Re: [rfc][patch 1/3] fs: new truncate sequence

[Nick Piggin <npiggin@suse.de>]

On Mon, Jul 06, 2009 at 08:00:07PM +0200, Miklos Szeredi wrote:
> > Index: linux-2.6/mm/truncate.c
> > ===================================================================
> > --- linux-2.6.orig/mm/truncate.c
> > +++ linux-2.6/mm/truncate.c
> > @@ -465,3 +465,79 @@ int invalidate_inode_pages2(struct addre
> >  	return invalidate_inode_pages2_range(mapping, 0, -1);
> >  }
> >  EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
> > +
> > +/**
> > + * truncate_pagecache - unmap mappings "freed" by truncate() syscall
> > + * @inode: inode
> > + * @old: old file offset
> > + * @new: new file offset
> > + *
> > + * inode's new i_size must already be written before truncate_pagecache
> > + * is called.
> > + */
> > +void truncate_pagecache(struct inode * inode, loff_t old, loff_t new)
> > +{
> > +	VM_BUG_ON(inode->i_size != new);
> 
> This is not true for fuse (and NFS?) as i_size isn't protected by
> i_mutex during attribute revalidation, and so it can change during the
> truncate.

Hmm, that's probably OK now. filemap_fault has some tricky code
to avoid faulting in pages past i_size, but since that has been
changed to use page lock a while back, the i_size checks can
probably go away.

So long as your filesystems obviously have to ensure the truncate
will not truncate the wrong pages, I can remove the VM_BUG_ON
just fine.

Thanks,
Nick