Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* rwsem.c: R/W semaphores: contention handling functions |
2 | * | |
3 | * Written by David Howells (dhowells@redhat.com). | |
4 | * Derived from arch/i386/kernel/semaphore.c | |
ce6711f3 AS |
5 | * |
6 | * Writer lock-stealing by Alex Shi <alex.shi@intel.com> | |
fe6e674c | 7 | * and Michel Lespinasse <walken@google.com> |
4fc828e2 DB |
8 | * |
9 | * Optimistic spinning by Tim Chen <tim.c.chen@intel.com> | |
10 | * and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes. | |
1da177e4 LT |
11 | */ |
12 | #include <linux/rwsem.h> | |
13 | #include <linux/sched.h> | |
14 | #include <linux/init.h> | |
8bc3bcc9 | 15 | #include <linux/export.h> |
4fc828e2 | 16 | #include <linux/sched/rt.h> |
7a215f89 | 17 | #include <linux/osq_lock.h> |
4fc828e2 | 18 | |
7a215f89 | 19 | #include "rwsem.h" |
1da177e4 | 20 | |
3cf2f34e TC |
21 | /* |
22 | * Guide to the rw_semaphore's count field for common values. | |
23 | * (32-bit case illustrated, similar for 64-bit) | |
24 | * | |
25 | * 0x0000000X (1) X readers active or attempting lock, no writer waiting | |
26 | * X = #active_readers + #readers attempting to lock | |
27 | * (X*ACTIVE_BIAS) | |
28 | * | |
29 | * 0x00000000 rwsem is unlocked, and no one is waiting for the lock or | |
30 | * attempting to read lock or write lock. | |
31 | * | |
32 | * 0xffff000X (1) X readers active or attempting lock, with waiters for lock | |
33 | * X = #active readers + # readers attempting lock | |
34 | * (X*ACTIVE_BIAS + WAITING_BIAS) | |
35 | * (2) 1 writer attempting lock, no waiters for lock | |
36 | * X-1 = #active readers + #readers attempting lock | |
37 | * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS) | |
38 | * (3) 1 writer active, no waiters for lock | |
39 | * X-1 = #active readers + #readers attempting lock | |
40 | * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS) | |
41 | * | |
42 | * 0xffff0001 (1) 1 reader active or attempting lock, waiters for lock | |
43 | * (WAITING_BIAS + ACTIVE_BIAS) | |
44 | * (2) 1 writer active or attempting lock, no waiters for lock | |
45 | * (ACTIVE_WRITE_BIAS) | |
46 | * | |
47 | * 0xffff0000 (1) There are writers or readers queued but none active | |
48 | * or in the process of attempting lock. | |
49 | * (WAITING_BIAS) | |
50 | * Note: writer can attempt to steal lock for this count by adding | |
51 | * ACTIVE_WRITE_BIAS in cmpxchg and checking the old count | |
52 | * | |
53 | * 0xfffe0001 (1) 1 writer active, or attempting lock. Waiters on queue. | |
54 | * (ACTIVE_WRITE_BIAS + WAITING_BIAS) | |
55 | * | |
56 | * Note: Readers attempt to lock by adding ACTIVE_BIAS in down_read and checking | |
57 | * the count becomes more than 0 for successful lock acquisition, | |
58 | * i.e. the case where there are only readers or nobody has lock. | |
59 | * (1st and 2nd case above). | |
60 | * | |
61 | * Writers attempt to lock by adding ACTIVE_WRITE_BIAS in down_write and | |
62 | * checking the count becomes ACTIVE_WRITE_BIAS for successful lock | |
63 | * acquisition (i.e. nobody else has lock or attempts lock). If | |
64 | * unsuccessful, in rwsem_down_write_failed, we'll check to see if there | |
65 | * are only waiters but none active (5th case above), and attempt to | |
66 | * steal the lock. | |
