1 | /* $NetBSD: kern_rwlock.c,v 1.45 2014/11/28 08:28:17 uebayasi Exp $ */ |
2 | |
3 | /*- |
4 | * Copyright (c) 2002, 2006, 2007, 2008, 2009 The NetBSD Foundation, Inc. |
5 | * All rights reserved. |
6 | * |
7 | * This code is derived from software contributed to The NetBSD Foundation |
8 | * by Jason R. Thorpe and Andrew Doran. |
9 | * |
10 | * Redistribution and use in source and binary forms, with or without |
11 | * modification, are permitted provided that the following conditions |
12 | * are met: |
13 | * 1. Redistributions of source code must retain the above copyright |
14 | * notice, this list of conditions and the following disclaimer. |
15 | * 2. Redistributions in binary form must reproduce the above copyright |
16 | * notice, this list of conditions and the following disclaimer in the |
17 | * documentation and/or other materials provided with the distribution. |
18 | * |
19 | * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS |
20 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED |
21 | * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
22 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS |
23 | * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
24 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
25 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
26 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
27 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
28 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
29 | * POSSIBILITY OF SUCH DAMAGE. |
30 | */ |
31 | |
32 | /* |
33 | * Kernel reader/writer lock implementation, modeled after those |
34 | * found in Solaris, a description of which can be found in: |
35 | * |
36 | * Solaris Internals: Core Kernel Architecture, Jim Mauro and |
37 | * Richard McDougall. |
38 | */ |
39 | |
40 | #include <sys/cdefs.h> |
41 | __KERNEL_RCSID(0, "$NetBSD: kern_rwlock.c,v 1.45 2014/11/28 08:28:17 uebayasi Exp $" ); |
42 | |
43 | #define __RWLOCK_PRIVATE |
44 | |
45 | #include <sys/param.h> |
46 | #include <sys/proc.h> |
47 | #include <sys/rwlock.h> |
48 | #include <sys/sched.h> |
49 | #include <sys/sleepq.h> |
50 | #include <sys/systm.h> |
51 | #include <sys/lockdebug.h> |
52 | #include <sys/cpu.h> |
53 | #include <sys/atomic.h> |
54 | #include <sys/lock.h> |
55 | |
56 | #include <dev/lockstat.h> |
57 | |
58 | /* |
59 | * LOCKDEBUG |
60 | */ |
61 | |
62 | #if defined(LOCKDEBUG) |
63 | |
64 | #define RW_WANTLOCK(rw, op) \ |
65 | LOCKDEBUG_WANTLOCK(RW_DEBUG_P(rw), (rw), \ |
66 | (uintptr_t)__builtin_return_address(0), op == RW_READER); |
67 | #define RW_LOCKED(rw, op) \ |
68 | LOCKDEBUG_LOCKED(RW_DEBUG_P(rw), (rw), NULL, \ |
69 | (uintptr_t)__builtin_return_address(0), op == RW_READER); |
70 | #define RW_UNLOCKED(rw, op) \ |
71 | LOCKDEBUG_UNLOCKED(RW_DEBUG_P(rw), (rw), \ |
72 | (uintptr_t)__builtin_return_address(0), op == RW_READER); |
73 | #define RW_DASSERT(rw, cond) \ |
74 | do { \ |
75 | if (!(cond)) \ |
