xref: /netbsd-src/sys/kern/kern_turnstile.c (revision cee67a0a7118f2ce9b95a1a8121764ba0705e7c4)
1 /*	$NetBSD: kern_turnstile.c,v 1.55 2023/10/15 10:30:20 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2009, 2019, 2020, 2023
5  *     The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Jason R. Thorpe and Andrew Doran.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 /*
34  * Turnstiles are described in detail in:
35  *
36  *	Solaris Internals: Core Kernel Architecture, Jim Mauro and
37  *	    Richard McDougall.
38  *
39  * Turnstiles are kept in a hash table.  There are likely to be many more
40  * synchronisation objects than there are threads.  Since a thread can block
41  * on only one lock at a time, we only need one turnstile per thread, and
42  * so they are allocated at thread creation time.
43  *
44  * When a thread decides it needs to block on a lock, it looks up the
45  * active turnstile for that lock.  If no active turnstile exists, then
46  * the process lends its turnstile to the lock.  If there is already an
47  * active turnstile for the lock, the thread places its turnstile on a
48  * list of free turnstiles, and references the active one instead.
49  *
50  * The act of looking up the turnstile acquires an interlock on the sleep
51  * queue.  If a thread decides it doesn't need to block after all, then this
52  * interlock must be released by explicitly aborting the turnstile
53  * operation.
54  *
55  * When a thread is awakened, it needs to get its turnstile back.  If there
56  * are still other threads waiting in the active turnstile, the thread
57  * grabs a free turnstile off the free list.  Otherwise, it can take back
58  * the active turnstile from the lock (thus deactivating the turnstile).
59  *
60  * Turnstiles are where we do priority inheritence.
61  */
62 
63 #include <sys/cdefs.h>
64 __KERNEL_RCSID(0, "$NetBSD: kern_turnstile.c,v 1.55 2023/10/15 10:30:20 riastradh Exp $");
65 
66 #include <sys/param.h>
67 
68 #include <sys/lockdebug.h>
69 #include <sys/lwp.h>
70 #include <sys/proc.h>
71 #include <sys/sleepq.h>
72 #include <sys/sleeptab.h>
73 #include <sys/syncobj.h>
74 #include <sys/systm.h>
75 
76 /*
77  * Shift of 6 aligns to typical cache line size of 64 bytes;  there's no
78  * point having two turnstile locks to back two lock objects that share one
79  * cache line.
80  */
81 #define	TS_HASH_SIZE	128
82 #define	TS_HASH_MASK	(TS_HASH_SIZE - 1)
83 #define	TS_HASH(obj)	(((uintptr_t)(obj) >> 6) & TS_HASH_MASK)
84 
85 static tschain_t	turnstile_chains[TS_HASH_SIZE] __cacheline_aligned;
86 
87 static union {
88 	kmutex_t	lock;
89 	uint8_t		pad[COHERENCY_UNIT];
90 } turnstile_locks[TS_HASH_SIZE] __cacheline_aligned;
91 
92 /*
93  * turnstile_init:
94  *
95  *	Initialize the turnstile mechanism.
96  */
97 void
turnstile_init(void)98 turnstile_init(void)
99 {
100 	int i;
101 
102 	for (i = 0; i < TS_HASH_SIZE; i++) {
103 		LIST_INIT(&turnstile_chains[i]);
104 		mutex_init(&turnstile_locks[i].lock, MUTEX_DEFAULT, IPL_SCHED);
105 	}
106 
107 	turnstile_ctor(&turnstile0);
108 }
109 
110 /*
111  * turnstile_ctor:
112  *
113  *	Constructor for turnstiles.
114  */
115 void
turnstile_ctor(turnstile_t * ts)116 turnstile_ctor(turnstile_t *ts)
117 {
118 
119 	memset(ts, 0, sizeof(*ts));
120 	sleepq_init(&ts->ts_sleepq[TS_READER_Q]);
121 	sleepq_init(&ts->ts_sleepq[TS_WRITER_Q]);
122 }
123 
124 /*
125  * turnstile_remove:
126  *
127  *	Remove an LWP from a turnstile sleep queue and wake it.
