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                 if (!rw_recursed(rw))
 304                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 305                             rw, 0, 0, file, line, LOCKSTAT_WRITER);
 306                 curthread->td_locks++;
 307         }
 308         return (rval);
 309 }
 310
 311 void
 312 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 313 {
 314         struct rwlock *rw;
 315
 316         if (SCHEDULER_STOPPED())
 317                 return;
 318
 319         rw = rwlock2rw(c);
 320
 321         KASSERT(rw->rw_lock != RW_DESTROYED,
 322             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 323         __rw_assert(c, RA_WLOCKED, file, line);
 324         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 325         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 326             line);
 327         __rw_wunlock(rw, curthread, file, line);
 328         curthread->td_locks--;
 329 }
 330 /*
 331  * Determines whether a new reader can acquire a lock.  Succeeds if the
 332  * reader already owns a read lock and the lock is locked for read to
 333  * prevent deadlock from reader recursion.  Also succeeds if the lock
 334  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 335  * prioritizes writers before readers.
 336  */
 337 #define RW_CAN_READ(_rw)                                                \
 338     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 339     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 340     RW_LOCK_READ)
 341
 342 void
 343 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 344 {
 345         struct rwlock *rw;
 346         struct turnstile *ts;
 347 #ifdef ADAPTIVE_RWLOCKS
 348         volatile struct thread *owner;
 349         int spintries = 0;
 350         int i;
 351 #endif
 352 #ifdef LOCK_PROFILING
 353         uint64_t waittime = 0;
 354         int contested = 0;
 355 #endif
 356         uintptr_t v;
 357 #ifdef KDTRACE_HOOKS
 358         uintptr_t state;
 359         uint64_t spin_cnt = 0;
 360         uint64_t sleep_cnt = 0;
 361         int64_t sleep_time = 0;
 362         int64_t all_time = 0;
 363 #endif
 364
 365         if (SCHEDULER_STOPPED())
 366                 return;
 367
 368         rw = rwlock2rw(c);
 369
 370         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 371             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 372             curthread, rw->lock_object.lo_name, file, line));
 373         KASSERT(rw->rw_lock != RW_DESTROYED,
 374             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 375         KASSERT(rw_wowner(rw) != curthread,
 376             ("rw_rlock: wlock already held for %s @ %s:%d",
 377             rw->lock_object.lo_name, file, line));
 378         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 379
 380 #ifdef KDTRACE_HOOKS
 381         all_time -= lockstat_nsecs(&rw->lock_object);
 382         state = rw->rw_lock;
 383 #endif
 384         for (;;) {
 385                 /*
 386                  * Handle the easy case.  If no other thread has a write
 387                  * lock, then try to bump up the count of read locks.  Note
 388                  * that we have to preserve the current state of the
 389                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 390                  * read lock, then rw_lock must have changed, so restart
 391                  * the loop.  Note that this handles the case of a
 392                  * completely unlocked rwlock since such a lock is encoded
 393                  * as a read lock with no waiters.
 394                  */
 395                 v = rw->rw_lock;
 396                 if (RW_CAN_READ(v)) {
 397                         /*
 398                          * The RW_LOCK_READ_WAITERS flag should only be set
 399                          * if the lock has been unlocked and write waiters
 400                          * were present.
 401                          */
 402                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 403                             v + RW_ONE_READER)) {
 404                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 405                                         CTR4(KTR_LOCK,
 406                                             "%s: %p succeed %p -> %p", __func__,
 407                                             rw, (void *)v,
 408                                             (void *)(v + RW_ONE_READER));
 409                                 break;
 410                         }
 411                         continue;
 412                 }
 413 #ifdef KDTRACE_HOOKS
 414                 spin_cnt++;
 415 #endif
 416 #ifdef HWPMC_HOOKS
 417                 PMC_SOFT_CALL( , , lock, failed);
 418 #endif
 419                 lock_profile_obtain_lock_failed(&rw->lock_object,
 420                     &contested, &waittime);
 421
 422 #ifdef ADAPTIVE_RWLOCKS
 423                 /*
 424                  * If the owner is running on another CPU, spin until
 425                  * the owner stops running or the state of the lock
 426                  * changes.
