1 /* $NetBSD: kern_turnstile.c,v 1.45 2022/10/26 23:27:16 riastradh Exp $ */
2
3 /*-
4 * Copyright (c) 2002, 2006, 2007, 2009, 2019, 2020
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 1.45 2022/10/26 23:27:16 riastradh Exp $");
65
66 #include <sys/param.h>
67 #include <sys/lockdebug.h>
68 #include <sys/pool.h>
69 #include <sys/proc.h>
70 #include <sys/sleepq.h>
71 #include <sys/sleeptab.h>
72 #include <sys/systm.h>
73
74 /*
75 * Shift of 6 aligns to typical cache line size of 64 bytes; there's no
76 * point having two turnstile locks to back two lock objects that share one
77 * cache line.
78 */
79 #define TS_HASH_SIZE 128
80 #define TS_HASH_MASK (TS_HASH_SIZE - 1)
81 #define TS_HASH(obj) (((uintptr_t)(obj) >> 6) & TS_HASH_MASK)
82
83 static tschain_t turnstile_chains[TS_HASH_SIZE] __cacheline_aligned;
84 struct pool turnstile_pool;
85
86 static union {
87 kmutex_t lock;
88 uint8_t pad[COHERENCY_UNIT];
89 } turnstile_locks[TS_HASH_SIZE] __cacheline_aligned;
90
91 /*
92 * turnstile_init:
93 *
94 * Initialize the turnstile mechanism.
95 */
96 void
97 turnstile_init(void)
98 {
99 int i;
100
101 for (i = 0; i < TS_HASH_SIZE; i++) {
102 LIST_INIT(&turnstile_chains[i]);
103 mutex_init(&turnstile_locks[i].lock, MUTEX_DEFAULT, IPL_SCHED);
104 }
105
106 pool_init(&turnstile_pool, sizeof(turnstile_t), coherency_unit,
107 0, 0, "tstile", NULL, IPL_NONE);
108
109 turnstile_ctor(&turnstile0);
110 }
111
112 /*
113 * turnstile_ctor:
114 *
115 * Constructor for turnstiles.
116 */
117 void
118 turnstile_ctor(turnstile_t *ts)
119 {
120
121 memset(ts, 0, sizeof(*ts));
122 sleepq_init(&ts->ts_sleepq[TS_READER_Q]);
123 sleepq_init(&ts->ts_sleepq[TS_WRITER_Q]);
124 }
125
126 /*
127 * turnstile_remove:
128 *
129 * Remove an LWP from a turnstile sleep queue and wake it.
130 */
131 static inline void
132 turnstile_remove(turnstile_t *ts, lwp_t *l, int q)
133 {
134 turnstile_t *nts;
135
136 KASSERT(l->l_ts == ts);
137
138 /*
139 * This process is no longer using the active turnstile.
140 * Find an inactive one on the free list to give to it.
141 */
142 if ((nts = ts->ts_free) != NULL) {
143 KASSERT(TS_ALL_WAITERS(ts) > 1);
144 l->l_ts = nts;
145 ts->ts_free = nts->ts_free;
146 nts->ts_free = NULL;
147 } else {
148 /*
149 * If the free list is empty, this is the last
150 * waiter.
151 */
152 KASSERT(TS_ALL_WAITERS(ts) == 1);
153 LIST_REMOVE(ts, ts_chain);
154 }
155
156 ts->ts_waiters[q]--;
157 sleepq_remove(&ts->ts_sleepq[q], l);
158 }
159
160 /*
161 * turnstile_lookup:
162 *
163 * Look up the turnstile for the specified lock. This acquires and
164 * holds the turnstile chain lock (sleep queue interlock).
165 */
166 turnstile_t *
167 turnstile_lookup(wchan_t obj)
168 {
169 turnstile_t *ts;
170 tschain_t *tc;
171 u_int hash;
172
173 hash = TS_HASH(obj);
174 tc = &turnstile_chains[hash];
175 mutex_spin_enter(&turnstile_locks[hash].lock);
176
177 LIST_FOREACH(ts, tc, ts_chain)
178 if (ts->ts_obj == obj)
179 return (ts);
180
181 /*
182 * No turnstile yet for this lock. No problem, turnstile_block()
183 * handles this by fetching the turnstile from the blocking thread.
