Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 |
Alex Shi | ce6711f | 2013-02-05 21:11:55 +0800 | [diff] [blame] | 5 | * |
| 6 | * Writer lock-stealing by Alex Shi <alex.shi@intel.com> |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 7 | * and Michel Lespinasse <walken@google.com> |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 8 | * |
| 9 | * Optimistic spinning by Tim Chen <tim.c.chen@intel.com> |
| 10 | * and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 11 | */ |
| 12 | #include <linux/rwsem.h> |
| 13 | #include <linux/sched.h> |
| 14 | #include <linux/init.h> |
Paul Gortmaker | 8bc3bcc | 2011-11-16 21:29:17 -0500 | [diff] [blame] | 15 | #include <linux/export.h> |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 16 | #include <linux/sched/rt.h> |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 17 | #include <linux/osq_lock.h> |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 18 | |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 19 | #include "rwsem.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 20 | |
Ingo Molnar | 4ea2176 | 2006-07-03 00:24:53 -0700 | [diff] [blame] | 21 | /* |
Tim Chen | 3cf2f34 | 2014-05-02 12:53:57 -0700 | [diff] [blame] | 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 | |
| 70 | /* |
Ingo Molnar | 4ea2176 | 2006-07-03 00:24:53 -0700 | [diff] [blame] | 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)); |
Peter Zijlstra | 4dfbb9d | 2006-10-11 01:45:14 -0400 | [diff] [blame] | 81 | lockdep_init_map(&sem->dep_map, name, key, 0); |
Ingo Molnar | 4ea2176 | 2006-07-03 00:24:53 -0700 | [diff] [blame] | 82 | #endif |
| 83 | sem->count = RWSEM_UNLOCKED_VALUE; |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 84 | raw_spin_lock_init(&sem->wait_lock); |
Ingo Molnar | 4ea2176 | 2006-07-03 00:24:53 -0700 | [diff] [blame] | 85 | INIT_LIST_HEAD(&sem->wait_list); |
Davidlohr Bueso | 5db6c6f | 2014-07-11 14:00:06 -0700 | [diff] [blame] | 86 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 87 | sem->owner = NULL; |
Jason Low | 4d9d951 | 2014-07-14 10:27:50 -0700 | [diff] [blame] | 88 | osq_lock_init(&sem->osq); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 89 | #endif |
Ingo Molnar | 4ea2176 | 2006-07-03 00:24:53 -0700 | [diff] [blame] | 90 | } |
| 91 | |
| 92 | EXPORT_SYMBOL(__init_rwsem); |
| 93 | |
Michel Lespinasse | e2d57f7 | 2013-05-07 06:45:49 -0700 | [diff] [blame] | 94 | enum rwsem_waiter_type { |
| 95 | RWSEM_WAITING_FOR_WRITE, |
| 96 | RWSEM_WAITING_FOR_READ |
| 97 | }; |
| 98 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 99 | struct rwsem_waiter { |
| 100 | struct list_head list; |
| 101 | struct task_struct *task; |
Michel Lespinasse | e2d57f7 | 2013-05-07 06:45:49 -0700 | [diff] [blame] | 102 | enum rwsem_waiter_type type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | }; |
| 104 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 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 | }; |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 110 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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) |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 116 | * - there must be someone on the queue |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 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 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 121 | * - woken process blocks are discarded from the list after having task zeroed |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 122 | * - writers are only marked woken if downgrading is false |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | */ |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 124 | static struct rw_semaphore * |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 125 | __rwsem_mark_wake(struct rw_semaphore *sem, |
| 126 | enum rwsem_wake_type wake_type, struct wake_q_head *wake_q) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 127 | { |
| 128 | struct rwsem_waiter *waiter; |
| 129 | struct task_struct *tsk; |
| 130 | struct list_head *next; |
Davidlohr Bueso | b5f5418 | 2013-05-07 06:46:02 -0700 | [diff] [blame] | 131 | long oldcount, woken, loop, adjustment; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 132 | |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 133 | waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list); |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 134 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) { |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 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. |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 142 | */ |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 143 | wake_q_add(wake_q, waiter->task); |
| 144 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 145 | goto out; |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 146 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 147 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 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. |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 151 | */ |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 152 | adjustment = 0; |
