgr_wait lwa;
1138         struct lock_class *class;
1139         const char *iwmesg;
1140         uintptr_t tid, v, x;
1141         u_int op, realexslp;
1142         int error, ipri, itimo, queue, wakeup_swapper;
1143 #ifdef LOCK_PROFILING
1144         uint64_t waittime = 0;
1145         int contested = 0;
1146 #endif
1147
1148         if (panicstr != NULL)
1149                 return (0);
1150
1151         error = 0;
1152         tid = (uintptr_t)curthread;
1153         op = (flags & LK_TYPE_MASK);
1154         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
1155         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
1156         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
1157
1158         lwa.iwmesg = iwmesg;
1159         lwa.ipri = ipri;
1160         lwa.itimo = itimo;
1161
1162         MPASS((flags & ~LK_TOTAL_MASK) == 0);
1163         KASSERT((op & (op - 1)) == 0,
1164             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
1165         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
1166             (op != LK_DOWNGRADE && op != LK_RELEASE),
1167             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
1168             __func__, file, line));
1169         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
1170             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
1171             __func__, file, line));
1172         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
1173             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
1174             lk->lock_object.lo_name, file, line));
1175
1176         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
1177
1178         if (lk->lock_object.lo_flags & LK_NOSHARE) {
1179                 switch (op) {
1180                 case LK_SHARED:
1181                         op = LK_EXCLUSIVE;
1182                         break;
1183                 case LK_UPGRADE:
1184                 case LK_TRYUPGRADE:
1185                 case LK_DOWNGRADE:
1186                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
1187                             file, line);
1188                         if (flags & LK_INTERLOCK)
1189                                 class->lc_unlock(ilk);
1190                         return (0);
1191                 }
1192         }
1193
1194         wakeup_swapper = 0;
1195         switch (op) {
1196         case LK_SHARED:
1197                 return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa));
1198                 break;
1199         case LK_UPGRADE:
1200         case LK_TRYUPGRADE:
1201                 return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa));
1202                 break;
1203         case LK_EXCLUSIVE:
1204                 return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa));
1205                 break;
1206         case LK_DOWNGRADE:
1207                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
1208                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
1209                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
1210
1211                 /*
1212                  * Panic if the lock is recursed.
1213                  */
1214                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1215                         if (flags & LK_INTERLOCK)
1216                                 class->lc_unlock(ilk);
1217                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
1218                             __func__, iwmesg, file, line);
1219                 }
1220                 TD_SLOCKS_INC(curthread);
1221
1222                 /*
1223                  * In order to preserve waiters flags, just spin.
1224                  */
1225                 for (;;) {
1226                         x = lk->lk_lock;
1227                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1228                         x &= LK_ALL_WAITERS;
1229                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1230                             LK_SHARERS_LOCK(1) | x))
1231                                 break;
1232                         cpu_spinwait();
1233                 }
1234                 break;
1235         case LK_RELEASE:
1236                 _lockmgr_assert(lk, KA_LOCKED, file, line);
1237                 x = lk->lk_lock;
1238
1239                 if (__predict_true(x & LK_SHARE) != 0) {
1240                         return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line));
1241                 } else {
1242                         return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line));
1243                 }
1244                 break;
1245         case LK_DRAIN:
1246                 if (LK_CAN_WITNESS(flags))
1247                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1248                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1249                             ilk : NULL);
1250
1251                 /*
1252                  * Trying to drain a lock we already own will result in a
1253                  * deadlock.
1254                  */
1255                 if (lockmgr_xlocked(lk)) {
1256                         if (flags & LK_INTERLOCK)
1257                                 class->lc_unlock(ilk);
1258                         panic("%s: draining %s with the lock held @ %s:%d\n",
1259                             __func__, iwmesg, file, line);
1260                 }
1261
1262                 for (;;) {
1263                         if (lk->lk_lock == LK_UNLOCKED &&
1264                             atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1265                                 break;
1266
1267 #ifdef HWPMC_HOOKS
1268                         PMC_SOFT_CALL( , , lock, failed);
1269 #endif
1270                         lock_profile_obtain_lock_failed(&lk->lock_object,
1271                             &contested, &waittime);
1272
1273                         /*
1274                          * If the lock is expected to not sleep just give up
1275                          * and return.
1276                          */
1277                         if (LK_TRYOP(flags)) {
1278                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1279                                     __func__, lk);
1280                                 error = EBUSY;
1281                                 break;
1282                         }
1283
1284                         /*
1285                          * Acquire the sleepqueue chain lock because we
1286                          * probabilly will need to manipulate waiters flags.
