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 |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 90 | #ifdef CONFIG_RWSEM_PRIO_AWARE |
| 91 | sem->m_count = 0; |
| 92 | #endif |
Ingo Molnar | 4ea2176 | 2006-07-03 00:24:53 -0700 | [diff] [blame] | 93 | } |
| 94 | |
| 95 | EXPORT_SYMBOL(__init_rwsem); |
| 96 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 97 | enum rwsem_wake_type { |
| 98 | RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */ |
| 99 | RWSEM_WAKE_READERS, /* Wake readers only */ |
| 100 | RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */ |
| 101 | }; |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 102 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | /* |
| 104 | * handle the lock release when processes blocked on it that can now run |
| 105 | * - if we come here from up_xxxx(), then: |
| 106 | * - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed) |
| 107 | * - 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] | 108 | * - there must be someone on the queue |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 109 | * - the wait_lock must be held by the caller |
| 110 | * - tasks are marked for wakeup, the caller must later invoke wake_up_q() |
| 111 | * to actually wakeup the blocked task(s) and drop the reference count, |
| 112 | * preferably when the wait_lock is released |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 113 | * - woken process blocks are discarded from the list after having task zeroed |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 114 | * - writers are only marked woken if downgrading is false |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 115 | */ |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 116 | static void __rwsem_mark_wake(struct rw_semaphore *sem, |
| 117 | enum rwsem_wake_type wake_type, |
| 118 | struct wake_q_head *wake_q) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 119 | { |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 120 | struct rwsem_waiter *waiter, *tmp; |
| 121 | long oldcount, woken = 0, adjustment = 0; |
Waiman Long | e85fab7 | 2019-04-28 17:25:38 -0400 | [diff] [blame] | 122 | struct list_head wlist; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 124 | /* |
| 125 | * Take a peek at the queue head waiter such that we can determine |
| 126 | * the wakeup(s) to perform. |
| 127 | */ |
| 128 | waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list); |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 129 | |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 130 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) { |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 131 | if (wake_type == RWSEM_WAKE_ANY) { |
| 132 | /* |
| 133 | * Mark writer at the front of the queue for wakeup. |
| 134 | * Until the task is actually later awoken later by |
| 135 | * the caller, other writers are able to steal it. |
| 136 | * Readers, on the other hand, will block as they |
| 137 | * will notice the queued writer. |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 138 | */ |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 139 | wake_q_add(wake_q, waiter->task); |
| 140 | } |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 141 | |
| 142 | return; |
Michel Lespinasse | 8cf5322 | 2013-05-07 06:45:58 -0700 | [diff] [blame] | 143 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 144 | |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 145 | /* |
| 146 | * 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] | 147 | * We prefer to do the first reader grant before counting readers |
| 148 | * so we can bail out early if a writer stole the lock. |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 149 | */ |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 150 | if (wake_type != RWSEM_WAKE_READ_OWNED) { |
| 151 | adjustment = RWSEM_ACTIVE_READ_BIAS; |
| 152 | try_reader_grant: |
Peter Zijlstra | 86a3b5f | 2016-05-18 12:42:21 +0200 | [diff] [blame] | 153 | oldcount = atomic_long_fetch_add(adjustment, &sem->count); |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 154 | if (unlikely(oldcount < RWSEM_WAITING_BIAS)) { |
Waiman Long | bf7b4c4 | 2016-05-17 21:26:22 -0400 | [diff] [blame] | 155 | /* |
| 156 | * If the count is still less than RWSEM_WAITING_BIAS |
| 157 | * after removing the adjustment, it is assumed that |
| 158 | * a writer has stolen the lock. We have to undo our |
| 159 | * reader grant. |
| 160 | */ |
| 161 | if (atomic_long_add_return(-adjustment, &sem->count) < |
| 162 | RWSEM_WAITING_BIAS) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 163 | return; |
| 164 | |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 165 | /* Last active locker left. Retry waking readers. */ |
| 166 | goto try_reader_grant; |
| 167 | } |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 168 | /* |