| 153 | if (wake_type != RWSEM_WAKE_READ_OWNED) { |
| 154 | adjustment = RWSEM_ACTIVE_READ_BIAS; |
| 155 | try_reader_grant: |
| 156 | oldcount = rwsem_atomic_update(adjustment, sem) - adjustment; |
| 157 | if (unlikely(oldcount < RWSEM_WAITING_BIAS)) { |
| 158 | /* A writer stole the lock. Undo our reader grant. */ |
| 159 | if (rwsem_atomic_update(-adjustment, sem) & |
| 160 | RWSEM_ACTIVE_MASK) |
| 161 | goto out; |
| 162 | /* Last active locker left. Retry waking readers. */ |
| 163 | goto try_reader_grant; |
| 164 | } |
| 165 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 166 | |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 167 | /* Grant an infinite number of read locks to the readers at the front |
| 168 | * of the queue. Note we increment the 'active part' of the count by |
| 169 | * the number of readers before waking any processes up. |
| 170 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | woken = 0; |
| 172 | do { |
| 173 | woken++; |
| 174 | |
| 175 | if (waiter->list.next == &sem->wait_list) |
| 176 | break; |
| 177 | |
| 178 | waiter = list_entry(waiter->list.next, |
| 179 | struct rwsem_waiter, list); |
| 180 | |
Michel Lespinasse | e2d57f7 | 2013-05-07 06:45:49 -0700 | [diff] [blame] | 181 | } while (waiter->type != RWSEM_WAITING_FOR_WRITE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 182 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 183 | adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment; |
Michel Lespinasse | e2d57f7 | 2013-05-07 06:45:49 -0700 | [diff] [blame] | 184 | if (waiter->type != RWSEM_WAITING_FOR_WRITE) |
Michel Lespinasse | fd41b33 | 2010-08-09 17:21:18 -0700 | [diff] [blame] | 185 | /* hit end of list above */ |
| 186 | adjustment -= RWSEM_WAITING_BIAS; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 187 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 188 | if (adjustment) |
| 189 | rwsem_atomic_add(adjustment, sem); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 190 | |
| 191 | next = sem->wait_list.next; |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 192 | loop = woken; |
| 193 | do { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 194 | waiter = list_entry(next, struct rwsem_waiter, list); |
| 195 | next = waiter->list.next; |
| 196 | tsk = waiter->task; |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 197 | |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 198 | wake_q_add(wake_q, tsk); |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 199 | /* |
| 200 | * Ensure that the last operation is setting the reader |
| 201 | * waiter to nil such that rwsem_down_read_failed() cannot |
| 202 | * race with do_exit() by always holding a reference count |
| 203 | * to the task to wakeup. |
| 204 | */ |
| 205 | smp_store_release(&waiter->task, NULL); |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 206 | } while (--loop); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 207 | |
| 208 | sem->wait_list.next = next; |
| 209 | next->prev = &sem->wait_list; |
| 210 | |
| 211 | out: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 212 | return sem; |
Alex Shi | ce6711f | 2013-02-05 21:11:55 +0800 | [diff] [blame] | 213 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 214 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 215 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 216 | * Wait for the read lock to be granted |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 217 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 218 | __visible |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 219 | struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 220 | { |
Davidlohr Bueso | b5f5418 | 2013-05-07 06:46:02 -0700 | [diff] [blame] | 221 | long count, adjustment = -RWSEM_ACTIVE_READ_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 222 | struct rwsem_waiter waiter; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 223 | struct task_struct *tsk = current; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 224 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 225 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 226 | /* set up my own style of waitqueue */ |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 227 | waiter.task = tsk; |
Michel Lespinasse | da16922 | 2013-05-07 06:45:52 -0700 | [diff] [blame] | 228 | waiter.type = RWSEM_WAITING_FOR_READ; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 229 | |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 230 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | fd41b33 | 2010-08-09 17:21:18 -0700 | [diff] [blame] | 231 | if (list_empty(&sem->wait_list)) |
| 232 | adjustment += RWSEM_WAITING_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 233 | list_add_tail(&waiter.list, &sem->wait_list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 234 | |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 235 | /* we're now waiting on the lock, but no longer actively locking */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | count = rwsem_atomic_update(adjustment, sem); |
