Re: [PATCH] WorkStruct: Implement generic UP cmpxchg() where an arch doesn't support it

[Nick Piggin <nickpiggin@yahoo.com.au>]

Russell King wrote:
> On Thu, Dec 07, 2006 at 08:31:08PM +1100, Nick Piggin wrote:

>>>Implementing ll/sc based accessor macros allows both ll/sc _and_ cmpxchg
>>>architectures to produce optimal code.
>>>
>>>Implementing an cmpxchg based accessor macro allows cmpxchg architectures
>>>to produce optimal code and ll/sc non-optimal code.
>>>
>>>See my point?
>>
>>Wrong. Your ll/sc implementation with cmpxchg is buggy. The cmpxchg
>>load_locked is not locked at all,
> 
> 
> Intentional - cmpxchg architectures don't generally have a load locked.

Exactly, so it is wrong -- you can't implement that behaviour with
load + cmpxchg.

>>and there can be interleaving writes
>>between the load and cmpxchg which do not cause the store_conditional
>>to fail.
> 
> 
> In which case the cmpxchg fails and we do the atomic operation again,
> in exactly the same way that we do the operation again if the 'sc'
> fails in the ll/sc case.

Not if cmpxchg sees the same value, it won't fail, regardless of how
many writes have hit that memory address.

> I do not see any problem.

This was not the big problem -- as I said, if this was the only problem
we could opt for a "watered down" version that doesn't actually load
locked [the ll/sc interface would be much cooler than cmpxchg, IMO :)]

The main problem is the restrictions between the ll and sc. This is why
I implemented atomic_cmpxchg rather than atomic_ll/sc. However I agree
that in critical code, a higher level API should be used instead (eg.
see atomic_add_unless, which can be implemented optimally on RISCs).

Nick

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com