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 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 117 | * - the spinlock must be held by the caller |
| 118 | * - woken process blocks are discarded from the list after having task zeroed |
| 119 | * - writers are only woken if downgrading is false |
| 120 | */ |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 121 | static struct rw_semaphore * |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 122 | __rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | { |
| 124 | struct rwsem_waiter *waiter; |
| 125 | struct task_struct *tsk; |
| 126 | struct list_head *next; |
Davidlohr Bueso | b5f5418 | 2013-05-07 06:46:02 -0700 | [diff] [blame] | 127 | long oldcount, woken, loop, adjustment; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 128 | |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 129 | waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list); |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 130 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) { |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 131 | if (wake_type == RWSEM_WAKE_ANY) |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 132 | /* Wake writer at the front of the queue, but do not |
| 133 | * grant it the lock yet as we want other writers |
| 134 | * to be able to steal it. Readers, on the other hand, |
| 135 | * will block as they will notice the queued writer. |
| 136 | */ |
| 137 | wake_up_process(waiter->task); |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 138 | goto out; |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 139 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 140 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 141 | /* Writers might steal the lock before we grant it to the next reader. |
| 142 | * We prefer to do the first reader grant before counting readers |
| 143 | * so we can bail out early if a writer stole the lock. |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 144 | */ |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 145 | adjustment = 0; |
| 146 | if (wake_type != RWSEM_WAKE_READ_OWNED) { |
| 147 | adjustment = RWSEM_ACTIVE_READ_BIAS; |
| 148 | try_reader_grant: |
| 149 | oldcount = rwsem_atomic_update(adjustment, sem) - adjustment; |
| 150 | if (unlikely(oldcount < RWSEM_WAITING_BIAS)) { |
| 151 | /* A writer stole the lock. Undo our reader grant. */ |
| 152 | if (rwsem_atomic_update(-adjustment, sem) & |
| 153 | RWSEM_ACTIVE_MASK) |
| 154 | goto out; |
| 155 | /* Last active locker left. Retry waking readers. */ |
| 156 | goto try_reader_grant; |
| 157 | } |
| 158 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 159 | |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 160 | /* Grant an infinite number of read locks to the readers at the front |
| 161 | * of the queue. Note we increment the 'active part' of the count by |
| 162 | * the number of readers before waking any processes up. |
| 163 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 164 | woken = 0; |
| 165 | do { |
| 166 | woken++; |
| 167 | |
| 168 | if (waiter->list.next == &sem->wait_list) |
| 169 | break; |
| 170 | |
| 171 | waiter = list_entry(waiter->list.next, |
| 172 | struct rwsem_waiter, list); |
| 173 | |
Michel Lespinasse | e2d57f7 | 2013-05-07 06:45:49 -0700 | [diff] [blame] | 174 | } while (waiter->type != RWSEM_WAITING_FOR_WRITE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 175 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 176 | adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment; |
Michel Lespinasse | e2d57f7 | 2013-05-07 06:45:49 -0700 | [diff] [blame] | 177 | if (waiter->type != RWSEM_WAITING_FOR_WRITE) |
Michel Lespinasse | fd41b33 | 2010-08-09 17:21:18 -0700 | [diff] [blame] | 178 | /* hit end of list above */ |
| 179 | adjustment -= RWSEM_WAITING_BIAS; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 180 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 181 | if (adjustment) |
| 182 | rwsem_atomic_add(adjustment, sem); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 183 | |
| 184 | next = sem->wait_list.next; |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 185 | loop = woken; |
| 186 | do { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 187 | waiter = list_entry(next, struct rwsem_waiter, list); |
| 188 | next = waiter->list.next; |
| 189 | tsk = waiter->task; |
Davidlohr Bueso | 49e4b2b | 2015-01-30 01:14:24 -0800 | [diff] [blame] | 190 | /* |
| 191 | * Make sure we do not wakeup the next reader before |
| 192 | * setting the nil condition to grant the next reader; |
| 193 | * otherwise we could miss the wakeup on the other |
| 194 | * side and end up sleeping again. See the pairing |
| 195 | * in rwsem_down_read_failed(). |
| 196 | */ |
akpm@osdl.org | d59dd46 | 2005-05-01 08:58:47 -0700 | [diff] [blame] | 197 | smp_mb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 198 | waiter->task = NULL; |