 427                  */
 428                 if ((v & RW_LOCK_READ) == 0) {
 429                         owner = (struct thread *)RW_OWNER(v);
 430                         if (TD_IS_RUNNING(owner)) {
 431                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 432                                         CTR3(KTR_LOCK,
 433                                             "%s: spinning on %p held by %p",
 434                                             __func__, rw, owner);
 435                                 KTR_STATE1(KTR_SCHED, "thread",
 436                                     sched_tdname(curthread), "spinning",
 437                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 438                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 439                                     owner && TD_IS_RUNNING(owner)) {
 440                                         cpu_spinwait();
 441 #ifdef KDTRACE_HOOKS
 442                                         spin_cnt++;
 443 #endif
 444                                 }
 445                                 KTR_STATE0(KTR_SCHED, "thread",
 446                                     sched_tdname(curthread), "running");
 447                                 continue;
 448                         }
 449                 } else if (spintries < rowner_retries) {
 450                         spintries++;
 451                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 452                             "spinning", "lockname:\"%s\"",
 453                             rw->lock_object.lo_name);
 454                         for (i = 0; i < rowner_loops; i++) {
 455                                 v = rw->rw_lock;
 456                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 457                                         break;
 458                                 cpu_spinwait();
 459                         }
 460 #ifdef KDTRACE_HOOKS
 461                         spin_cnt += rowner_loops - i;
 462 #endif
 463                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 464                             "running");
 465                         if (i != rowner_loops)
 466                                 continue;
 467                 }
 468 #endif
 469
 470                 /*
 471                  * Okay, now it's the hard case.  Some other thread already
 472                  * has a write lock or there are write waiters present,
 473                  * acquire the turnstile lock so we can begin the process
 474                  * of blocking.
 475                  */
 476                 ts = turnstile_trywait(&rw->lock_object);
 477
 478                 /*
 479                  * The lock might have been released while we spun, so
 480                  * recheck its state and restart the loop if needed.
 481                  */
 482                 v = rw->rw_lock;
 483                 if (RW_CAN_READ(v)) {
 484                         turnstile_cancel(ts);
 485                         continue;
 486                 }
 487
 488 #ifdef ADAPTIVE_RWLOCKS
 489                 /*
 490                  * The current lock owner might have started executing
 491                  * on another CPU (or the lock could have changed
 492                  * owners) while we were waiting on the turnstile
 493                  * chain lock.  If so, drop the turnstile lock and try
 494                  * again.
 495                  */
 496                 if ((v & RW_LOCK_READ) == 0) {
 497                         owner = (struct thread *)RW_OWNER(v);
 498                         if (TD_IS_RUNNING(owner)) {
 499                                 turnstile_cancel(ts);
 500                                 continue;
 501                         }
 502                 }
 503 #endif
 504
 505                 /*
 506                  * The lock is held in write mode or it already has waiters.
 507                  */
 508                 MPASS(!RW_CAN_READ(v));
 509
 510                 /*
 511                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 512                  * we can go ahead and block.  If it is not set then try
 513                  * to set it.  If we fail to set it drop the turnstile
 514                  * lock and restart the loop.
 515                  */
 516                 if (!(v & RW_LOCK_READ_WAITERS)) {
 517                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 518                             v | RW_LOCK_READ_WAITERS)) {
 519                                 turnstile_cancel(ts);
 520                                 continue;
 521                         }
 522                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 523                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 524                                     __func__, rw);
 525                 }
 526
 527                 /*
 528                  * We were unable to acquire the lock and the read waiters
 529                  * flag is set, so we must block on the turnstile.
