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 (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25 /*
26 * Copyright (c) 2012 by Delphix. All rights reserved.
27 */
28
29 #include <sys/refcount.h>
30 #include <sys/rrwlock.h>
31
32 /*
33 * This file contains the implementation of a re-entrant read
34 * reader/writer lock (aka "rrwlock").
35 *
36 * This is a normal reader/writer lock with the additional feature
37 * of allowing threads who have already obtained a read lock to
38 * re-enter another read lock (re-entrant read) - even if there are
39 * waiting writers.
40 *
41 * Callers who have not obtained a read lock give waiting writers priority.
42 *
43 * The rrwlock_t lock does not allow re-entrant writers, nor does it
44 * allow a re-entrant mix of reads and writes (that is, it does not
45 * allow a caller who has already obtained a read lock to be able to
46 * then grab a write lock without first dropping all read locks, and
47 * vice versa).
48 *
49 * The rrwlock_t uses tsd (thread specific data) to keep a list of
50 * nodes (rrw_node_t), where each node keeps track of which specific
51 * lock (rrw_node_t::rn_rrl) the thread has grabbed. Since re-entering
52 * should be rare, a thread that grabs multiple reads on the same rrwlock_t
53 * will store multiple rrw_node_ts of the same 'rrn_rrl'. Nodes on the
54 * tsd list can represent a different rrwlock_t. This allows a thread
55 * to enter multiple and unique rrwlock_ts for read locks at the same time.
56 *
57 * Since using tsd exposes some overhead, the rrwlock_t only needs to
58 * keep tsd data when writers are waiting. If no writers are waiting, then
59 * a reader just bumps the anonymous read count (rr_anon_rcount) - no tsd
60 * is needed. Once a writer attempts to grab the lock, readers then
61 * keep tsd data and bump the linked readers count (rr_linked_rcount).
62 *
63 * If there are waiting writers and there are anonymous readers, then a
64 * reader doesn't know if it is a re-entrant lock. But since it may be one,
65 * we allow the read to proceed (otherwise it could deadlock). Since once
66 * waiting writers are active, readers no longer bump the anonymous count,
67 * the anonymous readers will eventually flush themselves out. At this point,
68 * readers will be able to tell if they are a re-entrant lock (have a
69 * rrw_node_t entry for the lock) or not. If they are a re-entrant lock, then
70 * we must let the proceed. If they are not, then the reader blocks for the
71 * waiting writers. Hence, we do not starve writers.
72 */
73
74 /* global key for TSD */
75 uint_t rrw_tsd_key;
76
77 typedef struct rrw_node {
78 struct rrw_node *rn_next;
79 rrwlock_t *rn_rrl;
80 void *rn_tag;
81 } rrw_node_t;
82
83 static rrw_node_t *
84 rrn_find(rrwlock_t *rrl)
85 {
86 rrw_node_t *rn;
87
88 if (refcount_count(&rrl->rr_linked_rcount) == 0)
89 return (NULL);
90
91 for (rn = tsd_get(rrw_tsd_key); rn != NULL; rn = rn->rn_next) {
92 if (rn->rn_rrl == rrl)
93 return (rn);
94 }
95 return (NULL);
96 }
97
98 /*
99 * Add a node to the head of the singly linked list.
100 */
101 static void
102 rrn_add(rrwlock_t *rrl, void *tag)
103 {
104 rrw_node_t *rn;
105
106 rn = kmem_alloc(sizeof (*rn), KM_SLEEP);
107 rn->rn_rrl = rrl;
108 rn->rn_next = tsd_get(rrw_tsd_key);
109 rn->rn_tag = tag;
110 VERIFY(tsd_set(rrw_tsd_key, rn) == 0);
111 }
112
113 /*
114 * If a node is found for 'rrl', then remove the node from this
115 * thread's list and return TRUE; otherwise return FALSE.
