xref: /illumos-gate/usr/src/uts/common/os/rwlock.c (revision 355b4669e025ff377602b6fc7caaf30dbc218371)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/param.h>
30 #include <sys/thread.h>
31 #include <sys/cmn_err.h>
32 #include <sys/debug.h>
33 #include <sys/cpuvar.h>
34 #include <sys/sobject.h>
35 #include <sys/turnstile.h>
36 #include <sys/rwlock.h>
37 #include <sys/rwlock_impl.h>
38 #include <sys/atomic.h>
39 #include <sys/lockstat.h>
40 
41 /*
42  * Big Theory Statement for readers/writer locking primitives.
43  *
44  * An rwlock provides exclusive access to a single thread ("writer") or
45  * concurrent access to multiple threads ("readers").  See rwlock(9F)
46  * for a full description of the interfaces and programming model.
47  * The rest of this comment describes the implementation.
48  *
49  * An rwlock is a single word with the following structure:
50  *
51  *	---------------------------------------------------------------------
52  *	| OWNER (writer) or HOLD COUNT (readers)   | WRLOCK | WRWANT | WAIT |
53  *	---------------------------------------------------------------------
54  *			63 / 31 .. 3			2	1	0
55  *
56  * The waiters bit (0) indicates whether any threads are blocked waiting
57  * for the lock.  The write-wanted bit (1) indicates whether any threads
58  * are blocked waiting for write access.  The write-locked bit (2) indicates
59  * whether the lock is held by a writer, which determines whether the upper
60  * bits (3..31 in ILP32, 3..63 in LP64) should be interpreted as the owner
61  * (thread pointer) or the hold count (number of readers).
62  *
63  * In the absence of any contention, a writer gets the lock by setting
64  * this word to (curthread | RW_WRITE_LOCKED); a reader gets the lock
65  * by incrementing the hold count (i.e. adding 8, aka RW_READ_LOCK).
66  *
67  * A writer will fail to acquire the lock if any other thread owns it.
68  * A reader will fail if the lock is either owned or wanted by a writer.
69  * rw_tryenter() returns 0 in these cases; rw_enter() blocks until the
70  * lock becomes available.
71  *
72  * When a thread blocks it acquires the rwlock's hashed turnstile lock and
73  * attempts to set RW_HAS_WAITERS (and RW_WRITE_WANTED in the writer case)
74  * atomically *only if the lock still appears busy*.  A thread must never
75  * accidentally block for an available lock since there would be no owner
76  * to awaken it.  casip() provides the required atomicity.  Once casip()
77  * succeeds, the decision to block becomes final and irreversible.  The
78  * thread will not become runnable again until it has been granted ownership
79  * of the lock via direct handoff from a former owner as described below.
80  *
81  * In the absence of any waiters, rw_exit() just clears the lock (if it
82  * is write-locked) or decrements the hold count (if it is read-locked).
83  * Note that even if waiters are present, decrementing the hold count
84  * to a non-zero value requires no special action since the lock is still
85  * held by at least one other thread.
86  *
87  * On the "final exit" (transition to unheld state) of a lock with waiters,
88  * rw_exit_wakeup() grabs the turnstile lock and transfers ownership directly
89  * to the next writer or set of readers.  There are several advantages to this
90  * approach: (1) it closes all windows for priority inversion (when a new
91  * writer has grabbed the lock but has not yet inherited from blocked readers);
92  * (2) it prevents starvation of equal-priority threads by granting the lock
93  * in FIFO order; (3) it eliminates the need for a write-wanted count -- a
94  * single bit suffices because the lock remains held until all waiting
95  * writers are gone; (4) when we awaken N readers we can perform a single
96  * "atomic_add(&x, N)" to set the total hold count rather than having all N
97  * threads fight for the cache to perform an "atomic_add(&x, 1)" upon wakeup.
98  *
99  * The most interesting policy decision in rw_exit_wakeup() is which thread
100  * to wake.  Starvation is always possible with priority-based scheduling,
101  * but any sane wakeup policy should at least satisfy these requirements:
102  *
103  * (1) The highest-priority thread in the system should not starve.
104  * (2) The highest-priority writer should not starve.
105  * (3) No writer should starve due to lower-priority threads.
106  * (4) No reader should starve due to lower-priority writers.
107  * (5) If all threads have equal priority, none of them should starve.
108  *
109  * We used to employ a writers-always-win policy, which doesn't even
110  * satisfy (1): a steady stream of low-priority writers can starve out
111  * a real-time reader!  This is clearly a broken policy -- it violates
112  * (1), (4), and (5) -- but it's how rwlocks always used to behave.
