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 #include "opt_kdtrace.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 #define TD_LOCKS_INC(td)
  74 #define TD_LOCKS_DEC(td)
  75 #else
  76 #define TD_LOCKS_INC(td)        ((td)->td_locks++)
  77 #define TD_LOCKS_DEC(td)        ((td)->td_locks--)
  78 #endif
  79 #define TD_SLOCKS_INC(td)       ((td)->td_lk_slocks++)
  80 #define TD_SLOCKS_DEC(td)       ((td)->td_lk_slocks--)
  81
  82 #ifndef DEBUG_LOCKS
  83 #define STACK_PRINT(lk)
  84 #define STACK_SAVE(lk)
  85 #define STACK_ZERO(lk)
  86 #else
  87 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
  88 #define STACK_SAVE(lk)  stack_save(&(lk)->lk_stack)
  89 #define STACK_ZERO(lk)  stack_zero(&(lk)->lk_stack)
  90 #endif
  91
  92 #define LOCK_LOG2(lk, string, arg1, arg2)                               \
  93         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  94                 CTR2(KTR_LOCK, (string), (arg1), (arg2))
  95 #define LOCK_LOG3(lk, string, arg1, arg2, arg3)                         \
  96         if (LOCK_LOG_TEST(&(lk)->lock_object, 0))                       \
  97                 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
  98
  99 #define GIANT_DECLARE                                                   \
 100         int _i = 0;                                                     \
 101         WITNESS_SAVE_DECL(Giant)
 102 #define GIANT_RESTORE() do {                                            \
 103         if (_i > 0) {                                                   \
 104                 while (_i--)                                            \
 105                         mtx_lock(&Giant);                               \
 106                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 107         }                                                               \
 108 } while (0)
 109 #define GIANT_SAVE() do {                                               \
 110         if (mtx_owned(&Giant)) {                                        \
 111                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
 112                 while (mtx_owned(&Giant)) {                             \
 113                         _i++;                                           \
 114                         mtx_unlock(&Giant);                             \
 115                 }                                                       \
 116         }                                                               \
 117 } while (0)
 118
 119 #define LK_CAN_SHARE(x)                                                 \
 120         (((x) & LK_SHARE) && (((x) & LK_EXCLUSIVE_WAITERS) == 0 ||      \
 121         ((x) & LK_EXCLUSIVE_SPINNERS) == 0 ||                           \
 122         curthread->td_lk_slocks || (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, int how);
 146 #ifdef KDTRACE_HOOKS
 147 static int      owner_lockmgr(const struct lock_object *lock,
 148                     struct thread **owner);
 149 #endif
 150 static int      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, int how)
 354 {
 355
 356         panic("lockmgr locks do not support sleep interlocking");
 357 }
 358
 359 static int
 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 lockallowrecurse(struct lock *lk)
 423 {
 424
 425         lockmgr_assert(lk, KA_XLOCKED);
 426         lk->lock_object.lo_flags |= LO_RECURSABLE;
 427 }
 428
 429 void
 430 lockdisablerecurse(struct lock *lk)
 431 {
 432
 433         lockmgr_assert(lk, KA_XLOCKED);
 434         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 435 }
 436
 437 void
 438 lockdestroy(struct lock *lk)
 439 {
 440
 441         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 442         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 443         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 444         lock_destroy(&lk->lock_object);
 445 }
 446
 447 int
 448 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 449     const char *wmesg, int pri, int timo, const char *file, int line)
 450 {
 451         GIANT_DECLARE;
 452         struct lock_class *class;
 453         const char *iwmesg;
 454         uintptr_t tid, v, x;
 455         u_int op, realexslp;
 456         int error, ipri, itimo, queue, wakeup_swapper;
 457 #ifdef LOCK_PROFILING
 458         uint64_t waittime = 0;
 459         int contested = 0;
 460 #endif
 461 #ifdef ADAPTIVE_LOCKMGRS
 462         volatile struct thread *owner;
 463         u_int i, spintries = 0;
 464 #endif
 465
 466         error = 0;
 467         tid = (uintptr_t)curthread;
 468         op = (flags & LK_TYPE_MASK);
 469         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 470         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 471         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 472
 473         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 474         KASSERT((op & (op - 1)) == 0,
 475             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 476         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 477             (op != LK_DOWNGRADE && op != LK_RELEASE),
 478             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 479             __func__, file, line));
 480         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 481             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 482             __func__, file, line));
 483         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 484             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
 485             lk->lock_object.lo_name, file, line));
 486
 487         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 488         if (panicstr != NULL) {
 489                 if (flags & LK_INTERLOCK)
 490                         class->lc_unlock(ilk);
 491                 return (0);
 492         }
 493
 494         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 495                 switch (op) {
 496                 case LK_SHARED:
 497                         op = LK_EXCLUSIVE;
 498                         break;
 499                 case LK_UPGRADE:
 500                 case LK_DOWNGRADE:
 501                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 502                             file, line);
 503                         if (flags & LK_INTERLOCK)
 504                                 class->lc_unlock(ilk);
 505                         return (0);
 506                 }
 507         }
 508
 509         wakeup_swapper = 0;
 510         switch (op) {
 511         case LK_SHARED:
 512                 if (LK_CAN_WITNESS(flags))
 513                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 514                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 515                 for (;;) {
 516                         x = lk->lk_lock;
 517
 518                         /*
 519                          * If no other thread has an exclusive lock, or
 520                          * no exclusive waiter is present, bump the count of
 521                          * sharers.  Since we have to preserve the state of
 522                          * waiters, if we fail to acquire the shared lock
 523                          * loop back and retry.
