sys/kern/kern_rwlock.c

   1 /*-
   2  * Copyright (c) 2006 John Baldwin <jhb@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, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 /*
  28  * Machine independent bits of reader/writer lock implementation.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include "opt_ddb.h"
  35 #include "opt_hwpmc_hooks.h"
  36 #include "opt_no_adaptive_rwlocks.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/kdb.h>
  40 #include <sys/ktr.h>
  41 #include <sys/kernel.h>
  42 #include <sys/lock.h>
  43 #include <sys/mutex.h>
  44 #include <sys/proc.h>
  45 #include <sys/rwlock.h>
  46 #include <sys/sched.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/systm.h>
  49 #include <sys/turnstile.h>
  50
  51 #include <machine/cpu.h>
  52
  53 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
  54 #define ADAPTIVE_RWLOCKS
  55 #endif
  56
  57 #ifdef HWPMC_HOOKS
  58 #include <sys/pmckern.h>
  59 PMC_SOFT_DECLARE( , , lock, failed);
  60 #endif
  61
  62 /*
  63  * Return the rwlock address when the lock cookie address is provided.
  64  * This functionality assumes that struct rwlock* have a member named rw_lock.
  65  */
  66 #define rwlock2rw(c)    (__containerof(c, struct rwlock, rw_lock))
  67
  68 #ifdef ADAPTIVE_RWLOCKS
  69 static int rowner_retries = 10;
  70 static int rowner_loops = 10000;
  71 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
  72     "rwlock debugging");
  73 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
  74 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
  75 #endif
  76
  77 #ifdef DDB
  78 #include <ddb/ddb.h>
  79
  80 static void     db_show_rwlock(const struct lock_object *lock);
  81 #endif
  82 static void     assert_rw(const struct lock_object *lock, int what);
  83 static void     lock_rw(struct lock_object *lock, uintptr_t how);
  84 #ifdef KDTRACE_HOOKS
  85 static int      owner_rw(const struct lock_object *lock, struct thread **owner);
  86 #endif
  87 static uintptr_t unlock_rw(struct lock_object *lock);
  88
  89 struct lock_class lock_class_rw = {
  90         .lc_name = "rw",
  91         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
  92         .lc_assert = assert_rw,
  93 #ifdef DDB
  94         .lc_ddb_show = db_show_rwlock,
  95 #endif
  96         .lc_lock = lock_rw,
  97         .lc_unlock = unlock_rw,
  98 #ifdef KDTRACE_HOOKS
  99         .lc_owner = owner_rw,
 100 #endif
 101 };
 102
 103 /*
 104  * Return a pointer to the owning thread if the lock is write-locked or
 105  * NULL if the lock is unlocked or read-locked.
 106  */
 107 #define rw_wowner(rw)                                                   \
 108         ((rw)->rw_lock & RW_LOCK_READ ? NULL :                          \
 109             (struct thread *)RW_OWNER((rw)->rw_lock))
 110
 111 /*
 112  * Returns if a write owner is recursed.  Write ownership is not assured
 113  * here and should be previously checked.
 114  */
 115 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 116
 117 /*
 118  * Return true if curthread helds the lock.
 119  */
 120 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 121
 122 /*
 123  * Return a pointer to the owning thread for this lock who should receive
 124  * any priority lent by threads that block on this lock.  Currently this
 125  * is identical to rw_wowner().
 126  */
 127 #define rw_owner(rw)            rw_wowner(rw)
 128
 129 #ifndef INVARIANTS
 130 #define __rw_assert(c, what, file, line)
 131 #endif
 132
 133 void
 134 assert_rw(const struct lock_object *lock, int what)
 135 {
 136
 137         rw_assert((const struct rwlock *)lock, what);
 138 }
 139
 140 void
 141 lock_rw(struct lock_object *lock, uintptr_t how)
 142 {
 143         struct rwlock *rw;
 144
 145         rw = (struct rwlock *)lock;
 146         if (how)
 147                 rw_rlock(rw);
 148         else
 149                 rw_wlock(rw);
 150 }
 151
 152 uintptr_t
 153 unlock_rw(struct lock_object *lock)
 154 {
 155         struct rwlock *rw;
 156
 157         rw = (struct rwlock *)lock;
 158         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 159         if (rw->rw_lock & RW_LOCK_READ) {
 160                 rw_runlock(rw);
 161                 return (1);
 162         } else {
 163                 rw_wunlock(rw);
 164                 return (0);
 165         }
 166 }
 167
 168 #ifdef KDTRACE_HOOKS
 169 int
 170 owner_rw(const struct lock_object *lock, struct thread **owner)
 171 {
 172         const struct rwlock *rw = (const struct rwlock *)lock;
 173         uintptr_t x = rw->rw_lock;
 174
 175         *owner = rw_wowner(rw);
 176         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 177             (*owner != NULL));
 178 }
 179 #endif
 180
 181 void
 182 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
 183 {
 184         struct rwlock *rw;
 185         int flags;
 186
