From vendor-sec:

Date: Tue, 13 Apr 2004 23:37:44 +0100
From: Dave Jones <davej@redhat.com>
To: vendor-sec@lst.de
Cc: brodo@linux.de
Subject: [vendor-sec] cpufreq (un)signed int wrap.

Folks,
 Brad Spender found an exploitable bug in the Linux kernel cpufreq code.
This is a part of 2.6, and various vendors have variants of it in their
2.4 trees (All that I checked had the same flaw).
                                                                                
I've not yet heard from Brad as to when he plans to announce this.
                                                                                
Dominik also did a quick grep of other proc_handler's and turned up a number of
other instances of what seems to be the same bug.
Certainly looks worthy of an audit.

Until I hear of a date of disclosure from Brad, I'm going to hold
off fixing this in the cpufreq bk tree.

    Dave

----- Forwarded message from spender@grsecurity.net -----

X-Original-To: brodo@localhost
Delivered-To: brodo@localhost.dominikbrodowski.de
X-Envelope-From: <spender@grsecurity.net>
X-Envelope-To: <linux@brodo.de>
X-Delivery-Time: 1081879425
Date: Tue, 13 Apr 2004 13:58:33 -0400
To: rmk@arm.linux.org.uk, linux@brodo.de
Subject: exploitable stack overflow in cpufreq
User-Agent: Mutt/1.5.4i
From: spender@grsecurity.net

just wanted to call your attention to the following:

static int
cpufreq_procctl(ctl_table *ctl, int write, struct file *filp,
                void __user *buffer, size_t *lenp)
{
        char buf[16], *p;
        int cpu = (int) ctl->extra1;
        int len, left = *lenp;
                len = left;
                if (left > sizeof(buf))
                        left = sizeof(buf);
                if (copy_from_user(buf, buffer, left))
                        return -EFAULT;


note that you cast a user-supplied unsigned int to a signed int.
you then only check the upper bound on the value of the signed int
you then send it to copy_from_user, which uses an unsigned int (long)
this is exploitable, though only as a root user.

It gets a bit more interesting here:

        } else {
                len = sprintf(buf, "%d\n", cpufreq_get(cpu));
                if (len > left)
                        len = left;
                if (copy_to_user(buffer, buf, len))
                        return -EFAULT;

here by passing a large value in to a read, which will wrap to a
negative value, and then back to the large value for copy_to_user,
we can read into kernel memory as a non-root user.

To fix, use unsigned ints instead of signed.

-Brad

-----------

Not clear which of our kernels are affected.