128  */
129 static inline void
turnstile_remove(turnstile_t * ts,lwp_t * l,int q)130 turnstile_remove(turnstile_t *ts, lwp_t *l, int q)
131 {
132 	turnstile_t *nts;
133 
134 	KASSERT(l->l_ts == ts);
135 
136 	/*
137 	 * This process is no longer using the active turnstile.
138 	 * Find an inactive one on the free list to give to it.
139 	 */
140 	if ((nts = ts->ts_free) != NULL) {
141 		KASSERT(TS_ALL_WAITERS(ts) > 1);
142 		l->l_ts = nts;
143 		ts->ts_free = nts->ts_free;
144 		nts->ts_free = NULL;
145 	} else {
146 		/*
147 		 * If the free list is empty, this is the last
148 		 * waiter.
149 		 */
150 		KASSERT(TS_ALL_WAITERS(ts) == 1);
151 		LIST_REMOVE(ts, ts_chain);
152 	}
153 
154 	ts->ts_waiters[q]--;
155 	sleepq_remove(&ts->ts_sleepq[q], l, true);
156 }
157 
158 /*
159  * turnstile_lookup:
160  *
161  *	Look up the turnstile for the specified lock.  This acquires and
162  *	holds the turnstile chain lock (sleep queue interlock).
163  */
164 turnstile_t *
turnstile_lookup(wchan_t obj)165 turnstile_lookup(wchan_t obj)
166 {
167 	turnstile_t *ts;
168 	tschain_t *tc;
169 	u_int hash;
170 
171 	hash = TS_HASH(obj);
172 	tc = &turnstile_chains[hash];
173 	mutex_spin_enter(&turnstile_locks[hash].lock);
174 
175 	LIST_FOREACH(ts, tc, ts_chain)
176 		if (ts->ts_obj == obj)
177 			return (ts);
178 
179 	/*
180 	 * No turnstile yet for this lock.  No problem, turnstile_block()
181 	 * handles this by fetching the turnstile from the blocking thread.
182 	 */
183 	return (NULL);
184 }
185 
186 /*
187  * turnstile_exit:
188  *
189  *	Abort a turnstile operation.
190  */
191 void
turnstile_exit(wchan_t obj)192 turnstile_exit(wchan_t obj)
193 {
194 
195 	mutex_spin_exit(&turnstile_locks[TS_HASH(obj)].lock);
196 }
197 
198 /*
199  * turnstile_lendpri:
200  *
201  *	Lend our priority to lwps on the blocking chain.
202  *
203  *	If the current owner of the lock (l->l_wchan, set by sleepq_enqueue)
204  *	has a priority lower than ours (lwp_eprio(l)), lend our priority to
205  *	him to avoid priority inversions.
206  */
207 
208 static void
turnstile_lendpri(lwp_t * cur)209 turnstile_lendpri(lwp_t *cur)
210 {
211 	lwp_t * l = cur;
212 	pri_t prio;
213 
214 	/*
215 	 * NOTE: if you get a panic in this code block, it is likely that
216 	 * a lock has been destroyed or corrupted while still in use.  Try
217 	 * compiling a kernel with LOCKDEBUG to pinpoint the problem.
218 	 */
219 
220 	LOCKDEBUG_BARRIER(l->l_mutex, 1);
221 	KASSERT(l == curlwp);
222 	prio = lwp_eprio(l);
223 	for (;;) {
224 		lwp_t *owner;
225 		turnstile_t *ts;
226 		bool dolock;
227 
228 		if (l->l_wchan == NULL)
229 			break;
230 
231 		/*
232 		 * Ask syncobj the owner of the lock.
233 		 */
234 		owner = (*l->l_syncobj->sobj_owner)(l->l_wchan);
235 		if (owner == NULL)
236 			break;
237 
238 		/*
239 		 * The owner may have changed as we have dropped the tc lock.
240 		 */
241 		if (cur == owner) {
242 			/*
243 			 * We own the lock: stop here, sleepq_block()
244 			 * should wake up immediately.
