Re: CVE-2018-1000204: Linux kernel 3.18 to 4.16 infoleak due to incorrect handling of SG_IO ioctl

[Vladis Dronov <vdronov () redhat com>]

Hello, Alexander,

Could you please, explain, why do you think CVE-2018-1000204 is a security
flaw?

> The problem has limited scope, as users don't usually have permissions
> to access SCSI devices. On the other hand, e.g. the Nero user manual
> suggests doing `chmod o+r+w /dev/sg*` to make the devices accessible.

There is a check in the kernel in sg_build_indirect() exactly for this
situation:

        [drivers/scsi/sg.c]
        if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RAWIO))
                gfp_mask |= __GFP_ZERO;

This means non-root user will get zero-ed pages even if it has o+rw access
to /dev/sg*. Tests of your reproducer on systems available to me confirm
this, i.e. non-root user gets a zero-ed out buffer even if it is able to
access /dev/sg*.

I may not got smth correctly, but for now I do not see CVE-2018-1000204
as a security flaw and I believe a reject request to MITRE should be
issued.

Best regards,
Vladis Dronov | Red Hat, Inc. | Product Security Engineer