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 |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 83 | atomic_long_set(&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 | */ |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 124 | static void __rwsem_mark_wake(struct rw_semaphore *sem, |
| 125 | enum rwsem_wake_type wake_type, |
| 126 | struct wake_q_head *wake_q) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 127 | { |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 128 | struct rwsem_waiter *waiter, *tmp; |
| 129 | long oldcount, woken = 0, adjustment = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 130 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 131 | /* |
| 132 | * Take a peek at the queue head waiter such that we can determine |
| 133 | * the wakeup(s) to perform. |
| 134 | */ |
| 135 | waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list); |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 136 | |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 137 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) { |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 138 | if (wake_type == RWSEM_WAKE_ANY) { |
| 139 | /* |
| 140 | * Mark writer at the front of the queue for wakeup. |
| 141 | * Until the task is actually later awoken later by |
| 142 | * the caller, other writers are able to steal it. |
| 143 | * Readers, on the other hand, will block as they |
| 144 | * will notice the queued writer. |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 145 | */ |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 146 | wake_q_add(wake_q, waiter->task); |
| 147 | } |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 148 | |
| 149 | return; |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 150 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 151 | |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 152 | /* |
| 153 | * Writers might steal the lock before we grant it to the next reader. |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 154 | * We prefer to do the first reader grant before counting readers |
| 155 | * so we can bail out early if a writer stole the lock. |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 156 | */ |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 157 | if (wake_type != RWSEM_WAKE_READ_OWNED) { |
| 158 | adjustment = RWSEM_ACTIVE_READ_BIAS; |
| 159 | try_reader_grant: |
Peter Zijlstra | 86a3b5f | 2016-05-18 12:42:21 +0200 | [diff] [blame] | 160 | oldcount = atomic_long_fetch_add(adjustment, &sem->count); |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 161 | if (unlikely(oldcount < RWSEM_WAITING_BIAS)) { |
Waiman Long | bf7b4c4 | 2016-05-17 21:26:22 -0400 | [diff] [blame] | 162 | /* |
| 163 | * If the count is still less than RWSEM_WAITING_BIAS |
| 164 | * after removing the adjustment, it is assumed that |
| 165 | * a writer has stolen the lock. We have to undo our |
| 166 | * reader grant. |
| 167 | */ |
| 168 | if (atomic_long_add_return(-adjustment, &sem->count) < |
| 169 | RWSEM_WAITING_BIAS) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 170 | return; |
| 171 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 172 | /* Last active locker left. Retry waking readers. */ |
| 173 | goto try_reader_grant; |
| 174 | } |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 175 | /* |
| 176 | * It is not really necessary to set it to reader-owned here, |
| 177 | * but it gives the spinners an early indication that the |
| 178 | * readers now have the lock. |
| 179 | */ |
| 180 | rwsem_set_reader_owned(sem); |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 181 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 182 | |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 183 | /* |
| 184 | * Grant an infinite number of read locks to the readers at the front |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 185 | * of the queue. We know that woken will be at least 1 as we accounted |
| 186 | * for above. Note we increment the 'active part' of the count by the |
| 187 | * number of readers before waking any processes up. |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 188 | */ |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 189 | list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) { |
| 190 | struct task_struct *tsk; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 191 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 192 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 193 | break; |
| 194 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 195 | woken++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 196 | tsk = waiter->task; |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 197 | |
Xie Yongji | b4df458 | 2018-11-29 20:50:30 +0800 | [diff] [blame^] | 198 | get_task_struct(tsk); |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 199 | list_del(&waiter->list); |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 200 | /* |
Xie Yongji | b4df458 | 2018-11-29 20:50:30 +0800 | [diff] [blame^] | 201 | * Ensure calling get_task_struct() before setting the reader |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 202 | * waiter to nil such that rwsem_down_read_failed() cannot |
| 203 | * race with do_exit() by always holding a reference count |
| 204 | * to the task to wakeup. |
| 205 | */ |
| 206 | smp_store_release(&waiter->task, NULL); |
Xie Yongji | b4df458 | 2018-11-29 20:50:30 +0800 | [diff] [blame^] | 207 | /* |
| 208 | * Ensure issuing the wakeup (either by us or someone else) |
| 209 | * after setting the reader waiter to nil. |
| 210 | */ |
| 211 | wake_q_add(wake_q, tsk); |
| 212 | /* wake_q_add() already take the task ref */ |
| 213 | put_task_struct(tsk); |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 214 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 215 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 216 | adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment; |
| 217 | if (list_empty(&sem->wait_list)) { |