 187         rw = rwlock2rw(c);
 188
 189         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 190             RW_RECURSE | RW_NEW)) == 0);
 191         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 192             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 193             &rw->rw_lock));
 194
 195         flags = LO_UPGRADABLE;
 196         if (opts & RW_DUPOK)
 197                 flags |= LO_DUPOK;
 198         if (opts & RW_NOPROFILE)
 199                 flags |= LO_NOPROFILE;
 200         if (!(opts & RW_NOWITNESS))
 201                 flags |= LO_WITNESS;
 202         if (opts & RW_RECURSE)
 203                 flags |= LO_RECURSABLE;
 204         if (opts & RW_QUIET)
 205                 flags |= LO_QUIET;
 206         if (opts & RW_NEW)
 207                 flags |= LO_NEW;
 208
 209         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 210         rw->rw_lock = RW_UNLOCKED;
 211         rw->rw_recurse = 0;
 212 }
 213
 214 void
 215 _rw_destroy(volatile uintptr_t *c)
 216 {
 217         struct rwlock *rw;
 218
 219         rw = rwlock2rw(c);
 220
 221         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 222         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 223         rw->rw_lock = RW_DESTROYED;
 224         lock_destroy(&rw->lock_object);
 225 }
 226
 227 void
 228 rw_sysinit(void *arg)
 229 {
 230         struct rw_args *args = arg;
 231
 232         rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
 233 }
 234
 235 void
 236 rw_sysinit_flags(void *arg)
 237 {
 238         struct rw_args_flags *args = arg;
 239
 240         rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
 241             args->ra_flags);
 242 }
 243
 244 int
 245 _rw_wowned(const volatile uintptr_t *c)
 246 {
 247
 248         return (rw_wowner(rwlock2rw(c)) == curthread);
 249 }
 250
 251 void
 252 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
 253 {
 254         struct rwlock *rw;
 255
 256         if (SCHEDULER_STOPPED())
 257                 return;
 258
 259         rw = rwlock2rw(c);
 260
 261         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 262             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
 263             curthread, rw->lock_object.lo_name, file, line));
 264         KASSERT(rw->rw_lock != RW_DESTROYED,
 265             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 266         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 267             line, NULL);
 268         __rw_wlock(rw, curthread, file, line);
 269         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 270         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 271         curthread->td_locks++;
 272 }
 273
 274 int
 275 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
 276 {
 277         struct rwlock *rw;
 278         int rval;
 279
 280         if (SCHEDULER_STOPPED())
 281                 return (1);
 282
 283         rw = rwlock2rw(c);
 284
 285         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 286             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
 287             curthread, rw->lock_object.lo_name, file, line));
 288         KASSERT(rw->rw_lock != RW_DESTROYED,
 289             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 290
 291         if (rw_wlocked(rw) &&
 292             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 293                 rw->rw_recurse++;
 294                 rval = 1;
 295         } else
 296                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 297                     (uintptr_t)curthread);
 298
 299         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 300         if (rval) {
 301                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 302                     file, line);
 303                 curthread->td_locks++;
 304         }
 305         return (rval);
 306 }
 307
 308 void
 309 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 310 {
 311         struct rwlock *rw;
 312
 313         if (SCHEDULER_STOPPED())
 314                 return;
 315
 316         rw = rwlock2rw(c);
 317
 318         KASSERT(rw->rw_lock != RW_DESTROYED,
 319             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 320         __rw_assert(c, RA_WLOCKED, file, line);
 321         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 322         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 323             line);
 324         __rw_wunlock(rw, curthread, file, line);
 325         curthread->td_locks--;
 326 }
 327 /*
 328  * Determines whether a new reader can acquire a lock.  Succeeds if the
 329  * reader already owns a read lock and the lock is locked for read to
 330  * prevent deadlock from reader recursion.  Also succeeds if the lock
 331  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 332  * prioritizes writers before readers.