245 			 */
246 			break;
247 		}
248 		/*
249 		 * Acquire owner->l_mutex if we don't have it yet.
250 		 * Because we already have another LWP lock (l->l_mutex) held,
251 		 * we need to play a try lock dance to avoid deadlock.
252 		 */
253 		dolock = l->l_mutex != atomic_load_relaxed(&owner->l_mutex);
254 		if (l == owner || (dolock && !lwp_trylock(owner))) {
255 			/*
256 			 * The owner was changed behind us or trylock failed.
257 			 * Restart from curlwp.
258 			 *
259 			 * Note that there may be a livelock here:
260 			 * the owner may try grabbing cur's lock (which is the
261 			 * tc lock) while we're trying to grab the owner's lock.
262 			 */
263 			lwp_unlock(l);
264 			l = cur;
265 			lwp_lock(l);
266 			prio = lwp_eprio(l);
267 			continue;
268 		}
269 		/*
270 		 * If the owner's priority is already higher than ours,
271 		 * there's nothing to do anymore.
272 		 */
273 		if (prio <= lwp_eprio(owner)) {
274 			if (dolock)
275 				lwp_unlock(owner);
276 			break;
277 		}
278 		/*
279 		 * Lend our priority to the 'owner' LWP.
280 		 *
281 		 * Update lenders info for turnstile_unlendpri.
282 		 */
283 		ts = l->l_ts;
284 		KASSERT(ts->ts_inheritor == owner || ts->ts_inheritor == NULL);
285 		if (ts->ts_inheritor == NULL) {
286 			ts->ts_inheritor = owner;
287 			ts->ts_eprio = prio;
288 			SLIST_INSERT_HEAD(&owner->l_pi_lenders, ts, ts_pichain);
289 			lwp_lendpri(owner, prio);
290 		} else if (prio > ts->ts_eprio) {
291 			ts->ts_eprio = prio;
292 			lwp_lendpri(owner, prio);
293 		}
294 		if (dolock)
295 			lwp_unlock(l);
296 		LOCKDEBUG_BARRIER(owner->l_mutex, 1);
297 		l = owner;
298 	}
299 	LOCKDEBUG_BARRIER(l->l_mutex, 1);
300 	if (cur->l_mutex != atomic_load_relaxed(&l->l_mutex)) {
301 		lwp_unlock(l);
302 		lwp_lock(cur);
303 	}
304 	LOCKDEBUG_BARRIER(cur->l_mutex, 1);
305 }
306 
307 /*
308  * turnstile_unlendpri: undo turnstile_lendpri
309  */
310 
311 static void
turnstile_unlendpri(turnstile_t * ts)312 turnstile_unlendpri(turnstile_t *ts)
313 {
314 	lwp_t * const l = curlwp;
315 	turnstile_t *iter;
316 	turnstile_t *next;
317 	turnstile_t *prev = NULL;
318 	pri_t prio;
319 	bool dolock;
320 
321 	KASSERT(ts->ts_inheritor != NULL);
322 	ts->ts_inheritor = NULL;
323 	dolock = (atomic_load_relaxed(&l->l_mutex) ==
324 	    l->l_cpu->ci_schedstate.spc_lwplock);
325 	if (dolock) {
326 		lwp_lock(l);
327 	}
328 
329 	/*
330 	 * the following loop does two things.
331 	 *
332 	 * - remove ts from the list.
333 	 *
334 	 * - from the rest of the list, find the highest priority.
335 	 */
336 
337 	prio = -1;
338 	KASSERT(!SLIST_EMPTY(&l->l_pi_lenders));
339 	for (iter = SLIST_FIRST(&l->l_pi_lenders);
340 	    iter != NULL; iter = next) {
341 		KASSERT(lwp_eprio(l) >= ts->ts_eprio);
342 		next = SLIST_NEXT(iter, ts_pichain);
343 		if (iter == ts) {
344 			if (prev == NULL) {
345 				SLIST_REMOVE_HEAD(&l->l_pi_lenders,
346 				    ts_pichain);
347 			} else {
348 				SLIST_REMOVE_AFTER(prev, ts_pichain);
349 			}
350 		} else if (prio < iter->ts_eprio) {
351 			prio = iter->ts_eprio;
352 		}
353 		prev = iter;
354 	}
355 
356 	lwp_lendpri(l, prio);
357 
358 	if (dolock) {
359 		lwp_unlock(l);
360 	}
361 }
362 
363 /*
364  * turnstile_block:
365  *
366  *	 Enter an object into the turnstile chain and prepare the current
367  *	 LWP for sleep.