| 218 | /* hit end of list above */ |
| 219 | adjustment -= RWSEM_WAITING_BIAS; |
| 220 | } |
| 221 | |
| 222 | if (adjustment) |
| 223 | atomic_long_add(adjustment, &sem->count); |
Alex Shi | ce6711f | 2013-02-05 21:11:55 +0800 | [diff] [blame] | 224 | } |
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 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 227 | * Wait for the read lock to be granted |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 228 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 229 | __visible |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 230 | struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 231 | { |
Davidlohr Bueso | b5f5418 | 2013-05-07 06:46:02 -0700 | [diff] [blame] | 232 | long count, adjustment = -RWSEM_ACTIVE_READ_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 233 | struct rwsem_waiter waiter; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 234 | struct task_struct *tsk = current; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 235 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 237 | waiter.task = tsk; |
Michel Lespinasse | da16922 | 2013-05-07 06:45:52 -0700 | [diff] [blame] | 238 | waiter.type = RWSEM_WAITING_FOR_READ; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 239 | |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 240 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | fd41b33 | 2010-08-09 17:21:18 -0700 | [diff] [blame] | 241 | if (list_empty(&sem->wait_list)) |
| 242 | adjustment += RWSEM_WAITING_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 243 | list_add_tail(&waiter.list, &sem->wait_list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 244 | |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 245 | /* we're now waiting on the lock, but no longer actively locking */ |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 246 | count = atomic_long_add_return(adjustment, &sem->count); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 247 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 248 | /* |
| 249 | * If there are no active locks, wake the front queued process(es). |
Michel Lespinasse | 25c3932 | 2013-05-07 06:46:00 -0700 | [diff] [blame] | 250 | * |
| 251 | * If there are no writers and we are first in the queue, |
| 252 | * wake our own waiter to join the existing active readers ! |
| 253 | */ |
| 254 | if (count == RWSEM_WAITING_BIAS || |
| 255 | (count > RWSEM_WAITING_BIAS && |
| 256 | adjustment != -RWSEM_ACTIVE_READ_BIAS)) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 257 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 258 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 259 | raw_spin_unlock_irq(&sem->wait_lock); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 260 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 261 | |
| 262 | /* wait to be given the lock */ |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 263 | while (true) { |
| 264 | set_task_state(tsk, TASK_UNINTERRUPTIBLE); |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 265 | if (!waiter.task) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 266 | break; |
| 267 | schedule(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 268 | } |
| 269 | |
Davidlohr Bueso | 7310599 | 2015-01-25 23:36:04 -0800 | [diff] [blame] | 270 | __set_task_state(tsk, TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 271 | return sem; |
| 272 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 273 | EXPORT_SYMBOL(rwsem_down_read_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 274 | |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame] | 275 | /* |
| 276 | * This function must be called with the sem->wait_lock held to prevent |
| 277 | * race conditions between checking the rwsem wait list and setting the |
| 278 | * sem->count accordingly. |
| 279 | */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 280 | static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem) |
| 281 | { |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 282 | /* |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame] | 283 | * 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] | 284 | */ |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame] | 285 | if (count != RWSEM_WAITING_BIAS) |
| 286 | return false; |
| 287 | |
| 288 | /* |
| 289 | * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there |
| 290 | * are other tasks on the wait list, we need to add on WAITING_BIAS. |
| 291 | */ |
| 292 | count = list_is_singular(&sem->wait_list) ? |
| 293 | RWSEM_ACTIVE_WRITE_BIAS : |
| 294 | RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS; |
| 295 | |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 296 | if (atomic_long_cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count) |
| 297 | == RWSEM_WAITING_BIAS) { |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 298 | rwsem_set_owner(sem); |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 299 | return true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 300 | } |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 301 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 302 | return false; |
| 303 | } |
| 304 | |
Davidlohr Bueso | 5db6c6f | 2014-07-11 14:00:06 -0700 | [diff] [blame] | 305 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 306 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 307 | * Try to acquire write lock before the writer has been put on wait queue. |
| 308 | */ |
| 309 | static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem) |
| 310 | { |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 311 | long old, count = atomic_long_read(&sem->count); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 312 | |
| 313 | while (true) { |
| 314 | if (!(count == 0 || count == RWSEM_WAITING_BIAS)) |
| 315 | return false; |
| 316 | |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 317 | old = atomic_long_cmpxchg_acquire(&sem->count, count, |
