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, flags)                                          \
 120         (((x) & LK_SHARE) &&                                            \
 121         (((x) & (LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 0 || \
 122         (curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) ||  \
 123         (curthread->td_pflags & TDP_DEADLKTREAT)))
 124 #define LK_TRYOP(x)                                                     \
 125         ((x) & LK_NOWAIT)
 126
 127 #define LK_CAN_WITNESS(x)                                               \
 128         (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
 129 #define LK_TRYWIT(x)                                                    \
 130         (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
 131
 132 #define LK_CAN_ADAPT(lk, f)                                             \
 133         (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 &&             \
 134         ((f) & LK_SLEEPFAIL) == 0)
 135
 136 #define lockmgr_disowned(lk)                                            \
 137         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
 138
 139 #define lockmgr_xlocked(lk)                                             \
 140         (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
 141
 142 static void     assert_lockmgr(const struct lock_object *lock, int how);
 143 #ifdef DDB
 144 static void     db_show_lockmgr(const struct lock_object *lock);
 145 #endif
 146 static void     lock_lockmgr(struct lock_object *lock, uintptr_t how);
 147 #ifdef KDTRACE_HOOKS
 148 static int      owner_lockmgr(const struct lock_object *lock,
 149                     struct thread **owner);
 150 #endif
 151 static uintptr_t unlock_lockmgr(struct lock_object *lock);
 152
 153 struct lock_class lock_class_lockmgr = {
 154         .lc_name = "lockmgr",
 155         .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
 156         .lc_assert = assert_lockmgr,
 157 #ifdef DDB
 158         .lc_ddb_show = db_show_lockmgr,
 159 #endif
 160         .lc_lock = lock_lockmgr,
 161         .lc_unlock = unlock_lockmgr,
 162 #ifdef KDTRACE_HOOKS
 163         .lc_owner = owner_lockmgr,
 164 #endif
 165 };
 166
 167 #ifdef ADAPTIVE_LOCKMGRS
 168 static u_int alk_retries = 10;
 169 static u_int alk_loops = 10000;
 170 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL,
 171     "lockmgr debugging");
 172 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
 173 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
 174 #endif
 175
 176 static __inline struct thread *
 177 lockmgr_xholder(const struct lock *lk)
 178 {
 179         uintptr_t x;
 180
 181         x = lk->lk_lock;
 182         return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
 183 }
 184
 185 /*
 186  * It assumes sleepq_lock held and returns with this one unheld.
 187  * It also assumes the generic interlock is sane and previously checked.
 188  * If LK_INTERLOCK is specified the interlock is not reacquired after the
 189  * sleep.
 190  */
 191 static __inline int
 192 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
 193     const char *wmesg, int pri, int timo, int queue)
 194 {
 195         GIANT_DECLARE;
 196         struct lock_class *class;
 197         int catch, error;
 198
 199         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 200         catch = pri & PCATCH;
 201         pri &= PRIMASK;
 202         error = 0;
 203
 204         LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
 205             (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
 206
 207         if (flags & LK_INTERLOCK)
 208                 class->lc_unlock(ilk);
 209         if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
 210                 lk->lk_exslpfail++;
 211         GIANT_SAVE();
 212         sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
 213             SLEEPQ_INTERRUPTIBLE : 0), queue);
 214         if ((flags & LK_TIMELOCK) && timo)
 215                 sleepq_set_timeout(&lk->lock_object, timo);
 216
 217         /*
 218          * Decisional switch for real sleeping.
 219          */
 220         if ((flags & LK_TIMELOCK) && timo && catch)
 221                 error = sleepq_timedwait_sig(&lk->lock_object, pri);
 222         else if ((flags & LK_TIMELOCK) && timo)
 223                 error = sleepq_timedwait(&lk->lock_object, pri);
 224         else if (catch)
 225                 error = sleepq_wait_sig(&lk->lock_object, pri);
 226         else
 227                 sleepq_wait(&lk->lock_object, pri);
 228         GIANT_RESTORE();
 229         if ((flags & LK_SLEEPFAIL) && error == 0)
 230                 error = ENOLCK;
 231
 232         return (error);
 233 }
 234
 235 static __inline int
 236 wakeupshlk(struct lock *lk, const char *file, int line)
 237 {
 238         uintptr_t v, x;
 239         u_int realexslp;
 240         int queue, wakeup_swapper;
 241
 242         WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
 243         LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
 244
 245         wakeup_swapper = 0;
 246         for (;;) {
 247                 x = lk->lk_lock;
 248
 249                 /*
 250                  * If there is more than one shared lock held, just drop one
 251                  * and return.
 252                  */
 253                 if (LK_SHARERS(x) > 1) {
 254                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
 255                             x - LK_ONE_SHARER))
 256                                 break;
 257                         continue;
 258                 }
 259
 260                 /*
 261                  * If there are not waiters on the exclusive queue, drop the
 262                  * lock quickly.
 263                  */
 264                 if ((x & LK_ALL_WAITERS) == 0) {
 265                         MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
 266                             LK_SHARERS_LOCK(1));
 267                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
 268                                 break;
 269                         continue;
 270                 }
 271
 272                 /*
 273                  * We should have a sharer with waiters, so enter the hard
 274                  * path in order to handle wakeups correctly.
 275                  */
 276                 sleepq_lock(&lk->lock_object);
 277                 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 278                 v = LK_UNLOCKED;
 279
 280                 /*
 281                  * If the lock has exclusive waiters, give them preference in
 282                  * order to avoid deadlock with shared runners up.
 283                  * If interruptible sleeps left the exclusive queue empty
 284                  * avoid a starvation for the threads sleeping on the shared
 285                  * queue by giving them precedence and cleaning up the
 286                  * exclusive waiters bit anyway.
 287                  * Please note that lk_exslpfail count may be lying about
 288                  * the real number of waiters with the LK_SLEEPFAIL flag on
 289                  * because they may be used in conjunction with interruptible
 290                  * sleeps so lk_exslpfail might be considered an 'upper limit'
 291                  * bound, including the edge cases.