76 | rw_abort(rw, __func__, "assertion failed: " #cond); \ |
77 | } while (/* CONSTCOND */ 0); |
78 | |
79 | #else /* LOCKDEBUG */ |
80 | |
81 | #define RW_WANTLOCK(rw, op) /* nothing */ |
82 | #define RW_LOCKED(rw, op) /* nothing */ |
83 | #define RW_UNLOCKED(rw, op) /* nothing */ |
84 | #define RW_DASSERT(rw, cond) /* nothing */ |
85 | |
86 | #endif /* LOCKDEBUG */ |
87 | |
88 | /* |
89 | * DIAGNOSTIC |
90 | */ |
91 | |
92 | #if defined(DIAGNOSTIC) |
93 | |
94 | #define RW_ASSERT(rw, cond) \ |
95 | do { \ |
96 | if (!(cond)) \ |
97 | rw_abort(rw, __func__, "assertion failed: " #cond); \ |
98 | } while (/* CONSTCOND */ 0) |
99 | |
100 | #else |
101 | |
102 | #define RW_ASSERT(rw, cond) /* nothing */ |
103 | |
104 | #endif /* DIAGNOSTIC */ |
105 | |
106 | #define RW_SETDEBUG(rw, on) ((rw)->rw_owner |= (on) ? 0 : RW_NODEBUG) |
107 | #define RW_DEBUG_P(rw) (((rw)->rw_owner & RW_NODEBUG) == 0) |
108 | #if defined(LOCKDEBUG) |
109 | #define RW_INHERITDEBUG(n, o) (n) |= (o) & RW_NODEBUG |
110 | #else /* defined(LOCKDEBUG) */ |
111 | #define RW_INHERITDEBUG(n, o) /* nothing */ |
112 | #endif /* defined(LOCKDEBUG) */ |
113 | |
114 | static void rw_abort(krwlock_t *, const char *, const char *); |
115 | static void rw_dump(volatile void *); |
116 | static lwp_t *rw_owner(wchan_t); |
117 | |
118 | static inline uintptr_t |
119 | rw_cas(krwlock_t *rw, uintptr_t o, uintptr_t n) |
120 | { |
121 | |
122 | RW_INHERITDEBUG(n, o); |
123 | return (uintptr_t)atomic_cas_ptr((volatile void *)&rw->rw_owner, |
124 | (void *)o, (void *)n); |
125 | } |
126 | |
127 | static inline void |
128 | rw_swap(krwlock_t *rw, uintptr_t o, uintptr_t n) |
129 | { |
130 | |
131 | RW_INHERITDEBUG(n, o); |
132 | n = (uintptr_t)atomic_swap_ptr((volatile void *)&rw->rw_owner, |
133 | (void *)n); |
134 | RW_DASSERT(rw, n == o); |
135 | } |
136 | |
137 | /* |
138 | * For platforms that do not provide stubs, or for the LOCKDEBUG case. |
139 | */ |
140 | #ifdef LOCKDEBUG |
141 | #undef __HAVE_RW_STUBS |
142 | #endif |
143 | |
144 | #ifndef __HAVE_RW_STUBS |
145 | __strong_alias(rw_enter,rw_vector_enter); |
146 | __strong_alias(rw_exit,rw_vector_exit); |
147 | __strong_alias(rw_tryenter,rw_vector_tryenter); |
148 | #endif |
149 | |
150 | lockops_t rwlock_lockops = { |
151 | "Reader / writer lock" , |
152 | LOCKOPS_SLEEP, |
153 | rw_dump |
154 | }; |
155 | |
156 | syncobj_t rw_syncobj = { |
157 | SOBJ_SLEEPQ_SORTED, |
158 | turnstile_unsleep, |
159 | turnstile_changepri, |
160 | sleepq_lendpri, |
161 | rw_owner, |
162 | }; |
163 | |
164 | /* |
165 | * rw_dump: |
166 | * |
167 | * Dump the contents of a rwlock structure. |
168 | */ |
169 | static void |
170 | rw_dump(volatile void *cookie) |
171 | { |
172 | volatile krwlock_t *rw = cookie; |
173 | |
174 | printf_nolog("owner/count : %#018lx flags : %#018x\n" , |
175 | (long)RW_OWNER(rw), (int)RW_FLAGS(rw)); |
176 | } |
177 | |
178 | /* |
179 | * rw_abort: |
180 | * |
181 | * Dump information about an error and panic the system. This |