| 169 | * It is not really necessary to set it to reader-owned here, |
| 170 | * but it gives the spinners an early indication that the |
| 171 | * readers now have the lock. |
| 172 | */ |
| 173 | rwsem_set_reader_owned(sem); |
Michel Lespinasse | fe6e674 | 2013-05-07 06:45:59 -0700 | [diff] [blame] | 174 | } |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 175 | |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 176 | /* |
| 177 | * 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] | 178 | * of the queue. We know that woken will be at least 1 as we accounted |
| 179 | * for above. Note we increment the 'active part' of the count by the |
| 180 | * number of readers before waking any processes up. |
Waiman Long | e85fab7 | 2019-04-28 17:25:38 -0400 | [diff] [blame] | 181 | * |
| 182 | * We have to do wakeup in 2 passes to prevent the possibility that |
| 183 | * the reader count may be decremented before it is incremented. It |
| 184 | * is because the to-be-woken waiter may not have slept yet. So it |
| 185 | * may see waiter->task got cleared, finish its critical section and |
| 186 | * do an unlock before the reader count increment. |
| 187 | * |
| 188 | * 1) Collect the read-waiters in a separate list, count them and |
| 189 | * fully increment the reader count in rwsem. |
| 190 | * 2) For each waiters in the new list, clear waiter->task and |
| 191 | * put them into wake_q to be woken up later. |
Michel Lespinasse | 345af7b | 2010-08-09 17:21:15 -0700 | [diff] [blame] | 192 | */ |
Waiman Long | e85fab7 | 2019-04-28 17:25:38 -0400 | [diff] [blame] | 193 | list_for_each_entry(waiter, &sem->wait_list, list) { |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 194 | if (waiter->type == RWSEM_WAITING_FOR_WRITE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 195 | break; |
| 196 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 197 | woken++; |
Waiman Long | e85fab7 | 2019-04-28 17:25:38 -0400 | [diff] [blame] | 198 | } |
| 199 | list_cut_before(&wlist, &sem->wait_list, &waiter->list); |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 200 | |
Waiman Long | e85fab7 | 2019-04-28 17:25:38 -0400 | [diff] [blame] | 201 | adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment; |
| 202 | if (list_empty(&sem->wait_list)) { |
| 203 | /* hit end of list above */ |
| 204 | adjustment -= RWSEM_WAITING_BIAS; |
| 205 | } |
| 206 | |
| 207 | if (adjustment) |
| 208 | atomic_long_add(adjustment, &sem->count); |
| 209 | |
| 210 | /* 2nd pass */ |
| 211 | list_for_each_entry_safe(waiter, tmp, &wlist, list) { |
| 212 | struct task_struct *tsk; |
| 213 | |
| 214 | tsk = waiter->task; |
Xie Yongji | b4df458 | 2018-11-29 20:50:30 +0800 | [diff] [blame] | 215 | get_task_struct(tsk); |
Waiman Long | e85fab7 | 2019-04-28 17:25:38 -0400 | [diff] [blame] | 216 | |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 217 | /* |
Xie Yongji | b4df458 | 2018-11-29 20:50:30 +0800 | [diff] [blame] | 218 | * Ensure calling get_task_struct() before setting the reader |
Davidlohr Bueso | e385139 | 2016-05-13 11:56:27 -0700 | [diff] [blame] | 219 | * waiter to nil such that rwsem_down_read_failed() cannot |
| 220 | * race with do_exit() by always holding a reference count |
| 221 | * to the task to wakeup. |
| 222 | */ |
| 223 | smp_store_release(&waiter->task, NULL); |
Xie Yongji | b4df458 | 2018-11-29 20:50:30 +0800 | [diff] [blame] | 224 | /* |
| 225 | * Ensure issuing the wakeup (either by us or someone else) |
| 226 | * after setting the reader waiter to nil. |
| 227 | */ |
| 228 | wake_q_add(wake_q, tsk); |
| 229 | /* wake_q_add() already take the task ref */ |
| 230 | put_task_struct(tsk); |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 231 | } |
Alex Shi | ce6711f | 2013-02-05 21:11:55 +0800 | [diff] [blame] | 232 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 233 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 234 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 235 | * Wait for the read lock to be granted |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 237 | __visible |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 238 | struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 239 | { |
Davidlohr Bueso | b5f5418 | 2013-05-07 06:46:02 -0700 | [diff] [blame] | 240 | long count, adjustment = -RWSEM_ACTIVE_READ_BIAS; |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 241 | struct rwsem_waiter waiter; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 242 | struct task_struct *tsk = current; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 243 | WAKE_Q(wake_q); |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 244 | bool is_first_waiter = false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 245 | |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 246 | waiter.task = tsk; |
