sys/kern/kern_lock.c

   1 /*-
   2  * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice(s), this list of conditions and the following disclaimer as
  10  *    the first lines of this file unmodified other than the possible
  11  *    addition of one or more copyright notices.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice(s), this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  26  * DAMAGE.
  27  */
  28
  29 #include "opt_adaptive_lockmgrs.h"
  30 #include "opt_ddb.h"
  31 #include "opt_hwpmc_hooks.h"
  32
  33 #include <sys/cdefs.h>
  34 __FBSDID("$FreeBSD$");
  35
  36 #include <sys/param.h>
  37 #include <sys/kdb.h>
  38 #include <sys/ktr.h>
  39 #include <sys/lock.h>
  40 #include <sys/lock_profile.h>
  41 #include <sys/lockmgr.h>
  42 #include <sys/mutex.h>
  43 #include <sys/proc.h>
  44 #include <sys/sleepqueue.h>
  45 #ifdef DEBUG_LOCKS
  46 #include <sys/stack.h>
  47 #endif
  48 #include <sys/sysctl.h>
  49 #include <sys/systm.h>
  50
  51 #include <machine/cpu.h>
  52
  53 #ifdef DDB
  54 #include <ddb/ddb.h>
  55 #endif
  56
  57 #ifdef HWPMC_HOOKS
  58 #include <sys/pmckern.h>
  59 PMC_SOFT_DECLARE( , , lock, failed);
  60 #endif
  61
  62 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
  63     (LK_ADAPTIVE | LK_NOSHARE));
  64 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
  65     ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
  66
  67 #define SQ_EXCLUSIVE_QUEUE      0
  68 #define SQ_SHARED_QUEUE         1
  69
  70 #ifndef INVARIANTS
  71 #define _lockmgr_assert(lk, what, file, line)
  72 #define TD_LOCKS_INC(td)
  73 #define TD_LOCKS_DEC(td)
  74 #else
  75 #define TD_LOCKS_INC(td)        ((td)->td_locks++)
  76 #define TD_LOCKS_DEC(td)        ((td)->td_locks--)
  77 #endif
  78 #define TD_SLOCKS_INC(td)       ((td)->td_lk_slocks++)
  79 #define TD_SLOCKS_DEC(td)       ((td)->td_lk_slocks--)
  80
  81 #ifndef DEBUG_LOCKS
  82 #define STACK_PRINT(lk)
  83 #define STACK_SAVE(lk)
  84 #define STACK_ZERO(lk)
  85 #else
  86 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
  87 #define STACK_SAVE(lk)  stack_save(&(lk)->lk_stack)
  88 #define STACK_ZERO(lk)  stack_zero(&(lk)->lk_stack)
  89 #endif
  90
  91 #define LOCK_LOG2(lk, string, arg1, arg2)                               \
  92         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  93                 CTR2(KTR_LOCK, (string), (arg1), (arg2))
  94 #define LOCK_LOG3(lk, string, arg1, arg2, arg3)                         \
  95         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  96                 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
  97
  98 #define GIANT_DECLARE                                                   \
  99         int _i = 0;                                                     \
 100         WITNESS_SAVE_DECL(Giant)
 101 #define GIANT_RESTORE() do {                                            \
 102         if (_i > 0) {                                                   \
 103                 while (_i--)                                            \
 104                         mtx_lock(&Giant);                               \
 105                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 106         }                                                               \
 107 } while (0)
 108 #define GIANT_SAVE() do {                                               \
 109         if (mtx_owned(&Giant)) {                                        \
 110                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
 111                 while (mtx_owned(&Giant)) {                             \
 112                         _i++;                                           \
 113                         mtx_unlock(&Giant);                             \
 114                 }                                                       \
 115         }                                                               \
 116 } while (0)
 117
 118 #define LK_CAN_SHARE(x, flags)                                          \
 119         (((x) & LK_SHARE) &&                                            \
 120         (((x) & (LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 0 || \
 121         (curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||  \
 122         (curthread->td_pflags & TDP_DEADLKTREAT)))
 123 #define LK_TRYOP(x)                                                     \
 124         ((x) & LK_NOWAIT)
 125
 126 #define LK_CAN_WITNESS(x)                                               \
 127         (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
 128 #define LK_TRYWIT(x)                                                    \
 129         (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
 130
 131 #define LK_CAN_ADAPT(lk, f)                                             \
 132         (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&             \
 133         ((f) & LK_SLEEPFAIL) == 0)
 134
 135 #define lockmgr_disowned(lk)                                            \
 136         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
 137
 138 #define lockmgr_xlocked(lk)                                             \
 139         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
 140
 141 static void     assert_lockmgr(const struct lock_object *lock, int how);
 142 #ifdef DDB
 143 static void     db_show_lockmgr(const struct lock_object *lock);
 144 #endif
 145 static void     lock_lockmgr(struct lock_object *lock, uintptr_t how);
 146 #ifdef KDTRACE_HOOKS
 147 static int      owner_lockmgr(const struct lock_object *lock,
 148                     struct thread **owner);
 149 #endif
 150 static uintptr_t unlock_lockmgr(struct lock_object *lock);
 151
 152 struct lock_class lock_class_lockmgr = {
 153         .lc_name = "lockmgr",
 154         .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
 155         .lc_assert = assert_lockmgr,
 156 #ifdef DDB
 157         .lc_ddb_show = db_show_lockmgr,
 158 #endif
 159         .lc_lock = lock_lockmgr,
 160         .lc_unlock = unlock_lockmgr,
 161 #ifdef KDTRACE_HOOKS
 162         .lc_owner = owner_lockmgr,
 163 #endif
 164 };
 165
 166 #ifdef ADAPTIVE_LOCKMGRS
 167 static u_int alk_retries = 10;
 168 static u_int alk_loops = 10000;
 169 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL,
 170     "lockmgr debugging");
 171 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
 172 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
 173 #endif
 174
 175 static __inline struct thread *
 176 lockmgr_xholder(const struct lock *lk)
 177 {
 178         uintptr_t x;
 179
 180         x = lk->lk_lock;
 181         return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
 182 }
 183
 184 /*
 185  * It assumes sleepq_lock held and returns with this one unheld.
 186  * It also assumes the generic interlock is sane and previously checked.
 187  * If LK_INTERLOCK is specified the interlock is not reacquired after the
 188  * sleep.