 292                  */
 293                 realexslp = sleepq_sleepcnt(&lk->lock_object,
 294                     SQ_EXCLUSIVE_QUEUE);
 295                 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
 296                         if (lk->lk_exslpfail < realexslp) {
 297                                 lk->lk_exslpfail = 0;
 298                                 queue = SQ_EXCLUSIVE_QUEUE;
 299                                 v |= (x & LK_SHARED_WAITERS);
 300                         } else {
 301                                 lk->lk_exslpfail = 0;
 302                                 LOCK_LOG2(lk,
 303                                     "%s: %p has only LK_SLEEPFAIL sleepers",
 304                                     __func__, lk);
 305                                 LOCK_LOG2(lk,
 306                             "%s: %p waking up threads on the exclusive queue",
 307                                     __func__, lk);
 308                                 wakeup_swapper =
 309                                     sleepq_broadcast(&lk->lock_object,
 310                                     SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
 311                                 queue = SQ_SHARED_QUEUE;
 312                         }
 313
 314                 } else {
 315
 316                         /*
 317                          * Exclusive waiters sleeping with LK_SLEEPFAIL on
 318                          * and using interruptible sleeps/timeout may have
 319                          * left spourious lk_exslpfail counts on, so clean
 320                          * it up anyway.
 321                          */
 322                         lk->lk_exslpfail = 0;
 323                         queue = SQ_SHARED_QUEUE;
 324                 }
 325
 326                 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
 327                     v)) {
 328                         sleepq_release(&lk->lock_object);
 329                         continue;
 330                 }
 331                 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
 332                     __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
 333                     "exclusive");
 334                 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
 335                     0, queue);
 336                 sleepq_release(&lk->lock_object);
 337                 break;
 338         }
 339
 340         lock_profile_release_lock(&lk->lock_object);
 341         TD_LOCKS_DEC(curthread);
 342         TD_SLOCKS_DEC(curthread);
 343         return (wakeup_swapper);
 344 }
 345
 346 static void
 347 assert_lockmgr(const struct lock_object *lock, int what)
 348 {
 349
 350         panic("lockmgr locks do not support assertions");
 351 }
 352
 353 static void
 354 lock_lockmgr(struct lock_object *lock, uintptr_t how)
 355 {
 356
 357         panic("lockmgr locks do not support sleep interlocking");
 358 }
 359
 360 static uintptr_t
 361 unlock_lockmgr(struct lock_object *lock)
 362 {
 363
 364         panic("lockmgr locks do not support sleep interlocking");
 365 }
 366
 367 #ifdef KDTRACE_HOOKS
 368 static int
 369 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
 370 {
 371
 372         panic("lockmgr locks do not support owner inquiring");
 373 }
 374 #endif
 375
 376 void
 377 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 378 {
 379         int iflags;
 380
 381         MPASS((flags & ~LK_INIT_MASK) == 0);
 382         ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
 383             ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
 384             &lk->lk_lock));
 385
 386         iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 387         if (flags & LK_CANRECURSE)
 388                 iflags |= LO_RECURSABLE;
 389         if ((flags & LK_NODUP) == 0)
 390                 iflags |= LO_DUPOK;
 391         if (flags & LK_NOPROFILE)
 392                 iflags |= LO_NOPROFILE;
 393         if ((flags & LK_NOWITNESS) == 0)
 394                 iflags |= LO_WITNESS;
 395         if (flags & LK_QUIET)
 396                 iflags |= LO_QUIET;
 397         if (flags & LK_IS_VNODE)
 398                 iflags |= LO_IS_VNODE;
 399         iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
 400
 401         lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
 402         lk->lk_lock = LK_UNLOCKED;
 403         lk->lk_recurse = 0;
 404         lk->lk_exslpfail = 0;
 405         lk->lk_timo = timo;
 406         lk->lk_pri = pri;
 407         STACK_ZERO(lk);
 408 }
 409
 410 /*
 411  * XXX: Gross hacks to manipulate external lock flags after
 412  * initialization.  Used for certain vnode and buf locks.
 413  */
 414 void
 415 lockallowshare(struct lock *lk)
 416 {
 417
 418         lockmgr_assert(lk, KA_XLOCKED);
 419         lk->lock_object.lo_flags &= ~LK_NOSHARE;
 420 }
 421
 422 void
 423 lockdisableshare(struct lock *lk)
 424 {
 425
 426         lockmgr_assert(lk, KA_XLOCKED);
 427         lk->lock_object.lo_flags |= LK_NOSHARE;
 428 }
 429
 430 void
 431 lockallowrecurse(struct lock *lk)
 432 {
 433
 434         lockmgr_assert(lk, KA_XLOCKED);
 435         lk->lock_object.lo_flags |= LO_RECURSABLE;
 436 }
 437
 438 void
 439 lockdisablerecurse(struct lock *lk)
 440 {
 441
 442         lockmgr_assert(lk, KA_XLOCKED);
 443         lk->lock_object.lo_flags &= ~LO_RECURSABLE;
 444 }
 445
 446 void
 447 lockdestroy(struct lock *lk)
 448 {
 449
 450         KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
 451         KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
 452         KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
 453         lock_destroy(&lk->lock_object);
 454 }
 455
 456 int
 457 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 458     const char *wmesg, int pri, int timo, const char *file, int line)
 459 {
 460         GIANT_DECLARE;
 461         struct lock_class *class;
 462         const char *iwmesg;
 463         uintptr_t tid, v, x;
 464         u_int op, realexslp;
 465         int error, ipri, itimo, queue, wakeup_swapper;
 466 #ifdef LOCK_PROFILING
 467         uint64_t waittime = 0;
 468         int contested = 0;
 469 #endif
 470 #ifdef ADAPTIVE_LOCKMGRS
 471         volatile struct thread *owner;
 472         u_int i, spintries = 0;
 473 #endif
 474
 475         error = 0;
 476         tid = (uintptr_t)curthread;
 477         op = (flags & LK_TYPE_MASK);
 478         iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
 479         ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
 480         itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
 481
 482         MPASS((flags & ~LK_TOTAL_MASK) == 0);
 483         KASSERT((op & (op - 1)) == 0,
 484             ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
 485         KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
 486             (op != LK_DOWNGRADE && op != LK_RELEASE),
 487             ("%s: Invalid flags in regard of the operation desired @ %s:%d",
 488             __func__, file, line));
 489         KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
 490             ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
 491             __func__, file, line));
 492         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 493             ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
 494             lk->lock_object.lo_name, file, line));
 495
 496         class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
 497         if (panicstr != NULL) {
 498                 if (flags & LK_INTERLOCK)
 499                         class->lc_unlock(ilk);
 500                 return (0);
 501         }
 502
 503         if (lk->lock_object.lo_flags & LK_NOSHARE) {
 504                 switch (op) {
 505                 case LK_SHARED:
 506                         op = LK_EXCLUSIVE;
 507                         break;
 508                 case LK_UPGRADE:
 509                 case LK_TRYUPGRADE:
 510                 case LK_DOWNGRADE:
 511                         _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
 512                             file, line);
 513                         if (flags & LK_INTERLOCK)
 514                                 class->lc_unlock(ilk);
 515                         return (0);
 516                 }
 517         }
 518
 519         wakeup_swapper = 0;
 520         switch (op) {
 521         case LK_SHARED:
 522                 if (LK_CAN_WITNESS(flags))
 523                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 524                             file, line, flags & LK_INTERLOCK ? ilk : NULL);
 525                 for (;;) {
 526                         x = lk->lk_lock;
 527
 528                         /*
 529                          * If no other thread has an exclusive lock, or
 530                          * no exclusive waiter is present, bump the count of
 531                          * sharers.  Since we have to preserve the state of
 532                          * waiters, if we fail to acquire the shared lock
 533                          * loop back and retry.
