Re: [PATCH 2.6.13] lockless pagecache 2/7

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

Alan Cox wrote:
> On Sul, 2005-09-04 at 11:01 +1000, Nick Piggin wrote:
> 
>>I would be surprised if it was a big loss... but I'm assuming
>>a locked cmpxchg isn't outlandishly expensive. Basically:
>>
>>   read_lock_irqsave(cacheline1);
>>   atomic_inc_return(cacheline2);
>>   read_unlock_irqrestore(cacheline1);
>>
>>Turns into
>>
>>   atomic_cmpxchg();
>>
>>I'll do some microbenchmarks and get back to you. I'm quite
>>interested now ;) What sort of AMDs did you have in mind,
> 
> 
> 
> Athlon or higher give very different atomic numbers to P4. If you are
> losing the read_lock/unlock then the atomic_cmpxchg should be faster on
> all I agree.
> 

Phew! I'll test them anyway, however.

> One question however - atomic_foo operations are not store barriers so
> you might need mb() and friends for PPC ?
> 

Dave's documented that nicely in Documentation/atomic_ops.txt

In general, atomic ops that do not return a value are not barriers,
while operations that do return a value are.

So I think we can define the atomic_cmpxchg as providing a barrier.

Thanks,
Nick

-- 
SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com