| 199 | wake_up_process(tsk); |
| 200 | put_task_struct(tsk); |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 201 | } while (--loop); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 202 | |
| 203 | sem->wait_list.next = next; |
| 204 | next->prev = &sem->wait_list; |
| 205 | |
| 206 | out: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 207 | return sem; |
Alex Shi | ce6711f | 2013-02-05 21:11:55 +0800 | [diff] [blame] | 208 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 209 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 210 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 211 | * Wait for the read lock to be granted |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 212 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 213 | __visible |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 214 | struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 215 | { |
Davidlohr Bueso | b5f5418 | 2013-05-07 06:46:02 -0700 | [diff] [blame] | 216 | long count, adjustment = -RWSEM_ACTIVE_READ_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 217 | struct rwsem_waiter waiter; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 218 | struct task_struct *tsk = current; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 219 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 220 | /* set up my own style of waitqueue */ |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 221 | waiter.task = tsk; |
Michel Lespinasse | da16922 | 2013-05-07 06:45:52 -0700 | [diff] [blame] | 222 | waiter.type = RWSEM_WAITING_FOR_READ; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 223 | get_task_struct(tsk); |
| 224 | |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 225 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | fd41b33 | 2010-08-09 17:21:18 -0700 | [diff] [blame] | 226 | if (list_empty(&sem->wait_list)) |
| 227 | adjustment += RWSEM_WAITING_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 228 | list_add_tail(&waiter.list, &sem->wait_list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 229 | |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 230 | /* we're now waiting on the lock, but no longer actively locking */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 231 | count = rwsem_atomic_update(adjustment, sem); |
| 232 | |
Michel Lespinasse | 25c3932 | 2013-05-07 06:46:00 -0700 | [diff] [blame] | 233 | /* If there are no active locks, wake the front queued process(es). |
| 234 | * |
| 235 | * If there are no writers and we are first in the queue, |
| 236 | * wake our own waiter to join the existing active readers ! |
| 237 | */ |
| 238 | if (count == RWSEM_WAITING_BIAS || |
| 239 | (count > RWSEM_WAITING_BIAS && |
| 240 | adjustment != -RWSEM_ACTIVE_READ_BIAS)) |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 241 | sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 242 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 243 | raw_spin_unlock_irq(&sem->wait_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 244 | |
| 245 | /* wait to be given the lock */ |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 246 | while (true) { |
| 247 | set_task_state(tsk, TASK_UNINTERRUPTIBLE); |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 248 | if (!waiter.task) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 249 | break; |
| 250 | schedule(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 251 | } |
| 252 | |
Davidlohr Bueso | 7310599 | 2015-01-25 23:36:04 -0800 | [diff] [blame] | 253 | __set_task_state(tsk, TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 254 | return sem; |
| 255 | } |
Davidlohr Bueso | db0e716a | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 256 | EXPORT_SYMBOL(rwsem_down_read_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 257 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 258 | static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem) |
| 259 | { |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 260 | /* |
| 261 | * Try acquiring the write lock. Check count first in order |
| 262 | * to reduce unnecessary expensive cmpxchg() operations. |
| 263 | */ |
| 264 | if (count == RWSEM_WAITING_BIAS && |
| 265 | cmpxchg(&sem->count, RWSEM_WAITING_BIAS, |
| 266 | RWSEM_ACTIVE_WRITE_BIAS) == RWSEM_WAITING_BIAS) { |
| 267 | if (!list_is_singular(&sem->wait_list)) |
| 268 | rwsem_atomic_update(RWSEM_WAITING_BIAS, sem); |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 269 | rwsem_set_owner(sem); |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 270 | return true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 271 | } |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 272 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 273 | return false; |
| 274 | } |
| 275 | |