 534                          */
 535                         if (LK_CAN_SHARE(x, flags)) {
 536                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 537                                     x + LK_ONE_SHARER))
 538                                         break;
 539                                 continue;
 540                         }
 541 #ifdef HWPMC_HOOKS
 542                         PMC_SOFT_CALL( , , lock, failed);
 543 #endif
 544                         lock_profile_obtain_lock_failed(&lk->lock_object,
 545                             &contested, &waittime);
 546
 547                         /*
 548                          * If the lock is already held by curthread in
 549                          * exclusive way avoid a deadlock.
 550                          */
 551                         if (LK_HOLDER(x) == tid) {
 552                                 LOCK_LOG2(lk,
 553                                     "%s: %p already held in exclusive mode",
 554                                     __func__, lk);
 555                                 error = EDEADLK;
 556                                 break;
 557                         }
 558
 559                         /*
 560                          * If the lock is expected to not sleep just give up
 561                          * and return.
 562                          */
 563                         if (LK_TRYOP(flags)) {
 564                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 565                                     __func__, lk);
 566                                 error = EBUSY;
 567                                 break;
 568                         }
 569
 570 #ifdef ADAPTIVE_LOCKMGRS
 571                         /*
 572                          * If the owner is running on another CPU, spin until
 573                          * the owner stops running or the state of the lock
 574                          * changes.  We need a double-state handle here
 575                          * because for a failed acquisition the lock can be
 576                          * either held in exclusive mode or shared mode
 577                          * (for the writer starvation avoidance technique).
 578                          */
 579                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 580                             LK_HOLDER(x) != LK_KERNPROC) {
 581                                 owner = (struct thread *)LK_HOLDER(x);
 582                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 583                                         CTR3(KTR_LOCK,
 584                                             "%s: spinning on %p held by %p",
 585                                             __func__, lk, owner);
 586                                 KTR_STATE1(KTR_SCHED, "thread",
 587                                     sched_tdname(td), "spinning",
 588                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 589
 590                                 /*
 591                                  * If we are holding also an interlock drop it
 592                                  * in order to avoid a deadlock if the lockmgr
 593                                  * owner is adaptively spinning on the
 594                                  * interlock itself.
 595                                  */
 596                                 if (flags & LK_INTERLOCK) {
 597                                         class->lc_unlock(ilk);
 598                                         flags &= ~LK_INTERLOCK;
 599                                 }
 600                                 GIANT_SAVE();
 601                                 while (LK_HOLDER(lk->lk_lock) ==
 602                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 603                                         cpu_spinwait();
 604                                 KTR_STATE0(KTR_SCHED, "thread",
 605                                     sched_tdname(td), "running");
 606                                 GIANT_RESTORE();
 607                                 continue;
 608                         } else if (LK_CAN_ADAPT(lk, flags) &&
 609                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 610                             spintries < alk_retries) {
 611                                 KTR_STATE1(KTR_SCHED, "thread",
 612                                     sched_tdname(td), "spinning",
 613                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 614                                 if (flags & LK_INTERLOCK) {
 615                                         class->lc_unlock(ilk);
 616                                         flags &= ~LK_INTERLOCK;
 617                                 }
 618                                 GIANT_SAVE();
 619                                 spintries++;
 620                                 for (i = 0; i < alk_loops; i++) {
 621                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 622                                                 CTR4(KTR_LOCK,
 623                                     "%s: shared spinning on %p with %u and %u",
 624                                                     __func__, lk, spintries, i);
 625                                         x = lk->lk_lock;
 626                                         if ((x & LK_SHARE) == 0 ||
 627                                             LK_CAN_SHARE(x, flags) != 0)
 628                                                 break;
 629                                         cpu_spinwait();
 630                                 }
 631                                 KTR_STATE0(KTR_SCHED, "thread",
 632                                     sched_tdname(td), "running");
 633                                 GIANT_RESTORE();
 634                                 if (i != alk_loops)
 635                                         continue;
 636                         }
 637 #endif
 638
 639                         /*
 640                          * Acquire the sleepqueue chain lock because we
 641                          * probabilly will need to manipulate waiters flags.
 642                          */
 643                         sleepq_lock(&lk->lock_object);
 644                         x = lk->lk_lock;
 645
 646                         /*
 647                          * if the lock can be acquired in shared mode, try
 648                          * again.
 649                          */
 650                         if (LK_CAN_SHARE(x, flags)) {
 651                                 sleepq_release(&lk->lock_object);
 652                                 continue;
 653                         }
 654
 655 #ifdef ADAPTIVE_LOCKMGRS
 656                         /*
 657                          * The current lock owner might have started executing
 658                          * on another CPU (or the lock could have changed
 659                          * owner) while we were waiting on the turnstile
 660                          * chain lock.  If so, drop the turnstile lock and try
 661                          * again.
 662                          */
 663                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 664                             LK_HOLDER(x) != LK_KERNPROC) {
 665                                 owner = (struct thread *)LK_HOLDER(x);
 666                                 if (TD_IS_RUNNING(owner)) {
 667                                         sleepq_release(&lk->lock_object);
 668                                         continue;
 669                                 }
 670                         }
 671 #endif
 672
 673                         /*
 674                          * Try to set the LK_SHARED_WAITERS flag.  If we fail,
 675                          * loop back and retry.
