Re: [GIT PULL] Device tree updates for v3.12

[Linus Torvalds <torvalds@linux-foundation.org>]

On Tue, Sep 10, 2013 at 11:40 AM, Tim Bird <tbird20d@gmail.com> wrote:
> How much time does it add to boot to feed the device tree into the
> random number pool.
>
> Some of the device trees are expected to get pretty big.  If it's over
> a millisecond, IMHO, it should be configurable (but this is not).

It's detinitely not a very fast operation. "add_device_randomness()"
does four full "mix_pool_bytes()" operations, and those each iterate
over the input set one byte at a time.

It was kind of designed for things like mixing in ethernet MAC
addresses etc, so at the time that was written, the thinking was that
it would be just a few bytes, maybe tens of bytes.

I don't know how big flattened device trees can be, but I guess we're
talking a couple of kB?

So it might even be a better idea to feed the device tree to a hashing
function (eg SHA1 or even just MD5), and then just mix in the hash. At
least most block hash functions do things a word at a time. It does
*not* need to be cryptographically secure, so MD5 would be plenty good
enough - the only point of the hash would be to give a meaningful
number of result bits from the source array.

Of course, maybe even the stupid add_device_randomness() is fast
enough. I just wanted to point out that it definitely isn't some
optimized thing.

                Linus