 333  */
 334 #define RW_CAN_READ(_rw)                                                \
 335     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 336     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 337     RW_LOCK_READ)
 338
 339 void
 340 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 341 {
 342         struct rwlock *rw;
 343         struct turnstile *ts;
 344 #ifdef ADAPTIVE_RWLOCKS
 345         volatile struct thread *owner;
 346         int spintries = 0;
 347         int i;
 348 #endif
 349 #ifdef LOCK_PROFILING
 350         uint64_t waittime = 0;
 351         int contested = 0;
 352 #endif
 353         uintptr_t v;
 354 #ifdef KDTRACE_HOOKS
 355         uint64_t spin_cnt = 0;
 356         uint64_t sleep_cnt = 0;
 357         int64_t sleep_time = 0;
 358 #endif
 359
 360         if (SCHEDULER_STOPPED())
 361                 return;
 362
 363         rw = rwlock2rw(c);
 364
 365         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 366             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 367             curthread, rw->lock_object.lo_name, file, line));
 368         KASSERT(rw->rw_lock != RW_DESTROYED,
 369             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 370         KASSERT(rw_wowner(rw) != curthread,
 371             ("rw_rlock: wlock already held for %s @ %s:%d",
 372             rw->lock_object.lo_name, file, line));
 373         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 374
 375         for (;;) {
 376 #ifdef KDTRACE_HOOKS
 377                 spin_cnt++;
 378 #endif
 379                 /*
 380                  * Handle the easy case.  If no other thread has a write
 381                  * lock, then try to bump up the count of read locks.  Note
 382                  * that we have to preserve the current state of the
 383                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 384                  * read lock, then rw_lock must have changed, so restart
 385                  * the loop.  Note that this handles the case of a
 386                  * completely unlocked rwlock since such a lock is encoded
 387                  * as a read lock with no waiters.
 388                  */
 389                 v = rw->rw_lock;
 390                 if (RW_CAN_READ(v)) {
 391                         /*
 392                          * The RW_LOCK_READ_WAITERS flag should only be set
 393                          * if the lock has been unlocked and write waiters
 394                          * were present.
 395                          */
 396                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 397                             v + RW_ONE_READER)) {
 398                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 399                                         CTR4(KTR_LOCK,
 400                                             "%s: %p succeed %p -> %p", __func__,
 401                                             rw, (void *)v,
 402                                             (void *)(v + RW_ONE_READER));
 403                                 break;
 404                         }
 405                         continue;
 406                 }
 407 #ifdef HWPMC_HOOKS
 408                 PMC_SOFT_CALL( , , lock, failed);
 409 #endif
 410                 lock_profile_obtain_lock_failed(&rw->lock_object,
 411                     &contested, &waittime);
 412
 413 #ifdef ADAPTIVE_RWLOCKS
 414                 /*
 415                  * If the owner is running on another CPU, spin until
 416                  * the owner stops running or the state of the lock
 417                  * changes.
 418                  */
 419                 if ((v & RW_LOCK_READ) == 0) {
 420                         owner = (struct thread *)RW_OWNER(v);
 421                         if (TD_IS_RUNNING(owner)) {
 422                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 423                                         CTR3(KTR_LOCK,
 424                                             "%s: spinning on %p held by %p",
 425                                             __func__, rw, owner);
 426                                 KTR_STATE1(KTR_SCHED, "thread",
 427                                     sched_tdname(curthread), "spinning",
 428                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 429                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 430                                     owner && TD_IS_RUNNING(owner)) {
 431                                         cpu_spinwait();
 432 #ifdef KDTRACE_HOOKS
 433                                         spin_cnt++;
 434 #endif
 435                                 }
 436                                 KTR_STATE0(KTR_SCHED, "thread",
 437                                     sched_tdname(curthread), "running");
 438                                 continue;
 439                         }
 440                 } else if (spintries < rowner_retries) {
 441                         spintries++;
 442                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 443                             "spinning", "lockname:\"%s\"",
 444                             rw->lock_object.lo_name);
 445                         for (i = 0; i < rowner_loops; i++) {
 446                                 v = rw->rw_lock;
 447                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 448                                         break;
 449                                 cpu_spinwait();
 450                         }
 451 #ifdef KDTRACE_HOOKS
 452                         spin_cnt += rowner_loops - i;
 453 #endif
 454                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 455                             "running");
 456                         if (i != rowner_loops)
 457                                 continue;
 458                 }
 459 #endif
 460
 461                 /*
 462                  * Okay, now it's the hard case.  Some other thread already
 463                  * has a write lock or there are write waiters present,
 464                  * acquire the turnstile lock so we can begin the process
 465                  * of blocking.
 466                  */
 467                 ts = turnstile_trywait(&rw->lock_object);
 468
 469                 /*
 470                  * The lock might have been released while we spun, so
 471                  * recheck its state and restart the loop if needed.
 472                  */
 473                 v = rw->rw_lock;
 474                 if (RW_CAN_READ(v)) {
 475                         turnstile_cancel(ts);
 476                         continue;
 477                 }
 478
 479 #ifdef ADAPTIVE_RWLOCKS
 480                 /*
 481                  * The current lock owner might have started executing
 482                  * on another CPU (or the lock could have changed
 483                  * owners) while we were waiting on the turnstile
 484                  * chain lock.  If so, drop the turnstile lock and try
 485                  * again.
 486                  */
 487                 if ((v & RW_LOCK_READ) == 0) {
 488                         owner = (struct thread *)RW_OWNER(v);
 489                         if (TD_IS_RUNNING(owner)) {
 490                                 turnstile_cancel(ts);
 491                                 continue;
 492                         }
 493                 }
 494 #endif
 495
 496                 /*
 497                  * The lock is held in write mode or it already has waiters.