Michel Lespinasse | da16922 | 2013-05-07 06:45:52 -0700 | [diff] [blame] | 247 | waiter.type = RWSEM_WAITING_FOR_READ; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 248 | |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 249 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | fd41b33 | 2010-08-09 17:21:18 -0700 | [diff] [blame] | 250 | if (list_empty(&sem->wait_list)) |
| 251 | adjustment += RWSEM_WAITING_BIAS; |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 252 | |
| 253 | /* is_first_waiter == true means we are first in the queue */ |
| 254 | is_first_waiter = rwsem_list_add_per_prio(&waiter, sem); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 255 | |
Michel Lespinasse | 70bdc6e | 2010-08-09 17:21:17 -0700 | [diff] [blame] | 256 | /* we're now waiting on the lock, but no longer actively locking */ |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 257 | count = atomic_long_add_return(adjustment, &sem->count); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 258 | |
Davidlohr Bueso | 70800c3 | 2016-08-05 01:04:45 -0700 | [diff] [blame] | 259 | /* |
| 260 | * If there are no active locks, wake the front queued process(es). |
Michel Lespinasse | 25c3932 | 2013-05-07 06:46:00 -0700 | [diff] [blame] | 261 | * |
| 262 | * If there are no writers and we are first in the queue, |
| 263 | * wake our own waiter to join the existing active readers ! |
| 264 | */ |
| 265 | if (count == RWSEM_WAITING_BIAS || |
| 266 | (count > RWSEM_WAITING_BIAS && |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 267 | (adjustment != -RWSEM_ACTIVE_READ_BIAS || |
| 268 | is_first_waiter))) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 269 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 270 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 271 | raw_spin_unlock_irq(&sem->wait_lock); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 272 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 273 | |
| 274 | /* wait to be given the lock */ |
Michel Lespinasse | f7dd1ce | 2013-05-07 06:45:50 -0700 | [diff] [blame] | 275 | while (true) { |
| 276 | set_task_state(tsk, TASK_UNINTERRUPTIBLE); |
Michel Lespinasse | a8618a0 | 2010-08-09 17:21:20 -0700 | [diff] [blame] | 277 | if (!waiter.task) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 278 | break; |
| 279 | schedule(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 280 | } |
| 281 | |
Davidlohr Bueso | 7310599 | 2015-01-25 23:36:04 -0800 | [diff] [blame] | 282 | __set_task_state(tsk, TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 283 | return sem; |
| 284 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 285 | EXPORT_SYMBOL(rwsem_down_read_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 286 | |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame] | 287 | /* |
| 288 | * This function must be called with the sem->wait_lock held to prevent |
| 289 | * race conditions between checking the rwsem wait list and setting the |
| 290 | * sem->count accordingly. |
| 291 | */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 292 | static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem) |
| 293 | { |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 294 | /* |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame] | 295 | * 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] | 296 | */ |
Jason Low | c0fcb6c | 2016-05-16 17:38:00 -0700 | [diff] [blame] | 297 | if (count != RWSEM_WAITING_BIAS) |
| 298 | return false; |
| 299 | |
| 300 | /* |
| 301 | * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there |
| 302 | * are other tasks on the wait list, we need to add on WAITING_BIAS. |
| 303 | */ |
| 304 | count = list_is_singular(&sem->wait_list) ? |
| 305 | RWSEM_ACTIVE_WRITE_BIAS : |
| 306 | RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS; |
| 307 | |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 308 | if (atomic_long_cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count) |
| 309 | == RWSEM_WAITING_BIAS) { |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 310 | rwsem_set_owner(sem); |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 311 | return true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 312 | } |
Jason Low | debfab7 | 2014-09-16 17:16:57 -0700 | [diff] [blame] | 313 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 314 | return false; |
| 315 | } |
| 316 | |
Davidlohr Bueso | 5db6c6f | 2014-07-11 14:00:06 -0700 | [diff] [blame] | 317 | #ifdef CONFIG_RWSEM_SPIN_ON_OWNER |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 318 | /* |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 319 | * Try to acquire write lock before the writer has been put on wait queue. |
| 320 | */ |
| 321 | static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem) |
| 322 | { |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 323 | long old, count = atomic_long_read(&sem->count); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 324 | |