368  */
369 void
turnstile_block(turnstile_t * ts,int q,wchan_t obj,syncobj_t * sobj)370 turnstile_block(turnstile_t *ts, int q, wchan_t obj, syncobj_t *sobj)
371 {
372 	lwp_t * const l = curlwp; /* cached curlwp */
373 	turnstile_t *ots;
374 	tschain_t *tc;
375 	kmutex_t *lock;
376 	sleepq_t *sq;
377 	u_int hash;
378 	int nlocks;
379 
380 	hash = TS_HASH(obj);
381 	tc = &turnstile_chains[hash];
382 	lock = &turnstile_locks[hash].lock;
383 
384 	KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
385 	KASSERT(mutex_owned(lock));
386 	KASSERT(l != NULL);
387 	KASSERT(l->l_ts != NULL);
388 
389 	if (ts == NULL) {
390 		/*
391 		 * We are the first thread to wait for this object;
392 		 * lend our turnstile to it.
393 		 */
394 		ts = l->l_ts;
395 		KASSERT(TS_ALL_WAITERS(ts) == 0);
396 		KASSERT(LIST_EMPTY(&ts->ts_sleepq[TS_READER_Q]));
397 		KASSERT(LIST_EMPTY(&ts->ts_sleepq[TS_WRITER_Q]));
398 		ts->ts_obj = obj;
399 		ts->ts_inheritor = NULL;
400 		LIST_INSERT_HEAD(tc, ts, ts_chain);
401 	} else {
402 		/*
403 		 * Object already has a turnstile.  Put our turnstile
404 		 * onto the free list, and reference the existing
405 		 * turnstile instead.
406 		 */
407 		ots = l->l_ts;
408 		KASSERT(ots->ts_free == NULL);
409 		ots->ts_free = ts->ts_free;
410 		ts->ts_free = ots;
411 		l->l_ts = ts;
412 
413 		KASSERT(ts->ts_obj == obj);
414 		KASSERT(TS_ALL_WAITERS(ts) != 0);
415 		KASSERT(!LIST_EMPTY(&ts->ts_sleepq[TS_READER_Q]) ||
416 			!LIST_EMPTY(&ts->ts_sleepq[TS_WRITER_Q]));
417 	}
418 
419 	sq = &ts->ts_sleepq[q];
420 	ts->ts_waiters[q]++;
421 	nlocks = sleepq_enter(sq, l, lock);
422 	LOCKDEBUG_BARRIER(lock, 1);
423 	sleepq_enqueue(sq, obj, sobj->sobj_name, sobj, false);
424 
425 	/*
426 	 * Disable preemption across this entire block, as we may drop
427 	 * scheduler locks (allowing preemption), and would prefer not
428 	 * to be interrupted while in a state of flux.
429 	 */
430 	KPREEMPT_DISABLE(l);
431 	KASSERT(lock == l->l_mutex);
432 	turnstile_lendpri(l);
433 	sleepq_block(0, false, sobj, nlocks);
434 	KPREEMPT_ENABLE(l);
435 }
436 
437 /*
438  * turnstile_wakeup:
439  *
440  *	Wake up the specified number of threads that are blocked
441  *	in a turnstile.