182 | * generates a lot of machine code in the DIAGNOSTIC case, so |
183 | * we ask the compiler to not inline it. |
184 | */ |
185 | static void __noinline |
186 | rw_abort(krwlock_t *rw, const char *func, const char *msg) |
187 | { |
188 | |
189 | if (panicstr != NULL) |
190 | return; |
191 | |
192 | LOCKDEBUG_ABORT(rw, &rwlock_lockops, func, msg); |
193 | } |
194 | |
195 | /* |
196 | * rw_init: |
197 | * |
198 | * Initialize a rwlock for use. |
199 | */ |
200 | void |
201 | rw_init(krwlock_t *rw) |
202 | { |
203 | bool dodebug; |
204 | |
205 | memset(rw, 0, sizeof(*rw)); |
206 | |
207 | dodebug = LOCKDEBUG_ALLOC(rw, &rwlock_lockops, |
208 | (uintptr_t)__builtin_return_address(0)); |
209 | RW_SETDEBUG(rw, dodebug); |
210 | } |
211 | |
212 | /* |
213 | * rw_destroy: |
214 | * |
215 | * Tear down a rwlock. |
216 | */ |
217 | void |
218 | rw_destroy(krwlock_t *rw) |
219 | { |
220 | |
221 | RW_ASSERT(rw, (rw->rw_owner & ~RW_NODEBUG) == 0); |
222 | LOCKDEBUG_FREE(RW_DEBUG_P(rw), rw); |
223 | } |
224 | |
225 | /* |
226 | * rw_oncpu: |
227 | * |
228 | * Return true if an rwlock owner is running on a CPU in the system. |
229 | * If the target is waiting on the kernel big lock, then we must |
230 | * release it. This is necessary to avoid deadlock. |
231 | */ |
232 | static bool |
233 | rw_oncpu(uintptr_t owner) |
234 | { |
235 | #ifdef MULTIPROCESSOR |
236 | struct cpu_info *ci; |
237 | lwp_t *l; |
238 | |
239 | KASSERT(kpreempt_disabled()); |
240 | |
241 | if ((owner & (RW_WRITE_LOCKED|RW_HAS_WAITERS)) != RW_WRITE_LOCKED) { |
242 | return false; |
243 | } |
244 | |
245 | /* |
246 | * See lwp_dtor() why dereference of the LWP pointer is safe. |
247 | * We must have kernel preemption disabled for that. |
248 | */ |
249 | l = (lwp_t *)(owner & RW_THREAD); |
250 | ci = l->l_cpu; |
251 | |
252 | if (ci && ci->ci_curlwp == l) { |
253 | /* Target is running; do we need to block? */ |
254 | return (ci->ci_biglock_wanted != l); |
255 | } |
256 | #endif |
257 | /* Not running. It may be safe to block now. */ |
258 | return false; |
259 | } |
260 | |
261 | /* |
262 | * rw_vector_enter: |
263 | * |
264 | * Acquire a rwlock. |
265 | */ |
266 | void |
267 | rw_vector_enter(krwlock_t *rw, const krw_t op) |
268 | { |
269 | uintptr_t owner, incr, need_wait, set_wait, curthread, next; |
270 | turnstile_t *ts; |
271 | int queue; |
272 | lwp_t *l; |
273 | LOCKSTAT_TIMER(slptime); |
274 | LOCKSTAT_TIMER(slpcnt); |
275 | LOCKSTAT_TIMER(spintime); |
276 | LOCKSTAT_COUNTER(spincnt); |
277 | LOCKSTAT_FLAG(lsflag); |
278 | |
279 | l = curlwp; |
280 | curthread = (uintptr_t)l; |
281 | |
282 | RW_ASSERT(rw, !cpu_intr_p()); |
283 | RW_ASSERT(rw, curthread != 0); |
284 | RW_WANTLOCK(rw, op); |
285 | |
286 | if (panicstr == NULL) { |
287 | LOCKDEBUG_BARRIER(&kernel_lock, 1); |
288 | } |
289 | |
290 | /* |
291 | * We play a slight trick here. If we're a reader, we want |
292 | * increment the read count. If we're a writer, we want to |
293 | * set the owner field and the WRITE_LOCKED bit. |