67 | * | |
68 | */ | |
69 | ||
4ea2176d IM |
70 | /* |
71 | * Initialize an rwsem: | |
72 | */ | |
73 | void __init_rwsem(struct rw_semaphore *sem, const char *name, | |
74 | struct lock_class_key *key) | |
75 | { | |
76 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
77 | /* | |
78 | * Make sure we are not reinitializing a held semaphore: | |
79 | */ | |
80 | debug_check_no_locks_freed((void *)sem, sizeof(*sem)); | |
4dfbb9d8 | 81 | lockdep_init_map(&sem->dep_map, name, key, 0); |
4ea2176d | 82 | #endif |
8ee62b18 | 83 | atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE); |
ddb6c9b5 | 84 | raw_spin_lock_init(&sem->wait_lock); |
4ea2176d | 85 | INIT_LIST_HEAD(&sem->wait_list); |
5db6c6fe | 86 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
4fc828e2 | 87 | sem->owner = NULL; |
4d9d951e | 88 | osq_lock_init(&sem->osq); |
4fc828e2 | 89 | #endif |
4ea2176d IM |
90 | } |
91 | ||
92 | EXPORT_SYMBOL(__init_rwsem); | |
93 | ||
e2d57f78 ML |
94 | enum rwsem_waiter_type { |
95 | RWSEM_WAITING_FOR_WRITE, | |
96 | RWSEM_WAITING_FOR_READ | |
97 | }; | |
98 | ||
1da177e4 LT |
99 | struct rwsem_waiter { |
100 | struct list_head list; | |
101 | struct task_struct *task; | |
e2d57f78 | 102 | enum rwsem_waiter_type type; |
1da177e4 LT |
103 | }; |
104 | ||
fe6e674c ML |
105 | enum rwsem_wake_type { |
106 | RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */ | |
107 | RWSEM_WAKE_READERS, /* Wake readers only */ | |
108 | RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */ | |
109 | }; | |
70bdc6e0 | 110 | |
1da177e4 LT |
111 | /* |
112 | * handle the lock release when processes blocked on it that can now run | |
113 | * - if we come here from up_xxxx(), then: | |
114 | * - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed) | |
115 | * - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so) | |
345af7bf | 116 | * - there must be someone on the queue |
133e89ef DB |
117 | * - the wait_lock must be held by the caller |
118 | * - tasks are marked for wakeup, the caller must later invoke wake_up_q() | |
119 | * to actually wakeup the blocked task(s) and drop the reference count, | |
120 | * preferably when the wait_lock is released | |
1da177e4 | 121 | * - woken process blocks are discarded from the list after having task zeroed |
133e89ef | 122 | * - writers are only marked woken if downgrading is false |
1da177e4 | 123 | */ |
70bdc6e0 | 124 | static struct rw_semaphore * |
133e89ef DB |
125 | __rwsem_mark_wake(struct rw_semaphore *sem, |
126 | enum rwsem_wake_type wake_type, struct wake_q_head *wake_q) | |
1da177e4 LT |
127 | { |
128 | struct rwsem_waiter *waiter; | |
129 | struct task_struct *tsk; | |
130 | struct list_head *next; | |
b5f54181 | 131 | long oldcount, woken, loop, adjustment; |
1da177e4 | 132 | |
345af7bf | 133 | waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list); |
8cf5322c | 134 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) { |
133e89ef DB |
135 | if (wake_type == RWSEM_WAKE_ANY) { |
136 | /* | |
137 | * Mark writer at the front of the queue for wakeup. | |
138 | * Until the task is actually later awoken later by | |
139 | * the caller, other writers are able to steal it. | |
140 | * Readers, on the other hand, will block as they | |
141 | * will notice the queued writer. | |
