Peter Zijlstra | 706eeb3 | 2017-06-12 14:50:27 +0200 | [diff] [blame] | 1 | |
| 2 | On atomic bitops. |
| 3 | |
| 4 | |
| 5 | While our bitmap_{}() functions are non-atomic, we have a number of operations |
| 6 | operating on single bits in a bitmap that are atomic. |
| 7 | |
| 8 | |
| 9 | API |
| 10 | --- |
| 11 | |
| 12 | The single bit operations are: |
| 13 | |
| 14 | Non-RMW ops: |
| 15 | |
| 16 | test_bit() |
| 17 | |
| 18 | RMW atomic operations without return value: |
| 19 | |
| 20 | {set,clear,change}_bit() |
| 21 | clear_bit_unlock() |
| 22 | |
| 23 | RMW atomic operations with return value: |
| 24 | |
| 25 | test_and_{set,clear,change}_bit() |
| 26 | test_and_set_bit_lock() |
| 27 | |
| 28 | Barriers: |
| 29 | |
| 30 | smp_mb__{before,after}_atomic() |
| 31 | |
| 32 | |
| 33 | All RMW atomic operations have a '__' prefixed variant which is non-atomic. |
| 34 | |
| 35 | |
| 36 | SEMANTICS |
| 37 | --------- |
| 38 | |
| 39 | Non-atomic ops: |
| 40 | |
| 41 | In particular __clear_bit_unlock() suffers the same issue as atomic_set(), |
| 42 | which is why the generic version maps to clear_bit_unlock(), see atomic_t.txt. |
| 43 | |
| 44 | |
| 45 | RMW ops: |
| 46 | |
| 47 | The test_and_{}_bit() operations return the original value of the bit. |
| 48 | |
| 49 | |
| 50 | ORDERING |
| 51 | -------- |
| 52 | |
| 53 | Like with atomic_t, the rule of thumb is: |
| 54 | |
| 55 | - non-RMW operations are unordered; |
| 56 | |
| 57 | - RMW operations that have no return value are unordered; |
| 58 | |
| 59 | - RMW operations that have a return value are fully ordered. |
| 60 | |
| 61 | Except for test_and_set_bit_lock() which has ACQUIRE semantics and |
| 62 | clear_bit_unlock() which has RELEASE semantics. |
| 63 | |
| 64 | Since a platform only has a single means of achieving atomic operations |
| 65 | the same barriers as for atomic_t are used, see atomic_t.txt. |
| 66 | |