Arjan van de Ven | 2af7f59 | 2006-01-09 15:59:18 -0800 | [diff] [blame] | 1 | /* |
| 2 | * Assembly implementation of the mutex fastpath, based on atomic |
| 3 | * decrement/increment. |
| 4 | * |
| 5 | * started by Ingo Molnar: |
| 6 | * |
| 7 | * Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> |
| 8 | */ |
| 9 | #ifndef _ASM_MUTEX_H |
| 10 | #define _ASM_MUTEX_H |
| 11 | |
| 12 | /** |
| 13 | * __mutex_fastpath_lock - try to take the lock by moving the count |
| 14 | * from 1 to a 0 value |
| 15 | * @count: pointer of type atomic_t |
| 16 | * @fn: function to call if the original value was not 1 |
| 17 | * |
| 18 | * Change the count from 1 to a value lower than 1, and call <fn> if it |
| 19 | * wasn't 1 originally. This function MUST leave the value lower than 1 |
| 20 | * even when the "1" assertion wasn't true. |
| 21 | */ |
| 22 | #define __mutex_fastpath_lock(count, fail_fn) \ |
| 23 | do { \ |
| 24 | unsigned int dummy; \ |
| 25 | \ |
| 26 | typecheck(atomic_t *, count); \ |
| 27 | typecheck_fn(fastcall void (*)(atomic_t *), fail_fn); \ |
| 28 | \ |
| 29 | __asm__ __volatile__( \ |
| 30 | LOCK " decl (%%eax) \n" \ |
Ingo Molnar | 73165b8 | 2006-01-10 22:07:44 +0100 | [diff] [blame] | 31 | " js 2f \n" \ |
| 32 | "1: \n" \ |
| 33 | \ |
| 34 | LOCK_SECTION_START("") \ |
| 35 | "2: call "#fail_fn" \n" \ |
| 36 | " jmp 1b \n" \ |
| 37 | LOCK_SECTION_END \ |
Arjan van de Ven | 2af7f59 | 2006-01-09 15:59:18 -0800 | [diff] [blame] | 38 | \ |
| 39 | :"=a" (dummy) \ |
| 40 | : "a" (count) \ |
| 41 | : "memory", "ecx", "edx"); \ |
| 42 | } while (0) |
| 43 | |
| 44 | |
| 45 | /** |
| 46 | * __mutex_fastpath_lock_retval - try to take the lock by moving the count |
| 47 | * from 1 to a 0 value |
| 48 | * @count: pointer of type atomic_t |
| 49 | * @fail_fn: function to call if the original value was not 1 |
| 50 | * |
| 51 | * Change the count from 1 to a value lower than 1, and call <fail_fn> if it |
| 52 | * wasn't 1 originally. This function returns 0 if the fastpath succeeds, |
| 53 | * or anything the slow path function returns |
| 54 | */ |
| 55 | static inline int |
| 56 | __mutex_fastpath_lock_retval(atomic_t *count, |
| 57 | int fastcall (*fail_fn)(atomic_t *)) |
| 58 | { |
| 59 | if (unlikely(atomic_dec_return(count) < 0)) |
| 60 | return fail_fn(count); |
| 61 | else |
| 62 | return 0; |
| 63 | } |
| 64 | |
| 65 | /** |
| 66 | * __mutex_fastpath_unlock - try to promote the mutex from 0 to 1 |
| 67 | * @count: pointer of type atomic_t |
| 68 | * @fail_fn: function to call if the original value was not 0 |
| 69 | * |
| 70 | * try to promote the mutex from 0 to 1. if it wasn't 0, call <fail_fn>. |
| 71 | * In the failure case, this function is allowed to either set the value |
| 72 | * to 1, or to set it to a value lower than 1. |
| 73 | * |
| 74 | * If the implementation sets it to a value of lower than 1, the |
| 75 | * __mutex_slowpath_needs_to_unlock() macro needs to return 1, it needs |
| 76 | * to return 0 otherwise. |
| 77 | */ |
| 78 | #define __mutex_fastpath_unlock(count, fail_fn) \ |
| 79 | do { \ |
| 80 | unsigned int dummy; \ |
| 81 | \ |
| 82 | typecheck(atomic_t *, count); \ |
| 83 | typecheck_fn(fastcall void (*)(atomic_t *), fail_fn); \ |
| 84 | \ |
| 85 | __asm__ __volatile__( \ |
| 86 | LOCK " incl (%%eax) \n" \ |
Ingo Molnar | 73165b8 | 2006-01-10 22:07:44 +0100 | [diff] [blame] | 87 | " jle 2f \n" \ |
| 88 | "1: \n" \ |
| 89 | \ |
| 90 | LOCK_SECTION_START("") \ |
| 91 | "2: call "#fail_fn" \n" \ |
| 92 | " jmp 1b \n" \ |
| 93 | LOCK_SECTION_END \ |
Arjan van de Ven | 2af7f59 | 2006-01-09 15:59:18 -0800 | [diff] [blame] | 94 | \ |
| 95 | :"=a" (dummy) \ |
| 96 | : "a" (count) \ |
| 97 | : "memory", "ecx", "edx"); \ |
| 98 | } while (0) |
| 99 | |
| 100 | #define __mutex_slowpath_needs_to_unlock() 1 |
| 101 | |
| 102 | /** |
| 103 | * __mutex_fastpath_trylock - try to acquire the mutex, without waiting |
| 104 | * |
| 105 | * @count: pointer of type atomic_t |
| 106 | * @fail_fn: fallback function |
| 107 | * |
| 108 | * Change the count from 1 to a value lower than 1, and return 0 (failure) |
| 109 | * if it wasn't 1 originally, or return 1 (success) otherwise. This function |
| 110 | * MUST leave the value lower than 1 even when the "1" assertion wasn't true. |
| 111 | * Additionally, if the value was < 0 originally, this function must not leave |
| 112 | * it to 0 on failure. |
| 113 | */ |
| 114 | static inline int |
| 115 | __mutex_fastpath_trylock(atomic_t *count, int (*fail_fn)(atomic_t *)) |
| 116 | { |
| 117 | /* |
| 118 | * We have two variants here. The cmpxchg based one is the best one |
| 119 | * because it never induce a false contention state. It is included |
| 120 | * here because architectures using the inc/dec algorithms over the |
| 121 | * xchg ones are much more likely to support cmpxchg natively. |
| 122 | * |
| 123 | * If not we fall back to the spinlock based variant - that is |
| 124 | * just as efficient (and simpler) as a 'destructive' probing of |
| 125 | * the mutex state would be. |
| 126 | */ |
| 127 | #ifdef __HAVE_ARCH_CMPXCHG |
Linus Torvalds | 4cec873 | 2006-01-11 15:50:47 -0800 | [diff] [blame] | 128 | if (likely(atomic_cmpxchg(count, 1, 0) == 1)) |
Arjan van de Ven | 2af7f59 | 2006-01-09 15:59:18 -0800 | [diff] [blame] | 129 | return 1; |
| 130 | return 0; |
| 131 | #else |
| 132 | return fail_fn(count); |
| 133 | #endif |
| 134 | } |
| 135 | |
| 136 | #endif |