kern_turnstile.c revision 1.6.2.3 1 /* $NetBSD: kern_turnstile.c,v 1.6.2.3 2007/04/10 18:34:05 ad Exp $ */
2
3 /*-
4 * Copyright (c) 2002, 2006, 2007 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 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the NetBSD
21 * Foundation, Inc. and its contributors.
22 * 4. Neither the name of The NetBSD Foundation nor the names of its
23 * contributors may be used to endorse or promote products derived
24 * from this software without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36 * POSSIBILITY OF SUCH DAMAGE.
37 */
38
39 /*
40 * Turnstiles are described in detail in:
41 *
42 * Solaris Internals: Core Kernel Architecture, Jim Mauro and
43 * Richard McDougall.
44 *
45 * Turnstiles are kept in a hash table. There are likely to be many more
46 * synchronisation objects than there are threads. Since a thread can block
47 * on only one lock at a time, we only need one turnstile per thread, and
48 * so they are allocated at thread creation time.
49 *
50 * When a thread decides it needs to block on a lock, it looks up the
51 * active turnstile for that lock. If no active turnstile exists, then
52 * the process lends its turnstile to the lock. If there is already an
53 * active turnstile for the lock, the thread places its turnstile on a
54 * list of free turnstiles, and references the active one instead.
55 *
56 * The act of looking up the turnstile acquires an interlock on the sleep
57 * queue. If a thread decides it doesn't need to block after all, then this
58 * interlock must be released by explicitly aborting the turnstile
59 * operation.
60 *
61 * When a thread is awakened, it needs to get its turnstile back. If there
62 * are still other threads waiting in the active turnstile, the the thread
63 * grabs a free turnstile off the free list. Otherwise, it can take back
64 * the active turnstile from the lock (thus deactivating the turnstile).
65 *
66 * Turnstiles are the place to do priority inheritence.
67 */
68
69 #include <sys/cdefs.h>
70 __KERNEL_RCSID(0, "$NetBSD: kern_turnstile.c,v 1.6.2.3 2007/04/10 18:34:05 ad Exp $");
71
72 #include "opt_lockdebug.h"
73 #include "opt_multiprocessor.h"
74 #include "opt_ktrace.h"
75 #include "opt_ddb.h"
76
77 #include <sys/param.h>
78 #include <sys/lock.h>
79 #include <sys/lockdebug.h>
80 #include <sys/pool.h>
81 #include <sys/proc.h>
82 #include <sys/sleepq.h>
83 #include <sys/systm.h>
84
85 #define TS_HASH_SIZE 64
86 #define TS_HASH_MASK (TS_HASH_SIZE - 1)
87 #define TS_HASH(obj) (((uintptr_t)(obj) >> 3) & TS_HASH_MASK)
88
89 tschain_t turnstile_tab[TS_HASH_SIZE];
90
91 struct pool turnstile_pool;
92 struct pool_cache turnstile_cache;
93
94 int turnstile_ctor(void *, void *, int);
95
96 extern turnstile_t turnstile0;
97
98 /*
99 * turnstile_init:
100 *
101 * Initialize the turnstile mechanism.
102 */
103 void
104 turnstile_init(void)
105 {
106 tschain_t *tc;
107 int i;
108
109 for (i = 0; i < TS_HASH_SIZE; i++) {
110 tc = &turnstile_tab[i];
111 LIST_INIT(&tc->tc_chain);
112 #if defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
113 mutex_init(&tc->tc_mutexstore, MUTEX_SPIN, IPL_SCHED);
114 tc->tc_mutex = &tc->tc_mutexstore;
115 #else
116 tc->tc_mutex = &sched_mutex;
117 #endif
118 }
119
120 pool_init(&turnstile_pool, sizeof(turnstile_t), 0, 0, 0,
121 "tstilepl", &pool_allocator_nointr, IPL_NONE);
122 pool_cache_init(&turnstile_cache, &turnstile_pool,
123 turnstile_ctor, NULL, NULL);
124
125 (void)turnstile_ctor(NULL, &turnstile0, 0);
126 }
127
128 /*
129 * turnstile_ctor:
130 *
131 * Constructor for turnstiles.