 676                          */
 677                         if ((x & LK_SHARED_WAITERS) == 0) {
 678                                 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 679                                     x | LK_SHARED_WAITERS)) {
 680                                         sleepq_release(&lk->lock_object);
 681                                         continue;
 682                                 }
 683                                 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
 684                                     __func__, lk);
 685                         }
 686
 687                         /*
 688                          * As far as we have been unable to acquire the
 689                          * shared lock and the shared waiters flag is set,
 690                          * we will sleep.
 691                          */
 692                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 693                             SQ_SHARED_QUEUE);
 694                         flags &= ~LK_INTERLOCK;
 695                         if (error) {
 696                                 LOCK_LOG3(lk,
 697                                     "%s: interrupted sleep for %p with %d",
 698                                     __func__, lk, error);
 699                                 break;
 700                         }
 701                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 702                             __func__, lk);
 703                 }
 704                 if (error == 0) {
 705                         lock_profile_obtain_lock_success(&lk->lock_object,
 706                             contested, waittime, file, line);
 707                         LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
 708                             line);
 709                         WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
 710                             line);
 711                         TD_LOCKS_INC(curthread);
 712                         TD_SLOCKS_INC(curthread);
 713                         STACK_SAVE(lk);
 714                 }
 715                 break;
 716         case LK_UPGRADE:
 717         case LK_TRYUPGRADE:
 718                 _lockmgr_assert(lk, KA_SLOCKED, file, line);
 719                 v = lk->lk_lock;
 720                 x = v & LK_ALL_WAITERS;
 721                 v &= LK_EXCLUSIVE_SPINNERS;
 722
 723                 /*
 724                  * Try to switch from one shared lock to an exclusive one.
 725                  * We need to preserve waiters flags during the operation.
 726                  */
 727                 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
 728                     tid | x)) {
 729                         LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
 730                             line);
 731                         WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
 732                             LK_TRYWIT(flags), file, line);
 733                         TD_SLOCKS_DEC(curthread);
 734                         break;
 735                 }
 736
 737                 /*
 738                  * In LK_TRYUPGRADE mode, do not drop the lock,
 739                  * returning EBUSY instead.
 740                  */
 741                 if (op == LK_TRYUPGRADE) {
 742                         LOCK_LOG2(lk, "%s: %p failed the nowait upgrade",
 743                             __func__, lk);
 744                         error = EBUSY;
 745                         break;
 746                 }
 747
 748                 /*
 749                  * We have been unable to succeed in upgrading, so just
 750                  * give up the shared lock.
 751                  */
 752                 wakeup_swapper |= wakeupshlk(lk, file, line);
 753
 754                 /* FALLTHROUGH */
 755         case LK_EXCLUSIVE:
 756                 if (LK_CAN_WITNESS(flags))
 757                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
 758                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
 759                             ilk : NULL);
 760
 761                 /*
 762                  * If curthread already holds the lock and this one is
 763                  * allowed to recurse, simply recurse on it.
 764                  */
 765                 if (lockmgr_xlocked(lk)) {
 766                         if ((flags & LK_CANRECURSE) == 0 &&
 767                             (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
 768
 769                                 /*
 770                                  * If the lock is expected to not panic just
 771                                  * give up and return.
 772                                  */
 773                                 if (LK_TRYOP(flags)) {
 774                                         LOCK_LOG2(lk,
 775                                             "%s: %p fails the try operation",
 776                                             __func__, lk);
 777                                         error = EBUSY;
 778                                         break;
 779                                 }
 780                                 if (flags & LK_INTERLOCK)
 781                                         class->lc_unlock(ilk);
 782                 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
 783                                     __func__, iwmesg, file, line);
 784                         }
 785                         lk->lk_recurse++;
 786                         LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
 787                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 788                             lk->lk_recurse, file, line);
 789                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 790                             LK_TRYWIT(flags), file, line);
 791                         TD_LOCKS_INC(curthread);
 792                         break;
 793                 }
 794
 795                 for (;;) {
 796                         if (lk->lk_lock == LK_UNLOCKED &&
 797                             atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
 798                                 break;
 799 #ifdef HWPMC_HOOKS
 800                         PMC_SOFT_CALL( , , lock, failed);
 801 #endif
 802                         lock_profile_obtain_lock_failed(&lk->lock_object,
 803                             &contested, &waittime);
 804
 805                         /*
 806                          * If the lock is expected to not sleep just give up
 807                          * and return.
 808                          */
 809                         if (LK_TRYOP(flags)) {
 810                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
 811                                     __func__, lk);
 812                                 error = EBUSY;
 813                                 break;
 814                         }
 815
 816 #ifdef ADAPTIVE_LOCKMGRS
 817                         /*
 818                          * If the owner is running on another CPU, spin until
 819                          * the owner stops running or the state of the lock
 820                          * changes.
 821                          */
 822                         x = lk->lk_lock;
 823                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 824                             LK_HOLDER(x) != LK_KERNPROC) {
 825                                 owner = (struct thread *)LK_HOLDER(x);
 826                                 if (LOCK_LOG_TEST(&lk->lock_object, 0))
 827                                         CTR3(KTR_LOCK,
 828                                             "%s: spinning on %p held by %p",
 829                                             __func__, lk, owner);
 830                                 KTR_STATE1(KTR_SCHED, "thread",
 831                                     sched_tdname(td), "spinning",
 832                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 833
 834                                 /*
 835                                  * If we are holding also an interlock drop it
 836                                  * in order to avoid a deadlock if the lockmgr
 837                                  * owner is adaptively spinning on the
 838                                  * interlock itself.