294 | * |
295 | * In the latter case, we expect those bits to be zero, |
296 | * therefore we can use an add operation to set them, which |
297 | * means an add operation for both cases. |
298 | */ |
299 | if (__predict_true(op == RW_READER)) { |
300 | incr = RW_READ_INCR; |
301 | set_wait = RW_HAS_WAITERS; |
302 | need_wait = RW_WRITE_LOCKED | RW_WRITE_WANTED; |
303 | queue = TS_READER_Q; |
304 | } else { |
305 | RW_DASSERT(rw, op == RW_WRITER); |
306 | incr = curthread | RW_WRITE_LOCKED; |
307 | set_wait = RW_HAS_WAITERS | RW_WRITE_WANTED; |
308 | need_wait = RW_WRITE_LOCKED | RW_THREAD; |
309 | queue = TS_WRITER_Q; |
310 | } |
311 | |
312 | LOCKSTAT_ENTER(lsflag); |
313 | |
314 | KPREEMPT_DISABLE(curlwp); |
315 | for (owner = rw->rw_owner; ;) { |
316 | /* |
317 | * Read the lock owner field. If the need-to-wait |
318 | * indicator is clear, then try to acquire the lock. |
319 | */ |
320 | if ((owner & need_wait) == 0) { |
321 | next = rw_cas(rw, owner, (owner + incr) & |
322 | ~RW_WRITE_WANTED); |
323 | if (__predict_true(next == owner)) { |
324 | /* Got it! */ |
325 | membar_enter(); |
326 | break; |
327 | } |
328 | |
329 | /* |
330 | * Didn't get it -- spin around again (we'll |
331 | * probably sleep on the next iteration). |
332 | */ |
333 | owner = next; |
334 | continue; |
335 | } |
336 | if (__predict_false(panicstr != NULL)) { |
337 | KPREEMPT_ENABLE(curlwp); |
338 | return; |
339 | } |
340 | if (__predict_false(RW_OWNER(rw) == curthread)) { |
341 | rw_abort(rw, __func__, "locking against myself" ); |
342 | } |
343 | /* |
344 | * If the lock owner is running on another CPU, and |
345 | * there are no existing waiters, then spin. |
346 | */ |
347 | if (rw_oncpu(owner)) { |
348 | LOCKSTAT_START_TIMER(lsflag, spintime); |
349 | u_int count = SPINLOCK_BACKOFF_MIN; |
350 | do { |
351 | KPREEMPT_ENABLE(curlwp); |
352 | SPINLOCK_BACKOFF(count); |
353 | KPREEMPT_DISABLE(curlwp); |
354 | owner = rw->rw_owner; |
355 | } while (rw_oncpu(owner)); |
356 | LOCKSTAT_STOP_TIMER(lsflag, spintime); |
357 | LOCKSTAT_COUNT(spincnt, 1); |
358 | if ((owner & need_wait) == 0) |
359 | continue; |
360 | } |
361 | |
362 | /* |
363 | * Grab the turnstile chain lock. Once we have that, we |
364 | * can adjust the waiter bits and sleep queue. |
365 | */ |
366 | ts = turnstile_lookup(rw); |
367 | |
368 | /* |
369 | * Mark the rwlock as having waiters. If the set fails, |
370 | * then we may not need to sleep and should spin again. |
371 | * Reload rw_owner because turnstile_lookup() may have |
372 | * spun on the turnstile chain lock. |
373 | */ |
374 | owner = rw->rw_owner; |
375 | if ((owner & need_wait) == 0 || rw_oncpu(owner)) { |
376 | turnstile_exit(rw); |
377 | continue; |
378 | } |
379 | next = rw_cas(rw, owner, owner | set_wait); |
380 | if (__predict_false(next != owner)) { |
381 | turnstile_exit(rw); |
382 | owner = next; |
383 | continue; |
384 | } |
385 | |
386 | LOCKSTAT_START_TIMER(lsflag, slptime); |