| 237 | |
Michel Lespinasse | 25c3932 | 2013-05-07 06:46:00 -0700 | [diff] [blame] | 238 | /* If there are no active locks, wake the front queued process(es). |
| 239 | * |
| 240 | * If there are no writers and we are first in the queue, |
| 241 | * wake our own waiter to join the existing active readers ! |
| 242 | */ |
| 243 | if (count == RWSEM_WAITING_BIAS || |
| 244 | (count > RWSEM_WAITING_BIAS && |
| 245 | adjustment != -RWSEM_ACTIVE_READ_BIAS)) |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 246 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 247 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 248 | raw_spin_unlock_irq(&sem->wait_lock); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 249 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 250 | |
| 251 | /* wait to be given the lock */ |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 252 | while (true) { |
| 253 | set_task_state(tsk, TASK_UNINTERRUPTIBLE); |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 254 | if (!waiter.task) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 255 | break; |
| 256 | schedule(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 257 | } |
| 258 | |
Davidlohr Bueso | 7310599 | 2015-01-25 23:36:04 -0800 | [diff] [blame] | 259 | __set_task_state(tsk, TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 260 | return sem; |
| 261 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 262 | EXPORT_SYMBOL(rwsem_down_read_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 263 | |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame^] | 264 | /* |
| 265 | * This function must be called with the sem->wait_lock held to prevent |
| 266 | * race conditions between checking the rwsem wait list and setting the |
| 267 | * sem->count accordingly. |
| 268 | */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 269 | static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem) |
| 270 | { |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 271 | /* |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame^] | 272 | * Avoid trying to acquire write lock if count isn't RWSEM_WAITING_BIAS. |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 273 | */ |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame^] | 274 | if (count != RWSEM_WAITING_BIAS) |
| 275 | return false; |
| 276 | |
| 277 | /* |
| 278 | * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there |
| 279 | * are other tasks on the wait list, we need to add on WAITING_BIAS. |
| 280 | */ |
| 281 | count = list_is_singular(&sem->wait_list) ? |
| 282 | RWSEM_ACTIVE_WRITE_BIAS : |
| 283 | RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS; |
| 284 | |
| 285 | if (cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count) == RWSEM_WAITING_BIAS) { |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 286 | rwsem_set_owner(sem); |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 287 | return true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 288 | } |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 289 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 290 | return false; |
| 291 | } |
| 292 | |
Davidlohr Bueso | 5db6c6f | 2014-07-11 14:00:06 -0700 | [diff] [blame] | 293 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 294 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 295 | * Try to acquire write lock before the writer has been put on wait queue. |
| 296 | */ |
| 297 | static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem) |
| 298 | { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 299 | long old, count = READ_ONCE(sem->count); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 300 | |
| 301 | while (true) { |
| 302 | if (!(count == 0 || count == RWSEM_WAITING_BIAS)) |
| 303 | return false; |
| 304 | |
Davidlohr Bueso | 00eb4ba | 2015-09-30 13:03:15 -0700 | [diff] [blame] | 305 | old = cmpxchg_acquire(&sem->count, count, |
| 306 | count + RWSEM_ACTIVE_WRITE_BIAS); |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 307 | if (old == count) { |
| 308 | rwsem_set_owner(sem); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 309 | return true; |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 310 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 311 | |
| 312 | count = old; |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem) |
| 317 | { |
| 318 | struct task_struct *owner; |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 319 | bool ret = true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 320 | |
| 321 | if (need_resched()) |
Jason Low | 37e9562 | 2014-07-04 20:49:32 -0700 | [diff] [blame] | 322 | return false; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 323 | |
| 324 | rcu_read_lock(); |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 325 | owner = READ_ONCE(sem->owner); |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 326 | if (!owner) { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 327 | long count = READ_ONCE(sem->count); |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 328 | /* |