Davidlohr Bueso | 00eb4ba | 2015-09-30 13:03:15 -0700 | [diff] [blame] | 318 | count + RWSEM_ACTIVE_WRITE_BIAS); |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 319 | if (old == count) { |
| 320 | rwsem_set_owner(sem); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 321 | return true; |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 322 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 323 | |
| 324 | count = old; |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem) |
| 329 | { |
| 330 | struct task_struct *owner; |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 331 | bool ret = true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 332 | |
| 333 | if (need_resched()) |
Jason Low | 37e9562 | 2014-07-04 20:49:32 -0700 | [diff] [blame] | 334 | return false; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 335 | |
| 336 | rcu_read_lock(); |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 337 | owner = READ_ONCE(sem->owner); |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 338 | if (!rwsem_owner_is_writer(owner)) { |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 339 | /* |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 340 | * Don't spin if the rwsem is readers owned. |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 341 | */ |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 342 | ret = !rwsem_owner_is_reader(owner); |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 343 | goto done; |
| 344 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 345 | |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 346 | ret = owner->on_cpu; |
| 347 | done: |
| 348 | rcu_read_unlock(); |
| 349 | return ret; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 350 | } |
| 351 | |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 352 | /* |
| 353 | * Return true only if we can still spin on the owner field of the rwsem. |
| 354 | */ |
| 355 | static noinline bool rwsem_spin_on_owner(struct rw_semaphore *sem) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 356 | { |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 357 | struct task_struct *owner = READ_ONCE(sem->owner); |
| 358 | |
| 359 | if (!rwsem_owner_is_writer(owner)) |
| 360 | goto out; |
| 361 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 362 | rcu_read_lock(); |
Jason Low | 9198f6e | 2015-03-06 23:45:31 -0800 | [diff] [blame] | 363 | while (sem->owner == owner) { |
| 364 | /* |
| 365 | * Ensure we emit the owner->on_cpu, dereference _after_ |
| 366 | * checking sem->owner still matches owner, if that fails, |
| 367 | * owner might point to free()d memory, if it still matches, |
| 368 | * the rcu_read_lock() ensures the memory stays valid. |
| 369 | */ |
| 370 | barrier(); |
| 371 | |
| 372 | /* abort spinning when need_resched or owner is not running */ |
| 373 | if (!owner->on_cpu || need_resched()) { |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 374 | rcu_read_unlock(); |
| 375 | return false; |
| 376 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 377 | |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 378 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 379 | } |
| 380 | rcu_read_unlock(); |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 381 | out: |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 382 | /* |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 383 | * If there is a new owner or the owner is not set, we continue |
| 384 | * spinning. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 385 | */ |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 386 | return !rwsem_owner_is_reader(READ_ONCE(sem->owner)); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 387 | } |
| 388 | |
| 389 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 390 | { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 391 | bool taken = false; |
| 392 | |
| 393 | preempt_disable(); |
| 394 | |
| 395 | /* sem->wait_lock should not be held when doing optimistic spinning */ |
| 396 | if (!rwsem_can_spin_on_owner(sem)) |
| 397 | goto done; |
| 398 | |
| 399 | if (!osq_lock(&sem->osq)) |
| 400 | goto done; |
| 401 | |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 402 | /* |
| 403 | * Optimistically spin on the owner field and attempt to acquire the |
| 404 | * lock whenever the owner changes. Spinning will be stopped when: |
| 405 | * 1) the owning writer isn't running; or |
| 406 | * 2) readers own the lock as we can't determine if they are |
| 407 | * actively running or not. |
| 408 | */ |
| 409 | while (rwsem_spin_on_owner(sem)) { |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 410 | /* |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 411 | * Try to acquire the lock |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 412 | */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 413 | if (rwsem_try_write_lock_unqueued(sem)) { |
| 414 | taken = true; |
| 415 | break; |
| 416 | } |
| 417 | |
| 418 | /* |
| 419 | * When there's no owner, we might have preempted between the |
| 420 | * owner acquiring the lock and setting the owner field. If |
| 421 | * we're an RT task that will live-lock because we won't let |
| 422 | * the owner complete. |
| 423 | */ |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 424 | if (!sem->owner && (need_resched() || rt_task(current))) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 425 | break; |
| 426 | |
| 427 | /* |
| 428 | * The cpu_relax() call is a compiler barrier which forces |
| 429 | * everything in this loop to be re-loaded. We don't need |
| 430 | * memory barriers as we'll eventually observe the right |
| 431 | * values at the cost of a few extra spins. |
| 432 | */ |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 433 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 434 | } |
| 435 | osq_unlock(&sem->osq); |
| 436 | done: |
| 437 | preempt_enable(); |
| 438 | return taken; |
| 439 | } |
| 440 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 441 | /* |
| 442 | * Return true if the rwsem has active spinner |