116 */
117 static boolean_t
118 rrn_find_and_remove(rrwlock_t *rrl, void *tag)
119 {
120 rrw_node_t *rn;
121 rrw_node_t *prev = NULL;
122
123 if (refcount_count(&rrl->rr_linked_rcount) == 0)
124 return (B_FALSE);
125
126 for (rn = tsd_get(rrw_tsd_key); rn != NULL; rn = rn->rn_next) {
127 if (rn->rn_rrl == rrl && rn->rn_tag == tag) {
128 if (prev)
129 prev->rn_next = rn->rn_next;
130 else
131 VERIFY(tsd_set(rrw_tsd_key, rn->rn_next) == 0);
132 kmem_free(rn, sizeof (*rn));
133 return (B_TRUE);
134 }
135 prev = rn;
136 }
137 return (B_FALSE);
138 }
139
140 void
141 rrw_init(rrwlock_t *rrl, boolean_t track_all)
142 {
143 mutex_init(&rrl->rr_lock, NULL, MUTEX_DEFAULT, NULL);
144 cv_init(&rrl->rr_cv, NULL, CV_DEFAULT, NULL);
145 rrl->rr_writer = NULL;
146 refcount_create(&rrl->rr_anon_rcount);
147 refcount_create(&rrl->rr_linked_rcount);
148 rrl->rr_writer_wanted = B_FALSE;
149 rrl->rr_track_all = track_all;
150 }
151
152 void
153 rrw_destroy(rrwlock_t *rrl)
154 {
155 mutex_destroy(&rrl->rr_lock);
156 cv_destroy(&rrl->rr_cv);
157 ASSERT(rrl->rr_writer == NULL);
158 refcount_destroy(&rrl->rr_anon_rcount);
159 refcount_destroy(&rrl->rr_linked_rcount);
160 }
161
162 void
163 rrw_enter_read(rrwlock_t *rrl, void *tag)
164 {
165 mutex_enter(&rrl->rr_lock);
166 #if !defined(DEBUG) && defined(_KERNEL)
167 if (rrl->rr_writer == NULL && !rrl->rr_writer_wanted &&
168 !rrl->rr_track_all) {
169 rrl->rr_anon_rcount.rc_count++;
170 mutex_exit(&rrl->rr_lock);
171 return;
172 }
173 DTRACE_PROBE(zfs__rrwfastpath__rdmiss);
174 #endif
175 ASSERT(rrl->rr_writer != curthread);
176 ASSERT(refcount_count(&rrl->rr_anon_rcount) >= 0);
177
178 while (rrl->rr_writer != NULL || (rrl->rr_writer_wanted &&
179 refcount_is_zero(&rrl->rr_anon_rcount) &&
180 rrn_find(rrl) == NULL))
181 cv_wait(&rrl->rr_cv, &rrl->rr_lock);
182
183 if (rrl->rr_writer_wanted || rrl->rr_track_all) {
184 /* may or may not be a re-entrant enter */
185 rrn_add(rrl, tag);
186 (void) refcount_add(&rrl->rr_linked_rcount, tag);
187 } else {
188 (void) refcount_add(&rrl->rr_anon_rcount, tag);
189 }
190 ASSERT(rrl->rr_writer == NULL);
191 mutex_exit(&rrl->rr_lock);
192 }
193
194 void
195 rrw_enter_write(rrwlock_t *rrl)
196 {
197 mutex_enter(&rrl->rr_lock);
198 ASSERT(rrl->rr_writer != curthread);
199
200 while (refcount_count(&rrl->rr_anon_rcount) > 0 ||
201 refcount_count(&rrl->rr_linked_rcount) > 0 ||
202 rrl->rr_writer != NULL) {
203 rrl->rr_writer_wanted = B_TRUE;
204 cv_wait(&rrl->rr_cv, &rrl->rr_lock);
205 }
206 rrl->rr_writer_wanted = B_FALSE;
207 rrl->rr_writer = curthread;
208 mutex_exit(&rrl->rr_lock);
209 }
210
211 void
212 rrw_enter(rrwlock_t *rrl, krw_t rw, void *tag)
213 {
214 if (rw == RW_READER)
215 rrw_enter_read(rrl, tag);
216 else
217 rrw_enter_write(rrl);
218 }
219
220 void
221 rrw_exit(rrwlock_t *rrl, void *tag)
222 {
223 mutex_enter(&rrl->rr_lock);
224 #if !defined(DEBUG) && defined(_KERNEL)
225 if (!rrl->rr_writer && rrl->rr_linked_rcount.rc_count == 0) {
226 rrl->rr_anon_rcount.rc_count--;
227 if (rrl->rr_anon_rcount.rc_count == 0)
228 cv_broadcast(&rrl->rr_cv);
229 mutex_exit(&rrl->rr_lock);
230 return;
231 }
232 DTRACE_PROBE(zfs__rrwfastpath__exitmiss);
233 #endif
234 ASSERT(!refcount_is_zero(&rrl->rr_anon_rcount) ||
235 !refcount_is_zero(&rrl->rr_linked_rcount) ||
236 rrl->rr_writer != NULL);
237
238 if (rrl->rr_writer == NULL) {
239 int64_t count;
240 if (rrn_find_and_remove(rrl, tag)) {
241 count = refcount_remove(&rrl->rr_linked_rcount, tag);
242 } else {
243 ASSERT(!rrl->rr_track_all);
244 count = refcount_remove(&rrl->rr_anon_rcount, tag);
245 }
246 if (count == 0)
247 cv_broadcast(&rrl->rr_cv);
248 } else {
249 ASSERT(rrl->rr_writer == curthread);
250 ASSERT(refcount_is_zero(&rrl->rr_anon_rcount) &&
251 refcount_is_zero(&rrl->rr_linked_rcount));
252 rrl->rr_writer = NULL;
253 cv_broadcast(&rrl->rr_cv);
254 }
255 mutex_exit(&rrl->rr_lock);
256 }
257
258 /*
259 * If the lock was created with track_all, rrw_held(RW_READER) will return
260 * B_TRUE iff the current thread has the lock for reader. Otherwise it may
261 * return B_TRUE if any thread has the lock for reader.