 189  */
 190 static __inline int
 191 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
 192     const char *wmesg, int pri, int timo, int queue)
 193 {
 194         GIANT_DECLARE;
 195         struct lock_class *class;
 196         int catch, error;
 197
 198         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 199         catch = pri & PCATCH;
 200         pri &= PRIMASK;
 201         error = 0;
 202
 203         LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
 204             (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
 205
 206         if (flags & LK_INTERLOCK)
 207                 class->lc_unlock(ilk);
 208         if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
 209                 lk->lk_exslpfail++;
 210         GIANT_SAVE();
 211         sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
 212             SLEEPQ_INTERRUPTIBLE : 0), queue);
 213         if ((flags & LK_TIMELOCK) && timo)
 214                 sleepq_set_timeout(&lk->lock_object, timo);
 215
 216         /*
 217          * Decisional switch for real sleeping.
 218          */
 219         if ((flags & LK_TIMELOCK) && timo && catch)
 220                 error = sleepq_timedwait_sig(&lk->lock_object, pri);
 221         else if ((flags & LK_TIMELOCK) && timo)
 222                 error = sleepq_timedwait(&lk->lock_object, pri);
 223         else if (catch)
 224                 error = sleepq_wait_sig(&lk->lock_object, pri);
 225         else
 226                 sleepq_wait(&lk->lock_object, pri);
 227         GIANT_RESTORE();
 228         if ((flags & LK_SLEEPFAIL) && error == 0)
 229                 error = ENOLCK;
 230
 231         return (error);
 232 }
 233
 234 static __inline int
 235 wakeupshlk(struct lock *lk, const char *file, int line)
 236 {
 237         uintptr_t v, x;
 238         u_int realexslp;
 239         int queue, wakeup_swapper;
 240
 241         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 242         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 243
 244         wakeup_swapper = 0;
 245         for (;;) {
 246                 x = lk->lk_lock;
 247
 248                 /*
 249                  * If there is more than one shared lock held, just drop one
 250                  * and return.
 251                  */
 252                 if (LK_SHARERS(x) > 1) {
 253                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
 254                             x - LK_ONE_SHARER))
 255                                 break;
 256                         continue;
 257                 }
 258
 259                 /*
 260                  * If there are not waiters on the exclusive queue, drop the
 261                  * lock quickly.
 262                  */
 263                 if ((x & LK_ALL_WAITERS) == 0) {
 264                         MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
 265                             LK_SHARERS_LOCK(1));
 266                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
 267                                 break;
 268                         continue;
 269                 }
 270
 271                 /*
 272                  * We should have a sharer with waiters, so enter the hard
 273                  * path in order to handle wakeups correctly.
 274                  */
 275                 sleepq_lock(&lk->lock_object);
 276                 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 277                 v = LK_UNLOCKED;
 278
 279                 /*
 280                  * If the lock has exclusive waiters, give them preference in
 281                  * order to avoid deadlock with shared runners up.
 282                  * If interruptible sleeps left the exclusive queue empty
 283                  * avoid a starvation for the threads sleeping on the shared
 284                  * queue by giving them precedence and cleaning up the
 285                  * exclusive waiters bit anyway.
 286                  * Please note that lk_exslpfail count may be lying about
 287                  * the real number of waiters with the LK_SLEEPFAIL flag on
 288                  * because they may be used in conjuction with interruptible
 289                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 290                  * bound, including the edge cases.
 291                  */
 292                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 293                     SQ_EXCLUSIVE_QUEUE);
 294                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 295                         if (lk->lk_exslpfail < realexslp) {
 296                                 lk->lk_exslpfail = 0;
 297                                 queue = SQ_EXCLUSIVE_QUEUE;
 298                                 v |= (x & LK_SHARED_WAITERS);
 299                         } else {
 300                                 lk->lk_exslpfail = 0;
 301                                 LOCK_LOG2(lk,
 302                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 303                                     __func__, lk);
 304                                 LOCK_LOG2(lk,
 305                             "%s: %p waking up threads on the exclusive queue",
 306                                     __func__, lk);
 307                                 wakeup_swapper =
 308                                     sleepq_broadcast(&lk->lock_object,
 309                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 310                                 queue = SQ_SHARED_QUEUE;
 311                         }
 312
 313                 } else {
 314
 315                         /*
 316                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 317                          * and using interruptible sleeps/timeout may have
 318                          * left spourious lk_exslpfail counts on, so clean
 319                          * it up anyway.
 320                          */
 321                         lk->lk_exslpfail = 0;
 322                         queue = SQ_SHARED_QUEUE;
 323                 }
 324
 325                 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
 326                     v)) {
 327                         sleepq_release(&lk->lock_object);
 328                         continue;
 329                 }
 330                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 331                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 332                     "exclusive");
 333                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 334                     0, queue);
 335                 sleepq_release(&lk->lock_object);
 336                 break;
 337         }
 338
 339         lock_profile_release_lock(&lk->lock_object);
 340         TD_LOCKS_DEC(curthread);
 341         TD_SLOCKS_DEC(curthread);
 342         return (wakeup_swapper);
 343 }
 344
 345 static void
 346 assert_lockmgr(const struct lock_object *lock, int what)
 347 {
 348
 349         panic("lockmgr locks do not support assertions");
 350 }
 351
 352 static void
 353 lock_lockmgr(struct lock_object *lock, uintptr_t how)
 354 {
 355
 356         panic("lockmgr locks do not support sleep interlocking");
 357 }
 358
 359 static uintptr_t
 360 unlock_lockmgr(struct lock_object *lock)
 361 {
 362
 363         panic("lockmgr locks do not support sleep interlocking");
 364 }
 365
 366 #ifdef KDTRACE_HOOKS
 367 static int
 368 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
 369 {
 370
 371         panic("lockmgr locks do not support owner inquiring");
 372 }
 373 #endif
 374
 375 void
 376 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 377 {
 378         int iflags;
 379
 380         MPASS((flags & ~LK_INIT_MASK) == 0);
 381         ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
 382             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
 383             &lk->lk_lock));
 384
 385         iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 386         if (flags & LK_CANRECURSE)
 387                 iflags |= LO_RECURSABLE;
 388         if ((flags & LK_NODUP) == 0)
 389                 iflags |= LO_DUPOK;
 390         if (flags & LK_NOPROFILE)
 391                 iflags |= LO_NOPROFILE;
 392         if ((flags & LK_NOWITNESS) == 0)
 393                 iflags |= LO_WITNESS;
 394         if (flags & LK_QUIET)
 395                 iflags |= LO_QUIET;
 396         if (flags & LK_IS_VNODE)
 397                 iflags |= LO_IS_VNODE;
 398         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 399
 400         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 401         lk->lk_lock = LK_UNLOCKED;
 402         lk->lk_recurse = 0;
 403         lk->lk_exslpfail = 0;
 404         lk->lk_timo = timo;
 405         lk->lk_pri = pri;
 406         STACK_ZERO(lk);
 407 }
 408
 409 /*
 410  * XXX: Gross hacks to manipulate external lock flags after
 411  * initialization.  Used for certain vnode and buf locks.
