On Fri, May 19, 2023 at 2:00 AM Dave Hansen dave.hansen@intel.com wrote:
On 5/18/23 15:51, Jeff Xu wrote:
Do you have a solid handle on all call paths that will reach __arch_check_vma_pkey_for_write() and can you ensure they are all non-remote?
Is this about the attack scenario where the attacker uses ptrace() into the chrome process ? if so it is not in our threat model, and that is more related to sandboxing on the host.
The attacker would use *some* remote interface. ptrace() is just one of those remote interfaces.
Or is this about io_uring? Yes, io_uring kernel thread breaks our expectations of PKRU & user space threads, however I thought the break is not just for this - any syscall involved in memory operation will break after into io_uring ?
I'm not quite following.
Please just do me a favor: have the io_uring maintainers look at your proposal. Make sure that the defenses you are building can work in a process where io_uring is in use by the benign threads.
Those same folks are pretty familiar with the other, more traditional I/O syscalls that have in-memory descriptors that control syscall behavior like readv/writev. Those also need a close look.
Other than those, yes, I try to ensure the check is only used at the beginning of syscall entry in all cases, which should be non-remote I hope.
You're right that synchronous, shallow syscall paths are usually non-remote. But those aren't the problem. The problem is that there *ARE* remote accesses and those are a potential hole for this whole mechanism.
Can they be closed? I don't know. I honestly don't have a great grasp on how widespread these things are. You'll need a much more complete grasp on them than I have before this thing can go forward.
I don't think the remote writes are a problem for us if they're initiated from the same process. It's a case of syscalls where we need to add special validation in userspace.