From: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Sent: Friday, October 23, 2020 9:25 PM
> To: Konstantin Komarov <almaz.alexandrovich@xxxxxxxxxxxxxxxxxxxx>
> Cc: linux-fsdevel@xxxxxxxxxxxxxxx; viro@xxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; pali@xxxxxxxxxx; dsterba@xxxxxxx;
> aaptel@xxxxxxxx; rdunlap@xxxxxxxxxxxxx; joe@xxxxxxxxxxx; mark@xxxxxxxxxxxxx; nborisov@xxxxxxxx; linux-ntfs-
> dev@xxxxxxxxxxxxxxxxxxxxx; anton@xxxxxxxxxx
> Subject: Re: [PATCH v10 02/10] fs/ntfs3: Add initialization of super block
>
> On Fri, Oct 23, 2020 at 06:44:23PM +0300, Konstantin Komarov wrote:
> > +
> > +/*ntfs_readpage*/
> > +/*ntfs_readpages*/
> > +/*ntfs_writepage*/
> > +/*ntfs_writepages*/
> > +/*ntfs_block_truncate_page*/
>
> What are these for?
>
> > +int ntfs_readpage(struct file *file, struct page *page)
> > +{
> > + int err;
> > + struct address_space *mapping = page->mapping;
> > + struct inode *inode = mapping->host;
> > + struct ntfs_inode *ni = ntfs_i(inode);
> > + u64 vbo = (u64)page->index << PAGE_SHIFT;
> > + u64 valid;
> > + struct ATTRIB *attr;
> > + const char *data;
> > + u32 data_size;
> > +
> [...]
> > +
> > + if (is_compressed(ni)) {
> > + if (PageUptodate(page)) {
> > + unlock_page(page);
> > + return 0;
> > + }
>
> You can skip this -- the readpage op won't be called for pages which
> are Uptodate.
>
> > + /* normal + sparse files */
> > + err = mpage_readpage(page, ntfs_get_block);
> > + if (err)
> > + goto out;
>
> It would be nice to use iomap instead of mpage, but that's a big ask.
>
> > + valid = ni->i_valid;
> > + if (vbo < valid && valid < vbo + PAGE_SIZE) {
> > + if (PageLocked(page))
> > + wait_on_page_bit(page, PG_locked);
> > + if (PageError(page)) {
> > + ntfs_inode_warn(inode, "file garbage at 0x%llx", valid);
> > + goto out;
> > + }
> > + zero_user_segment(page, valid & (PAGE_SIZE - 1), PAGE_SIZE);
>
> Nono, you can't zero data after the page has been unlocked. You can
> handle this case in ntfs_get_block(). If the block is entirely beyond
> i_size, returning a hole will cause mpage_readpage() to zero it. If it
> straddles i_size, you can either ensure that the on-media block contains
> zeroes after the EOF, or if you can't depend on that, you can read it
> in synchronously in your get_block() and then zero the tail and set the
> buffer Uptodate. Not the most appetising solution, but what you have here
> is racy with the user writing to it after reading.
Hello Matthew! Thanks a lot for this feedback. Fixed in v11, please check out.
Cheers!