| 443 | */ |
| 444 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 445 | { |
| 446 | return osq_is_locked(&sem->osq); |
| 447 | } |
| 448 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 449 | #else |
| 450 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 451 | { |
| 452 | return false; |
| 453 | } |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 454 | |
| 455 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 456 | { |
| 457 | return false; |
| 458 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 459 | #endif |
| 460 | |
| 461 | /* |
| 462 | * Wait until we successfully acquire the write lock |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 463 | */ |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 464 | static inline struct rw_semaphore * |
| 465 | __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 466 | { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 467 | long count; |
| 468 | bool waiting = true; /* any queued threads before us */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 469 | struct rwsem_waiter waiter; |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 470 | struct rw_semaphore *ret = sem; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 471 | WAKE_Q(wake_q); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 472 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 473 | /* undo write bias from down_write operation, stop active locking */ |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 474 | count = atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS, &sem->count); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 475 | |
| 476 | /* do optimistic spinning and steal lock if possible */ |
| 477 | if (rwsem_optimistic_spin(sem)) |
| 478 | return sem; |
| 479 | |
| 480 | /* |
| 481 | * Optimistic spinning failed, proceed to the slowpath |
| 482 | * and block until we can acquire the sem. |
| 483 | */ |
| 484 | waiter.task = current; |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 485 | waiter.type = RWSEM_WAITING_FOR_WRITE; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 486 | |
| 487 | raw_spin_lock_irq(&sem->wait_lock); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 488 | |
| 489 | /* account for this before adding a new element to the list */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 490 | if (list_empty(&sem->wait_list)) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 491 | waiting = false; |
| 492 | |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 493 | list_add_tail(&waiter.list, &sem->wait_list); |
| 494 | |
| 495 | /* we're now waiting on the lock, but no longer actively locking */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 496 | if (waiting) { |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 497 | count = atomic_long_read(&sem->count); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 498 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 499 | /* |
Andrew Morton | 0cc3d01 | 2014-06-04 20:19:48 +0200 | [diff] [blame] | 500 | * If there were already threads queued before us and there are |
| 501 | * no active writers, the lock must be read owned; so we try to |
| 502 | * wake any read locks that were queued ahead of us. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 503 | */ |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 504 | if (count > RWSEM_WAITING_BIAS) { |
| 505 | WAKE_Q(wake_q); |
| 506 | |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 507 | __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 508 | /* |
| 509 | * The wakeup is normally called _after_ the wait_lock |
| 510 | * is released, but given that we are proactively waking |
| 511 | * readers we can deal with the wake_q overhead as it is |
| 512 | * similar to releasing and taking the wait_lock again |
| 513 | * for attempting rwsem_try_write_lock(). |
| 514 | */ |
| 515 | wake_up_q(&wake_q); |
| 516 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 517 | |
| 518 | } else |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 519 | count = atomic_long_add_return(RWSEM_WAITING_BIAS, &sem->count); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 520 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 521 | /* wait until we successfully acquire the lock */ |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 522 | set_current_state(state); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 523 | while (true) { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 524 | if (rwsem_try_write_lock(count, sem)) |
| 525 | break; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 526 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 527 | |
| 528 | /* Block until there are no active lockers. */ |
| 529 | do { |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 530 | if (signal_pending_state(state, current)) |
| 531 | goto out_nolock; |
| 532 | |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 533 | schedule(); |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 534 | set_current_state(state); |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 535 | } while ((count = atomic_long_read(&sem->count)) & RWSEM_ACTIVE_MASK); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 536 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 537 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 538 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 539 | __set_current_state(TASK_RUNNING); |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 540 | list_del(&waiter.list); |
| 541 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 542 | |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 543 | return ret; |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 544 | |
| 545 | out_nolock: |
| 546 | __set_current_state(TASK_RUNNING); |
| 547 | raw_spin_lock_irq(&sem->wait_lock); |
| 548 | list_del(&waiter.list); |