 498                  */
 499                 MPASS(!RW_CAN_READ(v));
 500
 501                 /*
 502                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 503                  * we can go ahead and block.  If it is not set then try
 504                  * to set it.  If we fail to set it drop the turnstile
 505                  * lock and restart the loop.
 506                  */
 507                 if (!(v & RW_LOCK_READ_WAITERS)) {
 508                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 509                             v | RW_LOCK_READ_WAITERS)) {
 510                                 turnstile_cancel(ts);
 511                                 continue;
 512                         }
 513                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 514                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 515                                     __func__, rw);
 516                 }
 517
 518                 /*
 519                  * We were unable to acquire the lock and the read waiters
 520                  * flag is set, so we must block on the turnstile.
 521                  */
 522                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 523                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 524                             rw);
 525 #ifdef KDTRACE_HOOKS
 526                 sleep_time -= lockstat_nsecs();
 527 #endif
 528                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 529 #ifdef KDTRACE_HOOKS
 530                 sleep_time += lockstat_nsecs();
 531                 sleep_cnt++;
 532 #endif
 533                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 534                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 535                             __func__, rw);
 536         }
 537
 538         /*
 539          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 540          * however.  turnstiles don't like owners changing between calls to
 541          * turnstile_wait() currently.
 542          */
 543         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
 544             waittime, file, line);
 545         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 546         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 547         curthread->td_locks++;
 548         curthread->td_rw_rlocks++;
 549 #ifdef KDTRACE_HOOKS
 550         if (sleep_time)
 551                 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
 552
 553         /*
 554          * Record only the loops spinning and not sleeping.
 555          */
 556         if (spin_cnt > sleep_cnt)
 557                 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 558 #endif
 559 }
 560
 561 int
 562 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 563 {
 564         struct rwlock *rw;
 565         uintptr_t x;
 566
 567         if (SCHEDULER_STOPPED())
 568                 return (1);
 569
 570         rw = rwlock2rw(c);
 571
 572         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 573             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 574             curthread, rw->lock_object.lo_name, file, line));
 575
 576         for (;;) {
 577                 x = rw->rw_lock;
 578                 KASSERT(rw->rw_lock != RW_DESTROYED,
 579                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 580                 if (!(x & RW_LOCK_READ))
 581                         break;
 582                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 583                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 584                             line);
 585                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 586                         curthread->td_locks++;
 587                         curthread->td_rw_rlocks++;
 588                         return (1);
 589                 }
 590         }
 591
 592         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 593         return (0);
 594 }
 595
 596 void
 597 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 598 {
 599         struct rwlock *rw;
 600         struct turnstile *ts;
 601         uintptr_t x, v, queue;
 602
 603         if (SCHEDULER_STOPPED())
 604                 return;
 605
 606         rw = rwlock2rw(c);
 607
 608         KASSERT(rw->rw_lock != RW_DESTROYED,
 609             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 610         __rw_assert(c, RA_RLOCKED, file, line);
 611         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 612         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 613
 614         /* TODO: drop "owner of record" here. */
 615
 616         for (;;) {
 617                 /*
 618                  * See if there is more than one read lock held.  If so,
 619                  * just drop one and return.
 620                  */
 621                 x = rw->rw_lock;
 622                 if (RW_READERS(x) > 1) {
 623                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 624                             x - RW_ONE_READER)) {
 625                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 626                                         CTR4(KTR_LOCK,
 627                                             "%s: %p succeeded %p -> %p",
 628                                             __func__, rw, (void *)x,
 629                                             (void *)(x - RW_ONE_READER));
 630                                 break;
 631                         }
 632                         continue;
 633                 }
 634                 /*
 635                  * If there aren't any waiters for a write lock, then try
 636                  * to drop it quickly.
 637                  */
 638                 if (!(x & RW_LOCK_WAITERS)) {
 639                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 640                             RW_READERS_LOCK(1));
 641                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 642                             RW_UNLOCKED)) {
 643                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 644                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 645                                             __func__, rw);
 646                                 break;
 647                         }
 648                         continue;
 649                 }
 650                 /*
 651                  * Ok, we know we have waiters and we think we are the
 652                  * last reader, so grab the turnstile lock.
 653                  */
 654                 turnstile_chain_lock(&rw->lock_object);
 655                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 656                 MPASS(v & RW_LOCK_WAITERS);
 657
 658                 /*
 659                  * Try to drop our lock leaving the lock in a unlocked
 660                  * state.
 661                  *
 662                  * If you wanted to do explicit lock handoff you'd have to
 663                  * do it here.  You'd also want to use turnstile_signal()
 664                  * and you'd have to handle the race where a higher
 665                  * priority thread blocks on the write lock before the
 666                  * thread you wakeup actually runs and have the new thread
 667                  * "steal" the lock.  For now it's a lot simpler to just
 668                  * wakeup all of the waiters.