 524                          */
 525                         if (LK_CAN_SHARE(x)) {
 526                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 527                                     x + LK_ONE_SHARER))
 528                                         break;
 529                                 continue;
 530                         }
 531 #ifdef HWPMC_HOOKS
 532                         PMC_SOFT_CALL( , , lock, failed);
 533 #endif
 534                         lock_profile_obtain_lock_failed(&lk->lock_object,
 535                             &contested, &waittime);
 536
 537                         /*
 538                          * If the lock is already held by curthread in
 539                          * exclusive way avoid a deadlock.
 540                          */
 541                         if (LK_HOLDER(x) == tid) {
 542                                 LOCK_LOG2(lk,
 543                                     "%s: %p already held in exclusive mode",
 544                                     __func__, lk);
 545                                 error = EDEADLK;
 546                                 break;
 547                         }
 548
 549                         /*
 550                          * If the lock is expected to not sleep just give up
 551                          * and return.
 552                          */
 553                         if (LK_TRYOP(flags)) {
 554                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 555                                     __func__, lk);
 556                                 error = EBUSY;
 557                                 break;
 558                         }
 559
 560 #ifdef ADAPTIVE_LOCKMGRS
 561                         /*
 562                          * If the owner is running on another CPU, spin until
 563                          * the owner stops running or the state of the lock
 564                          * changes.  We need a double-state handle here
 565                          * because for a failed acquisition the lock can be
 566                          * either held in exclusive mode or shared mode
 567                          * (for the writer starvation avoidance technique).
 568                          */
 569                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 570                             LK_HOLDER(x) != LK_KERNPROC) {
 571                                 owner = (struct thread *)LK_HOLDER(x);
 572                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 573                                         CTR3(KTR_LOCK,
 574                                             "%s: spinning on %p held by %p",
 575                                             __func__, lk, owner);
 576
 577                                 /*
 578                                  * If we are holding also an interlock drop it
 579                                  * in order to avoid a deadlock if the lockmgr
 580                                  * owner is adaptively spinning on the
 581                                  * interlock itself.
 582                                  */
 583                                 if (flags & LK_INTERLOCK) {
 584                                         class->lc_unlock(ilk);
 585                                         flags &= ~LK_INTERLOCK;
 586                                 }
 587                                 GIANT_SAVE();
 588                                 while (LK_HOLDER(lk->lk_lock) ==
 589                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 590                                         cpu_spinwait();
 591                                 GIANT_RESTORE();
 592                                 continue;
 593                         } else if (LK_CAN_ADAPT(lk, flags) &&
 594                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 595                             spintries < alk_retries) {
 596                                 if (flags & LK_INTERLOCK) {
 597                                         class->lc_unlock(ilk);
 598                                         flags &= ~LK_INTERLOCK;
 599                                 }
 600                                 GIANT_SAVE();
 601                                 spintries++;
 602                                 for (i = 0; i < alk_loops; i++) {
 603                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 604                                                 CTR4(KTR_LOCK,
 605                                     "%s: shared spinning on %p with %u and %u",
 606                                                     __func__, lk, spintries, i);
 607                                         x = lk->lk_lock;
 608                                         if ((x & LK_SHARE) == 0 ||
 609                                             LK_CAN_SHARE(x) != 0)
 610                                                 break;
 611                                         cpu_spinwait();
 612                                 }
 613                                 GIANT_RESTORE();
 614                                 if (i != alk_loops)
 615                                         continue;
 616                         }
 617 #endif
 618
 619                         /*
 620                          * Acquire the sleepqueue chain lock because we
 621                          * probabilly will need to manipulate waiters flags.
 622                          */
 623                         sleepq_lock(&lk->lock_object);
 624                         x = lk->lk_lock;
 625
 626                         /*
 627                          * if the lock can be acquired in shared mode, try
 628                          * again.
 629                          */
 630                         if (LK_CAN_SHARE(x)) {
 631                                 sleepq_release(&lk->lock_object);
 632                                 continue;
 633                         }
 634
 635 #ifdef ADAPTIVE_LOCKMGRS
 636                         /*
 637                          * The current lock owner might have started executing
 638                          * on another CPU (or the lock could have changed
 639                          * owner) while we were waiting on the turnstile
 640                          * chain lock.  If so, drop the turnstile lock and try
 641                          * again.