 412  */
 413 void
 414 lockallowshare(struct lock *lk)
 415 {
 416
 417         lockmgr_assert(lk, KA_XLOCKED);
 418         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 419 }
 420
 421 void
 422 lockdisableshare(struct lock *lk)
 423 {
 424
 425         lockmgr_assert(lk, KA_XLOCKED);
 426         lk->lock_object.lo_flags |= LK_NOSHARE;
 427 }
 428
 429 void
 430 lockallowrecurse(struct lock *lk)
 431 {
 432
 433         lockmgr_assert(lk, KA_XLOCKED);
 434         lk->lock_object.lo_flags |= LO_RECURSABLE;
 435 }
 436
 437 void
 438 lockdisablerecurse(struct lock *lk)
 439 {
 440
 441         lockmgr_assert(lk, KA_XLOCKED);
 442         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 443 }
 444
 445 void
 446 lockdestroy(struct lock *lk)
 447 {
 448
 449         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 450         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 451         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 452         lock_destroy(&lk->lock_object);
 453 }
 454
 455 int
 456 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 457     const char *wmesg, int pri, int timo, const char *file, int line)
 458 {
 459         GIANT_DECLARE;
 460         struct lock_class *class;
 461         const char *iwmesg;
 462         uintptr_t tid, v, x;
 463         u_int op, realexslp;
 464         int error, ipri, itimo, queue, wakeup_swapper;
 465 #ifdef LOCK_PROFILING
 466         uint64_t waittime = 0;
 467         int contested = 0;
 468 #endif
 469 #ifdef ADAPTIVE_LOCKMGRS
 470         volatile struct thread *owner;
 471         u_int i, spintries = 0;
 472 #endif
 473
 474         error = 0;
 475         tid = (uintptr_t)curthread;
 476         op = (flags & LK_TYPE_MASK);
 477         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 478         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 479         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 480
 481         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 482         KASSERT((op & (op - 1)) == 0,
 483             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 484         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 485             (op != LK_DOWNGRADE && op != LK_RELEASE),
 486             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 487             __func__, file, line));
 488         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 489             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 490             __func__, file, line));
 491         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 492             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
 493             lk->lock_object.lo_name, file, line));
 494
 495         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 496         if (panicstr != NULL) {
 497                 if (flags & LK_INTERLOCK)
 498                         class->lc_unlock(ilk);
 499                 return (0);
 500         }
 501
 502         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 503                 switch (op) {
 504                 case LK_SHARED:
 505                         op = LK_EXCLUSIVE;
 506                         break;
 507                 case LK_UPGRADE:
 508                 case LK_TRYUPGRADE:
 509                 case LK_DOWNGRADE:
 510                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 511                             file, line);
 512                         if (flags & LK_INTERLOCK)
 513                                 class->lc_unlock(ilk);
 514                         return (0);
 515                 }
 516         }
 517
 518         wakeup_swapper = 0;
 519         switch (op) {
 520         case LK_SHARED:
 521                 if (LK_CAN_WITNESS(flags))
 522                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 523                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 524                 for (;;) {
 525                         x = lk->lk_lock;
 526
 527                         /*
 528                          * If no other thread has an exclusive lock, or
 529                          * no exclusive waiter is present, bump the count of
 530                          * sharers.  Since we have to preserve the state of
 531                          * waiters, if we fail to acquire the shared lock
 532                          * loop back and retry.
 533                          */
 534                         if (LK_CAN_SHARE(x, flags)) {
 535                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 536                                     x + LK_ONE_SHARER))
 537                                         break;
 538                                 continue;
 539                         }
 540 #ifdef HWPMC_HOOKS
 541                         PMC_SOFT_CALL( , , lock, failed);
 542 #endif
 543                         lock_profile_obtain_lock_failed(&lk->lock_object,
 544                             &contested, &waittime);
 545
 546                         /*
 547                          * If the lock is already held by curthread in
 548                          * exclusive way avoid a deadlock.
 549                          */
 550                         if (LK_HOLDER(x) == tid) {
 551                                 LOCK_LOG2(lk,
 552                                     "%s: %p already held in exclusive mode",
 553                                     __func__, lk);
 554                                 error = EDEADLK;
 555                                 break;
 556                         }
 557
 558                         /*
 559                          * If the lock is expected to not sleep just give up
 560                          * and return.
 561                          */
 562                         if (LK_TRYOP(flags)) {
 563                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 564                                     __func__, lk);
 565                                 error = EBUSY;
 566                                 break;
 567                         }
 568
 569 #ifdef ADAPTIVE_LOCKMGRS
 570                         /*
 571                          * If the owner is running on another CPU, spin until
 572                          * the owner stops running or the state of the lock
 573                          * changes.  We need a double-state handle here
 574                          * because for a failed acquisition the lock can be
 575                          * either held in exclusive mode or shared mode
 576                          * (for the writer starvation avoidance technique).
 577                          */
 578                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 579                             LK_HOLDER(x) != LK_KERNPROC) {
 580                                 owner = (struct thread *)LK_HOLDER(x);
 581                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 582                                         CTR3(KTR_LOCK,
 583                                             "%s: spinning on %p held by %p",
 584                                             __func__, lk, owner);
 585                                 KTR_STATE1(KTR_SCHED, "thread",
 586                                     sched_tdname(td), "spinning",
 587                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 588
 589                                 /*
 590                                  * If we are holding also an interlock drop it
 591                                  * in order to avoid a deadlock if the lockmgr
 592                                  * owner is adaptively spinning on the
 593                                  * interlock itself.
