Christoph Lameter wrote:
On Tue, 28 Mar 2006, Zoltan Menyhart wrote:Why not to use separate bit operations for different purposes? - e.g. "test_and_set_bit_N_acquire()" for lock acquisition - "test_and_set_bit()", "clear_bit()" as they are today - "release_N_clear_bit()"...That would force IA64 specifics onto all other architectures. Could we simply define these smb_mb__*_clear_bit to be noopsand then make the atomic bit ops to have full barriers? That would satisfy Nick's objections.
I do not want to force IA64 specifics... I think the lock - unlock are as they are because of some historical reasons. IA64 only helps me to see what is really necessary for the lock implementation and what is just a heritage of the PCs. I try to summarize my thoughts about the locks in an architecture independent way. If we can agree upon how it should work, then we can enter into the details later :-) 1. lock(): The lock acquisition semantics guarantees that the lock is acquired by its new owner and the new ownership is made globally visible prior to all subsequent orderable instructions (issued by the new owner) becoming visible to the others. Example: | v | store_X ----------+ | | v | | | ======== lock ====+======= | | ^ ^ v v ^ ^ | | | load_Y ---------------+ | store_Z ----------------+ | v | Note that "store_X" is not guaranteed to be globally performed before the lock. 2. unlock(): The lock release semantics guarantees that all previous orderable instructions (issued by the lock owner) are made globally visible prior to make visible to the others the lock becoming free. Example: | v | store_A ----------+ | | v v ^ | v | ======= unlock =======+=== | | v | | | load_B ---------------+ | v | Note that "load_B" can be globally performed before the unlock. The mutual exclusion means: - "store_A" is before unlocking - unlocking is before locking by the next owner - locking is before "load_Y" or "store_Z" therefore "store_A" is before "load_Y" or "store_Z". No relation is defined between "store_X" and any other operation, nor is between "load_B" and any other operation. Can we agree that locking and unlocking mean no more than what I described, especially, they do not imply bidirectional fencing? Should someone need a bidirectional fencing, s/he would have to use some additional shync. methods. In order to separate the theory (how is should work) and the practice (how is should be implemented), I put the bit-op stuff in another mail. Thanks. Zoltan - : send the line "unsubscribe linux-ia64" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
![]() |