Arnd, > Actually that still feels wrong: the annotation of the struct is to > pack every member, which causes the access to be done in byte units on > architectures that do not have hardware unaligned load/store > instructions, at least for things like atomic_read() that does not go > through a cmpxchg() or ll/sc cycle. > This change may fix itanium, but it's still not correct. Other > architectures would have already been broken before the recent change, > but that's not a reason against fixing them now. I agree. I understand why there are restrictions on fields consumed by the hardware. But for fields internal to the driver the packing doesn't make sense to me. -- Martin K. Petersen Oracle Linux Engineering