Davidlohr Bueso | 5db6c6f | 2014-07-11 14:00:06 -0700 | [diff] [blame] | 276 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 277 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 278 | * Try to acquire write lock before the writer has been put on wait queue. |
| 279 | */ |
| 280 | static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem) |
| 281 | { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 282 | long old, count = READ_ONCE(sem->count); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 283 | |
| 284 | while (true) { |
| 285 | if (!(count == 0 || count == RWSEM_WAITING_BIAS)) |
| 286 | return false; |
| 287 | |
| 288 | old = cmpxchg(&sem->count, count, count + RWSEM_ACTIVE_WRITE_BIAS); |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 289 | if (old == count) { |
| 290 | rwsem_set_owner(sem); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 291 | return true; |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 292 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 293 | |
| 294 | count = old; |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem) |
| 299 | { |
| 300 | struct task_struct *owner; |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 301 | bool ret = true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 302 | |
| 303 | if (need_resched()) |
Jason Low | 37e9562 | 2014-07-04 20:49:32 -0700 | [diff] [blame] | 304 | return false; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 305 | |
| 306 | rcu_read_lock(); |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 307 | owner = READ_ONCE(sem->owner); |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 308 | if (!owner) { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 309 | long count = READ_ONCE(sem->count); |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 310 | /* |
| 311 | * If sem->owner is not set, yet we have just recently entered the |
| 312 | * slowpath with the lock being active, then there is a possibility |
| 313 | * reader(s) may have the lock. To be safe, bail spinning in these |
| 314 | * situations. |
| 315 | */ |
| 316 | if (count & RWSEM_ACTIVE_MASK) |
| 317 | ret = false; |
| 318 | goto done; |
| 319 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 320 | |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 321 | ret = owner->on_cpu; |
| 322 | done: |
| 323 | rcu_read_unlock(); |
| 324 | return ret; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 325 | } |
| 326 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 327 | static noinline |
| 328 | bool rwsem_spin_on_owner(struct rw_semaphore *sem, struct task_struct *owner) |
| 329 | { |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 330 | long count; |
| 331 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 332 | rcu_read_lock(); |
Jason Low | 9198f6e | 2015-03-06 23:45:31 -0800 | [diff] [blame] | 333 | while (sem->owner == owner) { |
| 334 | /* |
| 335 | * Ensure we emit the owner->on_cpu, dereference _after_ |
| 336 | * checking sem->owner still matches owner, if that fails, |
| 337 | * owner might point to free()d memory, if it still matches, |
| 338 | * the rcu_read_lock() ensures the memory stays valid. |
| 339 | */ |
| 340 | barrier(); |
| 341 | |
| 342 | /* abort spinning when need_resched or owner is not running */ |
| 343 | if (!owner->on_cpu || need_resched()) { |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 344 | rcu_read_unlock(); |
| 345 | return false; |
| 346 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 347 | |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 348 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 349 | } |
| 350 | rcu_read_unlock(); |
| 351 | |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 352 | if (READ_ONCE(sem->owner)) |
| 353 | return true; /* new owner, continue spinning */ |
| 354 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 355 | /* |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 356 | * When the owner is not set, the lock could be free or |
| 357 | * held by readers. Check the counter to verify the |
| 358 | * state. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 359 | */ |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 360 | count = READ_ONCE(sem->count); |
| 361 | return (count == 0 || count == RWSEM_WAITING_BIAS); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 362 | } |
| 363 | |
| 364 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 365 | { |
| 366 | struct task_struct *owner; |
| 367 | bool taken = false; |
| 368 | |
| 369 | preempt_disable(); |
| 370 | |
| 371 | /* sem->wait_lock should not be held when doing optimistic spinning */ |
| 372 | if (!rwsem_can_spin_on_owner(sem)) |
| 373 | goto done; |
| 374 | |
| 375 | if (!osq_lock(&sem->osq)) |
| 376 | goto done; |
| 377 | |
| 378 | while (true) { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 379 | owner = READ_ONCE(sem->owner); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 380 | if (owner && !rwsem_spin_on_owner(sem, owner)) |