113  *
114  * A round-robin policy (exiting readers grant the lock to blocked writers
115  * and vice versa) satisfies all but (3): a single high-priority writer
116  * and many low-priority readers can starve out medium-priority writers.
117  *
118  * A strict priority policy (grant the lock to the highest priority blocked
119  * thread) satisfies everything but (2): a steady stream of high-priority
120  * readers can permanently starve the highest-priority writer.
121  *
122  * The reason we care about (2) is that it's important to process writers
123  * reasonably quickly -- even if they're low priority -- because their very
124  * presence causes all readers to take the slow (blocking) path through this
125  * code.  There is also a general sense that writers deserve some degree of
126  * deference because they're updating the data upon which all readers act.
127  * Presumably this data should not be allowed to become arbitrarily stale
128  * due to writer starvation.  Finally, it seems reasonable to level the
129  * playing field a bit to compensate for the fact that it's so much harder
130  * for a writer to get in when there are already many readers present.
131  *
132  * A hybrid of round-robin and strict priority can be made to satisfy
133  * all five criteria.  In this "writer priority policy" exiting readers
134  * always grant the lock to waiting writers, but exiting writers only
135  * grant the lock to readers of the same or higher priority than the
136  * highest-priority blocked writer.  Thus requirement (2) is satisfied,
137  * necessarily, by a willful act of priority inversion: an exiting reader
138  * will grant the lock to a blocked writer even if there are blocked
139  * readers of higher priority.  The situation is mitigated by the fact
140  * that writers always inherit priority from blocked readers, and the
141  * writer will awaken those readers as soon as it exits the lock.
142  *
143  * rw_downgrade() follows the same wakeup policy as an exiting writer.
144  *
145  * rw_tryupgrade() has the same failure mode as rw_tryenter() for a
146  * write lock.  Both honor the WRITE_WANTED bit by specification.
147  *
148  * The following rules apply to manipulation of rwlock internal state:
149  *
150  * (1) The rwlock is only modified via the atomic primitives casip()
151  *     and atomic_add_ip().
152  *
153  * (2) The waiters bit and write-wanted bit are only modified under
154  *     turnstile_lookup().  This ensures that the turnstile is consistent
155  *     with the rwlock.
156  *
157  * (3) Waiters receive the lock by direct handoff from the previous
158  *     owner.  Therefore, waiters *always* wake up holding the lock.
159  */
160 
161 /*
162  * The sobj_ops vector exports a set of functions needed when a thread
163  * is asleep on a synchronization object of a given type.
164  */
165 static sobj_ops_t rw_sobj_ops = {
166 	SOBJ_RWLOCK, rw_owner, turnstile_stay_asleep, turnstile_change_pri
167 };
168 
169 /*
170  * If the system panics on an rwlock, save the address of the offending
171  * rwlock in panic_rwlock_addr, and save the contents in panic_rwlock.
172  */
173 static rwlock_impl_t panic_rwlock;
174 static rwlock_impl_t *panic_rwlock_addr;
175 
176 static void
177 rw_panic(char *msg, rwlock_impl_t *lp)
178 {
179 	if (panicstr)
180 		return;
181 
182 	if (casptr(&panic_rwlock_addr, NULL, lp) == NULL)
183 		panic_rwlock = *lp;
184 
185 	panic("%s, lp=%p wwwh=%lx thread=%p",
186 	    msg, lp, panic_rwlock.rw_wwwh, curthread);
187 }
188 
189 /* ARGSUSED */
190 void
191 rw_init(krwlock_t *rwlp, char *name, krw_type_t type, void *arg)
192 {
193 	((rwlock_impl_t *)rwlp)->rw_wwwh = 0;
194 }
195 
196 void
197 rw_destroy(krwlock_t *rwlp)
198 {
199 	rwlock_impl_t *lp = (rwlock_impl_t *)rwlp;
200 
201 	if (lp->rw_wwwh != 0) {
202 		if ((lp->rw_wwwh & RW_DOUBLE_LOCK) == RW_DOUBLE_LOCK)
203 			rw_panic("rw_destroy: lock already destroyed", lp);
204 		else
205 			rw_panic("rw_destroy: lock still active", lp);
206 	}
207 
208 	lp->rw_wwwh = RW_DOUBLE_LOCK;
209 }
210 
211 /*
212  * Verify that an rwlock is held correctly.
