xref: /netbsd-src/sys/kern/kern_turnstile.c (revision 8ac07aec990b9d2e483062509d0a9fa5b4f57cf2)
1 /*	$NetBSD: kern_turnstile.c,v 1.18 2008/04/24 23:26:00 alc 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 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 1.18 2008/04/24 23:26:00 alc Exp $");
71 
72 #include <sys/param.h>
73 #include <sys/lockdebug.h>
74 #include <sys/pool.h>
75 #include <sys/proc.h>
76 #include <sys/sleepq.h>
77 #include <sys/systm.h>
78 
79 #include <uvm/uvm_extern.h>
80 
81 #define	TS_HASH_SIZE	64
82 #define	TS_HASH_MASK	(TS_HASH_SIZE - 1)
83 #define	TS_HASH(obj)	(((uintptr_t)(obj) >> 3) & TS_HASH_MASK)
84 
85 tschain_t	turnstile_tab[TS_HASH_SIZE];
86 pool_cache_t	turnstile_cache;
87 
88 int	turnstile_ctor(void *, void *, int);
89 
90 extern turnstile_t turnstile0;
91 
92 /*
93  * turnstile_init:
94  *
95  *	Initialize the turnstile mechanism.
96  */
97 void
98 turnstile_init(void)
99 {
100 	tschain_t *tc;
101 	int i;
102 
103 	for (i = 0; i < TS_HASH_SIZE; i++) {
104 		tc = &turnstile_tab[i];
105 		LIST_INIT(&tc->tc_chain);
106 		mutex_init(&tc->tc_mutex, MUTEX_DEFAULT, IPL_SCHED);
107 	}
108 
109 	turnstile_cache = pool_cache_init(sizeof(turnstile_t), 0, 0, 0,
110 	    "tstilepl", NULL, IPL_NONE, turnstile_ctor, NULL, NULL);
111 	KASSERT(turnstile_cache != NULL);
112 
113 	(void)turnstile_ctor(NULL, &turnstile0, 0);
114 }
115 
116 /*
117  * turnstile_ctor:
118  *
119  *	Constructor for turnstiles.
120  */
121 int
122 turnstile_ctor(void *arg, void *obj, int flags)
123 {
124 	turnstile_t *ts = obj;
125 
126 	memset(ts, 0, sizeof(*ts));
127 	sleepq_init(&ts->ts_sleepq[TS_READER_Q], NULL);
128 	sleepq_init(&ts->ts_sleepq[TS_WRITER_Q], NULL);
129 	return (0);
130 }
131 
132 /*
133  * turnstile_remove:
134  *
135  *	Remove an LWP from a turnstile sleep queue and wake it.
136  */
137 static inline void
138 turnstile_remove(turnstile_t *ts, lwp_t *l, sleepq_t *sq)
139 {
140 	turnstile_t *nts;
141 
142 	KASSERT(l->l_ts == ts);
143 
144 	/*
145 	 * This process is no longer using the active turnstile.
146 	 * Find an inactive one on the free list to give to it.
147 	 */
148 	if ((nts = ts->ts_free) != NULL) {
149 		KASSERT(TS_ALL_WAITERS(ts) > 1);
150 		l->l_ts = nts;
151 		ts->ts_free = nts->ts_free;
152 		nts->ts_free = NULL;
153 	} else {
154 		/*
155 		 * If the free list is empty, this is the last
156 		 * waiter.
157 		 */
158 		KASSERT(TS_ALL_WAITERS(ts) == 1);
159 		LIST_REMOVE(ts, ts_chain);
160 	}
161 
162 	(void)sleepq_remove(sq, l);
163 }
164 
165 /*
166  * turnstile_lookup:
167  *
168  *	Look up the turnstile for the specified lock.  This acquires and
169  *	holds the turnstile chain lock (sleep queue interlock).