8cf5322c | 142 | */ |
133e89ef DB |
143 | wake_q_add(wake_q, waiter->task); |
144 | } | |
345af7bf | 145 | goto out; |
8cf5322c | 146 | } |
1da177e4 | 147 | |
fe6e674c ML |
148 | /* Writers might steal the lock before we grant it to the next reader. |
149 | * We prefer to do the first reader grant before counting readers | |
150 | * so we can bail out early if a writer stole the lock. | |
70bdc6e0 | 151 | */ |
fe6e674c ML |
152 | adjustment = 0; |
153 | if (wake_type != RWSEM_WAKE_READ_OWNED) { | |
154 | adjustment = RWSEM_ACTIVE_READ_BIAS; | |
155 | try_reader_grant: | |
8ee62b18 JL |
156 | oldcount = atomic_long_add_return(adjustment, &sem->count) - adjustment; |
157 | ||
fe6e674c | 158 | if (unlikely(oldcount < RWSEM_WAITING_BIAS)) { |
bf7b4c47 WL |
159 | /* |
160 | * If the count is still less than RWSEM_WAITING_BIAS | |
161 | * after removing the adjustment, it is assumed that | |
162 | * a writer has stolen the lock. We have to undo our | |
163 | * reader grant. | |
164 | */ | |
165 | if (atomic_long_add_return(-adjustment, &sem->count) < | |
166 | RWSEM_WAITING_BIAS) | |
fe6e674c ML |
167 | goto out; |
168 | /* Last active locker left. Retry waking readers. */ | |
169 | goto try_reader_grant; | |
170 | } | |
19c5d690 WL |
171 | /* |
172 | * It is not really necessary to set it to reader-owned here, | |
173 | * but it gives the spinners an early indication that the | |
174 | * readers now have the lock. | |
175 | */ | |
176 | rwsem_set_reader_owned(sem); | |
fe6e674c | 177 | } |
1da177e4 | 178 | |
345af7bf ML |
179 | /* Grant an infinite number of read locks to the readers at the front |
180 | * of the queue. Note we increment the 'active part' of the count by | |
181 | * the number of readers before waking any processes up. | |
1da177e4 | 182 | */ |
1da177e4 LT |
183 | woken = 0; |
184 | do { | |
185 | woken++; | |
186 | ||
187 | if (waiter->list.next == &sem->wait_list) | |
188 | break; | |
189 | ||
190 | waiter = list_entry(waiter->list.next, | |
191 | struct rwsem_waiter, list); | |
192 | ||
e2d57f78 | 193 | } while (waiter->type != RWSEM_WAITING_FOR_WRITE); |
1da177e4 | 194 | |
fe6e674c | 195 | adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment; |
e2d57f78 | 196 | if (waiter->type != RWSEM_WAITING_FOR_WRITE) |
fd41b334 ML |
197 | /* hit end of list above */ |
198 | adjustment -= RWSEM_WAITING_BIAS; | |
1da177e4 | 199 | |
fe6e674c | 200 | if (adjustment) |
8ee62b18 | 201 | atomic_long_add(adjustment, &sem->count); |
1da177e4 LT |
202 | |
203 | next = sem->wait_list.next; | |
8cf5322c ML |
204 | loop = woken; |
205 | do { | |
1da177e4 LT |
206 | waiter = list_entry(next, struct rwsem_waiter, list); |
207 | next = waiter->list.next; | |
208 | tsk = waiter->task; | |
e3851390 DB |
209 | |
210 | wake_q_add(wake_q, tsk); | |
49e4b2bc | 211 | /* |
e3851390 DB |
212 | * Ensure that the last operation is setting the reader |
213 | * waiter to nil such that rwsem_down_read_failed() cannot | |
214 | * race with do_exit() by always holding a reference count | |
215 | * to the task to wakeup. | |
49e4b2bc | 216 | */ |
e3851390 | 217 | smp_store_release(&waiter->task, NULL); |
8cf5322c | 218 | } while (--loop); |
1da177e4 LT |
219 | |
220 | sem->wait_list.next = next; | |
221 | next->prev = &sem->wait_list; | |
222 | ||
223 | out: | |