1287                          */
1288                         sleepq_lock(&lk->lock_object);
1289                         x = lk->lk_lock;
1290
1291                         /*
1292                          * if the lock has been released while we spun on
1293                          * the sleepqueue chain lock just try again.
1294                          */
1295                         if (x == LK_UNLOCKED) {
1296                                 sleepq_release(&lk->lock_object);
1297                                 continue;
1298                         }
1299
1300                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1301                         if ((x & ~v) == LK_UNLOCKED) {
1302                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1303
1304                                 /*
1305                                  * If interruptible sleeps left the exclusive
1306                                  * queue empty avoid a starvation for the
1307                                  * threads sleeping on the shared queue by
1308                                  * giving them precedence and cleaning up the
1309                                  * exclusive waiters bit anyway.
1310                                  * Please note that lk_exslpfail count may be
1311                                  * lying about the real number of waiters with
1312                                  * the LK_SLEEPFAIL flag on because they may
1313                                  * be used in conjunction with interruptible
1314                                  * sleeps so lk_exslpfail might be considered
1315                                  * an 'upper limit' bound, including the edge
1316                                  * cases.
1317                                  */
1318                                 if (v & LK_EXCLUSIVE_WAITERS) {
1319                                         queue = SQ_EXCLUSIVE_QUEUE;
1320                                         v &= ~LK_EXCLUSIVE_WAITERS;
1321                                 } else {
1322
1323                                         /*
1324                                          * Exclusive waiters sleeping with
1325                                          * LK_SLEEPFAIL on and using
1326                                          * interruptible sleeps/timeout may
1327                                          * have left spourious lk_exslpfail
1328                                          * counts on, so clean it up anyway.
1329                                          */
1330                                         MPASS(v & LK_SHARED_WAITERS);
1331                                         lk->lk_exslpfail = 0;
1332                                         queue = SQ_SHARED_QUEUE;
1333                                         v &= ~LK_SHARED_WAITERS;
1334                                 }
1335                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1336                                         realexslp =
1337                                             sleepq_sleepcnt(&lk->lock_object,
1338                                             SQ_EXCLUSIVE_QUEUE);
1339                                         if (lk->lk_exslpfail >= realexslp) {
1340                                                 lk->lk_exslpfail = 0;
1341                                                 queue = SQ_SHARED_QUEUE;
1342                                                 v &= ~LK_SHARED_WAITERS;
1343                                                 if (realexslp != 0) {
1344                                                         LOCK_LOG2(lk,
1345                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1346                                                             __func__, lk);
1347                                                         LOCK_LOG2(lk,
1348                         "%s: %p waking up threads on the exclusive queue",
1349                                                             __func__, lk);
1350                                                         wakeup_swapper =
1351                                                             sleepq_broadcast(
1352                                                             &lk->lock_object,
1353                                                             SLEEPQ_LK, 0,
1354                                                             SQ_EXCLUSIVE_QUEUE);
1355                                                 }
1356                                         } else
1357                                                 lk->lk_exslpfail = 0;
1358                                 }
1359                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1360                                         sleepq_release(&lk->lock_object);
1361                                         continue;
1362                                 }
1363                                 LOCK_LOG3(lk,
1364                                 "%s: %p waking up all threads on the %s queue",
1365                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1366                                     "shared" : "exclusive");
1367                                 wakeup_swapper |= sleepq_broadcast(
1368                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1369
1370                                 /*
1371                                  * If shared waiters have been woken up we need
1372                                  * to wait for one of them to acquire the lock
1373                                  * before to set the exclusive waiters in
1374                                  * order to avoid a deadlock.
1375                                  */
1376                                 if (queue == SQ_SHARED_QUEUE) {
1377                                         for (v = lk->lk_lock;
1378                                             (v & LK_SHARE) && !LK_SHARERS(v);
1379                                             v = lk->lk_lock)
1380                                                 cpu_spinwait();
1381                                 }
1382                         }
1383
1384                         /*
1385                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1386                          * fail, loop back and retry.
1387                          */
1388                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1389                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1390                                     x | LK_EXCLUSIVE_WAITERS)) {
1391                                         sleepq_release(&lk->lock_object);
1392                                         continue;
1393                                 }
1394                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1395                                     __func__, lk);
1396                         }
1397
1398                         /*
1399                          * As far as we have been unable to acquire the
1400                          * exclusive lock and the exclusive waiters flag
1401                          * is set, we will sleep.