 669                  *
 670                  * As above, if we fail, then another thread might have
 671                  * acquired a read lock, so drop the turnstile lock and
 672                  * restart.
 673                  */
 674                 x = RW_UNLOCKED;
 675                 if (v & RW_LOCK_WRITE_WAITERS) {
 676                         queue = TS_EXCLUSIVE_QUEUE;
 677                         x |= (v & RW_LOCK_READ_WAITERS);
 678                 } else
 679                         queue = TS_SHARED_QUEUE;
 680                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 681                     x)) {
 682                         turnstile_chain_unlock(&rw->lock_object);
 683                         continue;
 684                 }
 685                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 686                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 687                             __func__, rw);
 688
 689                 /*
 690                  * Ok.  The lock is released and all that's left is to
 691                  * wake up the waiters.  Note that the lock might not be
 692                  * free anymore, but in that case the writers will just
 693                  * block again if they run before the new lock holder(s)
 694                  * release the lock.
 695                  */
 696                 ts = turnstile_lookup(&rw->lock_object);
 697                 MPASS(ts != NULL);
 698                 turnstile_broadcast(ts, queue);
 699                 turnstile_unpend(ts, TS_SHARED_LOCK);
 700                 turnstile_chain_unlock(&rw->lock_object);
 701                 break;
 702         }
 703         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
 704         curthread->td_locks--;
 705         curthread->td_rw_rlocks--;
 706 }
 707
 708 /*
 709  * This function is called when we are unable to obtain a write lock on the
 710  * first try.  This means that at least one other thread holds either a
 711  * read or write lock.
 712  */
 713 void
 714 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 715     int line)
 716 {
 717         struct rwlock *rw;
 718         struct turnstile *ts;
 719 #ifdef ADAPTIVE_RWLOCKS
 720         volatile struct thread *owner;
 721         int spintries = 0;
 722         int i;
 723 #endif
 724         uintptr_t v, x;
 725 #ifdef LOCK_PROFILING
 726         uint64_t waittime = 0;
 727         int contested = 0;
 728 #endif
 729 #ifdef KDTRACE_HOOKS
 730         uint64_t spin_cnt = 0;
 731         uint64_t sleep_cnt = 0;
 732         int64_t sleep_time = 0;
 733 #endif
 734
 735         if (SCHEDULER_STOPPED())
 736                 return;
 737
 738         rw = rwlock2rw(c);
 739
 740         if (rw_wlocked(rw)) {
 741                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 742                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 743                     __func__, rw->lock_object.lo_name, file, line));
 744                 rw->rw_recurse++;
 745                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 746                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 747                 return;
 748         }
 749
 750         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 751                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 752                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 753
 754         while (!_rw_write_lock(rw, tid)) {
 755 #ifdef KDTRACE_HOOKS
 756                 spin_cnt++;
 757 #endif
 758 #ifdef HWPMC_HOOKS
 759                 PMC_SOFT_CALL( , , lock, failed);
 760 #endif
 761                 lock_profile_obtain_lock_failed(&rw->lock_object,
 762                     &contested, &waittime);
 763 #ifdef ADAPTIVE_RWLOCKS
 764                 /*
 765                  * If the lock is write locked and the owner is
 766                  * running on another CPU, spin until the owner stops
 767                  * running or the state of the lock changes.
 768                  */
 769                 v = rw->rw_lock;
 770                 owner = (struct thread *)RW_OWNER(v);
 771                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 772                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 773                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 774                                     __func__, rw, owner);
 775                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 776                             "spinning", "lockname:\"%s\"",
 777                             rw->lock_object.lo_name);
 778                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 779                             TD_IS_RUNNING(owner)) {
 780                                 cpu_spinwait();
 781 #ifdef KDTRACE_HOOKS
 782                                 spin_cnt++;
 783 #endif
 784                         }
 785                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 786                             "running");
 787                         continue;
 788                 }
 789                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 790                     spintries < rowner_retries) {
 791                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 792                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 793                                     v | RW_LOCK_WRITE_SPINNER)) {
 794                                         continue;
 795                                 }
 796                         }
 797                         spintries++;
 798                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 799                             "spinning", "lockname:\"%s\"",
 800                             rw->lock_object.lo_name);
 801                         for (i = 0; i < rowner_loops; i++) {
 802                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 803                                         break;
 804                                 cpu_spinwait();
 805                         }
 806                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 807                             "running");
 808 #ifdef KDTRACE_HOOKS
 809                         spin_cnt += rowner_loops - i;
 810 #endif
 811                         if (i != rowner_loops)
 812                                 continue;
 813                 }
 814 #endif
 815                 ts = turnstile_trywait(&rw->lock_object);
 816                 v = rw->rw_lock;
 817
 818 #ifdef ADAPTIVE_RWLOCKS
 819                 /*
 820                  * The current lock owner might have started executing
 821                  * on another CPU (or the lock could have changed
 822                  * owners) while we were waiting on the turnstile
 823                  * chain lock.  If so, drop the turnstile lock and try
 824                  * again.