132 */
133 int
134 turnstile_ctor(void *arg, void *obj, int flags)
135 {
136 turnstile_t *ts = obj;
137
138 memset(ts, 0, sizeof(*ts));
139 sleepq_init(&ts->ts_sleepq[TS_READER_Q], NULL);
140 sleepq_init(&ts->ts_sleepq[TS_WRITER_Q], NULL);
141 return (0);
142 }
143
144 /*
145 * turnstile_remove:
146 *
147 * Remove an LWP from a turnstile sleep queue and wake it.
148 */
149 static inline void
150 turnstile_remove(turnstile_t *ts, struct lwp *l, sleepq_t *sq)
151 {
152 turnstile_t *nts;
153
154 KASSERT(l->l_ts == ts);
155
156 /*
157 * This process is no longer using the active turnstile.
158 * Find an inactive one on the free list to give to it.
159 */
160 if ((nts = ts->ts_free) != NULL) {
161 KASSERT(TS_ALL_WAITERS(ts) > 1);
162 l->l_ts = nts;
163 ts->ts_free = nts->ts_free;
164 nts->ts_free = NULL;
165 } else {
166 /*
167 * If the free list is empty, this is the last
168 * waiter.
169 */
170 KASSERT(TS_ALL_WAITERS(ts) == 1);
171 LIST_REMOVE(ts, ts_chain);
172 }
173
174 /*
175 * Note that LWPs blocked on a turnstile cannot be swapped
176 * out. This is necessary to avoid deadlock, and because
177 * the wait times will be short.
178 */
179 (void)sleepq_remove(sq, l);
180 }
181
182 /*
183 * turnstile_lookup:
184 *
185 * Look up the turnstile for the specified lock. This acquires and
186 * holds the turnstile chain lock (sleep queue interlock).
187 */
188 turnstile_t *
189 turnstile_lookup(wchan_t obj)
190 {
191 turnstile_t *ts;
192 tschain_t *tc;
193
194 tc = &turnstile_tab[TS_HASH(obj)];
195 mutex_spin_enter(tc->tc_mutex);
196
197 LIST_FOREACH(ts, &tc->tc_chain, ts_chain)
198 if (ts->ts_obj == obj)
199 return (ts);
200
201 /*
202 * No turnstile yet for this lock. No problem, turnstile_block()
203 * handles this by fetching the turnstile from the blocking thread.
204 */
205 return (NULL);
206 }
207
208 /*
209 * turnstile_exit:
210 *
211 * Abort a turnstile operation.
212 */
213 void
214 turnstile_exit(wchan_t obj)
215 {
216 tschain_t *tc;
217
218 tc = &turnstile_tab[TS_HASH(obj)];
219 mutex_spin_exit(tc->tc_mutex);
220 }
221
222 /*
223 * turnstile_block:
224 *
225 * Enter an object into the turnstile chain and prepare the current
226 * LWP for sleep.
227 */
228 void
229 turnstile_block(turnstile_t *ts, int q, wchan_t obj, syncobj_t *sobj)
230 {
231 struct lwp *l;
232 struct lwp *cur; /* cached curlwp */
233 struct lwp *owner;
234 turnstile_t *ots;
235 tschain_t *tc;
236 sleepq_t *sq;
237 pri_t prio;
238
239 tc = &turnstile_tab[TS_HASH(obj)];
240 l = cur = curlwp;
241
242 KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
243 KASSERT(mutex_owned(tc->tc_mutex));
244 KASSERT(l != NULL && l->l_ts != NULL);
245
246 if (ts == NULL) {
247 /*
248 * We are the first thread to wait for this object;
249 * lend our turnstile to it.
250 */
251 ts = l->l_ts;
252 KASSERT(TS_ALL_WAITERS(ts) == 0);
253 KASSERT(TAILQ_EMPTY(&ts->ts_sleepq[TS_READER_Q].sq_queue) &&
254 TAILQ_EMPTY(&ts->ts_sleepq[TS_WRITER_Q].sq_queue));
255 ts->ts_obj = obj;
256 ts->ts_inheritor = NULL;
257 ts->ts_sleepq[TS_READER_Q].sq_mutex = tc->tc_mutex;
258 ts->ts_sleepq[TS_WRITER_Q].sq_mutex = tc->tc_mutex;
259 LIST_INSERT_HEAD(&tc->tc_chain, ts, ts_chain);
260 } else {
261 /*
262 * Object already has a turnstile. Put our turnstile
263 * onto the free list, and reference the existing
264 * turnstile instead.