 839                                  */
 840                                 if (flags & LK_INTERLOCK) {
 841                                         class->lc_unlock(ilk);
 842                                         flags &= ~LK_INTERLOCK;
 843                                 }
 844                                 GIANT_SAVE();
 845                                 while (LK_HOLDER(lk->lk_lock) ==
 846                                     (uintptr_t)owner && TD_IS_RUNNING(owner))
 847                                         cpu_spinwait();
 848                                 KTR_STATE0(KTR_SCHED, "thread",
 849                                     sched_tdname(td), "running");
 850                                 GIANT_RESTORE();
 851                                 continue;
 852                         } else if (LK_CAN_ADAPT(lk, flags) &&
 853                             (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
 854                             spintries < alk_retries) {
 855                                 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
 856                                     !atomic_cmpset_ptr(&lk->lk_lock, x,
 857                                     x | LK_EXCLUSIVE_SPINNERS))
 858                                         continue;
 859                                 KTR_STATE1(KTR_SCHED, "thread",
 860                                     sched_tdname(td), "spinning",
 861                                     "lockname:\"%s\"", lk->lock_object.lo_name);
 862                                 if (flags & LK_INTERLOCK) {
 863                                         class->lc_unlock(ilk);
 864                                         flags &= ~LK_INTERLOCK;
 865                                 }
 866                                 GIANT_SAVE();
 867                                 spintries++;
 868                                 for (i = 0; i < alk_loops; i++) {
 869                                         if (LOCK_LOG_TEST(&lk->lock_object, 0))
 870                                                 CTR4(KTR_LOCK,
 871                                     "%s: shared spinning on %p with %u and %u",
 872                                                     __func__, lk, spintries, i);
 873                                         if ((lk->lk_lock &
 874                                             LK_EXCLUSIVE_SPINNERS) == 0)
 875                                                 break;
 876                                         cpu_spinwait();
 877                                 }
 878                                 KTR_STATE0(KTR_SCHED, "thread",
 879                                     sched_tdname(td), "running");
 880                                 GIANT_RESTORE();
 881                                 if (i != alk_loops)
 882                                         continue;
 883                         }
 884 #endif
 885
 886                         /*
 887                          * Acquire the sleepqueue chain lock because we
 888                          * probabilly will need to manipulate waiters flags.
 889                          */
 890                         sleepq_lock(&lk->lock_object);
 891                         x = lk->lk_lock;
 892
 893                         /*
 894                          * if the lock has been released while we spun on
 895                          * the sleepqueue chain lock just try again.
 896                          */
 897                         if (x == LK_UNLOCKED) {
 898                                 sleepq_release(&lk->lock_object);
 899                                 continue;
 900                         }
 901
 902 #ifdef ADAPTIVE_LOCKMGRS
 903                         /*
 904                          * The current lock owner might have started executing
 905                          * on another CPU (or the lock could have changed
 906                          * owner) while we were waiting on the turnstile
 907                          * chain lock.  If so, drop the turnstile lock and try
 908                          * again.
 909                          */
 910                         if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
 911                             LK_HOLDER(x) != LK_KERNPROC) {
 912                                 owner = (struct thread *)LK_HOLDER(x);
 913                                 if (TD_IS_RUNNING(owner)) {
 914                                         sleepq_release(&lk->lock_object);
 915                                         continue;
 916                                 }
 917                         }
 918 #endif
 919
 920                         /*
 921                          * The lock can be in the state where there is a
 922                          * pending queue of waiters, but still no owner.
 923                          * This happens when the lock is contested and an
 924                          * owner is going to claim the lock.
 925                          * If curthread is the one successfully acquiring it
 926                          * claim lock ownership and return, preserving waiters
 927                          * flags.
 928                          */
 929                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
 930                         if ((x & ~v) == LK_UNLOCKED) {
 931                                 v &= ~LK_EXCLUSIVE_SPINNERS;
 932                                 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
 933                                     tid | v)) {
 934                                         sleepq_release(&lk->lock_object);
 935                                         LOCK_LOG2(lk,
 936                                             "%s: %p claimed by a new writer",
 937                                             __func__, lk);
 938                                         break;
 939                                 }
 940                                 sleepq_release(&lk->lock_object);
 941                                 continue;
 942                         }
 943
 944                         /*
 945                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
 946                          * fail, loop back and retry.
 947                          */
 948                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
 949                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
 950                                     x | LK_EXCLUSIVE_WAITERS)) {
 951                                         sleepq_release(&lk->lock_object);
 952                                         continue;
 953                                 }
 954                                 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
 955                                     __func__, lk);
 956                         }
 957
 958                         /*
 959                          * As far as we have been unable to acquire the
 960                          * exclusive lock and the exclusive waiters flag
 961                          * is set, we will sleep.
 962                          */
 963                         error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
 964                             SQ_EXCLUSIVE_QUEUE);
 965                         flags &= ~LK_INTERLOCK;
 966                         if (error) {
 967                                 LOCK_LOG3(lk,
 968                                     "%s: interrupted sleep for %p with %d",
 969                                     __func__, lk, error);
 970                                 break;
 971                         }
 972                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 973                             __func__, lk);
 974                 }
 975                 if (error == 0) {
 976                         lock_profile_obtain_lock_success(&lk->lock_object,
 977                             contested, waittime, file, line);
 978                         LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
 979                             lk->lk_recurse, file, line);
 980                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
 981                             LK_TRYWIT(flags), file, line);
 982                         TD_LOCKS_INC(curthread);
 983                         STACK_SAVE(lk);
 984                 }
 985                 break;
 986         case LK_DOWNGRADE:
 987                 _lockmgr_assert(lk, KA_XLOCKED, file, line);
 988                 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
 989                 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
 990
 991                 /*
 992                  * Panic if the lock is recursed.
 993                  */
 994                 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
 995                         if (flags & LK_INTERLOCK)
 996                                 class->lc_unlock(ilk);
 997                         panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
 998                             __func__, iwmesg, file, line);
 999                 }
1000                 TD_SLOCKS_INC(curthread);
1001
1002                 /*
1003                  * In order to preserve waiters flags, just spin.
1004                  */
1005                 for (;;) {
1006                         x = lk->lk_lock;
1007                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1008                         x &= LK_ALL_WAITERS;
1009                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1010                             LK_SHARERS_LOCK(1) | x))
1011                                 break;
1012                         cpu_spinwait();
1013                 }
1014                 break;
1015         case LK_RELEASE:
1016                 _lockmgr_assert(lk, KA_LOCKED, file, line);
1017                 x = lk->lk_lock;
1018
1019                 if ((x & LK_SHARE) == 0) {
1020
1021                         /*
1022                          * As first option, treact the lock as if it has not
1023                          * any waiter.
1024                          * Fix-up the tid var if the lock has been disowned.
1025                          */
1026                         if (LK_HOLDER(x) == LK_KERNPROC)
1027                                 tid = LK_KERNPROC;
1028                         else {
1029                                 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
1030                                     file, line);
1031                                 TD_LOCKS_DEC(curthread);
1032                         }
1033                         LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
1034                             lk->lk_recurse, file, line);
1035
1036                         /*
1037                          * The lock is held in exclusive mode.