 530                  */
 531                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 532                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 533                             rw);
 534 #ifdef KDTRACE_HOOKS
 535                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 536 #endif
 537                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 538 #ifdef KDTRACE_HOOKS
 539                 sleep_time += lockstat_nsecs(&rw->lock_object);
 540                 sleep_cnt++;
 541 #endif
 542                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 543                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 544                             __func__, rw);
 545         }
 546 #ifdef KDTRACE_HOOKS
 547         all_time += lockstat_nsecs(&rw->lock_object);
 548         if (sleep_time)
 549                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 550                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 551                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 552
 553         /* Record only the loops spinning and not sleeping. */
 554         if (spin_cnt > sleep_cnt)
 555                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 556                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 557                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 558 #endif
 559         /*
 560          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 561          * however.  turnstiles don't like owners changing between calls to
 562          * turnstile_wait() currently.
 563          */
 564         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 565             waittime, file, line, LOCKSTAT_READER);
 566         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 567         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 568         curthread->td_locks++;
 569         curthread->td_rw_rlocks++;
 570 }
 571
 572 int
 573 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 574 {
 575         struct rwlock *rw;
 576         uintptr_t x;
 577
 578         if (SCHEDULER_STOPPED())
 579                 return (1);
 580
 581         rw = rwlock2rw(c);
 582
 583         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 584             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 585             curthread, rw->lock_object.lo_name, file, line));
 586
 587         for (;;) {
 588                 x = rw->rw_lock;
 589                 KASSERT(rw->rw_lock != RW_DESTROYED,
 590                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 591                 if (!(x & RW_LOCK_READ))
 592                         break;
 593                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 594                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 595                             line);
 596                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 597                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 598                             rw, 0, 0, file, line, LOCKSTAT_READER);
 599                         curthread->td_locks++;
 600                         curthread->td_rw_rlocks++;
 601                         return (1);
 602                 }
 603         }
 604
 605         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 606         return (0);
 607 }
 608
 609 void
 610 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 611 {
 612         struct rwlock *rw;
 613         struct turnstile *ts;
 614         uintptr_t x, v, queue;
 615
 616         if (SCHEDULER_STOPPED())
 617                 return;
 618
 619         rw = rwlock2rw(c);
 620
 621         KASSERT(rw->rw_lock != RW_DESTROYED,
 622             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 623         __rw_assert(c, RA_RLOCKED, file, line);
 624         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 625         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 626
 627         /* TODO: drop "owner of record" here. */
 628
 629         for (;;) {
 630                 /*
 631                  * See if there is more than one read lock held.  If so,
 632                  * just drop one and return.
 633                  */
 634                 x = rw->rw_lock;
 635                 if (RW_READERS(x) > 1) {
 636                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 637                             x - RW_ONE_READER)) {
 638                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 639                                         CTR4(KTR_LOCK,
 640                                             "%s: %p succeeded %p -> %p",
 641                                             __func__, rw, (void *)x,
 642                                             (void *)(x - RW_ONE_READER));
 643                                 break;
 644                         }
 645                         continue;
 646                 }
 647                 /*
 648                  * If there aren't any waiters for a write lock, then try
 649                  * to drop it quickly.
 650                  */
 651                 if (!(x & RW_LOCK_WAITERS)) {
 652                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 653                             RW_READERS_LOCK(1));
 654                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 655                             RW_UNLOCKED)) {
 656                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 657                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 658                                             __func__, rw);
 659                                 break;
 660                         }
 661                         continue;
 662                 }
 663                 /*
 664                  * Ok, we know we have waiters and we think we are the
 665                  * last reader, so grab the turnstile lock.
 666                  */
 667                 turnstile_chain_lock(&rw->lock_object);
 668                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 669                 MPASS(v & RW_LOCK_WAITERS);
 670
 671                 /*
 672                  * Try to drop our lock leaving the lock in a unlocked
 673                  * state.
