Filling holes in ext2

[Adrian Cox <adrian@humboldt.co.uk>]

I've been looking at generic_file_write() a lot recently, and I'm a 
little bothered by this section, as mangled here by Mozilla:

status = mapping->a_ops->prepare_write(file, page, offset, offset+bytes);
if (status)
	goto unlock;
status = __copy_from_user(kaddr+offset, buf, bytes);
flush_dcache_page(page);
if (status)
	goto fail_write;
status = mapping->a_ops->commit_write(file, page, offset, offset+bytes);

If the __copy_from_user() does fail when writing to a hole or extending 
a file on ext2, disk blocks get added to the file, but are never 
cleared. The result is that data from a free block appears in the file.

I've not managed to trigger this in a real system, but I have explored 
the failure path by running UML under gdb. I filled the filesystem with 
data as root (yes > /mnt/test), deleted the files, then triggered this 
path on an application running as a normal user. The result was that 
root's old data appeared in the user file.

So:
Can this really happen on the mainstream kernel? (The kernel I tested on 
  was 2.4.7 with the corresponding UML patch.)

Can this actually be exploited?  I assume the test on __copy_from_user() 
is there in case another thread changes memory mappings while 
generic_file_write() is running. My attempts to do this haven't yet 
succeeded.

If this can happen, does it matter? Should ext2 have an abort_write() 
operation like ext3() has?

-- 
Adrian Cox   http://www.humboldt.co.uk/