|
41 | 41 | * |
42 | 42 | *int TAS_SPIN(slock_t *lock) |
43 | 43 | *Like TAS(), but this version is used when waiting for a lock |
44 | | - *previously found to be contended.Typically, this is the |
| 44 | + *previously found to be contended.By default, this is the |
45 | 45 | *same as TAS(), but on some architectures it's better to poll a |
46 | 46 | *contended lock using an unlocked instruction and retry the |
47 | 47 | *atomic test-and-set only when it appears free. |
|
54 | 54 | *on Alpha TAS() will "fail" if interrupted. Therefore a retry loop must |
55 | 55 | *always be used, even if you are certain the lock is free. |
56 | 56 | * |
57 | | - *ANOTHER CAUTION: be sure that TAS(), TAS_SPIN(), and S_UNLOCK() represent |
58 | | - *sequence points, ie, loads and stores of other values must not be moved |
59 | | - *across a lock or unlock. In most cases it suffices to make the operation |
60 | | - *be done through a "volatile" pointer. |
| 57 | + *Another caution for users of these macros is that it is the caller's |
| 58 | + *responsibility to ensure that the compiler doesn't re-order accesses |
| 59 | + *to shared memory to precede the actual lock acquisition, or follow the |
| 60 | + *lock release. Typically we handle this by using volatile-qualified |
| 61 | + *pointers to refer to both the spinlock itself and the shared data |
| 62 | + *structure being accessed within the spinlocked critical section. |
| 63 | + *That fixes it because compilers are not allowed to re-order accesses |
| 64 | + *to volatile objects relative to other such accesses. |
| 65 | + * |
| 66 | + *On platforms with weak memory ordering, the TAS(), TAS_SPIN(), and |
| 67 | + *S_UNLOCK() macros must further include hardware-level memory fence |
| 68 | + *instructions to prevent similar re-ordering at the hardware level. |
| 69 | + *TAS() and TAS_SPIN() must guarantee that loads and stores issued after |
| 70 | + *the macro are not executed until the lock has been obtained. Conversely, |
| 71 | + *S_UNLOCK() must guarantee that loads and stores issued before the macro |
| 72 | + *have been executed before the lock is released. |
61 | 73 | * |
62 | 74 | *On most supported platforms, TAS() uses a tas() function written |
63 | 75 | *in assembly language to execute a hardware atomic-test-and-set |
@@ -229,6 +241,9 @@ typedef unsigned int slock_t; |
229 | 241 |
|
230 | 242 | #defineTAS(lock) tas(lock) |
231 | 243 |
|
| 244 | +/* On IA64, it's a win to use a non-locking test before the xchg proper */ |
| 245 | +#defineTAS_SPIN(lock)(*(lock) ? 1 : TAS(lock)) |
| 246 | + |
232 | 247 | #ifndef__INTEL_COMPILER |
233 | 248 |
|
234 | 249 | static __inline__int |
@@ -735,7 +750,8 @@ typedef unsigned int slock_t; |
735 | 750 |
|
736 | 751 | #include<ia64/sys/inline.h> |
737 | 752 | #defineTAS(lock) _Asm_xchg(_SZ_W, lock, 1, _LDHINT_NONE) |
738 | | -#defineTAS_SPIN(lock) (*(lock) ? 1 : TAS(lock)) |
| 753 | +/* On IA64, it's a win to use a non-locking test before the xchg proper */ |
| 754 | +#defineTAS_SPIN(lock)(*(lock) ? 1 : TAS(lock)) |
739 | 755 |
|
740 | 756 | #endif/* HPUX on IA64, non gcc */ |
741 | 757 |
|
|