Re: [PATCH] mm/vmalloc: randomize vmalloc() allocations

From: Topi Miettinen
Date: Thu Dec 10 2020 - 15:01:30 EST


On 3.12.2020 8.58, Mike Rapoport wrote:
On Wed, Dec 02, 2020 at 08:49:06PM +0200, Topi Miettinen wrote:
On 1.12.2020 23.45, Topi Miettinen wrote:
Memory mappings inside kernel allocated with vmalloc() are in
predictable order and packed tightly toward the low addresses. With
new kernel boot parameter 'randomize_vmalloc=1', the entire area is
used randomly to make the allocations less predictable and harder to
guess for attackers.


This also seems to randomize module addresses. I was going to check that
next, so nice surprise!

Heh, that's because module_alloc() uses vmalloc() in that way or another :)

I got a bit further with really using vmalloc with [VMALLOC_START..VMALLOC_END] for modules, but then inserting a module fails because of the relocations:
[ 9.202856] module: overflow in relocation type 11 val ffffe1950e27f080

Type 11 is R_X86_64_32S expecting a 32 bits signed offset, so the loader obviously can't fit the relocation from the highest 2GB to somewhere 32 TB lower.

The problem seems to be that the modules aren't really built as position-independent shared objects with -fPIE/-fPIC, but instead there's explicit -fno-PIE. I guess the modules also shouldn't use -mcmodel=kernel. Though tweaking the flags shows that some combinations aren't well supported (like ’-mindirect-branch=thunk-extern’ and ‘-mcmodel=large’ are not compatible) and the handwritten assembly code also assumes 32 bit offsets.

A different approach could be to make the entire kernel relocatable to lower addresses and then the modules could stay close nearby. I guess the asm files aren't written with position independence in mind either.

But it seems that I'm finding and breaking lots of assumptions built in to the system. What's the experts' opinion, is full module/kernel randomization ever going to fly?

-Topi