1402                          */
1403                         if (flags & LK_INTERLOCK) {
1404                                 class->lc_unlock(ilk);
1405                                 flags &= ~LK_INTERLOCK;
1406                         }
1407                         GIANT_SAVE();
1408                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1409                             SQ_EXCLUSIVE_QUEUE);
1410                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1411                         GIANT_RESTORE();
1412                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1413                             __func__, lk);
1414                 }
1415
1416                 if (error == 0) {
1417                         lock_profile_obtain_lock_success(&lk->lock_object,
1418                             contested, waittime, file, line);
1419                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1420                             lk->lk_recurse, file, line);
1421                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1422                             LK_TRYWIT(flags), file, line);
1423                         TD_LOCKS_INC(curthread);
1424                         STACK_SAVE(lk);
1425                 }
1426                 break;
1427         default:
1428                 if (flags & LK_INTERLOCK)
1429                         class->lc_unlock(ilk);
1430                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1431         }
1432
1433         if (flags & LK_INTERLOCK)
1434                 class->lc_unlock(ilk);
1435         if (wakeup_swapper)
1436                 kick_proc0();
1437
1438         return (error);
1439 }
1440
1441 void
1442 _lockmgr_disown(struct lock *lk, const char *file, int line)
1443 {
1444         uintptr_t tid, x;
1445
1446         if (SCHEDULER_STOPPED())
1447                 return;
1448
1449         tid = (uintptr_t)curthread;
1450         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1451
1452         /*
1453          * Panic if the lock is recursed.
1454          */
1455         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1456                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1457                     __func__,  file, line);
1458
1459         /*
1460          * If the owner is already LK_KERNPROC just skip the whole operation.
1461          */
1462         if (LK_HOLDER(lk->lk_lock) != tid)
1463                 return;
1464         lock_profile_release_lock(&lk->lock_object);
1465         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1466         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1467         TD_LOCKS_DEC(curthread);
1468         STACK_SAVE(lk);
1469
1470         /*
1471          * In order to preserve waiters flags, just spin.
1472          */
1473         for (;;) {
1474                 x = lk->lk_lock;
1475                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1476                 x &= LK_ALL_WAITERS;
1477                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1478                     LK_KERNPROC | x))
1479                         return;
1480                 cpu_spinwait();
1481         }
1482 }
1483
1484 void
1485 lockmgr_printinfo(const struct lock *lk)
1486 {
1487         struct thread *td;
1488         uintptr_t x;
1489
1490         if (lk->lk_lock == LK_UNLOCKED)
1491                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1492         else if (lk->lk_lock & LK_SHARE)
1493                 printf("lock type %s: SHARED (count %ju)\n",
1494                     lk->lock_object.lo_name,
1495                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1496         else {
1497                 td = lockmgr_xholder(lk);
1498                 if (td == (struct thread *)LK_KERNPROC)
1499                         printf("lock type %s: EXCL by KERNPROC\n",
1500                             lk->lock_object.lo_name);
1501                 else
1502                         printf("lock type %s: EXCL by thread %p "
1503                             "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1504                             td, td->td_proc->p_pid, td->td_proc->p_comm,
1505                             td->td_tid);
1506         }
1507
1508         x = lk->lk_lock;
1509         if (x & LK_EXCLUSIVE_WAITERS)
1510                 printf(" with exclusive waiters pending\n");
1511         if (x & LK_SHARED_WAITERS)
1512                 printf(" with shared waiters pending\n");
1513         if (x & LK_EXCLUSIVE_SPINNERS)
1514                 printf(" with exclusive spinners pending\n");
1515
1516         STACK_PRINT(lk);
1517 }
1518
1519 int
1520 lockstatus(const struct lock *lk)
1521 {
1522         uintptr_t v, x;
1523         int ret;
1524
1525         ret = LK_SHARED;
1526         x = lk->lk_lock;
1527         v = LK_HOLDER(x);
1528
1529         if ((x & LK_SHARE) == 0) {
1530                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1531                         ret = LK_EXCLUSIVE;
1532                 else
1533                         ret = LK_EXCLOTHER;
1534         } else if (x == LK_UNLOCKED)
1535                 ret = 0;
1536
1537         return (ret);
1538 }
1539
1540 #ifdef INVARIANT_SUPPORT
1541
1542 FEATURE(invariant_support,
1543     "Support for modules compiled with INVARIANTS option");
1544
1545 #ifndef INVARIANTS
1546 #undef  _lockmgr_assert
1547 #endif
1548
1549 void
1550 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1551 {
1552         int slocked = 0;
1553
1554         if (panicstr != NULL)
1555                 return;
1556         switch (what) {
1557         case KA_SLOCKED:
1558         case KA_SLOCKED | KA_NOTRECURSED:
1559         case KA_SLOCKED | KA_RECURSED:
1560                 slocked = 1;
1561         case KA_LOCKED:
1562         case KA_LOCKED | KA_NOTRECURSED:
1563         case KA_LOCKED | KA_RECURSED:
1564 #ifdef WITNESS
1565
1566                 /*
1567                  * We cannot trust WITNESS if the lock is held in exclusive
1568                  * mode and a call to lockmgr_disown() happened.