170  */
171 turnstile_t *
172 turnstile_lookup(wchan_t obj)
173 {
174 	turnstile_t *ts;
175 	tschain_t *tc;
176 
177 	tc = &turnstile_tab[TS_HASH(obj)];
178 	mutex_spin_enter(&tc->tc_mutex);
179 
180 	LIST_FOREACH(ts, &tc->tc_chain, ts_chain)
181 		if (ts->ts_obj == obj)
182 			return (ts);
183 
184 	/*
185 	 * No turnstile yet for this lock.  No problem, turnstile_block()
186 	 * handles this by fetching the turnstile from the blocking thread.
187 	 */
188 	return (NULL);
189 }
190 
191 /*
192  * turnstile_exit:
193  *
194  *	Abort a turnstile operation.
195  */
196 void
197 turnstile_exit(wchan_t obj)
198 {
199 	tschain_t *tc;
200 
201 	tc = &turnstile_tab[TS_HASH(obj)];
202 	mutex_spin_exit(&tc->tc_mutex);
203 }
204 
205 /*
206  * turnstile_block:
207  *
208  *	 Enter an object into the turnstile chain and prepare the current
209  *	 LWP for sleep.
210  */
211 void
212 turnstile_block(turnstile_t *ts, int q, wchan_t obj, syncobj_t *sobj)
213 {
214 	lwp_t *l;
215 	lwp_t *cur; /* cached curlwp */
216 	lwp_t *owner;
217 	turnstile_t *ots;
218 	tschain_t *tc;
219 	sleepq_t *sq;
220 	pri_t prio, obase;
221 
222 	tc = &turnstile_tab[TS_HASH(obj)];
223 	l = cur = curlwp;
224 
225 	KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
226 	KASSERT(mutex_owned(&tc->tc_mutex));
227 	KASSERT(l != NULL && l->l_ts != NULL);
228 
229 	if (ts == NULL) {
230 		/*
231 		 * We are the first thread to wait for this object;
232 		 * lend our turnstile to it.
233 		 */
234 		ts = l->l_ts;
235 		KASSERT(TS_ALL_WAITERS(ts) == 0);
236 		KASSERT(TAILQ_EMPTY(&ts->ts_sleepq[TS_READER_Q].sq_queue) &&
237 			TAILQ_EMPTY(&ts->ts_sleepq[TS_WRITER_Q].sq_queue));
238 		ts->ts_obj = obj;
239 		ts->ts_inheritor = NULL;
240 		ts->ts_sleepq[TS_READER_Q].sq_mutex = &tc->tc_mutex;
241 		ts->ts_sleepq[TS_WRITER_Q].sq_mutex = &tc->tc_mutex;
242 		LIST_INSERT_HEAD(&tc->tc_chain, ts, ts_chain);
243 	} else {
244 		/*
245 		 * Object already has a turnstile.  Put our turnstile
246 		 * onto the free list, and reference the existing
247 		 * turnstile instead.
248 		 */
249 		ots = l->l_ts;
250 		ots->ts_free = ts->ts_free;
251 		ts->ts_free = ots;
252 		l->l_ts = ts;
253 
254 		KASSERT(ts->ts_obj == obj);
255 		KASSERT(TS_ALL_WAITERS(ts) != 0);
256 		KASSERT(!TAILQ_EMPTY(&ts->ts_sleepq[TS_READER_Q].sq_queue) ||
257 			!TAILQ_EMPTY(&ts->ts_sleepq[TS_WRITER_Q].sq_queue));
258 	}
259 
260 	sq = &ts->ts_sleepq[q];
261 	sleepq_enter(sq, l);
262 	LOCKDEBUG_BARRIER(&tc->tc_mutex, 1);
263 	l->l_kpriority = true;
264 	obase = l->l_kpribase;
265 	if (obase < PRI_KTHREAD)
266 		l->l_kpribase = PRI_KTHREAD;
267 	sleepq_enqueue(sq, obj, "tstile", sobj);
268 
269 	/*
270 	 * lend our priority to lwps on the blocking chain.