| 325 | while (true) { |
| 326 | if (!(count == 0 || count == RWSEM_WAITING_BIAS)) |
| 327 | return false; |
| 328 | |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 329 | old = atomic_long_cmpxchg_acquire(&sem->count, count, |
Davidlohr Bueso | 00eb4ba | 2015-09-30 13:03:15 -0700 | [diff] [blame] | 330 | count + RWSEM_ACTIVE_WRITE_BIAS); |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 331 | if (old == count) { |
| 332 | rwsem_set_owner(sem); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 333 | return true; |
Davidlohr Bueso | 7a215f8 | 2015-01-30 01:14:25 -0800 | [diff] [blame] | 334 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 335 | |
| 336 | count = old; |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem) |
| 341 | { |
| 342 | struct task_struct *owner; |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 343 | bool ret = true; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 344 | |
| 345 | if (need_resched()) |
Jason Low | 37e9562 | 2014-07-04 20:49:32 -0700 | [diff] [blame] | 346 | return false; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 347 | |
| 348 | rcu_read_lock(); |
Davidlohr Bueso | 4d3199e | 2015-02-22 19:31:41 -0800 | [diff] [blame] | 349 | owner = READ_ONCE(sem->owner); |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 350 | if (!rwsem_owner_is_writer(owner)) { |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 351 | /* |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 352 | * Don't spin if the rwsem is readers owned. |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 353 | */ |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 354 | ret = !rwsem_owner_is_reader(owner); |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 355 | goto done; |
| 356 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 357 | |
Davidlohr Bueso | 1a99367 | 2015-01-30 01:14:27 -0800 | [diff] [blame] | 358 | ret = owner->on_cpu; |
| 359 | done: |
| 360 | rcu_read_unlock(); |
| 361 | return ret; |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 362 | } |
| 363 | |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 364 | /* |
| 365 | * Return true only if we can still spin on the owner field of the rwsem. |
| 366 | */ |
| 367 | static noinline bool rwsem_spin_on_owner(struct rw_semaphore *sem) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 368 | { |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 369 | struct task_struct *owner = READ_ONCE(sem->owner); |
| 370 | |
| 371 | if (!rwsem_owner_is_writer(owner)) |
| 372 | goto out; |
| 373 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 374 | rcu_read_lock(); |
Jason Low | 9198f6e | 2015-03-06 23:45:31 -0800 | [diff] [blame] | 375 | while (sem->owner == owner) { |
| 376 | /* |
| 377 | * Ensure we emit the owner->on_cpu, dereference _after_ |
| 378 | * checking sem->owner still matches owner, if that fails, |
| 379 | * owner might point to free()d memory, if it still matches, |
| 380 | * the rcu_read_lock() ensures the memory stays valid. |
| 381 | */ |
| 382 | barrier(); |
| 383 | |
| 384 | /* abort spinning when need_resched or owner is not running */ |
| 385 | if (!owner->on_cpu || need_resched()) { |
Davidlohr Bueso | b3fd4f0 | 2015-01-30 01:14:26 -0800 | [diff] [blame] | 386 | rcu_read_unlock(); |
| 387 | return false; |
| 388 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 389 | |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 390 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 391 | } |
| 392 | rcu_read_unlock(); |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 393 | out: |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 394 | /* |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 395 | * If there is a new owner or the owner is not set, we continue |
| 396 | * spinning. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 397 | */ |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 398 | return !rwsem_owner_is_reader(READ_ONCE(sem->owner)); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 399 | } |
| 400 | |
| 401 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 402 | { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 403 | bool taken = false; |
| 404 | |
| 405 | preempt_disable(); |
| 406 | |
| 407 | /* sem->wait_lock should not be held when doing optimistic spinning */ |
| 408 | if (!rwsem_can_spin_on_owner(sem)) |
| 409 | goto done; |
| 410 | |
| 411 | if (!osq_lock(&sem->osq)) |
| 412 | goto done; |
| 413 | |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 414 | /* |
| 415 | * Optimistically spin on the owner field and attempt to acquire the |
| 416 | * lock whenever the owner changes. Spinning will be stopped when: |
| 417 | * 1) the owning writer isn't running; or |
| 418 | * 2) readers own the lock as we can't determine if they are |
| 419 | * actively running or not. |
