Re: If not readdir() then what?

[Neil Brown <neilb@suse.de>]

On Monday April 16, tytso@mit.edu wrote:
> 
> The challenge is making it be stable across inserts/deletes, never
> mind reboots.  And it's not a "little bit of cacheing"; in order to be
> correct we would have to cache *forever*, since at least in theory an
> NFS client could hold on to a cookie for an arbitrarily long period of
> time (weeks, months, years, decades), yes?

Yes.  But I think we've already establish that the on-disk structure
chosen by ext3/htree is not able to perfectly support NFS (which is a
pity given that it was written for Linux and Linux is thought to
support NFS).  Our goal is to find the best mapping possible and,
where caching can improve stability for real-world uses, use caching
to help stabilise that mapping.

> 
> You're welcome to try, but I suspect it won't take long before you'll
> see why I'm asserting that a directory fd cache in nfsd is *way* less
> work.  :-)

You have provided some very helpful insights into how ext3/htree
currently works - thanks for that.
I will definitely make a closer inspection of the code and so how
possible it is to realise by ideas.  I'll let you know how I go.

NeilBrown