 594                                  */
 595                                 if (flags & LK_INTERLOCK) {
 596                                         class->lc_unlock(ilk);
 597                                         flags &= ~LK_INTERLOCK;
 598                                 }
 599                                 GIANT_SAVE();
 600                                 while (LK_HOLDER(lk->lk_lock) ==
 601                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 602                                         cpu_spinwait();
 603                                 KTR_STATE0(KTR_SCHED, "thread",
 604                                     sched_tdname(td), "running");
 605                                 GIANT_RESTORE();
 606                                 continue;
 607                         } else if (LK_CAN_ADAPT(lk, flags) &&
 608                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 609                             spintries < alk_retries) {
 610                                 KTR_STATE1(KTR_SCHED, "thread",
 611                                     sched_tdname(td), "spinning",
 612                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 613                                 if (flags & LK_INTERLOCK) {
 614                                         class->lc_unlock(ilk);
 615                                         flags &= ~LK_INTERLOCK;
 616                                 }
 617                                 GIANT_SAVE();
 618                                 spintries++;
 619                                 for (i = 0; i < alk_loops; i++) {
 620                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 621                                                 CTR4(KTR_LOCK,
 622                                     "%s: shared spinning on %p with %u and %u",
 623                                                     __func__, lk, spintries, i);
 624                                         x = lk->lk_lock;
 625                                         if ((x & LK_SHARE) == 0 ||
 626                                             LK_CAN_SHARE(x, flags) != 0)
 627                                                 break;
 628                                         cpu_spinwait();
 629                                 }
 630                                 KTR_STATE0(KTR_SCHED, "thread",
 631                                     sched_tdname(td), "running");
 632                                 GIANT_RESTORE();
 633                                 if (i != alk_loops)
 634                                         continue;
 635                         }
 636 #endif
 637
 638                         /*
 639                          * Acquire the sleepqueue chain lock because we
 640                          * probabilly will need to manipulate waiters flags.
 641                          */
 642                         sleepq_lock(&lk->lock_object);
 643                         x = lk->lk_lock;
 644
 645                         /*
 646                          * if the lock can be acquired in shared mode, try
 647                          * again.
 648                          */
 649                         if (LK_CAN_SHARE(x, flags)) {
 650                                 sleepq_release(&lk->lock_object);
 651                                 continue;
 652                         }
 653
 654 #ifdef ADAPTIVE_LOCKMGRS
 655                         /*
 656                          * The current lock owner might have started executing
 657                          * on another CPU (or the lock could have changed
 658                          * owner) while we were waiting on the turnstile
 659                          * chain lock.  If so, drop the turnstile lock and try
 660                          * again.
 661                          */
 662                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 663                             LK_HOLDER(x) != LK_KERNPROC) {
 664                                 owner = (struct thread *)LK_HOLDER(x);
 665                                 if (TD_IS_RUNNING(owner)) {
 666                                         sleepq_release(&lk->lock_object);
 667                                         continue;
 668                                 }
 669                         }
 670 #endif
 671
 672                         /*
 673                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 674                          * loop back and retry.
 675                          */
 676                         if ((x & LK_SHARED_WAITERS) == 0) {
 677                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 678                                     x | LK_SHARED_WAITERS)) {
 679                                         sleepq_release(&lk->lock_object);
 680                                         continue;
 681                                 }
 682                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 683                                     __func__, lk);
 684                         }
 685
 686                         /*
 687                          * As far as we have been unable to acquire the
 688                          * shared lock and the shared waiters flag is set,
 689                          * we will sleep.
 690                          */
 691                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 692                             SQ_SHARED_QUEUE);
 693                         flags &= ~LK_INTERLOCK;
 694                         if (error) {
 695                                 LOCK_LOG3(lk,
 696                                     "%s: interrupted sleep for %p with %d",
 697                                     __func__, lk, error);
 698                                 break;
 699                         }
 700                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 701                             __func__, lk);
 702                 }
 703                 if (error == 0) {
 704                         lock_profile_obtain_lock_success(&lk->lock_object,
 705                             contested, waittime, file, line);
 706                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 707                             line);
 708                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 709                             line);
 710                         TD_LOCKS_INC(curthread);
 711                         TD_SLOCKS_INC(curthread);
 712                         STACK_SAVE(lk);
 713                 }
 714                 break;
 715         case LK_UPGRADE:
 716         case LK_TRYUPGRADE:
 717                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 718                 v = lk->lk_lock;
 719                 x = v & LK_ALL_WAITERS;
 720                 v &= LK_EXCLUSIVE_SPINNERS;
 721
 722                 /*
 723                  * Try to switch from one shared lock to an exclusive one.
 724                  * We need to preserve waiters flags during the operation.
 725                  */
 726                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 727                     tid | x)) {
 728                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 729                             line);
 730                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 731                             LK_TRYWIT(flags), file, line);
 732                         TD_SLOCKS_DEC(curthread);
 733                         break;
 734                 }
 735
 736                 /*
 737                  * In LK_TRYUPGRADE mode, do not drop the lock,
 738                  * returning EBUSY instead.
 739                  */
 740                 if (op == LK_TRYUPGRADE) {
 741                         LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
 742                             __func__, lk);
 743                         error = EBUSY;
 744                         break;
 745                 }
 746
 747                 /*
 748                  * We have been unable to succeed in upgrading, so just
 749                  * give up the shared lock.
 750                  */
 751                 wakeup_swapper |= wakeupshlk(lk, file, line);
 752
 753                 /* FALLTHROUGH */
 754         case LK_EXCLUSIVE:
 755                 if (LK_CAN_WITNESS(flags))
 756                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 757                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 758                             ilk : NULL);
 759
 760                 /*
 761                  * If curthread already holds the lock and this one is
 762                  * allowed to recurse, simply recurse on it.
 763                  */
 764                 if (lockmgr_xlocked(lk)) {
 765                         if ((flags & LK_CANRECURSE) == 0 &&
 766                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 767
 768                                 /*
 769                                  * If the lock is expected to not panic just
 770                                  * give up and return.
 771                                  */
 772                                 if (LK_TRYOP(flags)) {
 773                                         LOCK_LOG2(lk,
 774                                             "%s: %p fails the try operation",
 775                                             __func__, lk);
 776                                         error = EBUSY;
 777                                         break;
 778                                 }
 779                                 if (flags & LK_INTERLOCK)
 780                                         class->lc_unlock(ilk);
 781                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 782                                     __func__, iwmesg, file, line);
 783                         }
 784                         lk->lk_recurse++;
 785                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 786                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 787                             lk->lk_recurse, file, line);
 788                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 789                             LK_TRYWIT(flags), file, line);
 790                         TD_LOCKS_INC(curthread);
 791                         break;
 792                 }
 793
 794                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
 795                     tid)) {
 796 #ifdef HWPMC_HOOKS
 797                         PMC_SOFT_CALL( , , lock, failed);
 798 #endif
 799                         lock_profile_obtain_lock_failed(&lk->lock_object,
 800                             &contested, &waittime);
 801
 802                         /*
 803                          * If the lock is expected to not sleep just give up
 804                          * and return.