1569                  * Workaround this skipping the check if the lock is held in
1570                  * exclusive mode even for the KA_LOCKED case.
1571                  */
1572                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1573                         witness_assert(&lk->lock_object, what, file, line);
1574                         break;
1575                 }
1576 #endif
1577                 if (lk->lk_lock == LK_UNLOCKED ||
1578                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1579                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1580                         panic("Lock %s not %slocked @ %s:%d\n",
1581                             lk->lock_object.lo_name, slocked ? "share" : "",
1582                             file, line);
1583
1584                 if ((lk->lk_lock & LK_SHARE) == 0) {
1585                         if (lockmgr_recursed(lk)) {
1586                                 if (what & KA_NOTRECURSED)
1587                                         panic("Lock %s recursed @ %s:%d\n",
1588                                             lk->lock_object.lo_name, file,
1589                                             line);
1590                         } else if (what & KA_RECURSED)
1591                                 panic("Lock %s not recursed @ %s:%d\n",
1592                                     lk->lock_object.lo_name, file, line);
1593                 }
1594                 break;
1595         case KA_XLOCKED:
1596         case KA_XLOCKED | KA_NOTRECURSED:
1597         case KA_XLOCKED | KA_RECURSED:
1598                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1599                         panic("Lock %s not exclusively locked @ %s:%d\n",
1600                             lk->lock_object.lo_name, file, line);
1601                 if (lockmgr_recursed(lk)) {
1602                         if (what & KA_NOTRECURSED)
1603                                 panic("Lock %s recursed @ %s:%d\n",
1604                                     lk->lock_object.lo_name, file, line);
1605                 } else if (what & KA_RECURSED)
1606                         panic("Lock %s not recursed @ %s:%d\n",
1607                             lk->lock_object.lo_name, file, line);
1608                 break;
1609         case KA_UNLOCKED:
1610                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1611                         panic("Lock %s exclusively locked @ %s:%d\n",
1612                             lk->lock_object.lo_name, file, line);
1613                 break;
1614         default:
1615                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1616                     line);
1617         }
1618 }
1619 #endif
1620
1621 #ifdef DDB
1622 int
1623 lockmgr_chain(struct thread *td, struct thread **ownerp)
1624 {
1625         struct lock *lk;
1626
1627         lk = td->td_wchan;
1628
1629         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1630                 return (0);
1631         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1632         if (lk->lk_lock & LK_SHARE)
1633                 db_printf("SHARED (count %ju)\n",
1634                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1635         else
1636                 db_printf("EXCL\n");
1637         *ownerp = lockmgr_xholder(lk);
1638
1639         return (1);
1640 }
1641
1642 static void
1643 db_show_lockmgr(const struct lock_object *lock)
1644 {
1645         struct thread *td;
1646         const struct lock *lk;
1647
1648         lk = (const struct lock *)lock;
1649
1650         db_printf(" state: ");
1651         if (lk->lk_lock == LK_UNLOCKED)
1652                 db_printf("UNLOCKED\n");
1653         else if (lk->lk_lock & LK_SHARE)
1654                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1655         else {
1656                 td = lockmgr_xholder(lk);
1657                 if (td == (struct thread *)LK_KERNPROC)
1658                         db_printf("XLOCK: LK_KERNPROC\n");
1659                 else
1660                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1661                             td->td_tid, td->td_proc->p_pid,
1662                             td->td_proc->p_comm);
1663                 if (lockmgr_recursed(lk))
1664                         db_printf(" recursed: %d\n", lk->lk_recurse);
1665         }
1666         db_printf(" waiters: ");
1667         switch (lk->lk_lock & LK_ALL_WAITERS) {
1668         case LK_SHARED_WAITERS:
1669                 db_printf("shared\n");
1670                 break;
1671         case LK_EXCLUSIVE_WAITERS:
1672                 db_printf("exclusive\n");
1673                 break;
1674         case LK_ALL_WAITERS:
1675                 db_printf("shared and exclusive\n");
1676                 break;
1677         default:
1678                 db_printf("none\n");
1679         }
1680         db_printf(" spinners: ");
1681         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1682                 db_printf("exclusive\n");
1683         else
1684                 db_printf("none\n");
1685 }
1686 #endif