1da177e4 | 224 | return sem; |
ce6711f3 AS |
225 | } |
226 | ||
1da177e4 | 227 | /* |
4fc828e2 | 228 | * Wait for the read lock to be granted |
1da177e4 | 229 | */ |
3ebae4f3 | 230 | __visible |
1e78277c | 231 | struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem) |
1da177e4 | 232 | { |
b5f54181 | 233 | long count, adjustment = -RWSEM_ACTIVE_READ_BIAS; |
a8618a0e | 234 | struct rwsem_waiter waiter; |
1da177e4 | 235 | struct task_struct *tsk = current; |
133e89ef | 236 | WAKE_Q(wake_q); |
1da177e4 | 237 | |
1da177e4 | 238 | /* set up my own style of waitqueue */ |
a8618a0e | 239 | waiter.task = tsk; |
da16922c | 240 | waiter.type = RWSEM_WAITING_FOR_READ; |
1da177e4 | 241 | |
f7dd1cee | 242 | raw_spin_lock_irq(&sem->wait_lock); |
fd41b334 ML |
243 | if (list_empty(&sem->wait_list)) |
244 | adjustment += RWSEM_WAITING_BIAS; | |
a8618a0e | 245 | list_add_tail(&waiter.list, &sem->wait_list); |
1da177e4 | 246 | |
70bdc6e0 | 247 | /* we're now waiting on the lock, but no longer actively locking */ |
8ee62b18 | 248 | count = atomic_long_add_return(adjustment, &sem->count); |
1da177e4 | 249 | |
25c39325 ML |
250 | /* If there are no active locks, wake the front queued process(es). |
251 | * | |
252 | * If there are no writers and we are first in the queue, | |
253 | * wake our own waiter to join the existing active readers ! | |
254 | */ | |
255 | if (count == RWSEM_WAITING_BIAS || | |
256 | (count > RWSEM_WAITING_BIAS && | |
257 | adjustment != -RWSEM_ACTIVE_READ_BIAS)) | |
133e89ef | 258 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
1da177e4 | 259 | |
ddb6c9b5 | 260 | raw_spin_unlock_irq(&sem->wait_lock); |
133e89ef | 261 | wake_up_q(&wake_q); |
1da177e4 LT |
262 | |
263 | /* wait to be given the lock */ | |
f7dd1cee ML |
264 | while (true) { |
265 | set_task_state(tsk, TASK_UNINTERRUPTIBLE); | |
a8618a0e | 266 | if (!waiter.task) |
1da177e4 LT |
267 | break; |
268 | schedule(); | |
1da177e4 LT |
269 | } |
270 | ||
73105994 | 271 | __set_task_state(tsk, TASK_RUNNING); |
1da177e4 LT |
272 | return sem; |
273 | } | |
db0e716a | 274 | EXPORT_SYMBOL(rwsem_down_read_failed); |
1da177e4 | 275 | |
c0fcb6c2 JL |
276 | /* |
277 | * This function must be called with the sem->wait_lock held to prevent | |
278 | * race conditions between checking the rwsem wait list and setting the | |
279 | * sem->count accordingly. | |
280 | */ | |
4fc828e2 DB |
281 | static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem) |
282 | { | |
debfab74 | 283 | /* |
c0fcb6c2 | 284 | * Avoid trying to acquire write lock if count isn't RWSEM_WAITING_BIAS. |
debfab74 | 285 | */ |
c0fcb6c2 JL |
286 | if (count != RWSEM_WAITING_BIAS) |
287 | return false; | |
288 | ||
289 | /* | |
290 | * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there | |
291 | * are other tasks on the wait list, we need to add on WAITING_BIAS. | |
292 | */ | |
293 | count = list_is_singular(&sem->wait_list) ? | |
294 | RWSEM_ACTIVE_WRITE_BIAS : | |
295 | RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS; | |
296 | ||
8ee62b18 JL |
297 | if (atomic_long_cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count) |
298 | == RWSEM_WAITING_BIAS) { | |
7a215f89 | 299 | rwsem_set_owner(sem); |
debfab74 | 300 | return true; |
4fc828e2 | 301 | } |
debfab74 | 302 | |