| 329 | * If sem->owner is not set, yet we have just recently entered the |
| 330 | * slowpath with the lock being active, then there is a possibility |
| 331 | * reader(s) may have the lock. To be safe, bail spinning in these |
| 332 | * situations. |
| 333 | */ |
| 334 | if (count & RWSEM_ACTIVE_MASK) |
| 335 | ret = false; |
| 336 | goto done; |
| 337 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 338 | |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 339 | ret = owner->on_cpu; |
| 340 | done: |
| 341 | rcu_read_unlock(); |
| 342 | return ret; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 343 | } |
| 344 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 345 | static noinline |
| 346 | bool rwsem_spin_on_owner(struct rw_semaphore *sem, struct task_struct *owner) |
| 347 | { |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 348 | long count; |
| 349 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 350 | rcu_read_lock(); |
Jason Low | 9198f6e | 2015-03-06 23:45:31 -0800 | [diff] [blame] | 351 | while (sem->owner == owner) { |
| 352 | /* |
| 353 | * Ensure we emit the owner->on_cpu, dereference _after_ |
| 354 | * checking sem->owner still matches owner, if that fails, |
| 355 | * owner might point to free()d memory, if it still matches, |
| 356 | * the rcu_read_lock() ensures the memory stays valid. |
| 357 | */ |
| 358 | barrier(); |
| 359 | |
| 360 | /* abort spinning when need_resched or owner is not running */ |
| 361 | if (!owner->on_cpu || need_resched()) { |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 362 | rcu_read_unlock(); |
| 363 | return false; |
| 364 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 365 | |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 366 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 367 | } |
| 368 | rcu_read_unlock(); |
| 369 | |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 370 | if (READ_ONCE(sem->owner)) |
| 371 | return true; /* new owner, continue spinning */ |
| 372 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 373 | /* |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 374 | * When the owner is not set, the lock could be free or |
| 375 | * held by readers. Check the counter to verify the |
| 376 | * state. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 377 | */ |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 378 | count = READ_ONCE(sem->count); |
| 379 | return (count == 0 || count == RWSEM_WAITING_BIAS); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 380 | } |
| 381 | |
| 382 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 383 | { |
| 384 | struct task_struct *owner; |
| 385 | bool taken = false; |
| 386 | |
| 387 | preempt_disable(); |
| 388 | |
| 389 | /* sem->wait_lock should not be held when doing optimistic spinning */ |
| 390 | if (!rwsem_can_spin_on_owner(sem)) |
| 391 | goto done; |
| 392 | |
| 393 | if (!osq_lock(&sem->osq)) |
| 394 | goto done; |
| 395 | |
| 396 | while (true) { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 397 | owner = READ_ONCE(sem->owner); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 398 | if (owner && !rwsem_spin_on_owner(sem, owner)) |
| 399 | break; |
| 400 | |
| 401 | /* wait_lock will be acquired if write_lock is obtained */ |
| 402 | if (rwsem_try_write_lock_unqueued(sem)) { |
| 403 | taken = true; |
| 404 | break; |
| 405 | } |
| 406 | |
| 407 | /* |
| 408 | * When there's no owner, we might have preempted between the |
| 409 | * owner acquiring the lock and setting the owner field. If |
| 410 | * we're an RT task that will live-lock because we won't let |
| 411 | * the owner complete. |
| 412 | */ |
| 413 | if (!owner && (need_resched() || rt_task(current))) |
| 414 | break; |
| 415 | |
| 416 | /* |
| 417 | * The cpu_relax() call is a compiler barrier which forces |
| 418 | * everything in this loop to be re-loaded. We don't need |
| 419 | * memory barriers as we'll eventually observe the right |
| 420 | * values at the cost of a few extra spins. |
| 421 | */ |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 422 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 423 | } |
| 424 | osq_unlock(&sem->osq); |
| 425 | done: |
| 426 | preempt_enable(); |
| 427 | return taken; |
| 428 | } |
| 429 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 430 | /* |
| 431 | * Return true if the rwsem has active spinner |
| 432 | */ |
| 433 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 434 | { |
| 435 | return osq_is_locked(&sem->osq); |
| 436 | } |
| 437 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 438 | #else |
| 439 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 440 | { |
| 441 | return false; |
| 442 | } |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 443 | |
| 444 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 445 | { |
| 446 | return false; |
| 447 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 448 | #endif |
| 449 | |
| 450 | /* |