| 420 | */ |
| 421 | while (rwsem_spin_on_owner(sem)) { |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 422 | /* |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 423 | * Try to acquire the lock |
Waiman Long | 19c5d69 | 2016-05-17 21:26:19 -0400 | [diff] [blame] | 424 | */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 425 | if (rwsem_try_write_lock_unqueued(sem)) { |
| 426 | taken = true; |
| 427 | break; |
| 428 | } |
| 429 | |
| 430 | /* |
| 431 | * When there's no owner, we might have preempted between the |
| 432 | * owner acquiring the lock and setting the owner field. If |
| 433 | * we're an RT task that will live-lock because we won't let |
| 434 | * the owner complete. |
| 435 | */ |
Waiman Long | ddd0fa7 | 2016-05-17 21:26:23 -0400 | [diff] [blame] | 436 | if (!sem->owner && (need_resched() || rt_task(current))) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 437 | break; |
| 438 | |
| 439 | /* |
| 440 | * The cpu_relax() call is a compiler barrier which forces |
| 441 | * everything in this loop to be re-loaded. We don't need |
| 442 | * memory barriers as we'll eventually observe the right |
| 443 | * values at the cost of a few extra spins. |
| 444 | */ |
Davidlohr Bueso | 3a6bfbc | 2014-06-29 15:09:33 -0700 | [diff] [blame] | 445 | cpu_relax_lowlatency(); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 446 | } |
| 447 | osq_unlock(&sem->osq); |
| 448 | done: |
| 449 | preempt_enable(); |
| 450 | return taken; |
| 451 | } |
| 452 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 453 | /* |
| 454 | * Return true if the rwsem has active spinner |
| 455 | */ |
| 456 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 457 | { |
| 458 | return osq_is_locked(&sem->osq); |
| 459 | } |
| 460 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 461 | #else |
| 462 | static bool rwsem_optimistic_spin(struct rw_semaphore *sem) |
| 463 | { |
| 464 | return false; |
| 465 | } |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 466 | |
| 467 | static inline bool rwsem_has_spinner(struct rw_semaphore *sem) |
| 468 | { |
| 469 | return false; |
| 470 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 471 | #endif |
| 472 | |
| 473 | /* |
| 474 | * Wait until we successfully acquire the write lock |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 475 | */ |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 476 | static inline struct rw_semaphore * |
| 477 | __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 478 | { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 479 | long count; |
| 480 | bool waiting = true; /* any queued threads before us */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 481 | struct rwsem_waiter waiter; |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 482 | struct rw_semaphore *ret = sem; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 483 | WAKE_Q(wake_q); |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 484 | bool is_first_waiter = false; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 485 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 486 | /* undo write bias from down_write operation, stop active locking */ |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 487 | count = atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS, &sem->count); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 488 | |
| 489 | /* do optimistic spinning and steal lock if possible */ |
| 490 | if (rwsem_optimistic_spin(sem)) |
| 491 | return sem; |
| 492 | |
| 493 | /* |
| 494 | * Optimistic spinning failed, proceed to the slowpath |
| 495 | * and block until we can acquire the sem. |
| 496 | */ |
| 497 | waiter.task = current; |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 498 | waiter.type = RWSEM_WAITING_FOR_WRITE; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 499 | |
| 500 | raw_spin_lock_irq(&sem->wait_lock); |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 501 | |
| 502 | /* account for this before adding a new element to the list */ |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 503 | if (list_empty(&sem->wait_list)) |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 504 | waiting = false; |
| 505 | |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 506 | /* |
| 507 | * is_first_waiter == true means we are first in the queue, |
| 508 | * so there is no read locks that were queued ahead of us. |
| 509 | */ |
| 510 | is_first_waiter = rwsem_list_add_per_prio(&waiter, sem); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 511 | |
| 512 | /* we're now waiting on the lock, but no longer actively locking */ |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 513 | if (waiting) { |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 514 | count = atomic_long_read(&sem->count); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 515 | |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 516 | /* |