4fc828e2 DB |
303 | return false; |
304 | } | |
305 | ||
5db6c6fe | 306 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
1da177e4 | 307 | /* |
4fc828e2 DB |
308 | * Try to acquire write lock before the writer has been put on wait queue. |
309 | */ | |
310 | static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem) | |
311 | { | |
8ee62b18 | 312 | long old, count = atomic_long_read(&sem->count); |
4fc828e2 DB |
313 | |
314 | while (true) { | |
315 | if (!(count == 0 || count == RWSEM_WAITING_BIAS)) | |
316 | return false; | |
317 | ||
8ee62b18 | 318 | old = atomic_long_cmpxchg_acquire(&sem->count, count, |
00eb4bab | 319 | count + RWSEM_ACTIVE_WRITE_BIAS); |
7a215f89 DB |
320 | if (old == count) { |
321 | rwsem_set_owner(sem); | |
4fc828e2 | 322 | return true; |
7a215f89 | 323 | } |
4fc828e2 DB |
324 | |
325 | count = old; | |
326 | } | |
327 | } | |
328 | ||
329 | static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem) | |
330 | { | |
331 | struct task_struct *owner; | |
1a993670 | 332 | bool ret = true; |
4fc828e2 DB |
333 | |
334 | if (need_resched()) | |
37e95624 | 335 | return false; |
4fc828e2 DB |
336 | |
337 | rcu_read_lock(); | |
4d3199e4 | 338 | owner = READ_ONCE(sem->owner); |
19c5d690 | 339 | if (!rwsem_owner_is_writer(owner)) { |
1a993670 | 340 | /* |
19c5d690 | 341 | * Don't spin if the rwsem is readers owned. |
1a993670 | 342 | */ |
19c5d690 | 343 | ret = !rwsem_owner_is_reader(owner); |
1a993670 DB |
344 | goto done; |
345 | } | |
4fc828e2 | 346 | |
1a993670 DB |
347 | ret = owner->on_cpu; |
348 | done: | |
349 | rcu_read_unlock(); | |
350 | return ret; | |
4fc828e2 DB |
351 | } |
352 | ||
4fc828e2 DB |
353 | static noinline |
354 | bool rwsem_spin_on_owner(struct rw_semaphore *sem, struct task_struct *owner) | |
355 | { | |
356 | rcu_read_lock(); | |
9198f6ed JL |
357 | while (sem->owner == owner) { |
358 | /* | |
359 | * Ensure we emit the owner->on_cpu, dereference _after_ | |
360 | * checking sem->owner still matches owner, if that fails, | |
361 | * owner might point to free()d memory, if it still matches, | |
362 | * the rcu_read_lock() ensures the memory stays valid. | |
363 | */ | |
364 | barrier(); | |
365 | ||
366 | /* abort spinning when need_resched or owner is not running */ | |
367 | if (!owner->on_cpu || need_resched()) { | |
b3fd4f03 DB |
368 | rcu_read_unlock(); |
369 | return false; | |
370 | } | |
4fc828e2 | 371 | |
3a6bfbc9 | 372 | cpu_relax_lowlatency(); |
4fc828e2 DB |
373 | } |
374 | rcu_read_unlock(); | |
375 | ||
376 | /* | |
19c5d690 WL |
377 | * If there is a new owner or the owner is not set, we continue |
378 | * spinning. | |
4fc828e2 | 379 | */ |
19c5d690 | 380 | return !rwsem_owner_is_reader(READ_ONCE(sem->owner)); |
4fc828e2 DB |
381 | } |
382 | ||
383 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) | |
384 | { | |
385 | struct task_struct *owner; | |
386 | bool taken = false; | |
387 | ||
388 | preempt_disable(); | |
389 | ||
390 | /* sem->wait_lock should not be held when doing optimistic spinning */ | |
391 | if (!rwsem_can_spin_on_owner(sem)) | |
392 | goto done; | |
393 | ||
394 | if (!osq_lock(&sem->osq)) | |
395 | goto done; | |
396 | ||
397 | while (true) { | |
4d3199e4 | 398 | owner = READ_ONCE(sem->owner); |
19c5d690 WL |
399 | /* |
400 | * Don't spin if | |