213  */
214 static int
215 rw_locked(rwlock_impl_t *lp, krw_t rw)
216 {
217 	uintptr_t old = lp->rw_wwwh;
218 
219 	if (rw == RW_READER)
220 		return ((old & RW_LOCKED) && !(old & RW_WRITE_LOCKED));
221 
222 	if (rw == RW_WRITER)
223 		return ((old & RW_OWNER) == (uintptr_t)curthread);
224 
225 	return (0);
226 }
227 
228 /*
229  * Full-service implementation of rw_enter() to handle all the hard cases.
230  * Called from the assembly version if anything complicated is going on.
231  * The only semantic difference between calling rw_enter() and calling
232  * rw_enter_sleep() directly is that we assume the caller has already done
233  * a THREAD_KPRI_REQUEST() in the RW_READER case.
234  */
235 void
236 rw_enter_sleep(rwlock_impl_t *lp, krw_t rw)
237 {
238 	uintptr_t old, new, lock_value, lock_busy, lock_wait;
239 	hrtime_t sleep_time;
240 	turnstile_t *ts;
241 
242 	if (rw == RW_READER) {
243 		lock_value = RW_READ_LOCK;
244 		lock_busy = RW_WRITE_CLAIMED;
245 		lock_wait = RW_HAS_WAITERS;
246 	} else {
247 		lock_value = RW_WRITE_LOCK(curthread);
248 		lock_busy = (uintptr_t)RW_LOCKED;
249 		lock_wait = RW_HAS_WAITERS | RW_WRITE_WANTED;
250 	}
251 
252 	for (;;) {
253 		if (((old = lp->rw_wwwh) & lock_busy) == 0) {
254 			if (casip(&lp->rw_wwwh, old, old + lock_value) != old)
255 				continue;
256 			break;
257 		}
258 
259 		if (panicstr)
260 			return;
261 
262 		if ((old & RW_DOUBLE_LOCK) == RW_DOUBLE_LOCK) {
263 			rw_panic("rw_enter: bad rwlock", lp);
264 			return;
265 		}
266 
267 		if ((old & RW_OWNER) == (uintptr_t)curthread) {
268 			rw_panic("recursive rw_enter", lp);
269 			return;
270 		}
271 
272 		ts = turnstile_lookup(lp);
273 
274 		do {
275 			if (((old = lp->rw_wwwh) & lock_busy) == 0)
276 				break;
277 			new = old | lock_wait;
278 		} while (old != new && casip(&lp->rw_wwwh, old, new) != old);
279 
280 		if ((old & lock_busy) == 0) {
281 			/*
282 			 * The lock appears free now; try the dance again
283 			 */
284 			turnstile_exit(lp);
285 			continue;
286 		}
287 
288 		/*
289 		 * We really are going to block.  Bump the stats, and drop
290 		 * kpri if we're a reader.
291 		 */
292 		ASSERT(lp->rw_wwwh & lock_wait);
293 		ASSERT(lp->rw_wwwh & RW_LOCKED);
294 
295 		sleep_time = -gethrtime();
296 		if (rw == RW_READER) {
297 			THREAD_KPRI_RELEASE();
298 			CPU_STATS_ADDQ(CPU, sys, rw_rdfails, 1);
299 			(void) turnstile_block(ts, TS_READER_Q, lp,
300 			    &rw_sobj_ops, NULL, NULL);
301 		} else {
302 			CPU_STATS_ADDQ(CPU, sys, rw_wrfails, 1);
303 			(void) turnstile_block(ts, TS_WRITER_Q, lp,
304 			    &rw_sobj_ops, NULL, NULL);
305 		}
306 		sleep_time += gethrtime();
307 
308 		LOCKSTAT_RECORD4(LS_RW_ENTER_BLOCK, lp, sleep_time, rw,
309 		    (old & RW_WRITE_LOCKED) ? 1 : 0,
310 		    old >> RW_HOLD_COUNT_SHIFT);
311 
312 		/*
313 		 * We wake up holding the lock (and having kpri if we're
314 		 * a reader) via direct handoff from the previous owner.
315 		 */
316 		break;
317 	}
318 
319 	ASSERT(rw_locked(lp, rw));
320 
321 	membar_enter();
322 
323 	LOCKSTAT_RECORD(LS_RW_ENTER_ACQUIRE, lp, rw);
324 }
325 
326 /*
327  * Return the number of readers to wake, or zero if we should wake a writer.
328  * Called only by exiting/downgrading writers (readers don't wake readers).