 642                          */
 643                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 644                             LK_HOLDER(x) != LK_KERNPROC) {
 645                                 owner = (struct thread *)LK_HOLDER(x);
 646                                 if (TD_IS_RUNNING(owner)) {
 647                                         sleepq_release(&lk->lock_object);
 648                                         continue;
 649                                 }
 650                         }
 651 #endif
 652
 653                         /*
 654                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 655                          * loop back and retry.
 656                          */
 657                         if ((x & LK_SHARED_WAITERS) == 0) {
 658                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 659                                     x | LK_SHARED_WAITERS)) {
 660                                         sleepq_release(&lk->lock_object);
 661                                         continue;
 662                                 }
 663                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 664                                     __func__, lk);
 665                         }
 666
 667                         /*
 668                          * As far as we have been unable to acquire the
 669                          * shared lock and the shared waiters flag is set,
 670                          * we will sleep.
 671                          */
 672                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 673                             SQ_SHARED_QUEUE);
 674                         flags &= ~LK_INTERLOCK;
 675                         if (error) {
 676                                 LOCK_LOG3(lk,
 677                                     "%s: interrupted sleep for %p with %d",
 678                                     __func__, lk, error);
 679                                 break;
 680                         }
 681                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 682                             __func__, lk);
 683                 }
 684                 if (error == 0) {
 685                         lock_profile_obtain_lock_success(&lk->lock_object,
 686                             contested, waittime, file, line);
 687                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 688                             line);
 689                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 690                             line);
 691                         TD_LOCKS_INC(curthread);
 692                         TD_SLOCKS_INC(curthread);
 693                         STACK_SAVE(lk);
 694                 }
 695                 break;
 696         case LK_UPGRADE:
 697                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 698                 v = lk->lk_lock;
 699                 x = v & LK_ALL_WAITERS;
 700                 v &= LK_EXCLUSIVE_SPINNERS;
 701
 702                 /*
 703                  * Try to switch from one shared lock to an exclusive one.
 704                  * We need to preserve waiters flags during the operation.
 705                  */
 706                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 707                     tid | x)) {
 708                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 709                             line);
 710                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 711                             LK_TRYWIT(flags), file, line);
 712                         TD_SLOCKS_DEC(curthread);
 713                         break;
 714                 }
 715
 716                 /*
 717                  * We have been unable to succeed in upgrading, so just
 718                  * give up the shared lock.
 719                  */
 720                 wakeup_swapper |= wakeupshlk(lk, file, line);
 721
 722                 /* FALLTHROUGH */
 723         case LK_EXCLUSIVE:
 724                 if (LK_CAN_WITNESS(flags))
 725                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 726                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 727                             ilk : NULL);
 728
 729                 /*
 730                  * If curthread already holds the lock and this one is
 731                  * allowed to recurse, simply recurse on it.
 732                  */
 733                 if (lockmgr_xlocked(lk)) {
 734                         if ((flags & LK_CANRECURSE) == 0 &&
 735                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 736
 737                                 /*
 738                                  * If the lock is expected to not panic just
 739                                  * give up and return.
 740                                  */
 741                                 if (LK_TRYOP(flags)) {
 742                                         LOCK_LOG2(lk,
 743                                             "%s: %p fails the try operation",
 744                                             __func__, lk);
 745                                         error = EBUSY;
 746                                         break;
 747                                 }
 748                                 if (flags & LK_INTERLOCK)
 749                                         class->lc_unlock(ilk);
 750                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 751                                     __func__, iwmesg, file, line);
 752                         }
 753                         lk->lk_recurse++;
 754                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 755                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 756                             lk->lk_recurse, file, line);
 757                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 758                             LK_TRYWIT(flags), file, line);
 759                         TD_LOCKS_INC(curthread);
 760                         break;
 761                 }
 762
 763                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
 764                     tid)) {
 765 #ifdef HWPMC_HOOKS
 766                         PMC_SOFT_CALL( , , lock, failed);
 767 #endif
 768                         lock_profile_obtain_lock_failed(&lk->lock_object,
 769                             &contested, &waittime);
 770
 771                         /*
 772                          * If the lock is expected to not sleep just give up
 773                          * and return.
 774                          */
 775                         if (LK_TRYOP(flags)) {
 776                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 777                                     __func__, lk);
 778                                 error = EBUSY;
 779                                 break;
 780                         }
 781
 782 #ifdef ADAPTIVE_LOCKMGRS
 783                         /*
 784                          * If the owner is running on another CPU, spin until
 785                          * the owner stops running or the state of the lock
 786                          * changes.
