CVE-2019-15031: Linux kernel: powerpc: data leak with FP/VMX triggerable by interrupt in transaction

From: Michael Neuling
Date: Tue Sep 10 2019 - 09:16:57 EST


The Linux kernel for powerpc since v4.15 has a bug in it's TM handling during
interrupts where any user can read the FP/VMX registers of a difference user's
process. Users of TM + FP/VMX can also experience corruption of their FP/VMX
state.

To trigger the bug, a process starts a transaction with FP/VMX off and then
takes an interrupt. Due to the kernels incorrect handling of the interrupt,
FP/VMX is turned on but the checkpointed state is not updated. If this
transaction then rolls back, the checkpointed state may contain the state of a
different process. This checkpointed state can then be read by the process hence
leaking data from one process to another.

The trigger for this bug is an interrupt inside a transaction where FP/VMX is
off, hence the process needs FP/VMX off when starting the transaction. FP/VMX
availability is under the control of the kernel and is transparent to the user,
hence the user has to retry the transaction many times to trigger this bug. High
interrupt loads also help trigger this bug.

All 64-bit machines where TM is present are affected. This includes all POWER8
variants and POWER9 VMs under KVM or LPARs under PowerVM. POWER9 bare metal
doesn't support TM and hence is not affected.

The bug was introduced in commit:
fa7771176b439 ("powerpc: Don't enable FP/Altivec if not checkpointed")
Which was originally merged in v4.15

The upstream fix is here:
https://git.kernel.org/torvalds/c/a8318c13e79badb92bc6640704a64cc022a6eb97

The fix can be verified by running the tm-poison from the kernel selftests. This
test is in a patch here:
https://patchwork.ozlabs.org/patch/1157467/
which should eventually end up here:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/testing/selftests/powerpc/tm/tm-poison.c

cheers
Mikey