sys/kern/kern_lock.c

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