401 | * 1) the owner is a reader as we we can't determine if the | |
402 | * reader is actively running or not. | |
403 | * 2) The rwsem_spin_on_owner() returns false which means | |
404 | * the owner isn't running. | |
405 | */ | |
406 | if (rwsem_owner_is_reader(owner) || | |
407 | (rwsem_owner_is_writer(owner) && | |
408 | !rwsem_spin_on_owner(sem, owner))) | |
4fc828e2 DB |
409 | break; |
410 | ||
411 | /* wait_lock will be acquired if write_lock is obtained */ | |
412 | if (rwsem_try_write_lock_unqueued(sem)) { | |
413 | taken = true; | |
414 | break; | |
415 | } | |
416 | ||
417 | /* | |
418 | * When there's no owner, we might have preempted between the | |
419 | * owner acquiring the lock and setting the owner field. If | |
420 | * we're an RT task that will live-lock because we won't let | |
421 | * the owner complete. | |
422 | */ | |
423 | if (!owner && (need_resched() || rt_task(current))) | |
424 | break; | |
425 | ||
426 | /* | |
427 | * The cpu_relax() call is a compiler barrier which forces | |
428 | * everything in this loop to be re-loaded. We don't need | |
429 | * memory barriers as we'll eventually observe the right | |
430 | * values at the cost of a few extra spins. | |
431 | */ | |
3a6bfbc9 | 432 | cpu_relax_lowlatency(); |
4fc828e2 DB |
433 | } |
434 | osq_unlock(&sem->osq); | |
435 | done: | |
436 | preempt_enable(); | |
437 | return taken; | |
438 | } | |
439 | ||
59aabfc7 WL |
440 | /* |
441 | * Return true if the rwsem has active spinner | |
442 | */ | |
443 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) | |
444 | { | |
445 | return osq_is_locked(&sem->osq); | |
446 | } | |
447 | ||
4fc828e2 DB |
448 | #else |
449 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) | |
450 | { | |
451 | return false; | |
452 | } | |
59aabfc7 WL |
453 | |
454 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) | |
455 | { | |
456 | return false; | |
457 | } | |
4fc828e2 DB |
458 | #endif |
459 | ||
460 | /* | |
461 | * Wait until we successfully acquire the write lock | |
1da177e4 | 462 | */ |
d4799608 MH |
463 | static inline struct rw_semaphore * |
464 | __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state) | |
1da177e4 | 465 | { |
4fc828e2 DB |
466 | long count; |
467 | bool waiting = true; /* any queued threads before us */ | |
1e78277c | 468 | struct rwsem_waiter waiter; |
d4799608 | 469 | struct rw_semaphore *ret = sem; |
133e89ef | 470 | WAKE_Q(wake_q); |
1e78277c | 471 | |
4fc828e2 | 472 | /* undo write bias from down_write operation, stop active locking */ |
8ee62b18 | 473 | count = atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS, &sem->count); |
4fc828e2 DB |
474 | |
475 | /* do optimistic spinning and steal lock if possible */ | |
476 | if (rwsem_optimistic_spin(sem)) | |
477 | return sem; | |
478 | ||
479 | /* | |
480 | * Optimistic spinning failed, proceed to the slowpath | |
481 | * and block until we can acquire the sem. | |
482 | */ | |
483 | waiter.task = current; | |
023fe4f7 | 484 | waiter.type = RWSEM_WAITING_FOR_WRITE; |
1e78277c ML |
485 | |
486 | raw_spin_lock_irq(&sem->wait_lock); | |
4fc828e2 DB |
487 | |
488 | /* account for this before adding a new element to the list */ | |
1e78277c | 489 | if (list_empty(&sem->wait_list)) |
4fc828e2 DB |
490 | waiting = false; |