| 381 | break; |
| 382 | |
| 383 | /* wait_lock will be acquired if write_lock is obtained */ |
| 384 | if (rwsem_try_write_lock_unqueued(sem)) { |
| 385 | taken = true; |
| 386 | break; |
| 387 | } |
| 388 | |
| 389 | /* |
| 390 | * When there's no owner, we might have preempted between the |
| 391 | * owner acquiring the lock and setting the owner field. If |
| 392 | * we're an RT task that will live-lock because we won't let |
| 393 | * the owner complete. |
| 394 | */ |
| 395 | if (!owner && (need_resched() || rt_task(current))) |
| 396 | break; |
| 397 | |
| 398 | /* |
| 399 | * The cpu_relax() call is a compiler barrier which forces |
| 400 | * everything in this loop to be re-loaded. We don't need |
| 401 | * memory barriers as we'll eventually observe the right |
| 402 | * values at the cost of a few extra spins. |
| 403 | */ |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 404 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 405 | } |
| 406 | osq_unlock(&sem->osq); |
| 407 | done: |
| 408 | preempt_enable(); |
| 409 | return taken; |
| 410 | } |
| 411 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 412 | /* |
| 413 | * Return true if the rwsem has active spinner |
| 414 | */ |
| 415 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 416 | { |
| 417 | return osq_is_locked(&sem->osq); |
| 418 | } |
| 419 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 420 | #else |
| 421 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 422 | { |
| 423 | return false; |
| 424 | } |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 425 | |
| 426 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 427 | { |
| 428 | return false; |
| 429 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 430 | #endif |
| 431 | |
| 432 | /* |
| 433 | * Wait until we successfully acquire the write lock |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 434 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 435 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 436 | struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 437 | { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 438 | long count; |
| 439 | bool waiting = true; /* any queued threads before us */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 440 | struct rwsem_waiter waiter; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 441 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 442 | /* undo write bias from down_write operation, stop active locking */ |
| 443 | count = rwsem_atomic_update(-RWSEM_ACTIVE_WRITE_BIAS, sem); |
| 444 | |
| 445 | /* do optimistic spinning and steal lock if possible */ |
| 446 | if (rwsem_optimistic_spin(sem)) |
| 447 | return sem; |
| 448 | |
| 449 | /* |
| 450 | * Optimistic spinning failed, proceed to the slowpath |
| 451 | * and block until we can acquire the sem. |
| 452 | */ |
| 453 | waiter.task = current; |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 454 | waiter.type = RWSEM_WAITING_FOR_WRITE; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 455 | |
| 456 | raw_spin_lock_irq(&sem->wait_lock); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 457 | |
| 458 | /* account for this before adding a new element to the list */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 459 | if (list_empty(&sem->wait_list)) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 460 | waiting = false; |
| 461 | |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 462 | list_add_tail(&waiter.list, &sem->wait_list); |
| 463 | |
| 464 | /* we're now waiting on the lock, but no longer actively locking */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 465 | if (waiting) { |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 466 | count = READ_ONCE(sem->count); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 467 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 468 | /* |
Andrew Morton | 0cc3d01 | 2014-06-04 20:19:48 +0200 | [diff] [blame] | 469 | * If there were already threads queued before us and there are |
| 470 | * no active writers, the lock must be read owned; so we try to |
| 471 | * wake any read locks that were queued ahead of us. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 472 | */ |
| 473 | if (count > RWSEM_WAITING_BIAS) |
| 474 | sem = __rwsem_do_wake(sem, RWSEM_WAKE_READERS); |
| 475 | |
| 476 | } else |
| 477 | count = rwsem_atomic_update(RWSEM_WAITING_BIAS, sem); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 478 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 479 | /* wait until we successfully acquire the lock */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 480 | set_current_state(TASK_UNINTERRUPTIBLE); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 481 | while (true) { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 482 | if (rwsem_try_write_lock(count, sem)) |