329  */
330 static int
331 rw_readers_to_wake(turnstile_t *ts)
332 {
333 	kthread_t *next_writer = ts->ts_sleepq[TS_WRITER_Q].sq_first;
334 	kthread_t *next_reader = ts->ts_sleepq[TS_READER_Q].sq_first;
335 	pri_t wpri = (next_writer != NULL) ? DISP_PRIO(next_writer) : -1;
336 	int count = 0;
337 
338 	while (next_reader != NULL) {
339 		if (DISP_PRIO(next_reader) < wpri)
340 			break;
341 		next_reader->t_kpri_req++;
342 		next_reader = next_reader->t_link;
343 		count++;
344 	}
345 	return (count);
346 }
347 
348 /*
349  * Full-service implementation of rw_exit() to handle all the hard cases.
350  * Called from the assembly version if anything complicated is going on.
351  * There is no semantic difference between calling rw_exit() and calling
352  * rw_exit_wakeup() directly.
353  */
354 void
355 rw_exit_wakeup(rwlock_impl_t *lp)
356 {
357 	turnstile_t *ts;
358 	uintptr_t old, new, lock_value;
359 	kthread_t *next_writer;
360 	int nreaders;
361 
362 	membar_exit();
363 
364 	old = lp->rw_wwwh;
365 	if (old & RW_WRITE_LOCKED) {
366 		if ((old & RW_OWNER) != (uintptr_t)curthread) {
367 			rw_panic("rw_exit: not owner", lp);
368 			lp->rw_wwwh = 0;
369 			return;
370 		}
371 		lock_value = RW_WRITE_LOCK(curthread);
372 	} else {
373 		if ((old & RW_LOCKED) == 0) {
374 			rw_panic("rw_exit: lock not held", lp);
375 			return;
376 		}
377 		lock_value = RW_READ_LOCK;
378 	}
379 
380 	for (;;) {
381 		/*
382 		 * If this is *not* the final exit of a lock with waiters,
383 		 * just drop the lock -- there's nothing tricky going on.
384 		 */
385 		old = lp->rw_wwwh;
386 		new = old - lock_value;
387 		if ((new & (RW_LOCKED | RW_HAS_WAITERS)) != RW_HAS_WAITERS) {
388 			if (casip(&lp->rw_wwwh, old, new) != old)
389 				continue;
390 			break;
391 		}
392 
393 		/*
394 		 * Perform the final exit of a lock that has waiters.
395 		 */
396 		ts = turnstile_lookup(lp);
397 
398 		next_writer = ts->ts_sleepq[TS_WRITER_Q].sq_first;
399 
400 		if ((old & RW_WRITE_LOCKED) &&
401 		    (nreaders = rw_readers_to_wake(ts)) > 0) {
402 			/*
403 			 * Don't drop the lock -- just set the hold count
404 			 * such that we grant the lock to all readers at once.
405 			 */
406 			new = nreaders * RW_READ_LOCK;
407 			if (ts->ts_waiters > nreaders)
408 				new |= RW_HAS_WAITERS;
409 			if (next_writer)
410 				new |= RW_WRITE_WANTED;
411 			lp->rw_wwwh = new;
412 			membar_enter();
413 			turnstile_wakeup(ts, TS_READER_Q, nreaders, NULL);
414 		} else {
415 			/*
416 			 * Don't drop the lock -- just transfer ownership
417 			 * directly to next_writer.  Note that there must
418 			 * be at least one waiting writer, because we get
419 			 * here only if (A) the lock is read-locked or
420 			 * (B) there are no waiting readers.  In case (A),
421 			 * since the lock is read-locked there would be no
422 			 * reason for other readers to have blocked unless
423 			 * the RW_WRITE_WANTED bit was set.  In case (B),
424 			 * since there are waiters but no waiting readers,
425 			 * they must all be waiting writers.