 787                          */
 788                         x = lk->lk_lock;
 789                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 790                             LK_HOLDER(x) != LK_KERNPROC) {
 791                                 owner = (struct thread *)LK_HOLDER(x);
 792                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 793                                         CTR3(KTR_LOCK,
 794                                             "%s: spinning on %p held by %p",
 795                                             __func__, lk, owner);
 796
 797                                 /*
 798                                  * If we are holding also an interlock drop it
 799                                  * in order to avoid a deadlock if the lockmgr
 800                                  * owner is adaptively spinning on the
 801                                  * interlock itself.
 802                                  */
 803                                 if (flags & LK_INTERLOCK) {
 804                                         class->lc_unlock(ilk);
 805                                         flags &= ~LK_INTERLOCK;
 806                                 }
 807                                 GIANT_SAVE();
 808                                 while (LK_HOLDER(lk->lk_lock) ==
 809                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 810                                         cpu_spinwait();
 811                                 GIANT_RESTORE();
 812                                 continue;
 813                         } else if (LK_CAN_ADAPT(lk, flags) &&
 814                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 815                             spintries < alk_retries) {
 816                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 817                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 818                                     x | LK_EXCLUSIVE_SPINNERS))
 819                                         continue;
 820                                 if (flags & LK_INTERLOCK) {
 821                                         class->lc_unlock(ilk);
 822                                         flags &= ~LK_INTERLOCK;
 823                                 }
 824                                 GIANT_SAVE();
 825                                 spintries++;
 826                                 for (i = 0; i < alk_loops; i++) {
 827                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 828                                                 CTR4(KTR_LOCK,
 829                                     "%s: shared spinning on %p with %u and %u",
 830                                                     __func__, lk, spintries, i);
 831                                         if ((lk->lk_lock &
 832                                             LK_EXCLUSIVE_SPINNERS) == 0)
 833                                                 break;
 834                                         cpu_spinwait();
 835                                 }
 836                                 GIANT_RESTORE();
 837                                 if (i != alk_loops)
 838                                         continue;
 839                         }
 840 #endif
 841
 842                         /*
 843                          * Acquire the sleepqueue chain lock because we
 844                          * probabilly will need to manipulate waiters flags.
 845                          */
 846                         sleepq_lock(&lk->lock_object);
 847                         x = lk->lk_lock;
 848
 849                         /*
 850                          * if the lock has been released while we spun on
 851                          * the sleepqueue chain lock just try again.
 852                          */
 853                         if (x == LK_UNLOCKED) {
 854                                 sleepq_release(&lk->lock_object);
 855                                 continue;
 856                         }
 857
 858 #ifdef ADAPTIVE_LOCKMGRS
 859                         /*
 860                          * The current lock owner might have started executing
 861                          * on another CPU (or the lock could have changed
 862                          * owner) while we were waiting on the turnstile
 863                          * chain lock.  If so, drop the turnstile lock and try
 864                          * again.
 865                          */
 866                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 867                             LK_HOLDER(x) != LK_KERNPROC) {
 868                                 owner = (struct thread *)LK_HOLDER(x);
 869                                 if (TD_IS_RUNNING(owner)) {
 870                                         sleepq_release(&lk->lock_object);
 871                                         continue;
 872                                 }
 873                         }
 874 #endif
 875
 876                         /*
 877                          * The lock can be in the state where there is a
 878                          * pending queue of waiters, but still no owner.
 879                          * This happens when the lock is contested and an
 880                          * owner is going to claim the lock.
 881                          * If curthread is the one successfully acquiring it
 882                          * claim lock ownership and return, preserving waiters
 883                          * flags.
 884                          */
 885                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 886                         if ((x & ~v) == LK_UNLOCKED) {
 887                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 888                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 889                                     tid | v)) {
 890                                         sleepq_release(&lk->lock_object);
 891                                         LOCK_LOG2(lk,
 892                                             "%s: %p claimed by a new writer",
 893                                             __func__, lk);
 894                                         break;
 895                                 }
 896                                 sleepq_release(&lk->lock_object);
 897                                 continue;
 898                         }
 899
 900                         /*
 901                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 902                          * fail, loop back and retry.
 903                          */
 904                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 905                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 906                                     x | LK_EXCLUSIVE_WAITERS)) {
 907                                         sleepq_release(&lk->lock_object);
 908                                         continue;
 909                                 }
 910                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 911                                     __func__, lk);
 912                         }
 913
 914                         /*
 915                          * As far as we have been unable to acquire the
 916                          * exclusive lock and the exclusive waiters flag
 917                          * is set, we will sleep.
 918                          */
 919                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 920                             SQ_EXCLUSIVE_QUEUE);
 921                         flags &= ~LK_INTERLOCK;
 922                         if (error) {
 923                                 LOCK_LOG3(lk,
 924                                     "%s: interrupted sleep for %p with %d",
 925                                     __func__, lk, error);
 926                                 break;
 927                         }
 928                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 929                             __func__, lk);
 930                 }
 931                 if (error == 0) {
 932                         lock_profile_obtain_lock_success(&lk->lock_object,
 933                             contested, waittime, file, line);
 934                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 935                             lk->lk_recurse, file, line);
 936                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 937                             LK_TRYWIT(flags), file, line);
 938                         TD_LOCKS_INC(curthread);
 939                         STACK_SAVE(lk);
 940                 }
 941                 break;
 942         case LK_DOWNGRADE:
 943                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 944                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 945                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 946
 947                 /*
 948                  * Panic if the lock is recursed.