491 | ||
1e78277c ML |
492 | list_add_tail(&waiter.list, &sem->wait_list); |
493 | ||
494 | /* we're now waiting on the lock, but no longer actively locking */ | |
4fc828e2 | 495 | if (waiting) { |
8ee62b18 | 496 | count = atomic_long_read(&sem->count); |
1e78277c | 497 | |
4fc828e2 | 498 | /* |
0cc3d011 AM |
499 | * If there were already threads queued before us and there are |
500 | * no active writers, the lock must be read owned; so we try to | |
501 | * wake any read locks that were queued ahead of us. | |
4fc828e2 | 502 | */ |
133e89ef DB |
503 | if (count > RWSEM_WAITING_BIAS) { |
504 | WAKE_Q(wake_q); | |
505 | ||
506 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q); | |
507 | /* | |
508 | * The wakeup is normally called _after_ the wait_lock | |
509 | * is released, but given that we are proactively waking | |
510 | * readers we can deal with the wake_q overhead as it is | |
511 | * similar to releasing and taking the wait_lock again | |
512 | * for attempting rwsem_try_write_lock(). | |
513 | */ | |
514 | wake_up_q(&wake_q); | |
515 | } | |
4fc828e2 DB |
516 | |
517 | } else | |
8ee62b18 | 518 | count = atomic_long_add_return(RWSEM_WAITING_BIAS, &sem->count); |
1e78277c | 519 | |
023fe4f7 | 520 | /* wait until we successfully acquire the lock */ |
d4799608 | 521 | set_current_state(state); |
1e78277c | 522 | while (true) { |
4fc828e2 DB |
523 | if (rwsem_try_write_lock(count, sem)) |
524 | break; | |
1e78277c | 525 | raw_spin_unlock_irq(&sem->wait_lock); |
a7d2c573 ML |
526 | |
527 | /* Block until there are no active lockers. */ | |
528 | do { | |
04cafed7 PZ |
529 | if (signal_pending_state(state, current)) |
530 | goto out_nolock; | |
531 | ||
a7d2c573 | 532 | schedule(); |
d4799608 | 533 | set_current_state(state); |
8ee62b18 | 534 | } while ((count = atomic_long_read(&sem->count)) & RWSEM_ACTIVE_MASK); |
a7d2c573 | 535 | |
023fe4f7 | 536 | raw_spin_lock_irq(&sem->wait_lock); |
1e78277c | 537 | } |
4fc828e2 | 538 | __set_current_state(TASK_RUNNING); |
023fe4f7 ML |
539 | list_del(&waiter.list); |
540 | raw_spin_unlock_irq(&sem->wait_lock); | |
1e78277c | 541 | |
d4799608 | 542 | return ret; |
04cafed7 PZ |
543 | |
544 | out_nolock: | |
545 | __set_current_state(TASK_RUNNING); | |
546 | raw_spin_lock_irq(&sem->wait_lock); | |
547 | list_del(&waiter.list); | |
548 | if (list_empty(&sem->wait_list)) | |
8ee62b18 | 549 | atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count); |
04cafed7 | 550 | else |
133e89ef | 551 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
04cafed7 | 552 | raw_spin_unlock_irq(&sem->wait_lock); |
133e89ef | 553 | wake_up_q(&wake_q); |
04cafed7 PZ |
554 | |
555 | return ERR_PTR(-EINTR); | |
d4799608 MH |
556 | } |
557 | ||
558 | __visible struct rw_semaphore * __sched | |
559 | rwsem_down_write_failed(struct rw_semaphore *sem) | |
560 | { | |
561 | return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE); | |
1da177e4 | 562 | } |
db0e716a | 563 | EXPORT_SYMBOL(rwsem_down_write_failed); |
1da177e4 | 564 | |
d4799608 MH |
565 | __visible struct rw_semaphore * __sched |
566 | rwsem_down_write_failed_killable(struct rw_semaphore *sem) | |
567 | { | |
568 | return __rwsem_down_write_failed_common(sem, TASK_KILLABLE); | |
569 | } | |
570 | EXPORT_SYMBOL(rwsem_down_write_failed_killable); | |