184 */
185 return (NULL);
186 }
187
188 /*
189 * turnstile_exit:
190 *
191 * Abort a turnstile operation.
192 */
193 void
194 turnstile_exit(wchan_t obj)
195 {
196
197 mutex_spin_exit(&turnstile_locks[TS_HASH(obj)].lock);
198 }
199
200 /*
201 * turnstile_lendpri:
202 *
203 * Lend our priority to lwps on the blocking chain.
204 *
205 * If the current owner of the lock (l->l_wchan, set by sleepq_enqueue)
206 * has a priority lower than ours (lwp_eprio(l)), lend our priority to
207 * him to avoid priority inversions.
208 */
209
210 static void
211 turnstile_lendpri(lwp_t *cur)
212 {
213 lwp_t * l = cur;
214 pri_t prio;
215
216 /*
217 * NOTE: if you get a panic in this code block, it is likely that
218 * a lock has been destroyed or corrupted while still in use. Try
219 * compiling a kernel with LOCKDEBUG to pinpoint the problem.
220 */
221
222 LOCKDEBUG_BARRIER(l->l_mutex, 1);
223 KASSERT(l == curlwp);
224 prio = lwp_eprio(l);
225 for (;;) {
226 lwp_t *owner;
227 turnstile_t *ts;
228 bool dolock;
229
230 if (l->l_wchan == NULL)
231 break;
232
233 /*
234 * Ask syncobj the owner of the lock.
235 */
236 owner = (*l->l_syncobj->sobj_owner)(l->l_wchan);
237 if (owner == NULL)
238 break;
239
240 /*
241 * The owner may have changed as we have dropped the tc lock.
242 */
243 if (cur == owner) {
244 /*
245 * We own the lock: stop here, sleepq_block()
246 * should wake up immediately.
247 */
248 break;
249 }
250 /*
251 * Acquire owner->l_mutex if we don't have it yet.
252 * Because we already have another LWP lock (l->l_mutex) held,
253 * we need to play a try lock dance to avoid deadlock.
254 */
255 dolock = l->l_mutex != atomic_load_relaxed(&owner->l_mutex);
256 if (l == owner || (dolock && !lwp_trylock(owner))) {
257 /*
258 * The owner was changed behind us or trylock failed.
259 * Restart from curlwp.
260 *
261 * Note that there may be a livelock here:
262 * the owner may try grabbing cur's lock (which is the
263 * tc lock) while we're trying to grab the owner's lock.
264 */
265 lwp_unlock(l);
266 l = cur;
267 lwp_lock(l);
268 prio = lwp_eprio(l);
269 continue;
270 }
271 /*
272 * If the owner's priority is already higher than ours,
273 * there's nothing to do anymore.
274 */
275 if (prio <= lwp_eprio(owner)) {
276 if (dolock)
277 lwp_unlock(owner);
278 break;
279 }
280 /*
281 * Lend our priority to the 'owner' LWP.
282 *
283 * Update lenders info for turnstile_unlendpri.
284 */
285 ts = l->l_ts;
286 KASSERT(ts->ts_inheritor == owner || ts->ts_inheritor == NULL);
287 if (ts->ts_inheritor == NULL) {
288 ts->ts_inheritor = owner;
289 ts->ts_eprio = prio;
290 SLIST_INSERT_HEAD(&owner->l_pi_lenders, ts, ts_pichain);
291 lwp_lendpri(owner, prio);
292 } else if (prio > ts->ts_eprio) {
293 ts->ts_eprio = prio;
294 lwp_lendpri(owner, prio);
295 }
296 if (dolock)
297 lwp_unlock(l);
298 LOCKDEBUG_BARRIER(owner->l_mutex, 1);
299 l = owner;
300 }
301 LOCKDEBUG_BARRIER(l->l_mutex, 1);
302 if (cur->l_mutex != atomic_load_relaxed(&l->l_mutex)) {
303 lwp_unlock(l);
304 lwp_lock(cur);
305 }
306 LOCKDEBUG_BARRIER(cur->l_mutex, 1);
307 }
308
309 /*
310 * turnstile_unlendpri: undo turnstile_lendpri
311 */
312
313 static void
314 turnstile_unlendpri(turnstile_t *ts)
315 {
316 lwp_t * const l = curlwp;
317 turnstile_t *iter;
318 turnstile_t *next;
319 turnstile_t *prev = NULL;
320 pri_t prio;
321 bool dolock;
322
323 KASSERT(ts->ts_inheritor != NULL);
324 ts->ts_inheritor = NULL;
325 dolock = (atomic_load_relaxed(&l->l_mutex) ==
326 l->l_cpu->ci_schedstate.spc_lwplock);
327 if (dolock) {
328 lwp_lock(l);
329 }
330
331 /*
332 * the following loop does two things.