265 */
266 ots = l->l_ts;
267 ots->ts_free = ts->ts_free;
268 ts->ts_free = ots;
269 l->l_ts = ts;
270
271 KASSERT(ts->ts_obj == obj);
272 KASSERT(TS_ALL_WAITERS(ts) != 0);
273 KASSERT(!TAILQ_EMPTY(&ts->ts_sleepq[TS_READER_Q].sq_queue) ||
274 !TAILQ_EMPTY(&ts->ts_sleepq[TS_WRITER_Q].sq_queue));
275 }
276
277 sq = &ts->ts_sleepq[q];
278 sleepq_enter(sq, l);
279 LOCKDEBUG_BARRIER(tc->tc_mutex, 1);
280 sleepq_enqueue(sq, sched_kpri(l), obj, "tstile", sobj);
281 prio = lwp_eprio(l);
282
283 /*
284 * lend our priority to lwps on the blocking chain.
285 */
286
287 for (;;) {
288 bool dolock;
289
290 if (l->l_wchan == NULL)
291 break;
292
293 owner = (*l->l_syncobj->sobj_owner)(l->l_wchan);
294 if (owner == NULL)
295 break;
296
297 KASSERT(l != owner);
298 KASSERT(cur != owner);
299
300 if (l->l_mutex != owner->l_mutex)
301 dolock = true;
302 else
303 dolock = false;
304 if (dolock && !lwp_trylock(owner)) {
305 /*
306 * restart from curlwp.
307 */
308 lwp_unlock(l);
309 l = cur;
310 lwp_lock(l);
311 prio = lwp_eprio(l);
312 continue;
313 }
314 if (prio >= lwp_eprio(owner)) {
315 if (dolock)
316 lwp_unlock(owner);
317 break;
318 }
319 ts = l->l_ts;
320 KASSERT(ts->ts_inheritor == owner || ts->ts_inheritor == NULL);
321 if (ts->ts_inheritor == NULL) {
322 ts->ts_inheritor = owner;
323 ts->ts_eprio = prio;
324 SLIST_INSERT_HEAD(&owner->l_pi_lenders, ts, ts_pichain);
325 lwp_lendpri(owner, prio);
326 } else if (prio < ts->ts_eprio) {
327 ts->ts_eprio = prio;
328 lwp_lendpri(owner, prio);
329 }
330 if (dolock)
331 lwp_unlock(l);
332 l = owner;
333 }
334 LOCKDEBUG_BARRIER(l->l_mutex, 1);
335 if (cur->l_mutex != l->l_mutex) {
336 lwp_unlock(l);
337 lwp_lock(cur);
338 }
339 LOCKDEBUG_BARRIER(cur->l_mutex, 1);
340
341 (void)sleepq_block(0, false);
342 }
343
344 /*
345 * turnstile_wakeup:
346 *
347 * Wake up the specified number of threads that are blocked
348 * in a turnstile.
349 */
350 void
351 turnstile_wakeup(turnstile_t *ts, int q, int count, struct lwp *nl)
352 {
353 sleepq_t *sq;
354 tschain_t *tc;
355 struct lwp *l;
356
357 tc = &turnstile_tab[TS_HASH(ts->ts_obj)];
358 sq = &ts->ts_sleepq[q];
359
360 KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
361 KASSERT(count > 0 && count <= TS_WAITERS(ts, q));
362 KASSERT(mutex_owned(tc->tc_mutex) && sq->sq_mutex == tc->tc_mutex);
363 KASSERT(ts->ts_inheritor == curlwp || ts->ts_inheritor == NULL);
364
365 /*
366 * restore inherited priority if necessary.
367 */
368
369 if (ts->ts_inheritor != NULL) {
370 turnstile_t *iter;
371 turnstile_t *next;
372 turnstile_t *prev = NULL;
373 pri_t prio;
374
375 ts->ts_inheritor = NULL;
376 l = curlwp;
377
378 if (l->l_mutex == &sched_mutex) {
379 sleepq_lwp_lock(l);
380 }
381
382 /*
383 * the following loop does two things.
384 *
385 * - remove ts from the list.