 674                  *
 675                  * If you wanted to do explicit lock handoff you'd have to
 676                  * do it here.  You'd also want to use turnstile_signal()
 677                  * and you'd have to handle the race where a higher
 678                  * priority thread blocks on the write lock before the
 679                  * thread you wakeup actually runs and have the new thread
 680                  * "steal" the lock.  For now it's a lot simpler to just
 681                  * wakeup all of the waiters.
 682                  *
 683                  * As above, if we fail, then another thread might have
 684                  * acquired a read lock, so drop the turnstile lock and
 685                  * restart.
 686                  */
 687                 x = RW_UNLOCKED;
 688                 if (v & RW_LOCK_WRITE_WAITERS) {
 689                         queue = TS_EXCLUSIVE_QUEUE;
 690                         x |= (v & RW_LOCK_READ_WAITERS);
 691                 } else
 692                         queue = TS_SHARED_QUEUE;
 693                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 694                     x)) {
 695                         turnstile_chain_unlock(&rw->lock_object);
 696                         continue;
 697                 }
 698                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 699                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 700                             __func__, rw);
 701
 702                 /*
 703                  * Ok.  The lock is released and all that's left is to
 704                  * wake up the waiters.  Note that the lock might not be
 705                  * free anymore, but in that case the writers will just
 706                  * block again if they run before the new lock holder(s)
 707                  * release the lock.
 708                  */
 709                 ts = turnstile_lookup(&rw->lock_object);
 710                 MPASS(ts != NULL);
 711                 turnstile_broadcast(ts, queue);
 712                 turnstile_unpend(ts, TS_SHARED_LOCK);
 713                 turnstile_chain_unlock(&rw->lock_object);
 714                 break;
 715         }
 716         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
 717         curthread->td_locks--;
 718         curthread->td_rw_rlocks--;
 719 }
 720
 721 /*
 722  * This function is called when we are unable to obtain a write lock on the
 723  * first try.  This means that at least one other thread holds either a
 724  * read or write lock.
 725  */
 726 void
 727 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 728     int line)
 729 {
 730         struct rwlock *rw;
 731         struct turnstile *ts;
 732 #ifdef ADAPTIVE_RWLOCKS
 733         volatile struct thread *owner;
 734         int spintries = 0;
 735         int i;
 736 #endif
 737         uintptr_t v, x;
 738 #ifdef LOCK_PROFILING
 739         uint64_t waittime = 0;
 740         int contested = 0;
 741 #endif
 742 #ifdef KDTRACE_HOOKS
 743         uintptr_t state;
 744         uint64_t spin_cnt = 0;
 745         uint64_t sleep_cnt = 0;
 746         int64_t sleep_time = 0;
 747         int64_t all_time = 0;
 748 #endif
 749
 750         if (SCHEDULER_STOPPED())
 751                 return;
 752
 753         rw = rwlock2rw(c);
 754
 755         if (rw_wlocked(rw)) {
 756                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 757                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 758                     __func__, rw->lock_object.lo_name, file, line));
 759                 rw->rw_recurse++;
 760                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 761                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 762                 return;
 763         }
 764
 765         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 766                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 767                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 768
 769 #ifdef KDTRACE_HOOKS
 770         all_time -= lockstat_nsecs(&rw->lock_object);
 771         state = rw->rw_lock;
 772 #endif
 773         while (!_rw_write_lock(rw, tid)) {
 774 #ifdef KDTRACE_HOOKS
 775                 spin_cnt++;
 776 #endif
 777 #ifdef HWPMC_HOOKS
 778                 PMC_SOFT_CALL( , , lock, failed);
 779 #endif
 780                 lock_profile_obtain_lock_failed(&rw->lock_object,
 781                     &contested, &waittime);
 782 #ifdef ADAPTIVE_RWLOCKS
 783                 /*
 784                  * If the lock is write locked and the owner is
 785                  * running on another CPU, spin until the owner stops
 786                  * running or the state of the lock changes.
