Re: Re: [RFC/PATCH 1/2] stackfs: generic functions for obtaining hidden object

[Josef Sipek <jsipek@fsl.cs.sunysb.edu>]

On Fri, Oct 13, 2006 at 07:23:36PM +0300, Pekka Enberg wrote:
> On 10/13/06, Erez Zadok <ezk@cs.sunysb.edu> wrote:
> >I think we should do it right the first time (i.e., now :-)
> 
> I would much rather merge it now (assuming I didn't break ecryptfs)
> and have you unionfs developers fix it later :-).

Thanks :) As they say, it's the thought that counts, isn't it? ;)

> On 10/13/06, Erez Zadok <ezk@cs.sunysb.edu> wrote:
> >Why not make it something more dynamic, such as a mount-time option per sb?
> >Even at 8, you waste most of that space for non-fan-out stackable file
> >systems such as ecryptfs; and those unionfs users who want more, will have
> >to _recompile_ the code.
> 
> Yes, we discussed this with Jeff already. For unionfs, we must make it
> more dynamic. However, using slab unconditionally makes it totally
> unacceptable for ecryptfs. Therefore, we need a small static array
> that should satisfy most user (I think we can drop the static array
> size to three):

Nice, 3 pointers to inodes, and one to inode* = 4 pointers total, 128/256
bit struct on i386/x86_64.

> struct stackfs_inode_info {
>    struct inode **hidden_inodes;
>    struct inode *static_inodes[3];
> };
> 
> Initially, hidden_inodes can point to static_inodes which we can the
> replace with a dynamic array if required.

Hrm. You can have static store inodes {0,1,2} and the dynamic {3,4,5,...}
(this is what unionfs used to do - inline objects for performance). The
other way can be static array is ignored if dynamic array exists. In which
case, you effectively have {} in static, and {0,1,2,3,4,5,...} in dynamic. I
guess you could justify the wasting of the static array by arguing that if
the number of branches is << than number of static array elements, but I'm
afraid that that won't be the case most of the time.

> Btw, we probably want to do krealloc() for that in the slab allocator.

krealloc should be trivial to do (if the new size <= size of current slab,
do nothing, else alloc from a larger one).

Josef "Jeff" Sipek.

-- 
We have joy, we have fun, we have Linux on a Sun...