[PATCH 3/5] arm: use the spinlocked, generic atomic64 support

[Nicolas Pitre]

On Mon, 14 Dec 2009, Will Deacon wrote:

> Hi Nicolas,
> 
> *Nicolas Pitre wrote:
> 
> > Can't a variant of include/linux/cnt32_to_63.h be used here?> 
> > typedef struct {
> > 	atomic_t low;
> > 	u32 high;
> > } atomic64_t;
> > 
> > static inline void atomic64_set(atomic64_t *ptr, u64 new_val)
> > {
> > 	u32 low = new_val;
> > 	u32 high = new_val >> 32;
> > 	BUG_ON(high & 0x80000000);
> > 	atomic_set(&ptr->low, low);
> > 	ptr->high = (high & 0x7fffffff) | (low & 0x80000000);
> > }
> 
> How do you ensure that this is atomic? To me it looks like one CPU
> could write the lower 32-bits and another could write the upper 32,
> leaving the memory location in an inconsistent state.

This one is indeed not exactly atomic.  It isn't the interesting case 
either though.  Probably a strd could fix that.

> > static inline u64 atomic64_read(atomic64_t *ptr)
> > {
> > 	u32 high, low;
> > 	high = ptr->high;
> > 	smp_rmb();
> > 	low = atomic_read(&ptr->low);
> > 	if (unlikely((s32)(high ^ low) < 0))
> > 		ptr->high = high = (high ^ 0x80000000) + (high >> 31);
> > 	return ((u64)(high & 0x7fffffff) << 32) | low;
> > }
> > 
> > static inline u64 atomic64_inc_return(atomic64_t *ptr)
> > {
> > 	atomic_inc(&ptr->low);
> > 	return atomic64_read(ptr);
> > }
> > 
> > The atomic64_add_return() could be implemented the same way, however the
> > added value would have to be smaller than 31 bits for the algorythm to
> > work.
> 
> I posted a patch to this list on Friday which provides 64-bit atomic
> operations for ARM using exclusive loads and stores:
> 
> http://lists.infradead.org/pipermail/linux-arm-kernel/2009-December/005934.html

For ARMv7 (or ARMv6K), not for ARM.

> Once this patch has been successfully reviewed, these routines should be used
> instead. For now it makes sense to use the generic spinlocks version as a
> placeholder.

On ARMv6 the above is still faster.


Nicolas