 787                  */
 788                 v = rw->rw_lock;
 789                 owner = (struct thread *)RW_OWNER(v);
 790                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 791                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 792                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 793                                     __func__, rw, owner);
 794                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 795                             "spinning", "lockname:\"%s\"",
 796                             rw->lock_object.lo_name);
 797                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 798                             TD_IS_RUNNING(owner)) {
 799                                 cpu_spinwait();
 800 #ifdef KDTRACE_HOOKS
 801                                 spin_cnt++;
 802 #endif
 803                         }
 804                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 805                             "running");
 806                         continue;
 807                 }
 808                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 809                     spintries < rowner_retries) {
 810                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 811                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 812                                     v | RW_LOCK_WRITE_SPINNER)) {
 813                                         continue;
 814                                 }
 815                         }
 816                         spintries++;
 817                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 818                             "spinning", "lockname:\"%s\"",
 819                             rw->lock_object.lo_name);
 820                         for (i = 0; i < rowner_loops; i++) {
 821                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 822                                         break;
 823                                 cpu_spinwait();
 824                         }
 825                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 826                             "running");
 827 #ifdef KDTRACE_HOOKS
 828                         spin_cnt += rowner_loops - i;
 829 #endif
 830                         if (i != rowner_loops)
 831                                 continue;
 832                 }
 833 #endif
 834                 ts = turnstile_trywait(&rw->lock_object);
 835                 v = rw->rw_lock;
 836
 837 #ifdef ADAPTIVE_RWLOCKS
 838                 /*
 839                  * The current lock owner might have started executing
 840                  * on another CPU (or the lock could have changed
 841                  * owners) while we were waiting on the turnstile
 842                  * chain lock.  If so, drop the turnstile lock and try
 843                  * again.
 844                  */
 845                 if (!(v & RW_LOCK_READ)) {
 846                         owner = (struct thread *)RW_OWNER(v);
 847                         if (TD_IS_RUNNING(owner)) {
 848                                 turnstile_cancel(ts);
 849                                 continue;
 850                         }
 851                 }
 852 #endif
 853                 /*
 854                  * Check for the waiters flags about this rwlock.
 855                  * If the lock was released, without maintain any pending
 856                  * waiters queue, simply try to acquire it.
 857                  * If a pending waiters queue is present, claim the lock
 858                  * ownership and maintain the pending queue.
 859                  */
 860                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 861                 if ((v & ~x) == RW_UNLOCKED) {
 862                         x &= ~RW_LOCK_WRITE_SPINNER;
 863                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 864                                 if (x)
 865                                         turnstile_claim(ts);
 866                                 else
 867                                         turnstile_cancel(ts);
 868                                 break;
 869                         }
 870                         turnstile_cancel(ts);
 871                         continue;
 872                 }
 873                 /*
 874                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 875                  * set it.  If we fail to set it, then loop back and try
 876                  * again.
 877                  */
 878                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 879                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 880                             v | RW_LOCK_WRITE_WAITERS)) {
 881                                 turnstile_cancel(ts);
 882                                 continue;
 883                         }
 884                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 885                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 886                                     __func__, rw);
 887                 }
 888                 /*
 889                  * We were unable to acquire the lock and the write waiters
 890                  * flag is set, so we must block on the turnstile.
 891                  */
 892                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 893                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 894                             rw);
 895 #ifdef KDTRACE_HOOKS
 896                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 897 #endif
 898                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 899 #ifdef KDTRACE_HOOKS
 900                 sleep_time += lockstat_nsecs(&rw->lock_object);
 901                 sleep_cnt++;
 902 #endif
 903                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 904                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 905                             __func__, rw);
 906 #ifdef ADAPTIVE_RWLOCKS
 907                 spintries = 0;
 908 #endif
 909         }
 910 #ifdef KDTRACE_HOOKS
 911         all_time += lockstat_nsecs(&rw->lock_object);
 912         if (sleep_time)
 913                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 914                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
 915                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 916
 917         /* Record only the loops spinning and not sleeping. */
 918         if (spin_cnt > sleep_cnt)
 919                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 920                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 921                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 922 #endif
 923         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 924             waittime, file, line, LOCKSTAT_WRITER);
 925 }
 926
 927 /*
 928  * This function is called if the first try at releasing a write lock failed.
