Re: [PATCH 1/1] fs/ext2: Avoid page_address on pages returned by ext2_get_page

[Javier Pello <javier.pello@urjc.es>]

On Tue, 13 Jul 2021 18:30:18, Jan Kara wrote:
> On Tue 13-07-21 16:59:18, Javier Pello wrote:
> > Commit 782b76d7abdf02b12c46ed6f1e9bf715569027f7 ("fs/ext2: Replace
> > kmap() with kmap_local_page()") replaced the kmap/kunmap calls in
> > ext2_get_page/ext2_put_page with kmap_local_page/kunmap_local for
> > efficiency reasons. As a necessary side change, the commit also
> > made ext2_get_page (and ext2_find_entry and ext2_dotdot) return
> > the mapping address along with the page itself, as it is required
> > for kunmap_local, and converted uses of page_address on such pages
> > to use the newly returned address instead. However, uses of
> > page_address on such pages were missed in ext2_check_page and
> > ext2_delete_entry, which triggers oopses if kmap_local_page happens
> > to return an address from high memory. Fix this now by converting
> > the remaining uses of page_address to use the right address, as
> > returned by kmap_local_page.
> 
> Good catch. Thanks for the patch. Ira, can you please check the patch as
> well?
> 
> I have just a few nits below:
> 
> > Signed-off-by: Javier Pello <javier.pello@urjc.es>
> > Fixes: 782b76d7abdf fs/ext2: Replace kmap() with kmap_local_page()
> 
> Please wrap subject in Fixes tag into ("...").

Will do.

> > @@ -584,16 +584,16 @@ int ext2_add_link (struct dentry *dentry, struct inode *inode)
> >   * ext2_delete_entry deletes a directory entry by merging it with the
> >   * previous entry. Page is up-to-date.
> >   */
> > -int ext2_delete_entry (struct ext2_dir_entry_2 * dir, struct page * page )
> > +int ext2_delete_entry (struct ext2_dir_entry_2 *dir, struct page *page,
> > +                     void *kaddr)
> 
> Why not have 'kaddr' as char *. We type it to char * basically everywhere
> anyway.

I thought about that, as well, but in the end I leaned towards void *
because it is a generic pointer, conceptually. Would you rather have it
be char *?

> >  {
> >       struct inode *inode = page->mapping->host;
> > -     char *kaddr = page_address(page);
> > -     unsigned from = ((char*)dir - kaddr) & ~(ext2_chunk_size(inode)-1);
> > -     unsigned to = ((char *)dir - kaddr) +
> > -                             ext2_rec_len_from_disk(dir->rec_len);
> > +     unsigned int delta = (char *)dir - (char *)kaddr;
> 
> Maybe I'd call this 'offset' rather than 'delta'.

Fair enough.

> Also if kaddr will stay
> char *, you maybe just don't need to touch this...

Yes, if kaddr becomes char * then there is no need to touch these lines.

Javier