/dev/random entropy calcs - patch [not related to net devices]

/dev/random entropy calcs - patch [not related to net devices]

[Alex Bligh - linux-kernel]

1. Anyone have a problem with using xtime rather
   than jiffies on architectures which don't support cycle
   counters? (completely untested patch attached to illustrate)

2. Anyone have any problem changing fs/proc/proc_misc.c to
   register /proc/interrupts to be 0600 instead of 0644 to help
   prevent entropy attacks that way?

--
Alex Bligh

--- drivers/char/random.c.keep       Sat Feb 17 00:02:36 2001
+++ drivers/char/random.c     Tue Aug 21 23:19:17 2001
@@ -710,16 +710,27 @@
        int             entropy = 0;

 #if defined (__i386__)
+       /* If possible, use the clock cycle counter */
        if ( test_bit(X86_FEATURE_TSC, &boot_cpu_data.x86_capability) ) {
                __u32 high;
                __asm__(".byte 0x0f,0x31"
                        :"=a" (time), "=d" (high));
                num ^= high;
        } else {
-               time = jiffies;
+               time = (__u32)(xtime.tv_usec) ^ (__u32)(xtime.tv_sec);
        }
 #else
-       time = jiffies;
+       /* If we can't get the clock cycle counter, get the number
+        * of elapsed microseconds. Note:
+        * - we don't need the xtime spinlock as we are only reading
+        *   one half (even then, the odd SMP race is only going to
+        *   add to entropy)
+        * - this may indeed wrap; but so could the original cycle
+        *   counter, and as us occur less frequently than clock cycles,
+        *   it's less of a problem. In any case it would underestimate
+        *   entropy on a wrap.
+        */
+       time = (__u32)(xtime.tv_usec) ^ (__u32)(xtime.tv_sec);
 #endif

        /*

RE: /dev/random entropy calcs - patch [not related to net devices]

[David Schwartz]

> +       time = (__u32)(xtime.tv_usec) ^ (__u32)(xtime.tv_sec);

	I would prefer:

	time = (__u32)xtime.tv_usec | ( (__u32)(xtime.tv_sec)<<20) );

	The way you have it, you collide the two least-significant bits,
potentially losing some entropy.

	DS

Re: /dev/random entropy calcs - patch [not related to net devices]

[Ton Hospel]

In article <9547398.998437243@_169.254.198.40_>,
	Alex Bligh - linux-kernel <linux-kernel@alex.org.uk> writes:
> 2. Anyone have any problem changing fs/proc/proc_misc.c to
>    register /proc/interrupts to be 0600 instead of 0644 to help
>    prevent entropy attacks that way?

Yes, i don't want to have to switch to root to read /proc/interrupts
because some people have gone utterly paranoid about an attack that's not
feasable if you reseed your randum numbers normally at boot.