 929  * This means that one of the 2 waiter bits must be set indicating that at
 930  * least one thread is waiting on this lock.
 931  */
 932 void
 933 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 934     int line)
 935 {
 936         struct rwlock *rw;
 937         struct turnstile *ts;
 938         uintptr_t v;
 939         int queue;
 940
 941         if (SCHEDULER_STOPPED())
 942                 return;
 943
 944         rw = rwlock2rw(c);
 945
 946         if (rw_wlocked(rw) && rw_recursed(rw)) {
 947                 rw->rw_recurse--;
 948                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 949                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 950                 return;
 951         }
 952
 953         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 954             ("%s: neither of the waiter flags are set", __func__));
 955
 956         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 957                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 958
 959         turnstile_chain_lock(&rw->lock_object);
 960         ts = turnstile_lookup(&rw->lock_object);
 961         MPASS(ts != NULL);
 962
 963         /*
 964          * Use the same algo as sx locks for now.  Prefer waking up shared
 965          * waiters if we have any over writers.  This is probably not ideal.
 966          *
 967          * 'v' is the value we are going to write back to rw_lock.  If we
 968          * have waiters on both queues, we need to preserve the state of
 969          * the waiter flag for the queue we don't wake up.  For now this is
 970          * hardcoded for the algorithm mentioned above.
 971          *
 972          * In the case of both readers and writers waiting we wakeup the
 973          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
 974          * new writer comes in before a reader it will claim the lock up
 975          * above.  There is probably a potential priority inversion in
 976          * there that could be worked around either by waking both queues
 977          * of waiters or doing some complicated lock handoff gymnastics.
 978          */
 979         v = RW_UNLOCKED;
 980         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
 981                 queue = TS_EXCLUSIVE_QUEUE;
 982                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
 983         } else
 984                 queue = TS_SHARED_QUEUE;
 985
 986         /* Wake up all waiters for the specific queue. */
 987         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 988                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
 989                     queue == TS_SHARED_QUEUE ? "read" : "write");
 990         turnstile_broadcast(ts, queue);
 991         atomic_store_rel_ptr(&rw->rw_lock, v);
 992         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
 993         turnstile_chain_unlock(&rw->lock_object);
 994 }
 995
 996 /*
 997  * Attempt to do a non-blocking upgrade from a read lock to a write
 998  * lock.  This will only succeed if this thread holds a single read
 999  * lock.  Returns true if the upgrade succeeded and false otherwise.
1000  */
1001 int
1002 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1003 {
1004         struct rwlock *rw;
1005         uintptr_t v, x, tid;
1006         struct turnstile *ts;
1007         int success;
1008
1009         if (SCHEDULER_STOPPED())
1010                 return (1);
1011
1012         rw = rwlock2rw(c);
1013
1014         KASSERT(rw->rw_lock != RW_DESTROYED,
1015             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1016         __rw_assert(c, RA_RLOCKED, file, line);
1017
1018         /*
1019          * Attempt to switch from one reader to a writer.  If there
1020          * are any write waiters, then we will have to lock the
1021          * turnstile first to prevent races with another writer
1022          * calling turnstile_wait() before we have claimed this
1023          * turnstile.  So, do the simple case of no waiters first.
1024          */
1025         tid = (uintptr_t)curthread;
1026         success = 0;
1027         for (;;) {
1028                 v = rw->rw_lock;
1029                 if (RW_READERS(v) > 1)
1030                         break;
1031                 if (!(v & RW_LOCK_WAITERS)) {
1032                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
1033                         if (!success)
1034                                 continue;
1035                         break;
1036                 }
1037
1038                 /*
1039                  * Ok, we think we have waiters, so lock the turnstile.
