On Mon, Jul 20, 2020 at 07:43:22PM +0200, Christoph Hellwig wrote:

On Mon, Jul 20, 2020 at 09:37:48AM -0700, Eric Biggers wrote:

quoted

How does this not introduce a massive security hole when
CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE?

AFAICS, userspace can pass in a pointer >= TASK_SIZE,
and this code makes it be treated as a kernel pointer.

Yeah, we'll need to validate that before initializing the pointer.

But thinking this a little further:  doesn't this mean any
set_fs(KERNEL_DS) that has other user pointers than the one it is
intended for has the same issue?  Pretty much all of these are gone
in mainline now, but in older stable kernels there might be some
interesting cases, especially in the compat ioctl handlers.

Yes.  I thought that eliminating that class of bug is one of the main
motivations for your "remove set_fs" work.  See commit 128394eff343
("sg_write()/bsg_write() is not fit to be called under KERNEL_DS") for a case
where this type of bug was fixed.

Are you aware of any specific cases that weren't already fixed?  If there are
any, they need to be urgently fixed.

- Eric