Re: [PATC] small VFS change for JFFS2

[David Woodhouse <dwmw2@infradead.org>]

On Mon, 2005-04-18 at 16:31 +0400, Artem B. Bityuckiy wrote:
> Yes, exactly. VFS developers may always say "it is your problem -
> redesign JFFS2", but I think it is too late to redesign it.

Bear in mind that the reason I added the state tracking to the internal
jffs2_inode_cache state was specifically to deal with the fact that the
VFS can call read_inode() for an inode which was previously in core and
for which it hasn't yet called clear_inode().

I suspect that there's a way we could work around this problem in JFFS2,
and if not hch is probably right -- exporting the lock isn't good
enough; if we're going to change the VFS we should fix the actual
problem.

I'm confused by this though -- I thought we'd already _fixed_ this in
the VFS. What you describe shouldn't happen because I believe you're
missing a step:

> JFFS2 writer: looks at the inode state, realizes it is in state PRESENT,
> i.e. it is in the inode cache (which is wrong).
> JFFS2 writer: runs iget() to acquire a pointer to the struct inode of
> the inode 15601.

wait_for_freeing_inode() waits for the existing inode in state I_FREEING
to finish being cleared...

> JFFS2 writer: iget() calls ->read_inode(), i.e., jffs2_read_inode().
> JFFS2 writed: JFFS2 is surprised why read_inode() is called for the
> already built inode 15601.

... and this shouldn't happen.

-- 
dwmw2