442  */
443 void
turnstile_wakeup(turnstile_t * ts,int q,int count,lwp_t * nl)444 turnstile_wakeup(turnstile_t *ts, int q, int count, lwp_t *nl)
445 {
446 	sleepq_t *sq;
447 	kmutex_t *lock;
448 	u_int hash;
449 	lwp_t *l;
450 
451 	hash = TS_HASH(ts->ts_obj);
452 	lock = &turnstile_locks[hash].lock;
453 	sq = &ts->ts_sleepq[q];
454 
455 	KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
456 	KASSERT(count > 0);
457 	KASSERT(count <= TS_WAITERS(ts, q));
458 	KASSERT(mutex_owned(lock));
459 	KASSERT(ts->ts_inheritor == curlwp || ts->ts_inheritor == NULL);
460 
461 	/*
462 	 * restore inherited priority if necessary.
463 	 */
464 
465 	if (ts->ts_inheritor != NULL) {
466 		turnstile_unlendpri(ts);
467 	}
468 
469 	if (nl != NULL) {
470 #if defined(DEBUG) || defined(LOCKDEBUG)
471 		LIST_FOREACH(l, sq, l_sleepchain) {
472 			if (l == nl)
473 				break;
474 		}
475 		if (l == NULL)
476 			panic("turnstile_wakeup: nl not on sleepq");
477 #endif
478 		turnstile_remove(ts, nl, q);
479 	} else {
480 		while (count-- > 0) {
481 			l = LIST_FIRST(sq);
482 			KASSERT(l != NULL);
483 			turnstile_remove(ts, l, q);
484 		}
485 	}
486 	mutex_spin_exit(lock);
487 }
488 
489 /*
490  * turnstile_unsleep:
491  *
492  *	Remove an LWP from the turnstile.  This is called when the LWP has
493  *	not been awoken normally but instead interrupted: for example, if it
494  *	has received a signal.  It's not a valid action for turnstiles,
495  *	since LWPs blocking on a turnstile are not interruptable.
496  */
497 void
turnstile_unsleep(lwp_t * l,bool cleanup)498 turnstile_unsleep(lwp_t *l, bool cleanup)
499 {
500 
501 	lwp_unlock(l);
502 	panic("turnstile_unsleep");
503 }
504 
505 /*
506  * turnstile_changepri:
507  *
508  *	Adjust the priority of an LWP residing on a turnstile.
509  */
510 void
turnstile_changepri(lwp_t * l,pri_t pri)511 turnstile_changepri(lwp_t *l, pri_t pri)
512 {
513 
514 	/* XXX priority inheritance */
515 	sleepq_changepri(l, pri);
516 }
517 
518 #if defined(LOCKDEBUG)
519 /*
520  * turnstile_print:
521  *
522  *	Given the address of a lock object, print the contents of a
523  *	turnstile.
524  */
525 void
turnstile_print(volatile void * obj,void (* pr)(const char *,...))526 turnstile_print(volatile void *obj, void (*pr)(const char *, ...))
527 {
528 	turnstile_t *ts;
529 	tschain_t *tc;
530 	sleepq_t *rsq, *wsq;
531 	u_int hash;
532 	lwp_t *l;
533 
534 	hash = TS_HASH(obj);
535 	tc = &turnstile_chains[hash];
536 
537 	LIST_FOREACH(ts, tc, ts_chain)
538 		if (ts->ts_obj == obj)
539 			break;
540 
541 	if (ts == NULL) {
542 		(*pr)("Turnstile: no active turnstile for this lock.\n");
543 		return;
544 	}
545 
546 	rsq = &ts->ts_sleepq[TS_READER_Q];
547 	wsq = &ts->ts_sleepq[TS_WRITER_Q];
548 
549 	(*pr)("Turnstile:\n");
550 	(*pr)("=> %d waiting readers:", TS_WAITERS(ts, TS_READER_Q));
551 	LIST_FOREACH(l, rsq, l_sleepchain) {
552 		(*pr)(" %p", l);
553 	}
554 	(*pr)("\n");
555 
556 	(*pr)("=> %d waiting writers:", TS_WAITERS(ts, TS_WRITER_Q));
557 	LIST_FOREACH(l, wsq, l_sleepchain) {
558 		(*pr)(" %p", l);
559 	}
560 	(*pr)("\n");
561 }
562 #endif	/* LOCKDEBUG */
563