| 483 | break; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 484 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 485 | |
| 486 | /* Block until there are no active lockers. */ |
| 487 | do { |
| 488 | schedule(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 489 | set_current_state(TASK_UNINTERRUPTIBLE); |
Michel Lespinasse | 9b0fc9c | 2013-05-07 06:45:57 -0700 | [diff] [blame] | 490 | } while ((count = sem->count) & RWSEM_ACTIVE_MASK); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 491 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 492 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 493 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 494 | __set_current_state(TASK_RUNNING); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 495 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 496 | list_del(&waiter.list); |
| 497 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 498 | |
| 499 | return sem; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 500 | } |
Davidlohr Bueso | db0e716a | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 501 | EXPORT_SYMBOL(rwsem_down_write_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 502 | |
| 503 | /* |
| 504 | * handle waking up a waiter on the semaphore |
| 505 | * - up_read/up_write has decremented the active part of count if we come here |
| 506 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 507 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 508 | struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 509 | { |
| 510 | unsigned long flags; |
| 511 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 512 | /* |
| 513 | * If a spinner is present, it is not necessary to do the wakeup. |
| 514 | * Try to do wakeup only if the trylock succeeds to minimize |
| 515 | * spinlock contention which may introduce too much delay in the |
| 516 | * unlock operation. |
| 517 | * |
| 518 | * spinning writer up_write/up_read caller |
| 519 | * --------------- ----------------------- |
| 520 | * [S] osq_unlock() [L] osq |
| 521 | * MB RMB |
| 522 | * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock) |
| 523 | * |
| 524 | * Here, it is important to make sure that there won't be a missed |
| 525 | * wakeup while the rwsem is free and the only spinning writer goes |
| 526 | * to sleep without taking the rwsem. Even when the spinning writer |
| 527 | * is just going to break out of the waiting loop, it will still do |
| 528 | * a trylock in rwsem_down_write_failed() before sleeping. IOW, if |
| 529 | * rwsem_has_spinner() is true, it will guarantee at least one |
| 530 | * trylock attempt on the rwsem later on. |
| 531 | */ |
| 532 | if (rwsem_has_spinner(sem)) { |
| 533 | /* |
| 534 | * The smp_rmb() here is to make sure that the spinner |
| 535 | * state is consulted before reading the wait_lock. |
| 536 | */ |
| 537 | smp_rmb(); |
| 538 | if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags)) |
| 539 | return sem; |
| 540 | goto locked; |
| 541 | } |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 542 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 543 | locked: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 544 | |
| 545 | /* do nothing if list empty */ |
| 546 | if (!list_empty(&sem->wait_list)) |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 547 | sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 548 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 549 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 550 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 551 | return sem; |
| 552 | } |
Davidlohr Bueso | db0e716a | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 553 | EXPORT_SYMBOL(rwsem_wake); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 554 | |
| 555 | /* |
| 556 | * downgrade a write lock into a read lock |
| 557 | * - caller incremented waiting part of count and discovered it still negative |
| 558 | * - just wake up any readers at the front of the queue |
| 559 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 560 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 561 | struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 562 | { |
| 563 | unsigned long flags; |
| 564 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 565 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 566 | |
| 567 | /* do nothing if list empty */ |
| 568 | if (!list_empty(&sem->wait_list)) |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 569 | sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 570 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 571 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 572 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 573 | return sem; |
| 574 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 575 | EXPORT_SYMBOL(rwsem_downgrade_wake); |