333 *
334 * - remove ts from the list.
335 *
336 * - from the rest of the list, find the highest priority.
337 */
338
339 prio = -1;
340 KASSERT(!SLIST_EMPTY(&l->l_pi_lenders));
341 for (iter = SLIST_FIRST(&l->l_pi_lenders);
342 iter != NULL; iter = next) {
343 KASSERT(lwp_eprio(l) >= ts->ts_eprio);
344 next = SLIST_NEXT(iter, ts_pichain);
345 if (iter == ts) {
346 if (prev == NULL) {
347 SLIST_REMOVE_HEAD(&l->l_pi_lenders,
348 ts_pichain);
349 } else {
350 SLIST_REMOVE_AFTER(prev, ts_pichain);
351 }
352 } else if (prio < iter->ts_eprio) {
353 prio = iter->ts_eprio;
354 }
355 prev = iter;
356 }
357
358 lwp_lendpri(l, prio);
359
360 if (dolock) {
361 lwp_unlock(l);
362 }
363 }
364
365 /*
366 * turnstile_block:
367 *
368 * Enter an object into the turnstile chain and prepare the current
369 * LWP for sleep.
370 */
371 void
372 turnstile_block(turnstile_t *ts, int q, wchan_t obj, syncobj_t *sobj)
373 {
374 lwp_t * const l = curlwp; /* cached curlwp */
375 turnstile_t *ots;
376 tschain_t *tc;
377 kmutex_t *lock;
378 sleepq_t *sq;
379 pri_t obase;
380 u_int hash;
381
382 hash = TS_HASH(obj);
383 tc = &turnstile_chains[hash];
384 lock = &turnstile_locks[hash].lock;
385
386 KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
387 KASSERT(mutex_owned(lock));
388 KASSERT(l != NULL && l->l_ts != NULL);
389
390 if (ts == NULL) {
391 /*
392 * We are the first thread to wait for this object;
393 * lend our turnstile to it.
394 */
395 ts = l->l_ts;
396 KASSERT(TS_ALL_WAITERS(ts) == 0);
397 KASSERT(LIST_EMPTY(&ts->ts_sleepq[TS_READER_Q]) &&
398 LIST_EMPTY(&ts->ts_sleepq[TS_WRITER_Q]));
399 ts->ts_obj = obj;
400 ts->ts_inheritor = NULL;
401 LIST_INSERT_HEAD(tc, ts, ts_chain);
402 } else {
403 /*
404 * Object already has a turnstile. Put our turnstile
405 * onto the free list, and reference the existing
406 * turnstile instead.
407 */
408 ots = l->l_ts;
409 KASSERT(ots->ts_free == NULL);
410 ots->ts_free = ts->ts_free;
411 ts->ts_free = ots;
412 l->l_ts = ts;
413
414 KASSERT(ts->ts_obj == obj);
415 KASSERT(TS_ALL_WAITERS(ts) != 0);
416 KASSERT(!LIST_EMPTY(&ts->ts_sleepq[TS_READER_Q]) ||
417 !LIST_EMPTY(&ts->ts_sleepq[TS_WRITER_Q]));
418 }
419
420 sq = &ts->ts_sleepq[q];
421 ts->ts_waiters[q]++;
422 sleepq_enter(sq, l, lock);
423 LOCKDEBUG_BARRIER(lock, 1);
424 l->l_kpriority = true;
425 obase = l->l_kpribase;
426 if (obase < PRI_KTHREAD)
427 l->l_kpribase = PRI_KTHREAD;
428 sleepq_enqueue(sq, obj, "tstile", sobj, false);
429
430 /*
431 * Disable preemption across this entire block, as we may drop
432 * scheduler locks (allowing preemption), and would prefer not
433 * to be interrupted while in a state of flux.