 949                  */
 950                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 951                         if (flags & LK_INTERLOCK)
 952                                 class->lc_unlock(ilk);
 953                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 954                             __func__, iwmesg, file, line);
 955                 }
 956                 TD_SLOCKS_INC(curthread);
 957
 958                 /*
 959                  * In order to preserve waiters flags, just spin.
 960                  */
 961                 for (;;) {
 962                         x = lk->lk_lock;
 963                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
 964                         x &= LK_ALL_WAITERS;
 965                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
 966                             LK_SHARERS_LOCK(1) | x))
 967                                 break;
 968                         cpu_spinwait();
 969                 }
 970                 break;
 971         case LK_RELEASE:
 972                 _lockmgr_assert(lk, KA_LOCKED, file, line);
 973                 x = lk->lk_lock;
 974
 975                 if ((x & LK_SHARE) == 0) {
 976
 977                         /*
 978                          * As first option, treact the lock as if it has not
 979                          * any waiter.
 980                          * Fix-up the tid var if the lock has been disowned.
 981                          */
 982                         if (LK_HOLDER(x) == LK_KERNPROC)
 983                                 tid = LK_KERNPROC;
 984                         else {
 985                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
 986                                     file, line);
 987                                 TD_LOCKS_DEC(curthread);
 988                         }
 989                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
 990                             lk->lk_recurse, file, line);
 991
 992                         /*
 993                          * The lock is held in exclusive mode.
 994                          * If the lock is recursed also, then unrecurse it.
 995                          */
 996                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 997                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
 998                                     lk);
 999                                 lk->lk_recurse--;
1000                                 break;
1001                         }
1002                         if (tid != LK_KERNPROC)
1003                                 lock_profile_release_lock(&lk->lock_object);
1004
1005                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1006                             LK_UNLOCKED))
1007                                 break;
1008
1009                         sleepq_lock(&lk->lock_object);
1010                         x = lk->lk_lock;
1011                         v = LK_UNLOCKED;
1012
1013                         /*
1014                          * If the lock has exclusive waiters, give them
1015                          * preference in order to avoid deadlock with
1016                          * shared runners up.
1017                          * If interruptible sleeps left the exclusive queue
1018                          * empty avoid a starvation for the threads sleeping
1019                          * on the shared queue by giving them precedence
1020                          * and cleaning up the exclusive waiters bit anyway.
1021                          * Please note that lk_exslpfail count may be lying
1022                          * about the real number of waiters with the
1023                          * LK_SLEEPFAIL flag on because they may be used in
1024                          * conjuction with interruptible sleeps so
1025                          * lk_exslpfail might be considered an 'upper limit'
1026                          * bound, including the edge cases.
1027                          */
1028                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1029                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1030                             SQ_EXCLUSIVE_QUEUE);
1031                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1032                                 if (lk->lk_exslpfail < realexslp) {
1033                                         lk->lk_exslpfail = 0;
1034                                         queue = SQ_EXCLUSIVE_QUEUE;
1035                                         v |= (x & LK_SHARED_WAITERS);
1036                                 } else {
1037                                         lk->lk_exslpfail = 0;
1038                                         LOCK_LOG2(lk,
1039                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1040                                             __func__, lk);
1041                                         LOCK_LOG2(lk,
1042                         "%s: %p waking up threads on the exclusive queue",
1043                                             __func__, lk);
1044                                         wakeup_swapper =
1045                                             sleepq_broadcast(&lk->lock_object,
1046                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1047                                         queue = SQ_SHARED_QUEUE;
1048                                 }
1049                         } else {
1050
1051                                 /*
1052                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1053                                  * on and using interruptible sleeps/timeout
1054                                  * may have left spourious lk_exslpfail counts
1055                                  * on, so clean it up anyway.
1056                                  */
1057                                 lk->lk_exslpfail = 0;
1058                                 queue = SQ_SHARED_QUEUE;
1059                         }
1060
1061                         LOCK_LOG3(lk,
1062                             "%s: %p waking up threads on the %s queue",
1063                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1064                             "exclusive");
1065                         atomic_store_rel_ptr(&lk->lk_lock, v);
1066                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1067                             SLEEPQ_LK, 0, queue);
1068                         sleepq_release(&lk->lock_object);
1069                         break;
1070                 } else
1071                         wakeup_swapper = wakeupshlk(lk, file, line);
1072                 break;
1073         case LK_DRAIN:
1074                 if (LK_CAN_WITNESS(flags))
1075                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1076                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1077                             ilk : NULL);
1078
1079                 /*
1080                  * Trying to drain a lock we already own will result in a
1081                  * deadlock.