1038                          * If the lock is recursed also, then unrecurse it.
1039                          */
1040                         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
1041                                 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1042                                     lk);
1043                                 lk->lk_recurse--;
1044                                 break;
1045                         }
1046                         if (tid != LK_KERNPROC)
1047                                 lock_profile_release_lock(&lk->lock_object);
1048
1049                         if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1050                             LK_UNLOCKED))
1051                                 break;
1052
1053                         sleepq_lock(&lk->lock_object);
1054                         x = lk->lk_lock;
1055                         v = LK_UNLOCKED;
1056
1057                         /*
1058                          * If the lock has exclusive waiters, give them
1059                          * preference in order to avoid deadlock with
1060                          * shared runners up.
1061                          * If interruptible sleeps left the exclusive queue
1062                          * empty avoid a starvation for the threads sleeping
1063                          * on the shared queue by giving them precedence
1064                          * and cleaning up the exclusive waiters bit anyway.
1065                          * Please note that lk_exslpfail count may be lying
1066                          * about the real number of waiters with the
1067                          * LK_SLEEPFAIL flag on because they may be used in
1068                          * conjunction with interruptible sleeps so
1069                          * lk_exslpfail might be considered an 'upper limit'
1070                          * bound, including the edge cases.
1071                          */
1072                         MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1073                         realexslp = sleepq_sleepcnt(&lk->lock_object,
1074                             SQ_EXCLUSIVE_QUEUE);
1075                         if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1076                                 if (lk->lk_exslpfail < realexslp) {
1077                                         lk->lk_exslpfail = 0;
1078                                         queue = SQ_EXCLUSIVE_QUEUE;
1079                                         v |= (x & LK_SHARED_WAITERS);
1080                                 } else {
1081                                         lk->lk_exslpfail = 0;
1082                                         LOCK_LOG2(lk,
1083                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1084                                             __func__, lk);
1085                                         LOCK_LOG2(lk,
1086                         "%s: %p waking up threads on the exclusive queue",
1087                                             __func__, lk);
1088                                         wakeup_swapper =
1089                                             sleepq_broadcast(&lk->lock_object,
1090                                             SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1091                                         queue = SQ_SHARED_QUEUE;
1092                                 }
1093                         } else {
1094
1095                                 /*
1096                                  * Exclusive waiters sleeping with LK_SLEEPFAIL
1097                                  * on and using interruptible sleeps/timeout
1098                                  * may have left spourious lk_exslpfail counts
1099                                  * on, so clean it up anyway.
1100                                  */
1101                                 lk->lk_exslpfail = 0;
1102                                 queue = SQ_SHARED_QUEUE;
1103                         }
1104
1105                         LOCK_LOG3(lk,
1106                             "%s: %p waking up threads on the %s queue",
1107                             __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1108                             "exclusive");
1109                         atomic_store_rel_ptr(&lk->lk_lock, v);
1110                         wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1111                             SLEEPQ_LK, 0, queue);
1112                         sleepq_release(&lk->lock_object);
1113                         break;
1114                 } else
1115                         wakeup_swapper = wakeupshlk(lk, file, line);
1116                 break;
1117         case LK_DRAIN:
1118                 if (LK_CAN_WITNESS(flags))
1119                         WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1120                             LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1121                             ilk : NULL);
1122
1123                 /*
1124                  * Trying to drain a lock we already own will result in a
1125                  * deadlock.
1126                  */
1127                 if (lockmgr_xlocked(lk)) {
1128                         if (flags & LK_INTERLOCK)
1129                                 class->lc_unlock(ilk);
1130                         panic("%s: draining %s with the lock held @ %s:%d\n",
1131                             __func__, iwmesg, file, line);
1132                 }
1133
1134                 for (;;) {
1135                         if (lk->lk_lock == LK_UNLOCKED &&
1136                             atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
1137                                 break;
1138
1139 #ifdef HWPMC_HOOKS
1140                         PMC_SOFT_CALL( , , lock, failed);
1141 #endif
1142                         lock_profile_obtain_lock_failed(&lk->lock_object,
1143                             &contested, &waittime);
1144
1145                         /*
1146                          * If the lock is expected to not sleep just give up
1147                          * and return.
1148                          */
1149                         if (LK_TRYOP(flags)) {
1150                                 LOCK_LOG2(lk, "%s: %p fails the try operation",
1151                                     __func__, lk);
1152                                 error = EBUSY;
1153                                 break;
1154                         }
1155
1156                         /*
1157                          * Acquire the sleepqueue chain lock because we
1158                          * probabilly will need to manipulate waiters flags.
1159                          */
1160                         sleepq_lock(&lk->lock_object);
1161                         x = lk->lk_lock;
1162
1163                         /*
1164                          * if the lock has been released while we spun on
1165                          * the sleepqueue chain lock just try again.
1166                          */
1167                         if (x == LK_UNLOCKED) {
1168                                 sleepq_release(&lk->lock_object);
1169                                 continue;
1170                         }
1171
1172                         v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1173                         if ((x & ~v) == LK_UNLOCKED) {
1174                                 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1175
1176                                 /*
1177                                  * If interruptible sleeps left the exclusive
1178                                  * queue empty avoid a starvation for the
1179                                  * threads sleeping on the shared queue by
1180                                  * giving them precedence and cleaning up the
1181                                  * exclusive waiters bit anyway.
1182                                  * Please note that lk_exslpfail count may be
1183                                  * lying about the real number of waiters with
1184                                  * the LK_SLEEPFAIL flag on because they may
1185                                  * be used in conjunction with interruptible
1186                                  * sleeps so lk_exslpfail might be considered
1187                                  * an 'upper limit' bound, including the edge
1188                                  * cases.
1189                                  */
1190                                 if (v & LK_EXCLUSIVE_WAITERS) {
1191                                         queue = SQ_EXCLUSIVE_QUEUE;
1192                                         v &= ~LK_EXCLUSIVE_WAITERS;
1193                                 } else {
1194
1195                                         /*
1196                                          * Exclusive waiters sleeping with
1197                                          * LK_SLEEPFAIL on and using
1198                                          * interruptible sleeps/timeout may
1199                                          * have left spourious lk_exslpfail
1200                                          * counts on, so clean it up anyway.