 825                  */
 826                 if (!(v & RW_LOCK_READ)) {
 827                         owner = (struct thread *)RW_OWNER(v);
 828                         if (TD_IS_RUNNING(owner)) {
 829                                 turnstile_cancel(ts);
 830                                 continue;
 831                         }
 832                 }
 833 #endif
 834                 /*
 835                  * Check for the waiters flags about this rwlock.
 836                  * If the lock was released, without maintain any pending
 837                  * waiters queue, simply try to acquire it.
 838                  * If a pending waiters queue is present, claim the lock
 839                  * ownership and maintain the pending queue.
 840                  */
 841                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 842                 if ((v & ~x) == RW_UNLOCKED) {
 843                         x &= ~RW_LOCK_WRITE_SPINNER;
 844                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 845                                 if (x)
 846                                         turnstile_claim(ts);
 847                                 else
 848                                         turnstile_cancel(ts);
 849                                 break;
 850                         }
 851                         turnstile_cancel(ts);
 852                         continue;
 853                 }
 854                 /*
 855                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 856                  * set it.  If we fail to set it, then loop back and try
 857                  * again.
 858                  */
 859                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 860                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 861                             v | RW_LOCK_WRITE_WAITERS)) {
 862                                 turnstile_cancel(ts);
 863                                 continue;
 864                         }
 865                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 866                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 867                                     __func__, rw);
 868                 }
 869                 /*
 870                  * We were unable to acquire the lock and the write waiters
 871                  * flag is set, so we must block on the turnstile.
 872                  */
 873                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 874                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 875                             rw);
 876 #ifdef KDTRACE_HOOKS
 877                 sleep_time -= lockstat_nsecs();
 878 #endif
 879                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 880 #ifdef KDTRACE_HOOKS
 881                 sleep_time += lockstat_nsecs();
 882                 sleep_cnt++;
 883 #endif
 884                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 885                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 886                             __func__, rw);
 887 #ifdef ADAPTIVE_RWLOCKS
 888                 spintries = 0;
 889 #endif
 890         }
 891         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
 892             waittime, file, line);
 893 #ifdef KDTRACE_HOOKS
 894         if (sleep_time)
 895                 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
 896
 897         /*
 898          * Record only the loops spinning and not sleeping.
 899          */
 900         if (spin_cnt > sleep_cnt)
 901                 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 902 #endif
 903 }
 904
 905 /*
 906  * This function is called if the first try at releasing a write lock failed.
 907  * This means that one of the 2 waiter bits must be set indicating that at
 908  * least one thread is waiting on this lock.
 909  */
 910 void
 911 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 912     int line)
 913 {
 914         struct rwlock *rw;
 915         struct turnstile *ts;
 916         uintptr_t v;
 917         int queue;
 918
 919         if (SCHEDULER_STOPPED())
 920                 return;
 921
 922         rw = rwlock2rw(c);
 923
 924         if (rw_wlocked(rw) && rw_recursed(rw)) {
 925                 rw->rw_recurse--;
 926                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 927                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 928                 return;
 929         }
 930
 931         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 932             ("%s: neither of the waiter flags are set", __func__));
 933
 934         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 935                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 936
 937         turnstile_chain_lock(&rw->lock_object);
 938         ts = turnstile_lookup(&rw->lock_object);
 939         MPASS(ts != NULL);
 940
 941         /*
 942          * Use the same algo as sx locks for now.  Prefer waking up shared
 943          * waiters if we have any over writers.  This is probably not ideal.
 944          *
 945          * 'v' is the value we are going to write back to rw_lock.  If we
 946          * have waiters on both queues, we need to preserve the state of
 947          * the waiter flag for the queue we don't wake up.  For now this is
 948          * hardcoded for the algorithm mentioned above.
 949          *
 950          * In the case of both readers and writers waiting we wakeup the
 951          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
 952          * new writer comes in before a reader it will claim the lock up
 953          * above.  There is probably a potential priority inversion in
 954          * there that could be worked around either by waking both queues
 955          * of waiters or doing some complicated lock handoff gymnastics.