271 	 */
272 	prio = lwp_eprio(l);
273 	for (;;) {
274 		bool dolock;
275 
276 		if (l->l_wchan == NULL)
277 			break;
278 
279 		owner = (*l->l_syncobj->sobj_owner)(l->l_wchan);
280 		if (owner == NULL)
281 			break;
282 
283 		KASSERT(l != owner);
284 		KASSERT(cur != owner);
285 
286 		if (l->l_mutex != owner->l_mutex)
287 			dolock = true;
288 		else
289 			dolock = false;
290 		if (dolock && !lwp_trylock(owner)) {
291 			/*
292 			 * restart from curlwp.
293 			 */
294 			lwp_unlock(l);
295 			l = cur;
296 			lwp_lock(l);
297 			prio = lwp_eprio(l);
298 			continue;
299 		}
300 		if (prio <= lwp_eprio(owner)) {
301 			if (dolock)
302 				lwp_unlock(owner);
303 			break;
304 		}
305 		ts = l->l_ts;
306 		KASSERT(ts->ts_inheritor == owner || ts->ts_inheritor == NULL);
307 		if (ts->ts_inheritor == NULL) {
308 			ts->ts_inheritor = owner;
309 			ts->ts_eprio = prio;
310 			SLIST_INSERT_HEAD(&owner->l_pi_lenders, ts, ts_pichain);
311 			lwp_lendpri(owner, prio);
312 		} else if (prio > ts->ts_eprio) {
313 			ts->ts_eprio = prio;
314 			lwp_lendpri(owner, prio);
315 		}
316 		if (dolock)
317 			lwp_unlock(l);
318 		l = owner;
319 	}
320 	LOCKDEBUG_BARRIER(l->l_mutex, 1);
321 	if (cur->l_mutex != l->l_mutex) {
322 		lwp_unlock(l);
323 		lwp_lock(cur);
324 	}
325 	LOCKDEBUG_BARRIER(cur->l_mutex, 1);
326 
327 	sleepq_block(0, false);
328 	cur->l_kpribase = obase;
329 }
330 
331 /*
332  * turnstile_wakeup:
333  *
334  *	Wake up the specified number of threads that are blocked
335  *	in a turnstile.
336  */
337 void
338 turnstile_wakeup(turnstile_t *ts, int q, int count, lwp_t *nl)
339 {
340 	sleepq_t *sq;
341 	tschain_t *tc;
342 	lwp_t *l;
343 
344 	tc = &turnstile_tab[TS_HASH(ts->ts_obj)];
345 	sq = &ts->ts_sleepq[q];
346 
347 	KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
348 	KASSERT(count > 0 && count <= TS_WAITERS(ts, q));
349 	KASSERT(mutex_owned(&tc->tc_mutex) && sq->sq_mutex == &tc->tc_mutex);
350 	KASSERT(ts->ts_inheritor == curlwp || ts->ts_inheritor == NULL);
351 
352 	/*
353 	 * restore inherited priority if necessary.
354 	 */
355 
356 	if (ts->ts_inheritor != NULL) {
357 		turnstile_t *iter;
358 		turnstile_t *next;
359 		turnstile_t *prev = NULL;
360 		pri_t prio;
361 		bool dolock;
362 
363 		ts->ts_inheritor = NULL;
364 		l = curlwp;
365 
366 		dolock = l->l_mutex == l->l_cpu->ci_schedstate.spc_lwplock;
367 		if (dolock) {
368 			lwp_lock(l);
369 		}
370 
371 		/*
372 		 * the following loop does two things.
373 		 *
374 		 * - remove ts from the list.
375 		 *
376 		 * - from the rest of the list, find the highest priority.