387 | turnstile_block(ts, queue, rw, &rw_syncobj); |
388 | LOCKSTAT_STOP_TIMER(lsflag, slptime); |
389 | LOCKSTAT_COUNT(slpcnt, 1); |
390 | |
391 | /* |
392 | * No need for a memory barrier because of context switch. |
393 | * If not handed the lock, then spin again. |
394 | */ |
395 | if (op == RW_READER || (rw->rw_owner & RW_THREAD) == curthread) |
396 | break; |
397 | |
398 | owner = rw->rw_owner; |
399 | } |
400 | KPREEMPT_ENABLE(curlwp); |
401 | |
402 | LOCKSTAT_EVENT(lsflag, rw, LB_RWLOCK | |
403 | (op == RW_WRITER ? LB_SLEEP1 : LB_SLEEP2), slpcnt, slptime); |
404 | LOCKSTAT_EVENT(lsflag, rw, LB_RWLOCK | LB_SPIN, spincnt, spintime); |
405 | LOCKSTAT_EXIT(lsflag); |
406 | |
407 | RW_DASSERT(rw, (op != RW_READER && RW_OWNER(rw) == curthread) || |
408 | (op == RW_READER && RW_COUNT(rw) != 0)); |
409 | RW_LOCKED(rw, op); |
410 | } |
411 | |
412 | /* |
413 | * rw_vector_exit: |
414 | * |
415 | * Release a rwlock. |
416 | */ |
417 | void |
418 | rw_vector_exit(krwlock_t *rw) |
419 | { |
420 | uintptr_t curthread, owner, decr, newown, next; |
421 | turnstile_t *ts; |
422 | int rcnt, wcnt; |
423 | lwp_t *l; |
424 | |
425 | curthread = (uintptr_t)curlwp; |
426 | RW_ASSERT(rw, curthread != 0); |
427 | |
428 | if (__predict_false(panicstr != NULL)) |
429 | return; |
430 | |
431 | /* |
432 | * Again, we use a trick. Since we used an add operation to |
433 | * set the required lock bits, we can use a subtract to clear |
434 | * them, which makes the read-release and write-release path |
435 | * the same. |
436 | */ |
437 | owner = rw->rw_owner; |
438 | if (__predict_false((owner & RW_WRITE_LOCKED) != 0)) { |
439 | RW_UNLOCKED(rw, RW_WRITER); |
440 | RW_ASSERT(rw, RW_OWNER(rw) == curthread); |
441 | decr = curthread | RW_WRITE_LOCKED; |
442 | } else { |
443 | RW_UNLOCKED(rw, RW_READER); |
444 | RW_ASSERT(rw, RW_COUNT(rw) != 0); |
445 | decr = RW_READ_INCR; |
446 | } |
447 | |
448 | /* |
449 | * Compute what we expect the new value of the lock to be. Only |
450 | * proceed to do direct handoff if there are waiters, and if the |
451 | * lock would become unowned. |
452 | */ |
453 | membar_exit(); |
454 | for (;;) { |
455 | newown = (owner - decr); |
456 | if ((newown & (RW_THREAD | RW_HAS_WAITERS)) == RW_HAS_WAITERS) |
457 | break; |
458 | next = rw_cas(rw, owner, newown); |
459 | if (__predict_true(next == owner)) |
460 | return; |
461 | owner = next; |
462 | } |
463 | |
464 | /* |
465 | * Grab the turnstile chain lock. This gets the interlock |
466 | * on the sleep queue. Once we have that, we can adjust the |
467 | * waiter bits. |
468 | */ |
469 | ts = turnstile_lookup(rw); |
470 | owner = rw->rw_owner; |
471 | RW_DASSERT(rw, ts != NULL); |
472 | RW_DASSERT(rw, (owner & RW_HAS_WAITERS) != 0); |
473 | |
474 | wcnt = TS_WAITERS(ts, TS_WRITER_Q); |
475 | rcnt = TS_WAITERS(ts, TS_READER_Q); |
476 | |
477 | /* |
478 | * Give the lock away. |
479 | * |
480 | * If we are releasing a write lock, then prefer to wake all |
481 | * outstanding readers. Otherwise, wake one writer if there |