 956          */
 957         v = RW_UNLOCKED;
 958         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
 959                 queue = TS_EXCLUSIVE_QUEUE;
 960                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
 961         } else
 962                 queue = TS_SHARED_QUEUE;
 963
 964         /* Wake up all waiters for the specific queue. */
 965         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 966                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
 967                     queue == TS_SHARED_QUEUE ? "read" : "write");
 968         turnstile_broadcast(ts, queue);
 969         atomic_store_rel_ptr(&rw->rw_lock, v);
 970         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
 971         turnstile_chain_unlock(&rw->lock_object);
 972 }
 973
 974 /*
 975  * Attempt to do a non-blocking upgrade from a read lock to a write
 976  * lock.  This will only succeed if this thread holds a single read
 977  * lock.  Returns true if the upgrade succeeded and false otherwise.
 978  */
 979 int
 980 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
 981 {
 982         struct rwlock *rw;
 983         uintptr_t v, x, tid;
 984         struct turnstile *ts;
 985         int success;
 986
 987         if (SCHEDULER_STOPPED())
 988                 return (1);
 989
 990         rw = rwlock2rw(c);
 991
 992         KASSERT(rw->rw_lock != RW_DESTROYED,
 993             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
 994         __rw_assert(c, RA_RLOCKED, file, line);
 995
 996         /*
 997          * Attempt to switch from one reader to a writer.  If there
 998          * are any write waiters, then we will have to lock the
 999          * turnstile first to prevent races with another writer
1000          * calling turnstile_wait() before we have claimed this
1001          * turnstile.  So, do the simple case of no waiters first.
1002          */
1003         tid = (uintptr_t)curthread;
1004         success = 0;
1005         for (;;) {
1006                 v = rw->rw_lock;
1007                 if (RW_READERS(v) > 1)
1008                         break;
1009                 if (!(v & RW_LOCK_WAITERS)) {
1010                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
1011                         if (!success)
1012                                 continue;
1013                         break;
1014                 }
1015
1016                 /*
1017                  * Ok, we think we have waiters, so lock the turnstile.
1018                  */
1019                 ts = turnstile_trywait(&rw->lock_object);
1020                 v = rw->rw_lock;
1021                 if (RW_READERS(v) > 1) {
1022                         turnstile_cancel(ts);
1023                         break;
1024                 }
1025                 /*
1026                  * Try to switch from one reader to a writer again.  This time
1027                  * we honor the current state of the waiters flags.
1028                  * If we obtain the lock with the flags set, then claim
1029                  * ownership of the turnstile.
1030                  */
1031                 x = rw->rw_lock & RW_LOCK_WAITERS;
1032                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1033                 if (success) {
1034                         if (x)
1035                                 turnstile_claim(ts);
1036                         else
1037                                 turnstile_cancel(ts);
1038                         break;
1039                 }
1040                 turnstile_cancel(ts);
1041         }
1042         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1043         if (success) {
1044                 curthread->td_rw_rlocks--;
1045                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1046                     file, line);
1047                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
1048         }
1049         return (success);
1050 }
1051
1052 /*
1053  * Downgrade a write lock into a single read lock.
1054  */
1055 void
1056 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1057 {
1058         struct rwlock *rw;
1059         struct turnstile *ts;
1060         uintptr_t tid, v;
1061         int rwait, wwait;
1062
1063         if (SCHEDULER_STOPPED())
1064                 return;
1065
1066         rw = rwlock2rw(c);
1067
1068         KASSERT(rw->rw_lock != RW_DESTROYED,
1069             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1070         __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1071 #ifndef INVARIANTS
1072         if (rw_recursed(rw))
1073                 panic("downgrade of a recursed lock");
1074 #endif
1075
1076         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1077
1078         /*
1079          * Convert from a writer to a single reader.  First we handle
1080          * the easy case with no waiters.  If there are any waiters, we
1081          * lock the turnstile and "disown" the lock.
1082          */
1083         tid = (uintptr_t)curthread;
1084         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1085                 goto out;
1086
1087         /*
1088          * Ok, we think we have waiters, so lock the turnstile so we can
1089          * read the waiter flags without any races.
1090          */
1091         turnstile_chain_lock(&rw->lock_object);
1092         v = rw->rw_lock & RW_LOCK_WAITERS;
1093         rwait = v & RW_LOCK_READ_WAITERS;
1094         wwait = v & RW_LOCK_WRITE_WAITERS;
1095         MPASS(rwait | wwait);
1096
1097         /*
1098          * Downgrade from a write lock while preserving waiters flag
1099          * and give up ownership of the turnstile.
1100          */
1101         ts = turnstile_lookup(&rw->lock_object);
1102         MPASS(ts != NULL);
1103         if (!wwait)
1104                 v &= ~RW_LOCK_READ_WAITERS;
1105         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1106         /*
1107          * Wake other readers if there are no writers pending.  Otherwise they
1108          * won't be able to acquire the lock anyway.