377 		 */
378 
379 		prio = -1;
380 		KASSERT(!SLIST_EMPTY(&l->l_pi_lenders));
381 		for (iter = SLIST_FIRST(&l->l_pi_lenders);
382 		    iter != NULL; iter = next) {
383 			KASSERT(lwp_eprio(l) >= ts->ts_eprio);
384 			next = SLIST_NEXT(iter, ts_pichain);
385 			if (iter == ts) {
386 				if (prev == NULL) {
387 					SLIST_REMOVE_HEAD(&l->l_pi_lenders,
388 					    ts_pichain);
389 				} else {
390 					SLIST_REMOVE_AFTER(prev, ts_pichain);
391 				}
392 			} else if (prio < iter->ts_eprio) {
393 				prio = iter->ts_eprio;
394 			}
395 			prev = iter;
396 		}
397 
398 		lwp_lendpri(l, prio);
399 
400 		if (dolock) {
401 			lwp_unlock(l);
402 		}
403 	}
404 
405 	if (nl != NULL) {
406 #if defined(DEBUG) || defined(LOCKDEBUG)
407 		TAILQ_FOREACH(l, &sq->sq_queue, l_sleepchain) {
408 			if (l == nl)
409 				break;
410 		}
411 		if (l == NULL)
412 			panic("turnstile_wakeup: nl not on sleepq");
413 #endif
414 		turnstile_remove(ts, nl, sq);
415 	} else {
416 		while (count-- > 0) {
417 			l = TAILQ_FIRST(&sq->sq_queue);
418 			KASSERT(l != NULL);
419 			turnstile_remove(ts, l, sq);
420 		}
421 	}
422 	mutex_spin_exit(&tc->tc_mutex);
423 }
424 
425 /*
426  * turnstile_unsleep:
427  *
428  *	Remove an LWP from the turnstile.  This is called when the LWP has
429  *	not been awoken normally but instead interrupted: for example, if it
430  *	has received a signal.  It's not a valid action for turnstiles,
431  *	since LWPs blocking on a turnstile are not interruptable.
432  */
433 u_int
434 turnstile_unsleep(lwp_t *l, bool cleanup)
435 {
436 
437 	lwp_unlock(l);
438 	panic("turnstile_unsleep");
439 }
440 
441 /*
442  * turnstile_changepri:
443  *
444  *	Adjust the priority of an LWP residing on a turnstile.
445  */
446 void
447 turnstile_changepri(lwp_t *l, pri_t pri)
448 {
449 
450 	/* XXX priority inheritance */
451 	sleepq_changepri(l, pri);
452 }
453 
454 #if defined(LOCKDEBUG)
455 /*
456  * turnstile_print:
457  *
458  *	Given the address of a lock object, print the contents of a
459  *	turnstile.
460  */
461 void
462 turnstile_print(volatile void *obj, void (*pr)(const char *, ...))
463 {
464 	turnstile_t *ts;
465 	tschain_t *tc;
466 	sleepq_t *rsq, *wsq;
467 	lwp_t *l;
468 
469 	tc = &turnstile_tab[TS_HASH(obj)];
470 
471 	LIST_FOREACH(ts, &tc->tc_chain, ts_chain)
472 		if (ts->ts_obj == obj)
473 			break;
474 
475 	(*pr)("Turnstile chain at %p.\n", tc);
476 	if (ts == NULL) {
477 		(*pr)("=> No active turnstile for this lock.\n");
478 		return;
479 	}
480 
481 	rsq = &ts->ts_sleepq[TS_READER_Q];
482 	wsq = &ts->ts_sleepq[TS_WRITER_Q];
483 
484 	(*pr)("=> Turnstile at %p (wrq=%p, rdq=%p).\n", ts, rsq, wsq);
485 
486 	(*pr)("=> %d waiting readers:", rsq->sq_waiters);
487 	TAILQ_FOREACH(l, &rsq->sq_queue, l_sleepchain) {
488 		(*pr)(" %p", l);
489 	}
490 	(*pr)("\n");
491 
492 	(*pr)("=> %d waiting writers:", wsq->sq_waiters);
493 	TAILQ_FOREACH(l, &wsq->sq_queue, l_sleepchain) {
494 		(*pr)(" %p", l);
495 	}
496 	(*pr)("\n");
497 }
498 #endif	/* LOCKDEBUG */
499