434 */
435 KPREEMPT_DISABLE(l);
436 KASSERT(lock == l->l_mutex);
437 turnstile_lendpri(l);
438 sleepq_block(0, false, sobj);
439 l->l_kpribase = obase;
440 KPREEMPT_ENABLE(l);
441 }
442
443 /*
444 * turnstile_wakeup:
445 *
446 * Wake up the specified number of threads that are blocked
447 * in a turnstile.
448 */
449 void
450 turnstile_wakeup(turnstile_t *ts, int q, int count, lwp_t *nl)
451 {
452 sleepq_t *sq;
453 kmutex_t *lock;
454 u_int hash;
455 lwp_t *l;
456
457 hash = TS_HASH(ts->ts_obj);
458 lock = &turnstile_locks[hash].lock;
459 sq = &ts->ts_sleepq[q];
460
461 KASSERT(q == TS_READER_Q || q == TS_WRITER_Q);
462 KASSERT(count > 0 && count <= TS_WAITERS(ts, q));
463 KASSERT(mutex_owned(lock));
464 KASSERT(ts->ts_inheritor == curlwp || ts->ts_inheritor == NULL);
465
466 /*
467 * restore inherited priority if necessary.
468 */
469
470 if (ts->ts_inheritor != NULL) {
471 turnstile_unlendpri(ts);
472 }
473
474 if (nl != NULL) {
475 #if defined(DEBUG) || defined(LOCKDEBUG)
476 LIST_FOREACH(l, sq, l_sleepchain) {
477 if (l == nl)
478 break;
479 }
480 if (l == NULL)
481 panic("turnstile_wakeup: nl not on sleepq");
482 #endif
483 turnstile_remove(ts, nl, q);
484 } else {
485 while (count-- > 0) {
486 l = LIST_FIRST(sq);
487 KASSERT(l != NULL);
488 turnstile_remove(ts, l, q);
489 }
490 }
491 mutex_spin_exit(lock);
492 }
493
494 /*
495 * turnstile_unsleep:
496 *
497 * Remove an LWP from the turnstile. This is called when the LWP has
498 * not been awoken normally but instead interrupted: for example, if it
499 * has received a signal. It's not a valid action for turnstiles,
500 * since LWPs blocking on a turnstile are not interruptable.
501 */
502 void
503 turnstile_unsleep(lwp_t *l, bool cleanup)
504 {
505
506 lwp_unlock(l);
507 panic("turnstile_unsleep");
508 }
509
510 /*
511 * turnstile_changepri:
512 *
513 * Adjust the priority of an LWP residing on a turnstile.
514 */
515 void
516 turnstile_changepri(lwp_t *l, pri_t pri)
517 {
518
519 /* XXX priority inheritance */
520 sleepq_changepri(l, pri);
521 }
522
523 #if defined(LOCKDEBUG)
524 /*
525 * turnstile_print:
526 *
527 * Given the address of a lock object, print the contents of a
528 * turnstile.
529 */
530 void
531 turnstile_print(volatile void *obj, void (*pr)(const char *, ...))
532 {
533 turnstile_t *ts;
534 tschain_t *tc;
535 sleepq_t *rsq, *wsq;
536 u_int hash;
537 lwp_t *l;
538
539 hash = TS_HASH(obj);
540 tc = &turnstile_chains[hash];
541
542 LIST_FOREACH(ts, tc, ts_chain)
543 if (ts->ts_obj == obj)
544 break;
545
546 if (ts == NULL) {
547 (*pr)("Turnstile: no active turnstile for this lock.\n");
548 return;
549 }
550
551 rsq = &ts->ts_sleepq[TS_READER_Q];
552 wsq = &ts->ts_sleepq[TS_WRITER_Q];
553
554 (*pr)("Turnstile:\n");
555 (*pr)("=> %d waiting readers:", TS_WAITERS(ts, TS_READER_Q));
556 LIST_FOREACH(l, rsq, l_sleepchain) {
557 (*pr)(" %p", l);
558 }
559 (*pr)("\n");
560
561 (*pr)("=> %d waiting writers:", TS_WAITERS(ts, TS_WRITER_Q));
562 LIST_FOREACH(l, wsq, l_sleepchain) {
563 (*pr)(" %p", l);
564 }
565 (*pr)("\n");
566 }
567 #endif /* LOCKDEBUG */
Cache object: 127012c419dc7c5729ce7d7c0f564910
|