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