1040                  */
1041                 ts = turnstile_trywait(&rw->lock_object);
1042                 v = rw->rw_lock;
1043                 if (RW_READERS(v) > 1) {
1044                         turnstile_cancel(ts);
1045                         break;
1046                 }
1047                 /*
1048                  * Try to switch from one reader to a writer again.  This time
1049                  * we honor the current state of the waiters flags.
1050                  * If we obtain the lock with the flags set, then claim
1051                  * ownership of the turnstile.
1052                  */
1053                 x = rw->rw_lock & RW_LOCK_WAITERS;
1054                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1055                 if (success) {
1056                         if (x)
1057                                 turnstile_claim(ts);
1058                         else
1059                                 turnstile_cancel(ts);
1060                         break;
1061                 }
1062                 turnstile_cancel(ts);
1063         }
1064         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1065         if (success) {
1066                 curthread->td_rw_rlocks--;
1067                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1068                     file, line);
1069                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1070         }
1071         return (success);
1072 }
1073
1074 /*
1075  * Downgrade a write lock into a single read lock.
1076  */
1077 void
1078 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1079 {
1080         struct rwlock *rw;
1081         struct turnstile *ts;
1082         uintptr_t tid, v;
1083         int rwait, wwait;
1084
1085         if (SCHEDULER_STOPPED())
1086                 return;
1087
1088         rw = rwlock2rw(c);
1089
1090         KASSERT(rw->rw_lock != RW_DESTROYED,
1091             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1092         __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1093 #ifndef INVARIANTS
1094         if (rw_recursed(rw))
1095                 panic("downgrade of a recursed lock");
1096 #endif
1097
1098         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1099
1100         /*
1101          * Convert from a writer to a single reader.  First we handle
1102          * the easy case with no waiters.  If there are any waiters, we
1103          * lock the turnstile and "disown" the lock.
1104          */
1105         tid = (uintptr_t)curthread;
1106         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1107                 goto out;
1108
1109         /*
1110          * Ok, we think we have waiters, so lock the turnstile so we can
1111          * read the waiter flags without any races.
1112          */
1113         turnstile_chain_lock(&rw->lock_object);
1114         v = rw->rw_lock & RW_LOCK_WAITERS;
1115         rwait = v & RW_LOCK_READ_WAITERS;
1116         wwait = v & RW_LOCK_WRITE_WAITERS;
1117         MPASS(rwait | wwait);
1118
1119         /*
1120          * Downgrade from a write lock while preserving waiters flag
1121          * and give up ownership of the turnstile.
1122          */
1123         ts = turnstile_lookup(&rw->lock_object);
1124         MPASS(ts != NULL);
1125         if (!wwait)
1126                 v &= ~RW_LOCK_READ_WAITERS;
1127         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1128         /*
1129          * Wake other readers if there are no writers pending.  Otherwise they
1130          * won't be able to acquire the lock anyway.
1131          */
1132         if (rwait && !wwait) {
1133                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1134                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1135         } else
1136                 turnstile_disown(ts);
1137         turnstile_chain_unlock(&rw->lock_object);
1138 out:
1139         curthread->td_rw_rlocks++;
1140         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1141         LOCKSTAT_RECORD0(rw__downgrade, rw);
1142 }
1143
1144 #ifdef INVARIANT_SUPPORT
1145 #ifndef INVARIANTS
1146 #undef __rw_assert
1147 #endif
1148
1149 /*
1150  * In the non-WITNESS case, rw_assert() can only detect that at least
1151  * *some* thread owns an rlock, but it cannot guarantee that *this*
1152  * thread owns an rlock.
1153  */
1154 void
1155 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1156 {
1157         const struct rwlock *rw;
1158
1159         if (panicstr != NULL)
1160                 return;
1161
1162         rw = rwlock2rw(c);
1163
1164         switch (what) {
1165         case RA_LOCKED:
1166         case RA_LOCKED | RA_RECURSED:
1167         case RA_LOCKED | RA_NOTRECURSED:
1168         case RA_RLOCKED:
1169         case RA_RLOCKED | RA_RECURSED:
1170         case RA_RLOCKED | RA_NOTRECURSED:
1171 #ifdef WITNESS
1172                 witness_assert(&rw->lock_object, what, file, line);
1173 #else
1174                 /*
1175                  * If some other thread has a write lock or we have one
1176                  * and are asserting a read lock, fail.  Also, if no one
1177                  * has a lock at all, fail.