482 | * are outstanding readers, or all writers if there are no |
483 | * pending readers. If waking one specific writer, the writer |
484 | * is handed the lock here. If waking multiple writers, we |
485 | * set WRITE_WANTED to block out new readers, and let them |
486 | * do the work of acquiring the lock in rw_vector_enter(). |
487 | */ |
488 | if (rcnt == 0 || decr == RW_READ_INCR) { |
489 | RW_DASSERT(rw, wcnt != 0); |
490 | RW_DASSERT(rw, (owner & RW_WRITE_WANTED) != 0); |
491 | |
492 | if (rcnt != 0) { |
493 | /* Give the lock to the longest waiting writer. */ |
494 | l = TS_FIRST(ts, TS_WRITER_Q); |
495 | newown = (uintptr_t)l | RW_WRITE_LOCKED | RW_HAS_WAITERS; |
496 | if (wcnt > 1) |
497 | newown |= RW_WRITE_WANTED; |
498 | rw_swap(rw, owner, newown); |
499 | turnstile_wakeup(ts, TS_WRITER_Q, 1, l); |
500 | } else { |
501 | /* Wake all writers and let them fight it out. */ |
502 | rw_swap(rw, owner, RW_WRITE_WANTED); |
503 | turnstile_wakeup(ts, TS_WRITER_Q, wcnt, NULL); |
504 | } |
505 | } else { |
506 | RW_DASSERT(rw, rcnt != 0); |
507 | |
508 | /* |
509 | * Give the lock to all blocked readers. If there |
510 | * is a writer waiting, new readers that arrive |
511 | * after the release will be blocked out. |
512 | */ |
513 | newown = rcnt << RW_READ_COUNT_SHIFT; |
514 | if (wcnt != 0) |
515 | newown |= RW_HAS_WAITERS | RW_WRITE_WANTED; |
516 | |
517 | /* Wake up all sleeping readers. */ |
518 | rw_swap(rw, owner, newown); |
519 | turnstile_wakeup(ts, TS_READER_Q, rcnt, NULL); |
520 | } |
521 | } |
522 | |
523 | /* |
524 | * rw_vector_tryenter: |
525 | * |
526 | * Try to acquire a rwlock. |
527 | */ |
528 | int |
529 | rw_vector_tryenter(krwlock_t *rw, const krw_t op) |
530 | { |
531 | uintptr_t curthread, owner, incr, need_wait, next; |
532 | |
533 | curthread = (uintptr_t)curlwp; |
534 | |
535 | RW_ASSERT(rw, curthread != 0); |
536 | |
537 | if (op == RW_READER) { |
538 | incr = RW_READ_INCR; |
539 | need_wait = RW_WRITE_LOCKED | RW_WRITE_WANTED; |
540 | } else { |
541 | RW_DASSERT(rw, op == RW_WRITER); |
542 | incr = curthread | RW_WRITE_LOCKED; |
543 | need_wait = RW_WRITE_LOCKED | RW_THREAD; |
544 | } |
545 | |
546 | for (owner = rw->rw_owner;; owner = next) { |
547 | owner = rw->rw_owner; |
548 | if (__predict_false((owner & need_wait) != 0)) |
549 | return 0; |
550 | next = rw_cas(rw, owner, owner + incr); |
551 | if (__predict_true(next == owner)) { |
552 | /* Got it! */ |
553 | membar_enter(); |
554 | break; |
555 | } |
556 | } |
557 | |
558 | RW_WANTLOCK(rw, op); |
559 | RW_LOCKED(rw, op); |
560 | RW_DASSERT(rw, (op != RW_READER && RW_OWNER(rw) == curthread) || |
561 | (op == RW_READER && RW_COUNT(rw) != 0)); |
562 | |
563 | return 1; |
564 | } |
565 | |
566 | /* |
567 | * rw_downgrade: |
568 | * |
569 | * Downgrade a write lock to a read lock. |
570 | */ |
571 | void |
572 | rw_downgrade(krwlock_t *rw) |
573 | { |
574 | uintptr_t owner, curthread, newown, next; |
