Re: Bad virt_to_phys since commit 54c7a8916a887f35

[Mark Rutland]

On Thu, May 16, 2019 at 03:05:31PM +0100, Steven Price wrote:
> On 16/05/2019 14:41, Mark Rutland wrote:
> > On Thu, May 16, 2019 at 02:38:20PM +0100, Mark Rutland wrote:
> >> Hi,
> >>
> >> Since commit:
> >>
> >>   54c7a8916a887f35 ("initramfs: free initrd memory if opening /initrd.image fails")
> > 
> > Ugh, I dropped a paragarph here.
> > 
> > Since that commit, I'm seeing a boot-time splat on arm64 when using
> > CONFIG_DEBUG_VIRTUAL. I'm running an arm64 syzkaller instance, and this
> > kills the VM, preventing further testing, which is unfortunate.
> > 
> > Mark.
> > 
> >> IIUC prior to that commit, we'd only attempt to free an intird if we had
> >> one, whereas now we do so unconditionally. AFAICT, in this case
> >> initrd_start has not been initialized (I'm not using an initrd or
> >> initramfs on my system), so we end up trying virt_to_phys() on a bogus
> >> VA in free_initrd_mem().
> >>
> >> Any ideas on the right way to fix this?
> 
> Your analysis looks right to me. In my review I'd managed to spot the
> change in behaviour when CONFIG_INITRAMFS_FORCE is set (the initrd is
> freed), but I'd overlooked what happens if initrd_start == 0 (the
> non-existent initrd is attempted to be freed).
> 
> I suspect the following is sufficient to fix the problem:
> 
> ----8<-----
> diff --git a/init/initramfs.c b/init/initramfs.c
> index 435a428c2af1..178130fd61c2 100644
> --- a/init/initramfs.c
> +++ b/init/initramfs.c
> @@ -669,7 +669,7 @@ static int __init populate_rootfs(void)
>  	 * If the initrd region is overlapped with crashkernel reserved region,
>  	 * free only memory that is not part of crashkernel region.
>  	 */
> -	if (!do_retain_initrd && !kexec_free_initrd())
> +	if (!do_retain_initrd && initrd_start && !kexec_free_initrd())
>  		free_initrd_mem(initrd_start, initrd_end);
>  	initrd_start = 0;
>  	initrd_end = 0;

That works for me. If you spin this as a real patch:

Tested-by: Mark Rutland <mark.rutland@xxxxxxx>

As I mentioned, initrd_start has not been initialized at all, so I
suspect we should also update its declaration in init/do_mounts_initrd.c
such that it is guaranteed to be initialized to zero. We get away with
that today, but that won't necessarily hold with LTO and so on...

Thanks,
Mark.