 805                          */
 806                         if (LK_TRYOP(flags)) {
 807                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 808                                     __func__, lk);
 809                                 error = EBUSY;
 810                                 break;
 811                         }
 812
 813 #ifdef ADAPTIVE_LOCKMGRS
 814                         /*
 815                          * If the owner is running on another CPU, spin until
 816                          * the owner stops running or the state of the lock
 817                          * changes.
 818                          */
 819                         x = lk->lk_lock;
 820                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 821                             LK_HOLDER(x) != LK_KERNPROC) {
 822                                 owner = (struct thread *)LK_HOLDER(x);
 823                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 824                                         CTR3(KTR_LOCK,
 825                                             "%s: spinning on %p held by %p",
 826                                             __func__, lk, owner);
 827                                 KTR_STATE1(KTR_SCHED, "thread",
 828                                     sched_tdname(td), "spinning",
 829                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 830
 831                                 /*
 832                                  * If we are holding also an interlock drop it
 833                                  * in order to avoid a deadlock if the lockmgr
 834                                  * owner is adaptively spinning on the
 835                                  * interlock itself.
 836                                  */
 837                                 if (flags & LK_INTERLOCK) {
 838                                         class->lc_unlock(ilk);
 839                                         flags &= ~LK_INTERLOCK;
 840                                 }
 841                                 GIANT_SAVE();
 842                                 while (LK_HOLDER(lk->lk_lock) ==
 843                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 844                                         cpu_spinwait();
 845                                 KTR_STATE0(KTR_SCHED, "thread",
 846                                     sched_tdname(td), "running");
 847                                 GIANT_RESTORE();
 848                                 continue;
 849                         } else if (LK_CAN_ADAPT(lk, flags) &&
 850                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 851                             spintries < alk_retries) {
 852                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 853                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 854                                     x | LK_EXCLUSIVE_SPINNERS))
 855                                         continue;
 856                                 KTR_STATE1(KTR_SCHED, "thread",
 857                                     sched_tdname(td), "spinning",
 858                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 859                                 if (flags & LK_INTERLOCK) {
 860                                         class->lc_unlock(ilk);
 861                                         flags &= ~LK_INTERLOCK;
 862                                 }
 863                                 GIANT_SAVE();
 864                                 spintries++;
 865                                 for (i = 0; i < alk_loops; i++) {
 866                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 867                                                 CTR4(KTR_LOCK,
 868                                     "%s: shared spinning on %p with %u and %u",
 869                                                     __func__, lk, spintries, i);
 870                                         if ((lk->lk_lock &
 871                                             LK_EXCLUSIVE_SPINNERS) == 0)
 872                                                 break;
 873                                         cpu_spinwait();
 874                                 }
 875                                 KTR_STATE0(KTR_SCHED, "thread",
 876                                     sched_tdname(td), "running");
 877                                 GIANT_RESTORE();
 878                                 if (i != alk_loops)
 879                                         continue;
 880                         }
 881 #endif
 882
 883                         /*
 884                          * Acquire the sleepqueue chain lock because we
 885                          * probabilly will need to manipulate waiters flags.
 886                          */
 887                         sleepq_lock(&lk->lock_object);
 888                         x = lk->lk_lock;
 889
 890                         /*
 891                          * if the lock has been released while we spun on
 892                          * the sleepqueue chain lock just try again.
 893                          */
 894                         if (x == LK_UNLOCKED) {
 895                                 sleepq_release(&lk->lock_object);
 896                                 continue;
 897                         }
 898
 899 #ifdef ADAPTIVE_LOCKMGRS
 900                         /*
 901                          * The current lock owner might have started executing
 902                          * on another CPU (or the lock could have changed
 903                          * owner) while we were waiting on the turnstile
 904                          * chain lock.  If so, drop the turnstile lock and try
 905                          * again.
 906                          */
 907                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 908                             LK_HOLDER(x) != LK_KERNPROC) {
 909                                 owner = (struct thread *)LK_HOLDER(x);
 910                                 if (TD_IS_RUNNING(owner)) {
 911                                         sleepq_release(&lk->lock_object);
 912                                         continue;
 913                                 }
 914                         }
 915 #endif
 916
 917                         /*
 918                          * The lock can be in the state where there is a
 919                          * pending queue of waiters, but still no owner.
 920                          * This happens when the lock is contested and an
 921                          * owner is going to claim the lock.
 922                          * If curthread is the one successfully acquiring it
 923                          * claim lock ownership and return, preserving waiters
 924                          * flags.
 925                          */
 926                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 927                         if ((x & ~v) == LK_UNLOCKED) {
 928                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 929                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 930                                     tid | v)) {
 931                                         sleepq_release(&lk->lock_object);
 932                                         LOCK_LOG2(lk,
 933                                             "%s: %p claimed by a new writer",
 934                                             __func__, lk);
 935                                         break;
 936                                 }
 937                                 sleepq_release(&lk->lock_object);
 938                                 continue;
 939                         }
 940
 941                         /*
 942                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 943                          * fail, loop back and retry.
 944                          */
 945                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 946                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 947                                     x | LK_EXCLUSIVE_WAITERS)) {
 948                                         sleepq_release(&lk->lock_object);
 949                                         continue;
 950                                 }
 951                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 952                                     __func__, lk);
 953                         }
 954
 955                         /*
 956                          * As far as we have been unable to acquire the
 957                          * exclusive lock and the exclusive waiters flag
 958                          * is set, we will sleep.
 959                          */
 960                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 961                             SQ_EXCLUSIVE_QUEUE);
 962                         flags &= ~LK_INTERLOCK;
 963                         if (error) {
 964                                 LOCK_LOG3(lk,
 965                                     "%s: interrupted sleep for %p with %d",
 966                                     __func__, lk, error);
 967                                 break;
 968                         }
 969                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 970                             __func__, lk);
 971                 }
 972                 if (error == 0) {
 973                         lock_profile_obtain_lock_success(&lk->lock_object,
 974                             contested, waittime, file, line);
 975                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 976                             lk->lk_recurse, file, line);
 977                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 978                             LK_TRYWIT(flags), file, line);
 979                         TD_LOCKS_INC(curthread);
 980                         STACK_SAVE(lk);
 981                 }
 982                 break;
 983         case LK_DOWNGRADE:
 984                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 985                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 986                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 987
 988                 /*
 989                  * Panic if the lock is recursed.