262 */
263 boolean_t
264 rrw_held(rrwlock_t *rrl, krw_t rw)
265 {
266 boolean_t held;
267
268 mutex_enter(&rrl->rr_lock);
269 if (rw == RW_WRITER) {
270 held = (rrl->rr_writer == curthread);
271 } else {
272 held = (!refcount_is_zero(&rrl->rr_anon_rcount) ||
273 rrn_find(rrl) != NULL);
274 }
275 mutex_exit(&rrl->rr_lock);
276
277 return (held);
278 }
279
280 void
281 rrw_tsd_destroy(void *arg)
282 {
283 rrw_node_t *rn = arg;
284 if (rn != NULL) {
285 panic("thread %p terminating with rrw lock %p held",
286 (void *)curthread, (void *)rn->rn_rrl);
287 }
288 }
289
290 /*
291 * A reader-mostly lock implementation, tuning above reader-writer locks
292 * for hightly parallel read acquisitions, while pessimizing writes.
293 *
294 * The idea is to split single busy lock into array of locks, so that
295 * each reader can lock only one of them for read, depending on result
296 * of simple hash function. That proportionally reduces lock congestion.
297 * Writer same time has to sequentially aquire write on all the locks.
298 * That makes write aquisition proportionally slower, but in places where
299 * it is used (filesystem unmount) performance is not critical.
300 *
301 * All the functions below are direct wrappers around functions above.
302 */
303 void
304 rrm_init(rrmlock_t *rrl, boolean_t track_all)
305 {
306 int i;
307
308 for (i = 0; i < RRM_NUM_LOCKS; i++)
309 rrw_init(&rrl->locks[i], track_all);
310 }
311
312 void
313 rrm_destroy(rrmlock_t *rrl)
314 {
315 int i;
316
317 for (i = 0; i < RRM_NUM_LOCKS; i++)
318 rrw_destroy(&rrl->locks[i]);
319 }
320
321 void
322 rrm_enter(rrmlock_t *rrl, krw_t rw, void *tag)
323 {
324 if (rw == RW_READER)
325 rrm_enter_read(rrl, tag);
326 else
327 rrm_enter_write(rrl);
328 }
329
330 /*
331 * This maps the current thread to a specific lock. Note that the lock
332 * must be released by the same thread that acquired it. We do this
333 * mapping by taking the thread pointer mod a prime number. We examine
334 * only the low 32 bits of the thread pointer, because 32-bit division
335 * is faster than 64-bit division, and the high 32 bits have little
336 * entropy anyway.
337 */
338 #define RRM_TD_LOCK() (((uint32_t)(uintptr_t)(curthread)) % RRM_NUM_LOCKS)
339
340 void
341 rrm_enter_read(rrmlock_t *rrl, void *tag)
342 {
343 rrw_enter_read(&rrl->locks[RRM_TD_LOCK()], tag);
344 }
345
346 void
347 rrm_enter_write(rrmlock_t *rrl)
348 {
349 int i;
350
351 for (i = 0; i < RRM_NUM_LOCKS; i++)
352 rrw_enter_write(&rrl->locks[i]);
353 }
354
355 void
356 rrm_exit(rrmlock_t *rrl, void *tag)
357 {
358 int i;
359
360 if (rrl->locks[0].rr_writer == curthread) {
361 for (i = 0; i < RRM_NUM_LOCKS; i++)
362 rrw_exit(&rrl->locks[i], tag);
363 } else {
364 rrw_exit(&rrl->locks[RRM_TD_LOCK()], tag);
365 }
366 }
367
368 boolean_t
369 rrm_held(rrmlock_t *rrl, krw_t rw)
370 {
371 if (rw == RW_WRITER) {
372 return (rrw_held(&rrl->locks[0], rw));
373 } else {
374 return (rrw_held(&rrl->locks[RRM_TD_LOCK()], rw));
375 }
376 }