1109          */
1110         if (rwait && !wwait) {
1111                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1112                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1113         } else
1114                 turnstile_disown(ts);
1115         turnstile_chain_unlock(&rw->lock_object);
1116 out:
1117         curthread->td_rw_rlocks++;
1118         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1119         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1120 }
1121
1122 #ifdef INVARIANT_SUPPORT
1123 #ifndef INVARIANTS
1124 #undef __rw_assert
1125 #endif
1126
1127 /*
1128  * In the non-WITNESS case, rw_assert() can only detect that at least
1129  * *some* thread owns an rlock, but it cannot guarantee that *this*
1130  * thread owns an rlock.
1131  */
1132 void
1133 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1134 {
1135         const struct rwlock *rw;
1136
1137         if (panicstr != NULL)
1138                 return;
1139
1140         rw = rwlock2rw(c);
1141
1142         switch (what) {
1143         case RA_LOCKED:
1144         case RA_LOCKED | RA_RECURSED:
1145         case RA_LOCKED | RA_NOTRECURSED:
1146         case RA_RLOCKED:
1147         case RA_RLOCKED | RA_RECURSED:
1148         case RA_RLOCKED | RA_NOTRECURSED:
1149 #ifdef WITNESS
1150                 witness_assert(&rw->lock_object, what, file, line);
1151 #else
1152                 /*
1153                  * If some other thread has a write lock or we have one
1154                  * and are asserting a read lock, fail.  Also, if no one
1155                  * has a lock at all, fail.
1156                  */
1157                 if (rw->rw_lock == RW_UNLOCKED ||
1158                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1159                     rw_wowner(rw) != curthread)))
1160                         panic("Lock %s not %slocked @ %s:%d\n",
1161                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1162                             "read " : "", file, line);
1163
1164                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1165                         if (rw_recursed(rw)) {
1166                                 if (what & RA_NOTRECURSED)
1167                                         panic("Lock %s recursed @ %s:%d\n",
1168                                             rw->lock_object.lo_name, file,
1169                                             line);
1170                         } else if (what & RA_RECURSED)
1171                                 panic("Lock %s not recursed @ %s:%d\n",
1172                                     rw->lock_object.lo_name, file, line);
1173                 }
1174 #endif
1175                 break;
1176         case RA_WLOCKED:
1177         case RA_WLOCKED | RA_RECURSED:
1178         case RA_WLOCKED | RA_NOTRECURSED:
1179                 if (rw_wowner(rw) != curthread)
1180                         panic("Lock %s not exclusively locked @ %s:%d\n",
1181                             rw->lock_object.lo_name, file, line);
1182                 if (rw_recursed(rw)) {
1183                         if (what & RA_NOTRECURSED)
1184                                 panic("Lock %s recursed @ %s:%d\n",
1185                                     rw->lock_object.lo_name, file, line);
1186                 } else if (what & RA_RECURSED)
1187                         panic("Lock %s not recursed @ %s:%d\n",
1188                             rw->lock_object.lo_name, file, line);
1189                 break;
1190         case RA_UNLOCKED:
1191 #ifdef WITNESS
1192                 witness_assert(&rw->lock_object, what, file, line);
1193 #else
1194                 /*
1195                  * If we hold a write lock fail.  We can't reliably check
1196                  * to see if we hold a read lock or not.
1197                  */
1198                 if (rw_wowner(rw) == curthread)
1199                         panic("Lock %s exclusively locked @ %s:%d\n",
1200                             rw->lock_object.lo_name, file, line);
1201 #endif
1202                 break;
1203         default:
1204                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1205                     line);
1206         }
1207 }
1208 #endif /* INVARIANT_SUPPORT */
1209
1210 #ifdef DDB
1211 void
1212 db_show_rwlock(const struct lock_object *lock)
1213 {
1214         const struct rwlock *rw;
1215         struct thread *td;
1216
1217         rw = (const struct rwlock *)lock;
1218
1219         db_printf(" state: ");
1220         if (rw->rw_lock == RW_UNLOCKED)
1221                 db_printf("UNLOCKED\n");
1222         else if (rw->rw_lock == RW_DESTROYED) {
1223                 db_printf("DESTROYED\n");
1224                 return;
1225         } else if (rw->rw_lock & RW_LOCK_READ)
1226                 db_printf("RLOCK: %ju locks\n",
1227                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1228         else {
1229                 td = rw_wowner(rw);
1230                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1231                     td->td_tid, td->td_proc->p_pid, td->td_name);
1232                 if (rw_recursed(rw))
1233                         db_printf(" recursed: %u\n", rw->rw_recurse);
1234         }
1235         db_printf(" waiters: ");
1236         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1237         case RW_LOCK_READ_WAITERS:
1238                 db_printf("readers\n");
1239                 break;
1240         case RW_LOCK_WRITE_WAITERS:
1241                 db_printf("writers\n");
1242                 break;
1243         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1244                 db_printf("readers and writers\n");
1245                 break;
1246         default:
1247                 db_printf("none\n");
1248                 break;
1249         }
1250 }
1251
1252 #endif