1082                  */
1083                 if (lockmgr_xlocked(lk)) {
1084                         if (flags & LK_INTERLOCK)
1085                                 class->lc_unlock(ilk);
1086                         panic("%s: draining %s with the lock held @ %s:%d\n",
1087                             __func__, iwmesg, file, line);
1088                 }
1089
1090                 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1091 #ifdef HWPMC_HOOKS
1092                         PMC_SOFT_CALL( , , lock, failed);
1093 #endif
1094                         lock_profile_obtain_lock_failed(&lk->lock_object,
1095                             &contested, &waittime);
1096
1097                         /*
1098                          * If the lock is expected to not sleep just give up
1099                          * and return.
1100                          */
1101                         if (LK_TRYOP(flags)) {
1102                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1103                                     __func__, lk);
1104                                 error = EBUSY;
1105                                 break;
1106                         }
1107
1108                         /*
1109                          * Acquire the sleepqueue chain lock because we
1110                          * probabilly will need to manipulate waiters flags.
1111                          */
1112                         sleepq_lock(&lk->lock_object);
1113                         x = lk->lk_lock;
1114
1115                         /*
1116                          * if the lock has been released while we spun on
1117                          * the sleepqueue chain lock just try again.
1118                          */
1119                         if (x == LK_UNLOCKED) {
1120                                 sleepq_release(&lk->lock_object);
1121                                 continue;
1122                         }
1123
1124                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1125                         if ((x & ~v) == LK_UNLOCKED) {
1126                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1127
1128                                 /*
1129                                  * If interruptible sleeps left the exclusive
1130                                  * queue empty avoid a starvation for the
1131                                  * threads sleeping on the shared queue by
1132                                  * giving them precedence and cleaning up the
1133                                  * exclusive waiters bit anyway.
1134                                  * Please note that lk_exslpfail count may be
1135                                  * lying about the real number of waiters with
1136                                  * the LK_SLEEPFAIL flag on because they may
1137                                  * be used in conjuction with interruptible
1138                                  * sleeps so lk_exslpfail might be considered
1139                                  * an 'upper limit' bound, including the edge
1140                                  * cases.
1141                                  */
1142                                 if (v & LK_EXCLUSIVE_WAITERS) {
1143                                         queue = SQ_EXCLUSIVE_QUEUE;
1144                                         v &= ~LK_EXCLUSIVE_WAITERS;
1145                                 } else {
1146
1147                                         /*
1148                                          * Exclusive waiters sleeping with
1149                                          * LK_SLEEPFAIL on and using
1150                                          * interruptible sleeps/timeout may
1151                                          * have left spourious lk_exslpfail
1152                                          * counts on, so clean it up anyway.
1153                                          */
1154                                         MPASS(v & LK_SHARED_WAITERS);
1155                                         lk->lk_exslpfail = 0;
1156                                         queue = SQ_SHARED_QUEUE;
1157                                         v &= ~LK_SHARED_WAITERS;
1158                                 }
1159                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1160                                         realexslp =
1161                                             sleepq_sleepcnt(&lk->lock_object,
1162                                             SQ_EXCLUSIVE_QUEUE);
1163                                         if (lk->lk_exslpfail >= realexslp) {
1164                                                 lk->lk_exslpfail = 0;
1165                                                 queue = SQ_SHARED_QUEUE;
1166                                                 v &= ~LK_SHARED_WAITERS;
1167                                                 if (realexslp != 0) {
1168                                                         LOCK_LOG2(lk,
1169                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1170                                                             __func__, lk);
1171                                                         LOCK_LOG2(lk,
1172                         "%s: %p waking up threads on the exclusive queue",
1173                                                             __func__, lk);
1174                                                         wakeup_swapper =
1175                                                             sleepq_broadcast(
1176                                                             &lk->lock_object,
1177                                                             SLEEPQ_LK, 0,
1178                                                             SQ_EXCLUSIVE_QUEUE);
1179                                                 }
1180                                         } else
1181                                                 lk->lk_exslpfail = 0;
1182                                 }
1183                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1184                                         sleepq_release(&lk->lock_object);
1185                                         continue;
1186                                 }
1187                                 LOCK_LOG3(lk,
1188                                 "%s: %p waking up all threads on the %s queue",
1189                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1190                                     "shared" : "exclusive");
1191                                 wakeup_swapper |= sleepq_broadcast(
1192                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1193
1194                                 /*
1195                                  * If shared waiters have been woken up we need
1196                                  * to wait for one of them to acquire the lock
1197                                  * before to set the exclusive waiters in
1198                                  * order to avoid a deadlock.
1199                                  */
1200                                 if (queue == SQ_SHARED_QUEUE) {
1201                                         for (v = lk->lk_lock;
1202                                             (v & LK_SHARE) && !LK_SHARERS(v);
1203                                             v = lk->lk_lock)
1204                                                 cpu_spinwait();
1205                                 }
1206                         }
1207
1208                         /*
1209                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1210                          * fail, loop back and retry.