571 | ||
1da177e4 LT |
572 | /* |
573 | * handle waking up a waiter on the semaphore | |
574 | * - up_read/up_write has decremented the active part of count if we come here | |
575 | */ | |
3ebae4f3 | 576 | __visible |
d1233754 | 577 | struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem) |
1da177e4 LT |
578 | { |
579 | unsigned long flags; | |
133e89ef | 580 | WAKE_Q(wake_q); |
1da177e4 | 581 | |
59aabfc7 WL |
582 | /* |
583 | * If a spinner is present, it is not necessary to do the wakeup. | |
584 | * Try to do wakeup only if the trylock succeeds to minimize | |
585 | * spinlock contention which may introduce too much delay in the | |
586 | * unlock operation. | |
587 | * | |
588 | * spinning writer up_write/up_read caller | |
589 | * --------------- ----------------------- | |
590 | * [S] osq_unlock() [L] osq | |
591 | * MB RMB | |
592 | * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock) | |
593 | * | |
594 | * Here, it is important to make sure that there won't be a missed | |
595 | * wakeup while the rwsem is free and the only spinning writer goes | |
596 | * to sleep without taking the rwsem. Even when the spinning writer | |
597 | * is just going to break out of the waiting loop, it will still do | |
598 | * a trylock in rwsem_down_write_failed() before sleeping. IOW, if | |
599 | * rwsem_has_spinner() is true, it will guarantee at least one | |
600 | * trylock attempt on the rwsem later on. | |
601 | */ | |
602 | if (rwsem_has_spinner(sem)) { | |
603 | /* | |
604 | * The smp_rmb() here is to make sure that the spinner | |
605 | * state is consulted before reading the wait_lock. | |
606 | */ | |
607 | smp_rmb(); | |
608 | if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags)) | |
609 | return sem; | |
610 | goto locked; | |
611 | } | |
ddb6c9b5 | 612 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
59aabfc7 | 613 | locked: |
1da177e4 LT |
614 | |
615 | /* do nothing if list empty */ | |
616 | if (!list_empty(&sem->wait_list)) | |
133e89ef | 617 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
1da177e4 | 618 | |
ddb6c9b5 | 619 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
133e89ef | 620 | wake_up_q(&wake_q); |
1da177e4 | 621 | |
1da177e4 LT |
622 | return sem; |
623 | } | |
db0e716a | 624 | EXPORT_SYMBOL(rwsem_wake); |
1da177e4 LT |
625 | |
626 | /* | |
627 | * downgrade a write lock into a read lock | |
628 | * - caller incremented waiting part of count and discovered it still negative | |
629 | * - just wake up any readers at the front of the queue | |
630 | */ | |
3ebae4f3 | 631 | __visible |
d1233754 | 632 | struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem) |
1da177e4 LT |
633 | { |
634 | unsigned long flags; | |
133e89ef | 635 | WAKE_Q(wake_q); |
1da177e4 | 636 | |
ddb6c9b5 | 637 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
1da177e4 LT |
638 | |
639 | /* do nothing if list empty */ | |
640 | if (!list_empty(&sem->wait_list)) | |
133e89ef | 641 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q); |
1da177e4 | 642 | |
ddb6c9b5 | 643 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
133e89ef | 644 | wake_up_q(&wake_q); |
1da177e4 | 645 | |
1da177e4 LT |
646 | return sem; |
647 | } | |
1da177e4 | 648 | EXPORT_SYMBOL(rwsem_downgrade_wake); |