Re: [RFC][PATCH 2.4.17] Your suggestions for fast path walk

[Paul Menage]

>
>	Thank you for taking the time to look over my patch and sending
>your comments. In response, Yes, there is a reason why I can't implement 
>path_lookup by calling path_init. path_init calls dget which increments 
>the d_count. The function I wrote holds the dcache_lock instead of 
>incrementing d_count at that point.

OK, I see what you're doing now.

One obvious problem with it is that __emul_lookup_dentry()[1] calls
path_walk() internally, and the nd passed has uncounted references and
no LOOKUP_LOCKED flag - I suspect that this will cause reference counts
to get mucked up.

Also:

1) you really need to fix the patch whitespace/formatting

2) Please again consider calling it path_lookup() rather than 
path_init_walk() - the fact that since 2.3.x programmers have had to be 
aware of the separate path_init()/path_walk() stages is an 
implementation wrinkle that it would be nice to get rid of.

Paul

[1] Actually, __emul_lookup_dentry() probably ought to be simplified
somewhat - it seems to be duplicating a fair chunk of code from
path_init().