1178                  */
1179                 if (rw->rw_lock == RW_UNLOCKED ||
1180                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1181                     rw_wowner(rw) != curthread)))
1182                         panic("Lock %s not %slocked @ %s:%d\n",
1183                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1184                             "read " : "", file, line);
1185
1186                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1187                         if (rw_recursed(rw)) {
1188                                 if (what & RA_NOTRECURSED)
1189                                         panic("Lock %s recursed @ %s:%d\n",
1190                                             rw->lock_object.lo_name, file,
1191                                             line);
1192                         } else if (what & RA_RECURSED)
1193                                 panic("Lock %s not recursed @ %s:%d\n",
1194                                     rw->lock_object.lo_name, file, line);
1195                 }
1196 #endif
1197                 break;
1198         case RA_WLOCKED:
1199         case RA_WLOCKED | RA_RECURSED:
1200         case RA_WLOCKED | RA_NOTRECURSED:
1201                 if (rw_wowner(rw) != curthread)
1202                         panic("Lock %s not exclusively locked @ %s:%d\n",
1203                             rw->lock_object.lo_name, file, line);
1204                 if (rw_recursed(rw)) {
1205                         if (what & RA_NOTRECURSED)
1206                                 panic("Lock %s recursed @ %s:%d\n",
1207                                     rw->lock_object.lo_name, file, line);
1208                 } else if (what & RA_RECURSED)
1209                         panic("Lock %s not recursed @ %s:%d\n",
1210                             rw->lock_object.lo_name, file, line);
1211                 break;
1212         case RA_UNLOCKED:
1213 #ifdef WITNESS
1214                 witness_assert(&rw->lock_object, what, file, line);
1215 #else
1216                 /*
1217                  * If we hold a write lock fail.  We can't reliably check
1218                  * to see if we hold a read lock or not.
1219                  */
1220                 if (rw_wowner(rw) == curthread)
1221                         panic("Lock %s exclusively locked @ %s:%d\n",
1222                             rw->lock_object.lo_name, file, line);
1223 #endif
1224                 break;
1225         default:
1226                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1227                     line);
1228         }
1229 }
1230 #endif /* INVARIANT_SUPPORT */
1231
1232 #ifdef DDB
1233 void
1234 db_show_rwlock(const struct lock_object *lock)
1235 {
1236         const struct rwlock *rw;
1237         struct thread *td;
1238
1239         rw = (const struct rwlock *)lock;
1240
1241         db_printf(" state: ");
1242         if (rw->rw_lock == RW_UNLOCKED)
1243                 db_printf("UNLOCKED\n");
1244         else if (rw->rw_lock == RW_DESTROYED) {
1245                 db_printf("DESTROYED\n");
1246                 return;
1247         } else if (rw->rw_lock & RW_LOCK_READ)
1248                 db_printf("RLOCK: %ju locks\n",
1249                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1250         else {
1251                 td = rw_wowner(rw);
1252                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1253                     td->td_tid, td->td_proc->p_pid, td->td_name);
1254                 if (rw_recursed(rw))
1255                         db_printf(" recursed: %u\n", rw->rw_recurse);
1256         }
1257         db_printf(" waiters: ");
1258         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1259         case RW_LOCK_READ_WAITERS:
1260                 db_printf("readers\n");
1261                 break;
1262         case RW_LOCK_WRITE_WAITERS:
1263                 db_printf("writers\n");
1264                 break;
1265         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1266                 db_printf("readers and writers\n");
1267                 break;
1268         default:
1269                 db_printf("none\n");
1270                 break;
1271         }
1272 }
1273
1274 #endif