 990                  */
 991                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 992                         if (flags & LK_INTERLOCK)
 993                                 class->lc_unlock(ilk);
 994                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 995                             __func__, iwmesg, file, line);
 996                 }
 997                 TD_SLOCKS_INC(curthread);
 998
 999                 /*
1000                  * In order to preserve waiters flags, just spin.
1001                  */
1002                 for (;;) {
1003                         x = lk->lk_lock;
1004                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1005                         x &= LK_ALL_WAITERS;
1006                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1007                             LK_SHARERS_LOCK(1) | x))
1008                                 break;
1009                         cpu_spinwait();
1010                 }
1011                 break;
1012         case LK_RELEASE:
1013                 _lockmgr_assert(lk, KA_LOCKED, file, line);
1014                 x = lk->lk_lock;
1015
1016                 if ((x & LK_SHARE) == 0) {
1017
1018                         /*
1019                          * As first option, treact the lock as if it has not
1020                          * any waiter.
1021                          * Fix-up the tid var if the lock has been disowned.
1022                          */
1023                         if (LK_HOLDER(x) == LK_KERNPROC)
1024                                 tid = LK_KERNPROC;
1025                         else {
1026                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
1027                                     file, line);
1028                                 TD_LOCKS_DEC(curthread);
1029                         }
1030                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
1031                             lk->lk_recurse, file, line);
1032
1033                         /*
1034                          * The lock is held in exclusive mode.
1035                          * If the lock is recursed also, then unrecurse it.
1036                          */
1037                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1038                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1039                                     lk);
1040                                 lk->lk_recurse--;
1041                                 break;
1042                         }
1043                         if (tid != LK_KERNPROC)
1044                                 lock_profile_release_lock(&lk->lock_object);
1045
1046                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1047                             LK_UNLOCKED))
1048                                 break;
1049
1050                         sleepq_lock(&lk->lock_object);
1051                         x = lk->lk_lock;
1052                         v = LK_UNLOCKED;
1053
1054                         /*
1055                          * If the lock has exclusive waiters, give them
1056                          * preference in order to avoid deadlock with
1057                          * shared runners up.
1058                          * If interruptible sleeps left the exclusive queue
1059                          * empty avoid a starvation for the threads sleeping
1060                          * on the shared queue by giving them precedence
1061                          * and cleaning up the exclusive waiters bit anyway.
1062                          * Please note that lk_exslpfail count may be lying
1063                          * about the real number of waiters with the
1064                          * LK_SLEEPFAIL flag on because they may be used in
1065                          * conjuction with interruptible sleeps so
1066                          * lk_exslpfail might be considered an 'upper limit'
1067                          * bound, including the edge cases.
1068                          */
1069                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1070                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1071                             SQ_EXCLUSIVE_QUEUE);
1072                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1073                                 if (lk->lk_exslpfail < realexslp) {
1074                                         lk->lk_exslpfail = 0;
1075                                         queue = SQ_EXCLUSIVE_QUEUE;
1076                                         v |= (x & LK_SHARED_WAITERS);
1077                                 } else {
1078                                         lk->lk_exslpfail = 0;
1079                                         LOCK_LOG2(lk,
1080                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1081                                             __func__, lk);
1082                                         LOCK_LOG2(lk,
1083                         "%s: %p waking up threads on the exclusive queue",
1084                                             __func__, lk);
1085                                         wakeup_swapper =
1086                                             sleepq_broadcast(&lk->lock_object,
1087                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1088                                         queue = SQ_SHARED_QUEUE;
1089                                 }
1090                         } else {
1091
1092                                 /*
1093                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1094                                  * on and using interruptible sleeps/timeout
1095                                  * may have left spourious lk_exslpfail counts
1096                                  * on, so clean it up anyway.
1097                                  */
1098                                 lk->lk_exslpfail = 0;
1099                                 queue = SQ_SHARED_QUEUE;
1100                         }
1101
1102                         LOCK_LOG3(lk,
1103                             "%s: %p waking up threads on the %s queue",
1104                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1105                             "exclusive");
1106                         atomic_store_rel_ptr(&lk->lk_lock, v);
1107                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1108                             SLEEPQ_LK, 0, queue);
1109                         sleepq_release(&lk->lock_object);
1110                         break;
1111                 } else
1112                         wakeup_swapper = wakeupshlk(lk, file, line);
1113                 break;
1114         case LK_DRAIN:
1115                 if (LK_CAN_WITNESS(flags))
1116                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1117                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1118                             ilk : NULL);
1119
1120                 /*
1121                  * Trying to drain a lock we already own will result in a
1122                  * deadlock.
1123                  */
1124                 if (lockmgr_xlocked(lk)) {
1125                         if (flags & LK_INTERLOCK)
1126                                 class->lc_unlock(ilk);
1127                         panic("%s: draining %s with the lock held @ %s:%d\n",
1128                             __func__, iwmesg, file, line);
1129                 }
1130
1131                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1132 #ifdef HWPMC_HOOKS
1133                         PMC_SOFT_CALL( , , lock, failed);
1134 #endif
1135                         lock_profile_obtain_lock_failed(&lk->lock_object,
1136                             &contested, &waittime);
1137
1138                         /*
1139                          * If the lock is expected to not sleep just give up
1140                          * and return.
1141                          */
1142                         if (LK_TRYOP(flags)) {
1143                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1144                                     __func__, lk);
1145                                 error = EBUSY;
1146                                 break;
1147                         }
1148
1149                         /*
1150                          * Acquire the sleepqueue chain lock because we
1151                          * probabilly will need to manipulate waiters flags.
1152                          */
1153                         sleepq_lock(&lk->lock_object);
1154                         x = lk->lk_lock;
1155
1156                         /*
1157                          * if the lock has been released while we spun on
1158                          * the sleepqueue chain lock just try again.
1159                          */
1160                         if (x == LK_UNLOCKED) {
1161                                 sleepq_release(&lk->lock_object);
1162                                 continue;
1163                         }
1164
1165                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1166                         if ((x & ~v) == LK_UNLOCKED) {
1167                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1168
1169                                 /*
1170                                  * If interruptible sleeps left the exclusive
1171                                  * queue empty avoid a starvation for the
1172                                  * threads sleeping on the shared queue by
1173                                  * giving them precedence and cleaning up the
1174                                  * exclusive waiters bit anyway.