Andrew Morton | 0cc3d01 | 2014-06-04 20:19:48 +0200 | [diff] [blame] | 517 | * If there were already threads queued before us and there are |
| 518 | * no active writers, the lock must be read owned; so we try to |
| 519 | * wake any read locks that were queued ahead of us. |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 520 | */ |
Maria Yu | a5946df | 2018-11-29 13:57:53 +0800 | [diff] [blame] | 521 | if (!is_first_waiter && count > RWSEM_WAITING_BIAS) { |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 522 | WAKE_Q(wake_q); |
| 523 | |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 524 | __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 525 | /* |
| 526 | * The wakeup is normally called _after_ the wait_lock |
| 527 | * is released, but given that we are proactively waking |
| 528 | * readers we can deal with the wake_q overhead as it is |
| 529 | * similar to releasing and taking the wait_lock again |
| 530 | * for attempting rwsem_try_write_lock(). |
| 531 | */ |
| 532 | wake_up_q(&wake_q); |
| 533 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 534 | |
| 535 | } else |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 536 | count = atomic_long_add_return(RWSEM_WAITING_BIAS, &sem->count); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 537 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 538 | /* wait until we successfully acquire the lock */ |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 539 | set_current_state(state); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 540 | while (true) { |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 541 | if (rwsem_try_write_lock(count, sem)) |
| 542 | break; |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 543 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 544 | |
| 545 | /* Block until there are no active lockers. */ |
| 546 | do { |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 547 | if (signal_pending_state(state, current)) |
| 548 | goto out_nolock; |
| 549 | |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 550 | schedule(); |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 551 | set_current_state(state); |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 552 | } while ((count = atomic_long_read(&sem->count)) & RWSEM_ACTIVE_MASK); |
Michel Lespinasse | a7d2c57 | 2013-05-07 06:45:56 -0700 | [diff] [blame] | 553 | |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 554 | raw_spin_lock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 555 | } |
Davidlohr Bueso | 4fc828e | 2014-05-02 11:24:15 -0700 | [diff] [blame] | 556 | __set_current_state(TASK_RUNNING); |
Michel Lespinasse | 023fe4f | 2013-05-07 06:45:53 -0700 | [diff] [blame] | 557 | list_del(&waiter.list); |
| 558 | raw_spin_unlock_irq(&sem->wait_lock); |
Michel Lespinasse | 1e78277 | 2013-05-07 06:45:51 -0700 | [diff] [blame] | 559 | |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 560 | return ret; |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 561 | |
| 562 | out_nolock: |
| 563 | __set_current_state(TASK_RUNNING); |
| 564 | raw_spin_lock_irq(&sem->wait_lock); |
| 565 | list_del(&waiter.list); |
| 566 | if (list_empty(&sem->wait_list)) |
Jason Low | 8ee62b1 | 2016-06-03 22:26:02 -0700 | [diff] [blame] | 567 | atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 568 | else |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 569 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 570 | raw_spin_unlock_irq(&sem->wait_lock); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 571 | wake_up_q(&wake_q); |
Peter Zijlstra | 04cafed | 2016-05-12 13:57:45 +0200 | [diff] [blame] | 572 | |
| 573 | return ERR_PTR(-EINTR); |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 574 | } |
| 575 | |
| 576 | __visible struct rw_semaphore * __sched |
| 577 | rwsem_down_write_failed(struct rw_semaphore *sem) |
| 578 | { |
| 579 | return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 580 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 581 | EXPORT_SYMBOL(rwsem_down_write_failed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 582 | |
Michal Hocko | d479960 | 2016-04-07 17:12:26 +0200 | [diff] [blame] | 583 | __visible struct rw_semaphore * __sched |
| 584 | rwsem_down_write_failed_killable(struct rw_semaphore *sem) |
| 585 | { |
| 586 | return __rwsem_down_write_failed_common(sem, TASK_KILLABLE); |
| 587 | } |
| 588 | EXPORT_SYMBOL(rwsem_down_write_failed_killable); |
| 589 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 590 | /* |
| 591 | * handle waking up a waiter on the semaphore |
| 592 | * - up_read/up_write has decremented the active part of count if we come here |
| 593 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 594 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 595 | struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 596 | { |