386 *
387 * - from the rest of the list, find the highest priority.
388 */
389
390 prio = MAXPRI;
391 KASSERT(!SLIST_EMPTY(&l->l_pi_lenders));
392 for (iter = SLIST_FIRST(&l->l_pi_lenders);
393 iter != NULL; iter = next) {
394 KASSERT(lwp_eprio(l) <= ts->ts_eprio);
395 next = SLIST_NEXT(iter, ts_pichain);
396 if (iter == ts) {
397 if (prev == NULL) {
398 SLIST_REMOVE_HEAD(&l->l_pi_lenders,
399 ts_pichain);
400 } else {
401 SLIST_REMOVE_AFTER(prev, ts_pichain);
402 }
403 } else if (prio > iter->ts_eprio) {
404 prio = iter->ts_eprio;
405 }
406 prev = iter;
407 }
408
409 lwp_lendpri(l, prio);
410
411 if (l->l_mutex == &sched_mutex) {
412 sleepq_lwp_unlock(l);
413 }
414 }
415
416 if (nl != NULL) {
417 #if defined(DEBUG) || defined(LOCKDEBUG)
418 TAILQ_FOREACH(l, &sq->sq_queue, l_sleepchain) {
419 if (l == nl)
420 break;
421 }
422 if (l == NULL)
423 panic("turnstile_wakeup: nl not on sleepq");
424 #endif
425 turnstile_remove(ts, nl, sq);
426 } else {
427 while (count-- > 0) {
428 l = TAILQ_FIRST(&sq->sq_queue);
429 KASSERT(l != NULL);
430 turnstile_remove(ts, l, sq);
431 }
432 }
433 mutex_spin_exit(tc->tc_mutex);
434 }
435
436 /*
437 * turnstile_unsleep:
438 *
439 * Remove an LWP from the turnstile. This is called when the LWP has
440 * not been awoken normally but instead interrupted: for example, if it
441 * has received a signal. It's not a valid action for turnstiles,
442 * since LWPs blocking on a turnstile are not interruptable.
443 */
444 void
445 turnstile_unsleep(struct lwp *l)
446 {
447
448 lwp_unlock(l);
449 panic("turnstile_unsleep");
450 }
451
452 /*
453 * turnstile_changepri:
454 *
455 * Adjust the priority of an LWP residing on a turnstile.
456 */
457 void
458 turnstile_changepri(struct lwp *l, pri_t pri)
459 {
460
461 /* XXX priority inheritance */
462 sleepq_changepri(l, pri);
463 }
464
465 #if defined(LOCKDEBUG)
466 /*
467 * turnstile_print:
468 *
469 * Given the address of a lock object, print the contents of a
470 * turnstile.
471 */
472 void
473 turnstile_print(volatile void *obj, void (*pr)(const char *, ...))
474 {
475 turnstile_t *ts;
476 tschain_t *tc;
477 sleepq_t *rsq, *wsq;
478 struct lwp *l;
479
480 tc = &turnstile_tab[TS_HASH(obj)];
481
482 LIST_FOREACH(ts, &tc->tc_chain, ts_chain)
483 if (ts->ts_obj == obj)
484 break;
485
486 (*pr)("Turnstile chain at %p with tc_mutex at %p.\n", tc, tc->tc_mutex);
487 if (ts == NULL) {
488 (*pr)("=> No active turnstile for this lock.\n");
489 return;
490 }
491
492 rsq = &ts->ts_sleepq[TS_READER_Q];
493 wsq = &ts->ts_sleepq[TS_WRITER_Q];
494
495 (*pr)("=> Turnstile at %p (wrq=%p, rdq=%p).\n", ts, rsq, wsq);
496
497 (*pr)("=> %d waiting readers:", rsq->sq_waiters);
498 TAILQ_FOREACH(l, &rsq->sq_queue, l_sleepchain) {
499 (*pr)(" %p", l);
500 }
501 (*pr)("\n");
502
503 (*pr)("=> %d waiting writers:", wsq->sq_waiters);
504 TAILQ_FOREACH(l, &wsq->sq_queue, l_sleepchain) {
505 (*pr)(" %p", l);
506 }
507 (*pr)("\n");
508 }
509 #endif /* LOCKDEBUG */
510