Re: invalidate_inode_pages in 2.5.32/3

[Andrew Morton <akpm@digeo.com>]

Daniel Phillips wrote:
> 
> > I'm very unkeen about using the inaccurate invalidate_inode_pages
> > for anything which matters, really.   And the consistency of pagecache
> > data matters.
> >
> > NFS should be using something stronger.  And that's basically
> > vmtruncate() without the i_size manipulation.
> 
> Yes, that looks good.  Semantics are basically "and don't come back
> until every damm page is gone" which is enforced by the requirement
> that we hold the mapping->page_lock though one entire scan of the
> truncated region.  (Yes, I remember sweating this one out a year
> or two ago so it doesn't eat 100% CPU on regular occasions.)
> 
> So, specifically, we want:
> 
> void invalidate_inode_pages(struct inode *inode)
> {
>         truncate_inode_pages(mapping, 0);
> }
> 
> Is it any harder than that?

Pretty much - need to leave i_size where it was.  But there are
apparently reasons why NFS cannot sleepingly lock pages in this particular
context.

> By the way, now that we're all happy with the radix tree, we might
> as well just go traverse that instead of all the mapping->*_pages.
> (Not that I'm seriously suggesting rocking the boat that way just
> now, but it might yield some interesting de-crufting possibilities.)

Oh absolutely.

 unsigned long radix_tree_gang_lookup(void **pointers,
		unsiged long starting_from_here, unsigned long this_many);

could be used nicely in readahead, drop_behind, truncate, invalidate
and invalidate2.  But to use it in writeback (desirable) we would need
additional metadata in radix_tree_node.  One bit per page, which means
"this page is dirty" or "this subtree has dirty pages".

I keep saying this in the hope that someone will take pity and write it.

> ...
> Now, what is this invalidate_inode_pages2 seepage about?  Called from
> one place.  Sheesh.

heh.  We still do have some O_DIRECT/pagecache coherency problems.