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