575 | turnstile_t *ts; |
576 | int rcnt, wcnt; |
577 | |
578 | curthread = (uintptr_t)curlwp; |
579 | RW_ASSERT(rw, curthread != 0); |
580 | RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_LOCKED) != 0); |
581 | RW_ASSERT(rw, RW_OWNER(rw) == curthread); |
582 | RW_UNLOCKED(rw, RW_WRITER); |
583 | #if !defined(DIAGNOSTIC) |
584 | __USE(curthread); |
585 | #endif |
586 | |
587 | |
588 | membar_producer(); |
589 | owner = rw->rw_owner; |
590 | if ((owner & RW_HAS_WAITERS) == 0) { |
591 | /* |
592 | * There are no waiters, so we can do this the easy way. |
593 | * Try swapping us down to one read hold. If it fails, the |
594 | * lock condition has changed and we most likely now have |
595 | * waiters. |
596 | */ |
597 | next = rw_cas(rw, owner, RW_READ_INCR); |
598 | if (__predict_true(next == owner)) { |
599 | RW_LOCKED(rw, RW_READER); |
600 | RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_LOCKED) == 0); |
601 | RW_DASSERT(rw, RW_COUNT(rw) != 0); |
602 | return; |
603 | } |
604 | owner = next; |
605 | } |
606 | |
607 | /* |
608 | * Grab the turnstile chain lock. This gets the interlock |
609 | * on the sleep queue. Once we have that, we can adjust the |
610 | * waiter bits. |
611 | */ |
612 | for (;; owner = next) { |
613 | ts = turnstile_lookup(rw); |
614 | RW_DASSERT(rw, ts != NULL); |
615 | |
616 | rcnt = TS_WAITERS(ts, TS_READER_Q); |
617 | wcnt = TS_WAITERS(ts, TS_WRITER_Q); |
618 | |
619 | /* |
620 | * If there are no readers, just preserve the waiters |
621 | * bits, swap us down to one read hold and return. |
622 | */ |
623 | if (rcnt == 0) { |
624 | RW_DASSERT(rw, wcnt != 0); |
625 | RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_WANTED) != 0); |
626 | RW_DASSERT(rw, (rw->rw_owner & RW_HAS_WAITERS) != 0); |
627 | |
628 | newown = RW_READ_INCR | RW_HAS_WAITERS | RW_WRITE_WANTED; |
629 | next = rw_cas(rw, owner, newown); |
630 | turnstile_exit(rw); |
631 | if (__predict_true(next == owner)) |
632 | break; |
633 | } else { |
634 | /* |
635 | * Give the lock to all blocked readers. We may |
636 | * retain one read hold if downgrading. If there |
637 | * is a writer waiting, new readers will be blocked |
638 | * out. |
639 | */ |
640 | newown = (rcnt << RW_READ_COUNT_SHIFT) + RW_READ_INCR; |
641 | if (wcnt != 0) |
642 | newown |= RW_HAS_WAITERS | RW_WRITE_WANTED; |
643 | |
644 | next = rw_cas(rw, owner, newown); |
645 | if (__predict_true(next == owner)) { |
646 | /* Wake up all sleeping readers. */ |
647 | turnstile_wakeup(ts, TS_READER_Q, rcnt, NULL); |
648 | break; |
649 | } |
650 | turnstile_exit(rw); |
651 | } |
652 | } |
653 | |
654 | RW_WANTLOCK(rw, RW_READER); |
655 | RW_LOCKED(rw, RW_READER); |
656 | RW_DASSERT(rw, (rw->rw_owner & RW_WRITE_LOCKED) == 0); |
657 | RW_DASSERT(rw, RW_COUNT(rw) != 0); |
658 | } |
659 | |
660 | /* |
661 | * rw_tryupgrade: |
662 | * |
663 | * Try to upgrade a read lock to a write lock. We must be the |
664 | * only reader. |
665 | */ |
666 | int |
667 | rw_tryupgrade(krwlock_t *rw) |
