On 6/21/21 7:29 AM, Borislav Petkov wrote:
> On Mon, Jun 21, 2021 at 12:51:46PM +0200, gregkh@xxxxxxxxxxxxxxxxxxx wrote:
>>
>> The patch below does not apply to the 5.4-stable tree.
>> If someone wants it applied there, or to any other stable or longterm
>> tree, then please email the backport, including the original git commit
>> id to <stable@xxxxxxxxxxxxxxx>.
>>
>> thanks,
>>
>> greg k-h
>>
>> ------------------ original commit in Linus's tree ------------------
>>
>> From d8778e393afa421f1f117471144f8ce6deb6953a Mon Sep 17 00:00:00 2001
>> From: Andy Lutomirski <luto@xxxxxxxxxx>
>> Date: Tue, 8 Jun 2021 16:36:19 +0200
>> Subject: [PATCH] x86/fpu: Invalidate FPU state after a failed XRSTOR from a
>> user buffer
>>
>> Both Intel and AMD consider it to be architecturally valid for XRSTOR to
>> fail with #PF but nonetheless change the register state. The actual
>> conditions under which this might occur are unclear [1], but it seems
>> plausible that this might be triggered if one sibling thread unmaps a page
>> and invalidates the shared TLB while another sibling thread is executing
>> XRSTOR on the page in question.
>>
>> __fpu__restore_sig() can execute XRSTOR while the hardware registers
>> are preserved on behalf of a different victim task (using the
>> fpu_fpregs_owner_ctx mechanism), and, in theory, XRSTOR could fail but
>> modify the registers.
>>
>> If this happens, then there is a window in which __fpu__restore_sig()
>> could schedule out and the victim task could schedule back in without
>> reloading its own FPU registers. This would result in part of the FPU
>> state that __fpu__restore_sig() was attempting to load leaking into the
>> victim task's user-visible state.
>>
>> Invalidate preserved FPU registers on XRSTOR failure to prevent this
>> situation from corrupting any state.
>>
>> [1] Frequent readers of the errata lists might imagine "complex
>> microarchitectural conditions".
>>
>> Fixes: 1d731e731c4c ("x86/fpu: Add a fastpath to __fpu__restore_sig()")
>> Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
>> Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
>> Signed-off-by: Borislav Petkov <bp@xxxxxxx>
>> Acked-by: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
>> Acked-by: Rik van Riel <riel@xxxxxxxxxxx>
>> Cc: stable@xxxxxxxxxxxxxxx
>> Link: https://lkml.kernel.org/r/20210608144345.758116583@xxxxxxxxxxxxx
>>
>> diff --git a/arch/x86/kernel/fpu/signal.c b/arch/x86/kernel/fpu/signal.c
>> index d5bc96a536c2..4ab9aeb9a963 100644
>> --- a/arch/x86/kernel/fpu/signal.c
>> +++ b/arch/x86/kernel/fpu/signal.c
>> @@ -369,6 +369,25 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
>> fpregs_unlock();
>> return 0;
>> }
>> +
>> + /*
>> + * The above did an FPU restore operation, restricted to
>> + * the user portion of the registers, and failed, but the
>> + * microcode might have modified the FPU registers
>> + * nevertheless.
>> + *
>> + * If the FPU registers do not belong to current, then
>> + * invalidate the FPU register state otherwise the task might
>> + * preempt current and return to user space with corrupted
>> + * FPU registers.
>> + *
>> + * In case current owns the FPU registers then no further
>> + * action is required. The fixup below will handle it
>> + * correctly.
>> + */
>> + if (test_thread_flag(TIF_NEED_FPU_LOAD))
>> + __cpu_invalidate_fpregs_state();
>> +
>> fpregs_unlock();
>> } else {
>
> So I'm looking at this and 5.4.127 has:
>
> if (!ret) {
> fpregs_mark_activate();
> fpregs_unlock();
> return 0;
> }
> fpregs_deactivate(fpu); <---
> fpregs_unlock();
>
> i.e., an unconditional fpu invalidation there. Which got removed by:
>
> 98265c17efa9 ("x86/fpu/xstate: Preserve supervisor states for the slow path in __fpu__restore_sig()")
>
> in 5.7.
>
> so that Fixes: commit above which points to a 5.1 kernel is probably wrong-ish.
>
> amluto?
>
I agree. The fixes line is indeed wrong, and the (horribly misnamed)
fpu_deactivate() call did the right thing.