1201                                          */
1202                                         MPASS(v & LK_SHARED_WAITERS);
1203                                         lk->lk_exslpfail = 0;
1204                                         queue = SQ_SHARED_QUEUE;
1205                                         v &= ~LK_SHARED_WAITERS;
1206                                 }
1207                                 if (queue == SQ_EXCLUSIVE_QUEUE) {
1208                                         realexslp =
1209                                             sleepq_sleepcnt(&lk->lock_object,
1210                                             SQ_EXCLUSIVE_QUEUE);
1211                                         if (lk->lk_exslpfail >= realexslp) {
1212                                                 lk->lk_exslpfail = 0;
1213                                                 queue = SQ_SHARED_QUEUE;
1214                                                 v &= ~LK_SHARED_WAITERS;
1215                                                 if (realexslp != 0) {
1216                                                         LOCK_LOG2(lk,
1217                                         "%s: %p has only LK_SLEEPFAIL sleepers",
1218                                                             __func__, lk);
1219                                                         LOCK_LOG2(lk,
1220                         "%s: %p waking up threads on the exclusive queue",
1221                                                             __func__, lk);
1222                                                         wakeup_swapper =
1223                                                             sleepq_broadcast(
1224                                                             &lk->lock_object,
1225                                                             SLEEPQ_LK, 0,
1226                                                             SQ_EXCLUSIVE_QUEUE);
1227                                                 }
1228                                         } else
1229                                                 lk->lk_exslpfail = 0;
1230                                 }
1231                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1232                                         sleepq_release(&lk->lock_object);
1233                                         continue;
1234                                 }
1235                                 LOCK_LOG3(lk,
1236                                 "%s: %p waking up all threads on the %s queue",
1237                                     __func__, lk, queue == SQ_SHARED_QUEUE ?
1238                                     "shared" : "exclusive");
1239                                 wakeup_swapper |= sleepq_broadcast(
1240                                     &lk->lock_object, SLEEPQ_LK, 0, queue);
1241
1242                                 /*
1243                                  * If shared waiters have been woken up we need
1244                                  * to wait for one of them to acquire the lock
1245                                  * before to set the exclusive waiters in
1246                                  * order to avoid a deadlock.
1247                                  */
1248                                 if (queue == SQ_SHARED_QUEUE) {
1249                                         for (v = lk->lk_lock;
1250                                             (v & LK_SHARE) && !LK_SHARERS(v);
1251                                             v = lk->lk_lock)
1252                                                 cpu_spinwait();
1253                                 }
1254                         }
1255
1256                         /*
1257                          * Try to set the LK_EXCLUSIVE_WAITERS flag.  If we
1258                          * fail, loop back and retry.
1259                          */
1260                         if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1261                                 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1262                                     x | LK_EXCLUSIVE_WAITERS)) {
1263                                         sleepq_release(&lk->lock_object);
1264                                         continue;
1265                                 }
1266                                 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1267                                     __func__, lk);
1268                         }
1269
1270                         /*
1271                          * As far as we have been unable to acquire the
1272                          * exclusive lock and the exclusive waiters flag
1273                          * is set, we will sleep.
1274                          */
1275                         if (flags & LK_INTERLOCK) {
1276                                 class->lc_unlock(ilk);
1277                                 flags &= ~LK_INTERLOCK;
1278                         }
1279                         GIANT_SAVE();
1280                         sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1281                             SQ_EXCLUSIVE_QUEUE);
1282                         sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1283                         GIANT_RESTORE();
1284                         LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1285                             __func__, lk);
1286                 }
1287
1288                 if (error == 0) {
1289                         lock_profile_obtain_lock_success(&lk->lock_object,
1290                             contested, waittime, file, line);
1291                         LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1292                             lk->lk_recurse, file, line);
1293                         WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1294                             LK_TRYWIT(flags), file, line);
1295                         TD_LOCKS_INC(curthread);
1296                         STACK_SAVE(lk);
1297                 }
1298                 break;
1299         default:
1300                 if (flags & LK_INTERLOCK)
1301                         class->lc_unlock(ilk);
1302                 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1303         }
1304
1305         if (flags & LK_INTERLOCK)
1306                 class->lc_unlock(ilk);
1307         if (wakeup_swapper)
1308                 kick_proc0();
1309
1310         return (error);
1311 }
1312
1313 void
1314 _lockmgr_disown(struct lock *lk, const char *file, int line)
1315 {
1316         uintptr_t tid, x;
1317
1318         if (SCHEDULER_STOPPED())
1319                 return;
1320
1321         tid = (uintptr_t)curthread;
1322         _lockmgr_assert(lk, KA_XLOCKED, file, line);
1323
1324         /*
1325          * Panic if the lock is recursed.
1326          */
1327         if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1328                 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1329                     __func__,  file, line);
1330
1331         /*
1332          * If the owner is already LK_KERNPROC just skip the whole operation.
1333          */
1334         if (LK_HOLDER(lk->lk_lock) != tid)
1335                 return;
1336         lock_profile_release_lock(&lk->lock_object);
1337         LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1338         WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1339         TD_LOCKS_DEC(curthread);
1340         STACK_SAVE(lk);
1341
1342         /*
1343          * In order to preserve waiters flags, just spin.