668 | { |
669 | uintptr_t owner, curthread, newown, next; |
670 | |
671 | curthread = (uintptr_t)curlwp; |
672 | RW_ASSERT(rw, curthread != 0); |
673 | RW_ASSERT(rw, rw_read_held(rw)); |
674 | |
675 | for (owner = rw->rw_owner;; owner = next) { |
676 | RW_ASSERT(rw, (owner & RW_WRITE_LOCKED) == 0); |
677 | if (__predict_false((owner & RW_THREAD) != RW_READ_INCR)) { |
678 | RW_ASSERT(rw, (owner & RW_THREAD) != 0); |
679 | return 0; |
680 | } |
681 | newown = curthread | RW_WRITE_LOCKED | (owner & ~RW_THREAD); |
682 | next = rw_cas(rw, owner, newown); |
683 | if (__predict_true(next == owner)) { |
684 | membar_producer(); |
685 | break; |
686 | } |
687 | } |
688 | |
689 | RW_UNLOCKED(rw, RW_READER); |
690 | RW_WANTLOCK(rw, RW_WRITER); |
691 | RW_LOCKED(rw, RW_WRITER); |
692 | RW_DASSERT(rw, rw->rw_owner & RW_WRITE_LOCKED); |
693 | RW_DASSERT(rw, RW_OWNER(rw) == curthread); |
694 | |
695 | return 1; |
696 | } |
697 | |
698 | /* |
699 | * rw_read_held: |
700 | * |
701 | * Returns true if the rwlock is held for reading. Must only be |
702 | * used for diagnostic assertions, and never be used to make |
703 | * decisions about how to use a rwlock. |
704 | */ |
705 | int |
706 | rw_read_held(krwlock_t *rw) |
707 | { |
708 | uintptr_t owner; |
709 | |
710 | if (panicstr != NULL) |
711 | return 1; |
712 | if (rw == NULL) |
713 | return 0; |
714 | owner = rw->rw_owner; |
715 | return (owner & RW_WRITE_LOCKED) == 0 && (owner & RW_THREAD) != 0; |
716 | } |
717 | |
718 | /* |
719 | * rw_write_held: |
720 | * |
721 | * Returns true if the rwlock is held for writing. Must only be |
722 | * used for diagnostic assertions, and never be used to make |
723 | * decisions about how to use a rwlock. |
724 | */ |
725 | int |
726 | rw_write_held(krwlock_t *rw) |
727 | { |
728 | |
729 | if (panicstr != NULL) |
730 | return 1; |
731 | if (rw == NULL) |
732 | return 0; |
733 | return (rw->rw_owner & (RW_WRITE_LOCKED | RW_THREAD)) == |
734 | (RW_WRITE_LOCKED | (uintptr_t)curlwp); |
735 | } |
736 | |
737 | /* |
738 | * rw_lock_held: |
739 | * |
740 | * Returns true if the rwlock is held for reading or writing. Must |
741 | * only be used for diagnostic assertions, and never be used to make |
742 | * decisions about how to use a rwlock. |
743 | */ |
744 | int |
745 | rw_lock_held(krwlock_t *rw) |
746 | { |
747 | |
748 | if (panicstr != NULL) |
749 | return 1; |
750 | if (rw == NULL) |
751 | return 0; |
752 | return (rw->rw_owner & RW_THREAD) != 0; |
753 | } |
754 | |
755 | /* |
756 | * rw_owner: |
757 | * |
758 | * Return the current owner of an RW lock, but only if it is write |
759 | * held. Used for priority inheritance. |
760 | */ |
761 | static lwp_t * |
762 | rw_owner(wchan_t obj) |
763 | { |
764 | krwlock_t *rw = (void *)(uintptr_t)obj; /* discard qualifiers */ |
765 | uintptr_t owner = rw->rw_owner; |
766 | |
767 | if ((owner & RW_WRITE_LOCKED) == 0) |
768 | return NULL; |
769 | |
770 | return (void *)(owner & RW_THREAD); |
771 | } |
772 | |