1175                                  * Please note that lk_exslpfail count may be
1176                                  * lying about the real number of waiters with
1177                                  * the LK_SLEEPFAIL flag on because they may
1178                                  * be used in conjuction with interruptible
1179                                  * sleeps so lk_exslpfail might be considered
1180                                  * an 'upper limit' bound, including the edge
1181                                  * cases.
1182                                  */
1183                                 if (v & LK_EXCLUSIVE_WAITERS) {
1184                                         queue = SQ_EXCLUSIVE_QUEUE;
1185                                         v &= ~LK_EXCLUSIVE_WAITERS;
1186                                 } else {
1187
1188                                         /*
1189                                          * Exclusive waiters sleeping with
1190                                          * LK_SLEEPFAIL on and using
1191                                          * interruptible sleeps/timeout may
1192                                          * have left spourious lk_exslpfail
1193                                          * counts on, so clean it up anyway.
1194                                          */
1195                                         MPASS(v & LK_SHARED_WAITERS);
1196                                         lk->lk_exslpfail = 0;
1197                                         queue = SQ_SHARED_QUEUE;
1198                                         v &= ~LK_SHARED_WAITERS;
1199                                 }
1200                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1201                                         realexslp =
1202                                             sleepq_sleepcnt(&lk->lock_object,
1203                                             SQ_EXCLUSIVE_QUEUE);
1204                                         if (lk->lk_exslpfail >= realexslp) {
1205                                                 lk->lk_exslpfail = 0;
1206                                                 queue = SQ_SHARED_QUEUE;
1207                                                 v &= ~LK_SHARED_WAITERS;
1208                                                 if (realexslp != 0) {
1209                                                         LOCK_LOG2(lk,
1210                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1211                                                             __func__, lk);
1212                                                         LOCK_LOG2(lk,
1213                         "%s: %p waking up threads on the exclusive queue",
1214                                                             __func__, lk);
1215                                                         wakeup_swapper =
1216                                                             sleepq_broadcast(
1217                                                             &lk->lock_object,
1218                                                             SLEEPQ_LK, 0,
1219                                                             SQ_EXCLUSIVE_QUEUE);
1220                                                 }
1221                                         } else
1222                                                 lk->lk_exslpfail = 0;
1223                                 }
1224                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1225                                         sleepq_release(&lk->lock_object);
1226                                         continue;
1227                                 }
1228                                 LOCK_LOG3(lk,
1229                                 "%s: %p waking up all threads on the %s queue",
1230                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1231                                     "shared" : "exclusive");
1232                                 wakeup_swapper |= sleepq_broadcast(
1233                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1234
1235                                 /*
1236                                  * If shared waiters have been woken up we need
1237                                  * to wait for one of them to acquire the lock
1238                                  * before to set the exclusive waiters in
1239                                  * order to avoid a deadlock.
1240                                  */
1241                                 if (queue == SQ_SHARED_QUEUE) {
1242                                         for (v = lk->lk_lock;
1243                                             (v & LK_SHARE) && !LK_SHARERS(v);
1244                                             v = lk->lk_lock)
1245                                                 cpu_spinwait();
1246                                 }
1247                         }
1248
1249                         /*
1250                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1251                          * fail, loop back and retry.
1252                          */
1253                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1254                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1255                                     x | LK_EXCLUSIVE_WAITERS)) {
1256                                         sleepq_release(&lk->lock_object);
1257                                         continue;
1258                                 }
1259                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1260                                     __func__, lk);
1261                         }
1262
1263                         /*
1264                          * As far as we have been unable to acquire the
1265                          * exclusive lock and the exclusive waiters flag
1266                          * is set, we will sleep.
1267                          */
1268                         if (flags & LK_INTERLOCK) {
1269                                 class->lc_unlock(ilk);
1270                                 flags &= ~LK_INTERLOCK;
1271                         }
1272                         GIANT_SAVE();
1273                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1274                             SQ_EXCLUSIVE_QUEUE);
1275                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1276                         GIANT_RESTORE();
1277                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1278                             __func__, lk);
1279                 }
1280
1281                 if (error == 0) {
1282                         lock_profile_obtain_lock_success(&lk->lock_object,
1283                             contested, waittime, file, line);
1284                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1285                             lk->lk_recurse, file, line);
1286                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1287                             LK_TRYWIT(flags), file, line);
1288                         TD_LOCKS_INC(curthread);
1289                         STACK_SAVE(lk);
1290                 }
1291                 break;
1292         default:
1293                 if (flags & LK_INTERLOCK)
1294                         class->lc_unlock(ilk);
1295                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1296         }
1297
1298         if (flags & LK_INTERLOCK)
1299                 class->lc_unlock(ilk);
1300         if (wakeup_swapper)
1301                 kick_proc0();
1302
1303         return (error);
1304 }
1305
1306 void
1307 _lockmgr_disown(struct lock *lk, const char *file, int line)
1308 {
1309         uintptr_t tid, x;
1310
1311         if (SCHEDULER_STOPPED())
1312                 return;
1313
1314         tid = (uintptr_t)curthread;
1315         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1316
1317         /*
1318          * Panic if the lock is recursed.
1319          */
1320         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1321                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1322                     __func__,  file, line);
1323
1324         /*
1325          * If the owner is already LK_KERNPROC just skip the whole operation.
1326          */
1327         if (LK_HOLDER(lk->lk_lock) != tid)
1328                 return;
1329         lock_profile_release_lock(&lk->lock_object);
1330         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1331         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1332         TD_LOCKS_DEC(curthread);
1333         STACK_SAVE(lk);
1334
1335         /*
1336          * In order to preserve waiters flags, just spin.
