Re: [PATCH] 2.6.10 - direct-io async short read bug
From: Andrew Morton
Date: Thu Mar 10 2005 - 00:02:10 EST
Badari Pulavarty <pbadari@xxxxxxxxxx> wrote:
> On Wed, 2005-03-09 at 11:53, Andrew Morton wrote:
> > Suparna Bhattacharya <suparna@xxxxxxxxxx> wrote:
> > >
> > > > Solaris, which does forcedirectio as a mount option, actually
> > > > will do buffered I/O on the trailing part. Consider it like a bounce
> > > > buffer. That way they don't DMA the trailing data and succeed the I/O.
> > > > The I/O returns actual bytes till EOF, just like read(2) is supposed to.
> > > > Either this or a fully DMA'd number 4 is really what we should
> > > > do. If security can only be solved via a bounce buffer, who cares? If
> > > > the user created themselves a non-aligned file to open O_DIRECT, that's
> > > > their problem if the last part-sector is negligably slower.
> > >
> > > If writes/truncates take care of zeroing out the rest of the sector
> > > on disk, might we still be OK without having to do the bounce buffer
> > > thing ?
> > We can probably rely on the rest of the sector outside i_size being zeroed
> > anyway. Because if it contains non-zero gunk then the fs already has a
> > problem, and the user can get at that gunk with an expanding truncate and
> > mmap() anyway.
> Rest of the sector or rest of the block ?
The filesystem-sized block (1<<i_blkbits) which straddles i_size should
have zeroes outside i_size.
There's one situation where it might not be zeroed out, and that's when the
final page is mapped MAP_SHARED and the application modifies that page
outside i_size while writeout is actually in flight. We can't do much about
> Are you implying that, we
> already do this, so there is no problem reading beyond EOF to user
> buffer ? Or we need to zero out the userbuffer beyond EOF ?
It should be acceptable to assume that the final (1<<i_blkbits) block of
the file contains zeroes outside i_size.
And if it doesn't contain those zeroes, well, applications are able to read
that data already. Although I wouldn't count that as a security hole: that
data is something which an application wrote there while writing the file,
rather than being left-over uncontrolled stuff.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/