Re: [patch 3/3] use kref for bio

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

Al Viro wrote:
> On Wed, Apr 26, 2006 at 03:20:30PM +0800, Akinobu Mita wrote:
> 
>>If this is good one and the places where Al Viro pointed out really affect
>>performance, should we propagate this faster one by introducing helper
>>function like:
>>
>>static inline int refcount_test(atomic_t *refcount)
>>{
>>	return (atomic_read(refcount) == 1) || (atomic_dec_and_test(refcount));
>>}
>>
>>and replace atomic_dec_and_test with it?
> 
> 
> No.  It's obviously slower than atomic_dec_and_test() if refcount is
> greater than 1.  And I'm less than sure that you can show that benefits
> in case when it is 1 outweight that.

Especially with the indirect function call. Modern CPUs often won't load
the destructor pointer quickly enough to avoid the pipeline bubble.

>  Moreover, for dentries, inodes,
> superblocks and vfsmounts you'd have to pull spin_lock() in front of
> it, which would _definitely_ hurt (these are atomic_dec_and_lock()).

Also, it results in altered memory barrier semantics. Whether or not
this is actually an issue anywhere, any conversion would have to be
careful. If a memory barrier is required _anywhere_, it is likely to
be required on the final put.

With all those arguments against it, you need to demonstrate
improvements before it can be considered.

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