426 			 */
427 			ASSERT(lp->rw_wwwh & RW_WRITE_WANTED);
428 			new = RW_WRITE_LOCK(next_writer);
429 			if (ts->ts_waiters > 1)
430 				new |= RW_HAS_WAITERS;
431 			if (next_writer->t_link)
432 				new |= RW_WRITE_WANTED;
433 			lp->rw_wwwh = new;
434 			membar_enter();
435 			turnstile_wakeup(ts, TS_WRITER_Q, 1, next_writer);
436 		}
437 		break;
438 	}
439 
440 	if (lock_value == RW_READ_LOCK) {
441 		THREAD_KPRI_RELEASE();
442 		LOCKSTAT_RECORD(LS_RW_EXIT_RELEASE, lp, RW_READER);
443 	} else {
444 		LOCKSTAT_RECORD(LS_RW_EXIT_RELEASE, lp, RW_WRITER);
445 	}
446 }
447 
448 int
449 rw_tryenter(krwlock_t *rwlp, krw_t rw)
450 {
451 	rwlock_impl_t *lp = (rwlock_impl_t *)rwlp;
452 	uintptr_t old;
453 
454 	if (rw == RW_READER) {
455 		THREAD_KPRI_REQUEST();
456 		do {
457 			if ((old = lp->rw_wwwh) & RW_WRITE_CLAIMED) {
458 				THREAD_KPRI_RELEASE();
459 				return (0);
460 			}
461 		} while (casip(&lp->rw_wwwh, old, old + RW_READ_LOCK) != old);
462 		LOCKSTAT_RECORD(LS_RW_TRYENTER_ACQUIRE, lp, rw);
463 	} else {
464 		if (casip(&lp->rw_wwwh, 0, RW_WRITE_LOCK(curthread)) != 0)
465 			return (0);
466 		LOCKSTAT_RECORD(LS_RW_TRYENTER_ACQUIRE, lp, rw);
467 	}
468 	ASSERT(rw_locked(lp, rw));
469 	membar_enter();
470 	return (1);
471 }
472 
473 void
474 rw_downgrade(krwlock_t *rwlp)
475 {
476 	rwlock_impl_t *lp = (rwlock_impl_t *)rwlp;
477 
478 	THREAD_KPRI_REQUEST();
479 	membar_exit();
480 
481 	if ((lp->rw_wwwh & RW_OWNER) != (uintptr_t)curthread) {
482 		rw_panic("rw_downgrade: not owner", lp);
483 		return;
484 	}
485 
486 	if (atomic_add_ip_nv(&lp->rw_wwwh,
487 	    RW_READ_LOCK - RW_WRITE_LOCK(curthread)) & RW_HAS_WAITERS) {
488 		turnstile_t *ts = turnstile_lookup(lp);
489 		int nreaders = rw_readers_to_wake(ts);
490 		if (nreaders > 0) {
491 			uintptr_t delta = nreaders * RW_READ_LOCK;
492 			if (ts->ts_waiters == nreaders)
493 				delta -= RW_HAS_WAITERS;
494 			atomic_add_ip(&lp->rw_wwwh, delta);
495 		}
496 		turnstile_wakeup(ts, TS_READER_Q, nreaders, NULL);
497 	}
498 	ASSERT(rw_locked(lp, RW_READER));
499 	LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, lp);
500 }
501 
502 int
503 rw_tryupgrade(krwlock_t *rwlp)
504 {
505 	rwlock_impl_t *lp = (rwlock_impl_t *)rwlp;
506 	uintptr_t old, new;
507 
508 	ASSERT(rw_locked(lp, RW_READER));
509 
510 	do {
511 		if (((old = lp->rw_wwwh) & ~RW_HAS_WAITERS) != RW_READ_LOCK)
512 			return (0);
513 		new = old + RW_WRITE_LOCK(curthread) - RW_READ_LOCK;
514 	} while (casip(&lp->rw_wwwh, old, new) != old);
515 
516 	membar_enter();
517 	THREAD_KPRI_RELEASE();
518 	LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, lp);
519 	ASSERT(rw_locked(lp, RW_WRITER));
520 	return (1);
521 }
522 
523 int
524 rw_read_held(krwlock_t *rwlp)
525 {
526 	uintptr_t tmp;
527 
528 	return (_RW_READ_HELD(rwlp, tmp));
529 }
530 
531 int
532 rw_write_held(krwlock_t *rwlp)
533 {
534 	return (_RW_WRITE_HELD(rwlp));
535 }
536 
537 int
538 rw_lock_held(krwlock_t *rwlp)
539 {
540 	return (_RW_LOCK_HELD(rwlp));
541 }
542 
543 /*
544  * Like rw_read_held(), but ASSERTs that the lock is currently held
545  */
546 int
547 rw_read_locked(krwlock_t *rwlp)
548 {
549 	uintptr_t old = ((rwlock_impl_t *)rwlp)->rw_wwwh;
550 
551 	ASSERT(old & RW_LOCKED);
552 	return ((old & RW_LOCKED) && !(old & RW_WRITE_LOCKED));
553 }
554 
555 /*
556  * Returns non-zero if the lock is either held or desired by a writer
557  */
558 int
559 rw_iswriter(krwlock_t *rwlp)
560 {
561 	return (_RW_ISWRITER(rwlp));
562 }
563 
564 kthread_t *
565 rw_owner(krwlock_t *rwlp)
566 {
567 	uintptr_t old = ((rwlock_impl_t *)rwlp)->rw_wwwh;
568 
569 	return ((old & RW_WRITE_LOCKED) ? (kthread_t *)(old & RW_OWNER) : NULL);
570 }
571