1344          */
1345         for (;;) {
1346                 x = lk->lk_lock;
1347                 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1348                 x &= LK_ALL_WAITERS;
1349                 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1350                     LK_KERNPROC | x))
1351                         return;
1352                 cpu_spinwait();
1353         }
1354 }
1355
1356 void
1357 lockmgr_printinfo(const struct lock *lk)
1358 {
1359         struct thread *td;
1360         uintptr_t x;
1361
1362         if (lk->lk_lock == LK_UNLOCKED)
1363                 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1364         else if (lk->lk_lock & LK_SHARE)
1365                 printf("lock type %s: SHARED (count %ju)\n",
1366                     lk->lock_object.lo_name,
1367                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1368         else {
1369                 td = lockmgr_xholder(lk);
1370                 if (td == (struct thread *)LK_KERNPROC)
1371                         printf("lock type %s: EXCL by KERNPROC\n",
1372                             lk->lock_object.lo_name);
1373                 else
1374                         printf("lock type %s: EXCL by thread %p "
1375                             "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name,
1376                             td, td->td_proc->p_pid, td->td_proc->p_comm,
1377                             td->td_tid);
1378         }
1379
1380         x = lk->lk_lock;
1381         if (x & LK_EXCLUSIVE_WAITERS)
1382                 printf(" with exclusive waiters pending\n");
1383         if (x & LK_SHARED_WAITERS)
1384                 printf(" with shared waiters pending\n");
1385         if (x & LK_EXCLUSIVE_SPINNERS)
1386                 printf(" with exclusive spinners pending\n");
1387
1388         STACK_PRINT(lk);
1389 }
1390
1391 int
1392 lockstatus(const struct lock *lk)
1393 {
1394         uintptr_t v, x;
1395         int ret;
1396
1397         ret = LK_SHARED;
1398         x = lk->lk_lock;
1399         v = LK_HOLDER(x);
1400
1401         if ((x & LK_SHARE) == 0) {
1402                 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1403                         ret = LK_EXCLUSIVE;
1404                 else
1405                         ret = LK_EXCLOTHER;
1406         } else if (x == LK_UNLOCKED)
1407                 ret = 0;
1408
1409         return (ret);
1410 }
1411
1412 #ifdef INVARIANT_SUPPORT
1413
1414 FEATURE(invariant_support,
1415     "Support for modules compiled with INVARIANTS option");
1416
1417 #ifndef INVARIANTS
1418 #undef  _lockmgr_assert
1419 #endif
1420
1421 void
1422 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1423 {
1424         int slocked = 0;
1425
1426         if (panicstr != NULL)
1427                 return;
1428         switch (what) {
1429         case KA_SLOCKED:
1430         case KA_SLOCKED | KA_NOTRECURSED:
1431         case KA_SLOCKED | KA_RECURSED:
1432                 slocked = 1;
1433         case KA_LOCKED:
1434         case KA_LOCKED | KA_NOTRECURSED:
1435         case KA_LOCKED | KA_RECURSED:
1436 #ifdef WITNESS
1437
1438                 /*
1439                  * We cannot trust WITNESS if the lock is held in exclusive
1440                  * mode and a call to lockmgr_disown() happened.
1441                  * Workaround this skipping the check if the lock is held in
1442                  * exclusive mode even for the KA_LOCKED case.
1443                  */
1444                 if (slocked || (lk->lk_lock & LK_SHARE)) {
1445                         witness_assert(&lk->lock_object, what, file, line);
1446                         break;
1447                 }
1448 #endif
1449                 if (lk->lk_lock == LK_UNLOCKED ||
1450                     ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1451                     (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1452                         panic("Lock %s not %slocked @ %s:%d\n",
1453                             lk->lock_object.lo_name, slocked ? "share" : "",
1454                             file, line);
1455
1456                 if ((lk->lk_lock & LK_SHARE) == 0) {
1457                         if (lockmgr_recursed(lk)) {
1458                                 if (what & KA_NOTRECURSED)
1459                                         panic("Lock %s recursed @ %s:%d\n",
1460                                             lk->lock_object.lo_name, file,
1461                                             line);
1462                         } else if (what & KA_RECURSED)
1463                                 panic("Lock %s not recursed @ %s:%d\n",
1464                                     lk->lock_object.lo_name, file, line);
1465                 }
1466                 break;
1467         case KA_XLOCKED:
1468         case KA_XLOCKED | KA_NOTRECURSED:
1469         case KA_XLOCKED | KA_RECURSED:
1470                 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1471                         panic("Lock %s not exclusively locked @ %s:%d\n",
1472                             lk->lock_object.lo_name, file, line);
1473                 if (lockmgr_recursed(lk)) {
1474                         if (what & KA_NOTRECURSED)
1475                                 panic("Lock %s recursed @ %s:%d\n",
1476                                     lk->lock_object.lo_name, file, line);
1477                 } else if (what & KA_RECURSED)
1478                         panic("Lock %s not recursed @ %s:%d\n",
1479                             lk->lock_object.lo_name, file, line);
1480                 break;
1481         case KA_UNLOCKED:
1482                 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1483                         panic("Lock %s exclusively locked @ %s:%d\n",
1484                             lk->lock_object.lo_name, file, line);
1485                 break;
1486         default:
1487                 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1488                     line);
1489         }
1490 }
1491 #endif
1492
1493 #ifdef DDB
1494 int
1495 lockmgr_chain(struct thread *td, struct thread **ownerp)
1496 {
1497         struct lock *lk;
1498
1499         lk = td->td_wchan;
1500
1501         if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1502                 return (0);
1503         db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1504         if (lk->lk_lock & LK_SHARE)
1505                 db_printf("SHARED (count %ju)\n",
1506                     (uintmax_t)LK_SHARERS(lk->lk_lock));
1507         else
1508                 db_printf("EXCL\n");
1509         *ownerp = lockmgr_xholder(lk);
1510
1511         return (1);
1512 }
1513
1514 static void
1515 db_show_lockmgr(const struct lock_object *lock)
1516 {
1517         struct thread *td;
1518         const struct lock *lk;
1519
1520         lk = (const struct lock *)lock;
1521
1522         db_printf(" state: ");
1523         if (lk->lk_lock == LK_UNLOCKED)
1524                 db_printf("UNLOCKED\n");
1525         else if (lk->lk_lock & LK_SHARE)
1526                 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1527         else {
1528                 td = lockmgr_xholder(lk);
1529                 if (td == (struct thread *)LK_KERNPROC)
1530                         db_printf("XLOCK: LK_KERNPROC\n");
1531                 else
1532                         db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1533                             td->td_tid, td->td_proc->p_pid,
1534                             td->td_proc->p_comm);
1535                 if (lockmgr_recursed(lk))
1536                         db_printf(" recursed: %d\n", lk->lk_recurse);
1537         }
1538         db_printf(" waiters: ");
1539         switch (lk->lk_lock & LK_ALL_WAITERS) {
1540         case LK_SHARED_WAITERS:
1541                 db_printf("shared\n");
1542                 break;
1543         case LK_EXCLUSIVE_WAITERS:
1544                 db_printf("exclusive\n");
1545                 break;
1546         case LK_ALL_WAITERS:
1547                 db_printf("shared and exclusive\n");
1548                 break;
1549         default:
1550                 db_printf("none\n");
1551         }
1552         db_printf(" spinners: ");
1553         if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1554                 db_printf("exclusive\n");
1555         else
1556                 db_printf("none\n");
1557 }
1558 #endif