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

[Al Viro <viro@ftp.linux.org.uk>]

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.  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()).