| 549 | if (list_empty(&sem->wait_list)) |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 550 | atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 551 | else |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 552 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 553 | raw_spin_unlock_irq(&sem->wait_lock); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 554 | wake_up_q(&wake_q); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 555 | |
| 556 | return ERR_PTR(-EINTR); |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 557 | } |
| 558 | |
| 559 | __visible struct rw_semaphore * __sched |
| 560 | rwsem_down_write_failed(struct rw_semaphore *sem) |
| 561 | { |
| 562 | return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 563 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 564 | EXPORT_SYMBOL(rwsem_down_write_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 565 | |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 566 | __visible struct rw_semaphore * __sched |
| 567 | rwsem_down_write_failed_killable(struct rw_semaphore *sem) |
| 568 | { |
| 569 | return __rwsem_down_write_failed_common(sem, TASK_KILLABLE); |
| 570 | } |
| 571 | EXPORT_SYMBOL(rwsem_down_write_failed_killable); |
| 572 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 573 | /* |
| 574 | * handle waking up a waiter on the semaphore |
| 575 | * - up_read/up_write has decremented the active part of count if we come here |
| 576 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 577 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 578 | struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 579 | { |
| 580 | unsigned long flags; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 581 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 582 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 583 | /* |
Prateek Sood | 0cbde6c | 2017-09-07 20:00:58 +0530 | [diff] [blame] | 584 | * __rwsem_down_write_failed_common(sem) |
| 585 | * rwsem_optimistic_spin(sem) |
| 586 | * osq_unlock(sem->osq) |
| 587 | * ... |
| 588 | * atomic_long_add_return(&sem->count) |
| 589 | * |
| 590 | * - VS - |
| 591 | * |
| 592 | * __up_write() |
| 593 | * if (atomic_long_sub_return_release(&sem->count) < 0) |
| 594 | * rwsem_wake(sem) |
| 595 | * osq_is_locked(&sem->osq) |
| 596 | * |
| 597 | * And __up_write() must observe !osq_is_locked() when it observes the |
| 598 | * atomic_long_add_return() in order to not miss a wakeup. |
| 599 | * |
| 600 | * This boils down to: |
| 601 | * |
| 602 | * [S.rel] X = 1 [RmW] r0 = (Y += 0) |
| 603 | * MB RMB |
| 604 | * [RmW] Y += 1 [L] r1 = X |
| 605 | * |
| 606 | * exists (r0=1 /\ r1=0) |
| 607 | */ |
| 608 | smp_rmb(); |
| 609 | |
| 610 | /* |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 611 | * If a spinner is present, it is not necessary to do the wakeup. |
| 612 | * Try to do wakeup only if the trylock succeeds to minimize |
| 613 | * spinlock contention which may introduce too much delay in the |
| 614 | * unlock operation. |
| 615 | * |
| 616 | * spinning writer up_write/up_read caller |
| 617 | * --------------- ----------------------- |
| 618 | * [S] osq_unlock() [L] osq |
| 619 | * MB RMB |
| 620 | * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock) |
| 621 | * |
| 622 | * Here, it is important to make sure that there won't be a missed |
| 623 | * wakeup while the rwsem is free and the only spinning writer goes |
| 624 | * to sleep without taking the rwsem. Even when the spinning writer |
| 625 | * is just going to break out of the waiting loop, it will still do |
| 626 | * a trylock in rwsem_down_write_failed() before sleeping. IOW, if |
| 627 | * rwsem_has_spinner() is true, it will guarantee at least one |
| 628 | * trylock attempt on the rwsem later on. |
| 629 | */ |
| 630 | if (rwsem_has_spinner(sem)) { |
| 631 | /* |
| 632 | * The smp_rmb() here is to make sure that the spinner |
| 633 | * state is consulted before reading the wait_lock. |
| 634 | */ |
| 635 | smp_rmb(); |
| 636 | if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags)) |
| 637 | return sem; |
| 638 | goto locked; |
| 639 | } |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 640 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 641 | locked: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 642 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 643 | if (!list_empty(&sem->wait_list)) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 644 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 645 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 646 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 647 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 648 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 649 | return sem; |
| 650 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 651 | EXPORT_SYMBOL(rwsem_wake); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 652 | |
| 653 | /* |
| 654 | * downgrade a write lock into a read lock |
| 655 | * - caller incremented waiting part of count and discovered it still negative |
| 656 | * - just wake up any readers at the front of the queue |
| 657 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 658 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 659 | struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 660 | { |
| 661 | unsigned long flags; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 662 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 663 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 664 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 665 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 666 | if (!list_empty(&sem->wait_list)) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 667 | __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 668 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 669 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 670 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 671 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 672 | return sem; |
| 673 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 674 | EXPORT_SYMBOL(rwsem_downgrade_wake); |