On 28 Mar 2007 16:14:17 +0200
Andi Kleen [off-list ref] wrote:

TCP tends to be initialized early before there is anything
good in the entropy pool.

static void init_std_data(struct entropy_store *r)
{
        struct timeval tv;
        unsigned long flags;

        spin_lock_irqsave(&r->lock, flags);
        r->entropy_count = 0;
        spin_unlock_irqrestore(&r->lock, flags);

        do_gettimeofday(&tv);
        add_entropy_words(r, (__u32 *)&tv, sizeof(tv)/4);
        add_entropy_words(r, (__u32 *)utsname(),
                          sizeof(*(utsname()))/4);
}

utsname is useless here because it runs before user space has 
a chance to set it. The only truly variable thing is the 
boot time, which can be guessed with the ns part being brute forced.

To make it secure you would need to do regular rehash like
the routing cache which would pick up true randomness on the first
rehash.

Good point, but :

1) We can now use "struct timespec" to get more bits in init_std_data()

2) tcp ehash salt is initialized at first socket creation, not boot time. Maybe we have more available entropy at this point.

3) We dont want to be 'totally secure'. We only want to raise the level, and eventually see if we have to spend more time on this next year(s). AFAIK we had two different reports from people being hit by the flaw of previous hash. Not really a critical issue.

4) We could add a hard limit on the length of one chain. Even if the bad guys discover a flaw, it wont hurt too much.