[OpenAFS] Linux client and find command in AFS

[Jeffrey Hutzelman]

On Thursday, April 07, 2005 09:48:20 AM -0400 Jim Rees <rees@umich.edu> 
wrote:

> A bit off the subject, but the way to really speed up "find" is to store a
> flag in the directory that says whether each leaf is a file or directory.
> That way you avoid stat on regular files if there are no other predicates
> that require it.
>
> This can be done in afs by cheating, because the viceid is always even for
> files and odd for directories (or maybe it's the other way around).

Except that to AFS, mount points are files, not directories, and they have 
even vnode numbers and even inode numbers.  So, an application can't tell 
whether something with an even inode number is a file or a mount point 
without a stat.  Worse, this same property means that a _cache manager_ 
can't tell whether something is a file or a mount point without calling 
FetchStatus on it.  Of course, once that's done once, the answer will be in 
the vcache.

This actually means Garrance's proposal won't work, either -- we can't cons 
up a "correct" link count for a directory without checking to see whether 
each of the even-numbered items inside is a file or a mount point.  And I'm 
pretty sure the cache manager doesn't actually use the "files are even" 
trick anywhere, because I don't think that rule holds for DFS-translator 
volumes.  So the CM would have to do FetchStatus calls for each item in a 
directory (or more likely, one or more BulkStatus calls) before it could 
return the result of a stat() on the directory itself.  That would 
definitely _not_ make find faster, and would likely slow down a lot of 
cases where the thing calling stat couldn't care less about the link count.

-- Jeff