1211                          */
1212                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1213                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1214                                     x | LK_EXCLUSIVE_WAITERS)) {
1215                                         sleepq_release(&lk->lock_object);
1216                                         continue;
1217                                 }
1218                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1219                                     __func__, lk);
1220                         }
1221
1222                         /*
1223                          * As far as we have been unable to acquire the
1224                          * exclusive lock and the exclusive waiters flag
1225                          * is set, we will sleep.
1226                          */
1227                         if (flags & LK_INTERLOCK) {
1228                                 class->lc_unlock(ilk);
1229                                 flags &= ~LK_INTERLOCK;
1230                         }
1231                         GIANT_SAVE();
1232                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1233                             SQ_EXCLUSIVE_QUEUE);
1234                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1235                         GIANT_RESTORE();
1236                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1237                             __func__, lk);
1238                 }
1239
1240                 if (error == 0) {
1241                         lock_profile_obtain_lock_success(&lk->lock_object,
1242                             contested, waittime, file, line);
1243                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1244                             lk->lk_recurse, file, line);
1245                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1246                             LK_TRYWIT(flags), file, line);
1247                         TD_LOCKS_INC(curthread);
1248                         STACK_SAVE(lk);
1249                 }
1250                 break;
1251         default:
1252                 if (flags & LK_INTERLOCK)
1253                         class->lc_unlock(ilk);
1254                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1255         }
1256
1257         if (flags & LK_INTERLOCK)
1258                 class->lc_unlock(ilk);
1259         if (wakeup_swapper)
1260                 kick_proc0();
1261
1262         return (error);
1263 }
1264
1265 void
1266 _lockmgr_disown(struct lock *lk, const char *file, int line)
1267 {
1268         uintptr_t tid, x;
1269
1270         if (SCHEDULER_STOPPED())
1271                 return;
1272
1273         tid = (uintptr_t)curthread;
1274         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1275
1276         /*
1277          * Panic if the lock is recursed.
1278          */
1279         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1280                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1281                     __func__,  file, line);
1282
1283         /*
1284          * If the owner is already LK_KERNPROC just skip the whole operation.
1285          */
1286         if (LK_HOLDER(lk->lk_lock) != tid)
1287                 return;
1288         lock_profile_release_lock(&lk->lock_object);
1289         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1290         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1291         TD_LOCKS_DEC(curthread);
1292         STACK_SAVE(lk);
1293
1294         /*
1295          * In order to preserve waiters flags, just spin.
1296          */
1297         for (;;) {
1298                 x = lk->lk_lock;
1299                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1300                 x &= LK_ALL_WAITERS;
1301                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1302                     LK_KERNPROC | x))
1303                         return;
1304                 cpu_spinwait();
1305         }
1306 }
1307
1308 void
1309 lockmgr_printinfo(const struct lock *lk)
1310 {
1311         struct thread *td;
1312         uintptr_t x;
1313
1314         if (lk->lk_lock == LK_UNLOCKED)
1315                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1316         else if (lk->lk_lock & LK_SHARE)
1317                 printf("lock type %s: SHARED (count %ju)\n",
1318                     lk->lock_object.lo_name,
1319                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1320         else {
1321                 td = lockmgr_xholder(lk);
1322                 printf("lock type %s: EXCL by thread %p "
1323                     "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, td,
1324                     td->td_proc->p_pid, td->td_proc->p_comm, td->td_tid);
1325         }
1326
1327         x = lk->lk_lock;
1328         if (x & LK_EXCLUSIVE_WAITERS)
1329                 printf(" with exclusive waiters pending\n");
1330         if (x & LK_SHARED_WAITERS)
1331                 printf(" with shared waiters pending\n");
1332         if (x & LK_EXCLUSIVE_SPINNERS)
1333                 printf(" with exclusive spinners pending\n");
1334
1335         STACK_PRINT(lk);
1336 }
1337
1338 int
1339 lockstatus(const struct lock *lk)
1340 {
1341         uintptr_t v, x;
1342         int ret;
1343
1344         ret = LK_SHARED;
1345         x = lk->lk_lock;
1346         v = LK_HOLDER(x);
1347
1348         if ((x & LK_SHARE) == 0) {
1349                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1350                         ret = LK_EXCLUSIVE;
1351                 else
1352                         ret = LK_EXCLOTHER;
1353         } else if (x == LK_UNLOCKED)
1354                 ret = 0;
1355
1356         return (ret);
1357 }
1358
1359 #ifdef INVARIANT_SUPPORT
1360
1361 FEATURE(invariant_support,
1362     "Support for modules compiled with INVARIANTS option");
1363
1364 #ifndef INVARIANTS
1365 #undef  _lockmgr_assert
1366 #endif
1367
1368 void
1369 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1370 {
1371         int slocked = 0;
1372
1373         if (panicstr != NULL)
1374                 return;
1375         switch (what) {
1376         case KA_SLOCKED:
1377         case KA_SLOCKED | KA_NOTRECURSED:
1378         case KA_SLOCKED | KA_RECURSED:
1379                 slocked = 1;
1380         case KA_LOCKED:
1381         case KA_LOCKED | KA_NOTRECURSED:
1382         case KA_LOCKED | KA_RECURSED:
1383 #ifdef WITNESS
1384
1385                 /*
1386                  * We cannot trust WITNESS if the lock is held in exclusive
1387                  * mode and a call to lockmgr_disown() happened.