| 451 | * Wait until we successfully acquire the write lock |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 452 | */ |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 453 | static inline struct rw_semaphore * |
| 454 | __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 455 | { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 456 | long count; |
| 457 | bool waiting = true; /* any queued threads before us */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 458 | struct rwsem_waiter waiter; |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 459 | struct rw_semaphore *ret = sem; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 460 | WAKE_Q(wake_q); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 461 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 462 | /* undo write bias from down_write operation, stop active locking */ |
| 463 | count = rwsem_atomic_update(-RWSEM_ACTIVE_WRITE_BIAS, sem); |
| 464 | |
| 465 | /* do optimistic spinning and steal lock if possible */ |
| 466 | if (rwsem_optimistic_spin(sem)) |
| 467 | return sem; |
| 468 | |
| 469 | /* |
| 470 | * Optimistic spinning failed, proceed to the slowpath |
| 471 | * and block until we can acquire the sem. |
| 472 | */ |
| 473 | waiter.task = current; |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 474 | waiter.type = RWSEM_WAITING_FOR_WRITE; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 475 | |
| 476 | raw_spin_lock_irq(&sem->wait_lock); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 477 | |
| 478 | /* account for this before adding a new element to the list */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 479 | if (list_empty(&sem->wait_list)) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 480 | waiting = false; |
| 481 | |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 482 | list_add_tail(&waiter.list, &sem->wait_list); |
| 483 | |
| 484 | /* we're now waiting on the lock, but no longer actively locking */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 485 | if (waiting) { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 486 | count = READ_ONCE(sem->count); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 487 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 488 | /* |
Andrew Morton | 0cc3d01 | 2014-06-04 20:19:48 +0200 | [diff] [blame] | 489 | * If there were already threads queued before us and there are |
| 490 | * no active writers, the lock must be read owned; so we try to |
| 491 | * wake any read locks that were queued ahead of us. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 492 | */ |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 493 | if (count > RWSEM_WAITING_BIAS) { |
| 494 | WAKE_Q(wake_q); |
| 495 | |
| 496 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q); |
| 497 | /* |
| 498 | * The wakeup is normally called _after_ the wait_lock |
| 499 | * is released, but given that we are proactively waking |
| 500 | * readers we can deal with the wake_q overhead as it is |
| 501 | * similar to releasing and taking the wait_lock again |
| 502 | * for attempting rwsem_try_write_lock(). |
| 503 | */ |
| 504 | wake_up_q(&wake_q); |
| 505 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 506 | |
| 507 | } else |
| 508 | count = rwsem_atomic_update(RWSEM_WAITING_BIAS, sem); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 509 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 510 | /* wait until we successfully acquire the lock */ |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 511 | set_current_state(state); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 512 | while (true) { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 513 | if (rwsem_try_write_lock(count, sem)) |
| 514 | break; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 515 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 516 | |
| 517 | /* Block until there are no active lockers. */ |
| 518 | do { |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 519 | if (signal_pending_state(state, current)) |
| 520 | goto out_nolock; |
| 521 | |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 522 | schedule(); |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 523 | set_current_state(state); |
Michel Lespinasse | 9b0fc9c | 2013-05-07 06:45:57 -0700 | [diff] [blame] | 524 | } while ((count = sem->count) & RWSEM_ACTIVE_MASK); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 525 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 526 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 527 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 528 | __set_current_state(TASK_RUNNING); |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 529 | list_del(&waiter.list); |
| 530 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 531 | |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 532 | return ret; |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 533 | |
| 534 | out_nolock: |
| 535 | __set_current_state(TASK_RUNNING); |
| 536 | raw_spin_lock_irq(&sem->wait_lock); |
| 537 | list_del(&waiter.list); |
| 538 | if (list_empty(&sem->wait_list)) |
| 539 | rwsem_atomic_update(-RWSEM_WAITING_BIAS, sem); |