| 597 | unsigned long flags; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 598 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 599 | |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 600 | /* |
Prateek Sood | 67a288e | 2017-09-07 20:00:58 +0530 | [diff] [blame] | 601 | * __rwsem_down_write_failed_common(sem) |
| 602 | * rwsem_optimistic_spin(sem) |
| 603 | * osq_unlock(sem->osq) |
| 604 | * ... |
| 605 | * atomic_long_add_return(&sem->count) |
| 606 | * |
| 607 | * - VS - |
| 608 | * |
| 609 | * __up_write() |
| 610 | * if (atomic_long_sub_return_release(&sem->count) < 0) |
| 611 | * rwsem_wake(sem) |
| 612 | * osq_is_locked(&sem->osq) |
| 613 | * |
| 614 | * And __up_write() must observe !osq_is_locked() when it observes the |
| 615 | * atomic_long_add_return() in order to not miss a wakeup. |
| 616 | * |
| 617 | * This boils down to: |
| 618 | * |
| 619 | * [S.rel] X = 1 [RmW] r0 = (Y += 0) |
| 620 | * MB RMB |
| 621 | * [RmW] Y += 1 [L] r1 = X |
| 622 | * |
| 623 | * exists (r0=1 /\ r1=0) |
| 624 | */ |
| 625 | smp_rmb(); |
| 626 | |
| 627 | /* |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 628 | * If a spinner is present, it is not necessary to do the wakeup. |
| 629 | * Try to do wakeup only if the trylock succeeds to minimize |
| 630 | * spinlock contention which may introduce too much delay in the |
| 631 | * unlock operation. |
| 632 | * |
| 633 | * spinning writer up_write/up_read caller |
| 634 | * --------------- ----------------------- |
| 635 | * [S] osq_unlock() [L] osq |
| 636 | * MB RMB |
| 637 | * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock) |
| 638 | * |
| 639 | * Here, it is important to make sure that there won't be a missed |
| 640 | * wakeup while the rwsem is free and the only spinning writer goes |
| 641 | * to sleep without taking the rwsem. Even when the spinning writer |
| 642 | * is just going to break out of the waiting loop, it will still do |
| 643 | * a trylock in rwsem_down_write_failed() before sleeping. IOW, if |
| 644 | * rwsem_has_spinner() is true, it will guarantee at least one |
| 645 | * trylock attempt on the rwsem later on. |
| 646 | */ |
| 647 | if (rwsem_has_spinner(sem)) { |
| 648 | /* |
| 649 | * The smp_rmb() here is to make sure that the spinner |
| 650 | * state is consulted before reading the wait_lock. |
| 651 | */ |
| 652 | smp_rmb(); |
| 653 | if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags)) |
| 654 | return sem; |
| 655 | goto locked; |
| 656 | } |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 657 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Waiman Long | 59aabfc | 2015-04-30 17:12:16 -0400 | [diff] [blame] | 658 | locked: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 659 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 660 | if (!list_empty(&sem->wait_list)) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 661 | __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 662 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 663 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 664 | wake_up_q(&wake_q); |
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 | return sem; |
| 667 | } |
Davidlohr Bueso | db0e716 | 2014-09-11 22:34:25 -0700 | [diff] [blame] | 668 | EXPORT_SYMBOL(rwsem_wake); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 669 | |
| 670 | /* |
| 671 | * downgrade a write lock into a read lock |
| 672 | * - caller incremented waiting part of count and discovered it still negative |
| 673 | * - just wake up any readers at the front of the queue |
| 674 | */ |
Andi Kleen | 3ebae4f | 2014-02-08 08:52:05 +0100 | [diff] [blame] | 675 | __visible |
Thomas Gleixner | d123375 | 2011-01-26 21:32:01 +0100 | [diff] [blame] | 676 | struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 677 | { |
| 678 | unsigned long flags; |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 679 | WAKE_Q(wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 680 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 681 | raw_spin_lock_irqsave(&sem->wait_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 682 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 683 | if (!list_empty(&sem->wait_list)) |
Davidlohr Bueso | 84b23f9 | 2016-08-05 01:04:43 -0700 | [diff] [blame] | 684 | __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 685 | |
Thomas Gleixner | ddb6c9b | 2010-02-24 09:54:54 +0100 | [diff] [blame] | 686 | raw_spin_unlock_irqrestore(&sem->wait_lock, flags); |
Davidlohr Bueso | 133e89e | 2016-05-13 11:56:26 -0700 | [diff] [blame] | 687 | wake_up_q(&wake_q); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 688 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 689 | return sem; |
| 690 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 691 | EXPORT_SYMBOL(rwsem_downgrade_wake); |