1337          */
1338         for (;;) {
1339                 x = lk->lk_lock;
1340                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1341                 x &= LK_ALL_WAITERS;
1342                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1343                     LK_KERNPROC | x))
1344                         return;
1345                 cpu_spinwait();
1346         }
1347 }
1348
1349 void
1350 lockmgr_printinfo(const struct lock *lk)
1351 {
1352         struct thread *td;
1353         uintptr_t x;
1354
1355         if (lk->lk_lock == LK_UNLOCKED)
1356                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1357         else if (lk->lk_lock & LK_SHARE)
1358                 printf("lock type %s: SHARED (count %ju)\n",
1359                     lk->lock_object.lo_name,
1360                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1361         else {
1362                 td = lockmgr_xholder(lk);
1363                 if (td == (struct thread *)LK_KERNPROC)
1364                         printf("lock type %s: EXCL by KERNPROC\n",
1365                             lk->lock_object.lo_name);
1366                 else
1367                         printf("lock type %s: EXCL by thread %p "
1368                             "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1369                             td, td->td_proc->p_pid, td->td_proc->p_comm,
1370                             td->td_tid);
1371         }
1372
1373         x = lk->lk_lock;
1374         if (x & LK_EXCLUSIVE_WAITERS)
1375                 printf(" with exclusive waiters pending\n");
1376         if (x & LK_SHARED_WAITERS)
1377                 printf(" with shared waiters pending\n");
1378         if (x & LK_EXCLUSIVE_SPINNERS)
1379                 printf(" with exclusive spinners pending\n");
1380
1381         STACK_PRINT(lk);
1382 }
1383
1384 int
1385 lockstatus(const struct lock *lk)
1386 {
1387         uintptr_t v, x;
1388         int ret;
1389
1390         ret = LK_SHARED;
1391         x = lk->lk_lock;
1392         v = LK_HOLDER(x);
1393
1394         if ((x & LK_SHARE) == 0) {
1395                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1396                         ret = LK_EXCLUSIVE;
1397                 else
1398                         ret = LK_EXCLOTHER;
1399         } else if (x == LK_UNLOCKED)
1400                 ret = 0;
1401
1402         return (ret);
1403 }
1404
1405 #ifdef INVARIANT_SUPPORT
1406
1407 FEATURE(invariant_support,
1408     "Support for modules compiled with INVARIANTS option");
1409
1410 #ifndef INVARIANTS
1411 #undef  _lockmgr_assert
1412 #endif
1413
1414 void
1415 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1416 {
1417         int slocked = 0;
1418
1419         if (panicstr != NULL)
1420                 return;
1421         switch (what) {
1422         case KA_SLOCKED:
1423         case KA_SLOCKED | KA_NOTRECURSED:
1424         case KA_SLOCKED | KA_RECURSED:
1425                 slocked = 1;
1426         case KA_LOCKED:
1427         case KA_LOCKED | KA_NOTRECURSED:
1428         case KA_LOCKED | KA_RECURSED:
1429 #ifdef WITNESS
1430
1431                 /*
1432                  * We cannot trust WITNESS if the lock is held in exclusive
1433                  * mode and a call to lockmgr_disown() happened.
1434                  * Workaround this skipping the check if the lock is held in
1435                  * exclusive mode even for the KA_LOCKED case.
1436                  */
1437                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1438                         witness_assert(&lk->lock_object, what, file, line);
1439                         break;
1440                 }
1441 #endif
1442                 if (lk->lk_lock == LK_UNLOCKED ||
1443                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1444                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1445                         panic("Lock %s not %slocked @ %s:%d\n",
1446                             lk->lock_object.lo_name, slocked ? "share" : "",
1447                             file, line);
1448
1449                 if ((lk->lk_lock & LK_SHARE) == 0) {
1450                         if (lockmgr_recursed(lk)) {
1451                                 if (what & KA_NOTRECURSED)
1452                                         panic("Lock %s recursed @ %s:%d\n",
1453                                             lk->lock_object.lo_name, file,
1454                                             line);
1455                         } else if (what & KA_RECURSED)
1456                                 panic("Lock %s not recursed @ %s:%d\n",
1457                                     lk->lock_object.lo_name, file, line);
1458                 }
1459                 break;
1460         case KA_XLOCKED:
1461         case KA_XLOCKED | KA_NOTRECURSED:
1462         case KA_XLOCKED | KA_RECURSED:
1463                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1464                         panic("Lock %s not exclusively locked @ %s:%d\n",
1465                             lk->lock_object.lo_name, file, line);
1466                 if (lockmgr_recursed(lk)) {
1467                         if (what & KA_NOTRECURSED)
1468                                 panic("Lock %s recursed @ %s:%d\n",
1469                                     lk->lock_object.lo_name, file, line);
1470                 } else if (what & KA_RECURSED)
1471                         panic("Lock %s not recursed @ %s:%d\n",
1472                             lk->lock_object.lo_name, file, line);
1473                 break;
1474         case KA_UNLOCKED:
1475                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1476                         panic("Lock %s exclusively locked @ %s:%d\n",
1477                             lk->lock_object.lo_name, file, line);
1478                 break;
1479         default:
1480                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1481                     line);
1482         }
1483 }
1484 #endif
1485
1486 #ifdef DDB
1487 int
1488 lockmgr_chain(struct thread *td, struct thread **ownerp)
1489 {
1490         struct lock *lk;
1491
1492         lk = td->td_wchan;
1493
1494         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1495                 return (0);
1496         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1497         if (lk->lk_lock & LK_SHARE)
1498                 db_printf("SHARED (count %ju)\n",
1499                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1500         else
1501                 db_printf("EXCL\n");
1502         *ownerp = lockmgr_xholder(lk);
1503
1504         return (1);
1505 }
1506
1507 static void
1508 db_show_lockmgr(const struct lock_object *lock)
1509 {
1510         struct thread *td;
1511         const struct lock *lk;
1512
1513         lk = (const struct lock *)lock;
1514
1515         db_printf(" state: ");
1516         if (lk->lk_lock == LK_UNLOCKED)
1517                 db_printf("UNLOCKED\n");
1518         else if (lk->lk_lock & LK_SHARE)
1519                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1520         else {
1521                 td = lockmgr_xholder(lk);
1522                 if (td == (struct thread *)LK_KERNPROC)
1523                         db_printf("XLOCK: LK_KERNPROC\n");
1524                 else
1525                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1526                             td->td_tid, td->td_proc->p_pid,
1527                             td->td_proc->p_comm);
1528                 if (lockmgr_recursed(lk))
1529                         db_printf(" recursed: %d\n", lk->lk_recurse);
1530         }
1531         db_printf(" waiters: ");
1532         switch (lk->lk_lock & LK_ALL_WAITERS) {
1533         case LK_SHARED_WAITERS:
1534                 db_printf("shared\n");
1535                 break;
1536         case LK_EXCLUSIVE_WAITERS:
1537                 db_printf("exclusive\n");
1538                 break;
1539         case LK_ALL_WAITERS:
1540                 db_printf("shared and exclusive\n");
1541                 break;
1542         default:
1543                 db_printf("none\n");
1544         }
1545         db_printf(" spinners: ");
1546         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1547                 db_printf("exclusive\n");
1548         else
1549                 db_printf("none\n");
1550 }
1551 #endif