1388                  * Workaround this skipping the check if the lock is held in
1389                  * exclusive mode even for the KA_LOCKED case.
1390                  */
1391                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1392                         witness_assert(&lk->lock_object, what, file, line);
1393                         break;
1394                 }
1395 #endif
1396                 if (lk->lk_lock == LK_UNLOCKED ||
1397                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1398                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1399                         panic("Lock %s not %slocked @ %s:%d\n",
1400                             lk->lock_object.lo_name, slocked ? "share" : "",
1401                             file, line);
1402
1403                 if ((lk->lk_lock & LK_SHARE) == 0) {
1404                         if (lockmgr_recursed(lk)) {
1405                                 if (what & KA_NOTRECURSED)
1406                                         panic("Lock %s recursed @ %s:%d\n",
1407                                             lk->lock_object.lo_name, file,
1408                                             line);
1409                         } else if (what & KA_RECURSED)
1410                                 panic("Lock %s not recursed @ %s:%d\n",
1411                                     lk->lock_object.lo_name, file, line);
1412                 }
1413                 break;
1414         case KA_XLOCKED:
1415         case KA_XLOCKED | KA_NOTRECURSED:
1416         case KA_XLOCKED | KA_RECURSED:
1417                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1418                         panic("Lock %s not exclusively locked @ %s:%d\n",
1419                             lk->lock_object.lo_name, file, line);
1420                 if (lockmgr_recursed(lk)) {
1421                         if (what & KA_NOTRECURSED)
1422                                 panic("Lock %s recursed @ %s:%d\n",
1423                                     lk->lock_object.lo_name, file, line);
1424                 } else if (what & KA_RECURSED)
1425                         panic("Lock %s not recursed @ %s:%d\n",
1426                             lk->lock_object.lo_name, file, line);
1427                 break;
1428         case KA_UNLOCKED:
1429                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1430                         panic("Lock %s exclusively locked @ %s:%d\n",
1431                             lk->lock_object.lo_name, file, line);
1432                 break;
1433         default:
1434                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1435                     line);
1436         }
1437 }
1438 #endif
1439
1440 #ifdef DDB
1441 int
1442 lockmgr_chain(struct thread *td, struct thread **ownerp)
1443 {
1444         struct lock *lk;
1445
1446         lk = td->td_wchan;
1447
1448         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1449                 return (0);
1450         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1451         if (lk->lk_lock & LK_SHARE)
1452                 db_printf("SHARED (count %ju)\n",
1453                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1454         else
1455                 db_printf("EXCL\n");
1456         *ownerp = lockmgr_xholder(lk);
1457
1458         return (1);
1459 }
1460
1461 static void
1462 db_show_lockmgr(const struct lock_object *lock)
1463 {
1464         struct thread *td;
1465         const struct lock *lk;
1466
1467         lk = (const struct lock *)lock;
1468
1469         db_printf(" state: ");
1470         if (lk->lk_lock == LK_UNLOCKED)
1471                 db_printf("UNLOCKED\n");
1472         else if (lk->lk_lock & LK_SHARE)
1473                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1474         else {
1475                 td = lockmgr_xholder(lk);
1476                 if (td == (struct thread *)LK_KERNPROC)
1477                         db_printf("XLOCK: LK_KERNPROC\n");
1478                 else
1479                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1480                             td->td_tid, td->td_proc->p_pid,
1481                             td->td_proc->p_comm);
1482                 if (lockmgr_recursed(lk))
1483                         db_printf(" recursed: %d\n", lk->lk_recurse);
1484         }
1485         db_printf(" waiters: ");
1486         switch (lk->lk_lock & LK_ALL_WAITERS) {
1487         case LK_SHARED_WAITERS:
1488                 db_printf("shared\n");
1489                 break;
1490         case LK_EXCLUSIVE_WAITERS:
1491                 db_printf("exclusive\n");
1492                 break;
1493         case LK_ALL_WAITERS:
1494                 db_printf("shared and exclusive\n");
1495                 break;
1496         default:
1497                 db_printf("none\n");
1498         }
1499         db_printf(" spinners: ");
1500         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1501                 db_printf("exclusive\n");
1502         else
1503                 db_printf("none\n");
1504 }
1505 #endif