| 540 | else |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 541 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 542 | raw_spin_unlock_irq(&sem->wait_lock); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 543 | wake_up_q(&wake_q); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 544 | |
| 545 | return ERR_PTR(-EINTR); |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 546 | } |
| 547 | |
| 548 | __visible struct rw_semaphore * __sched |
| 549 | rwsem_down_write_failed(struct rw_semaphore *sem) |
| 550 | { |
| 551 | return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 552 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 553 | EXPORT_SYMBOL(rwsem_down_write_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 554 | |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 555 | __visible struct rw_semaphore * __sched |
| 556 | rwsem_down_write_failed_killable(struct rw_semaphore *sem) |
| 557 | { |
| 558 | return __rwsem_down_write_failed_common(sem, TASK_KILLABLE); |
| 559 | } |
| 560 | EXPORT_SYMBOL(rwsem_down_write_failed_killable); |
| 561 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 562 | /* |
| 563 | * handle waking up a waiter on the semaphore |
| 564 | * - up_read/up_write has decremented the active part of count if we come here |
| 565 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 566 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 567 | struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 568 | { |
| 569 | unsigned long flags; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 570 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 571 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 572 | /* |
| 573 | * If a spinner is present, it is not necessary to do the wakeup. |
| 574 | * Try to do wakeup only if the trylock succeeds to minimize |
| 575 | * spinlock contention which may introduce too much delay in the |
| 576 | * unlock operation. |
| 577 | * |
| 578 | * spinning writer up_write/up_read caller |
| 579 | * --------------- ----------------------- |
| 580 | * [S] osq_unlock() [L] osq |
| 581 | * MB RMB |
| 582 | * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock) |
| 583 | * |
| 584 | * Here, it is important to make sure that there won't be a missed |
| 585 | * wakeup while the rwsem is free and the only spinning writer goes |
| 586 | * to sleep without taking the rwsem. Even when the spinning writer |
| 587 | * is just going to break out of the waiting loop, it will still do |
| 588 | * a trylock in rwsem_down_write_failed() before sleeping. IOW, if |
| 589 | * rwsem_has_spinner() is true, it will guarantee at least one |
| 590 | * trylock attempt on the rwsem later on. |
| 591 | */ |
| 592 | if (rwsem_has_spinner(sem)) { |
| 593 | /* |
| 594 | * The smp_rmb() here is to make sure that the spinner |
| 595 | * state is consulted before reading the wait_lock. |
| 596 | */ |
| 597 | smp_rmb(); |
| 598 | if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags)) |
| 599 | return sem; |
| 600 | goto locked; |
| 601 | } |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 602 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 603 | locked: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 604 | |
| 605 | /* do nothing if list empty */ |
| 606 | if (!list_empty(&sem->wait_list)) |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 607 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 608 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 609 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 610 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 611 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 612 | return sem; |
| 613 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 614 | EXPORT_SYMBOL(rwsem_wake); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 615 | |
| 616 | /* |
| 617 | * downgrade a write lock into a read lock |
| 618 | * - caller incremented waiting part of count and discovered it still negative |
| 619 | * - just wake up any readers at the front of the queue |
| 620 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 621 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 622 | struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 623 | { |
| 624 | unsigned long flags; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 625 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 626 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 627 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 628 | |
| 629 | /* do nothing if list empty */ |
| 630 | if (!list_empty(&sem->wait_list)) |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 631 | sem = __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 632 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 633 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 634 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 635 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 636 | return sem; |
| 637 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 638 | EXPORT_SYMBOL(rwsem_downgrade_wake); |