Ben Laurie <[EMAIL PROTECTED]> writes:
> Ben Collins-Sussman wrote:
> > He did give a very detailed overview of his proposed API, if you go
> > back read his original response. The problem is that I haven't the
> > first idea of how to start implementing it. And his API seems to
>
> Well, I'll worry about how to implement it, but...
>
> > focus on passing parameters that tweak the quality of the randomness,
> > but for Subversion, what really matters in the non-blocking aspect.
>
> I'd really like to understand this - if you really don't care about
> quality of randomness, why aren't you using rand()? Or a fixed number?
>
> What is the randomness used for?
Allow me to foist Subversion's design issues on the APR folks. :-)
Maybe the great minds here can suggest a good solution to our problem.
Here's our situation.
* svn_repos_create() needs to generate a UUID that identifies the
newly created repository. This UUID becomes a permanent identifier
for the repository... so it calls apr_uuid_get().
* apr_uuid_get() calls apr_generate_random_bytes().
* apr_generate_random_bytes() can potentially block for entropy when
reading /dev/random.
And in practice, this *has* been biting a few svn users, including
me. When I run 'make check' (which creates a new repository for
each test), and go to bed for the night, we discover only a few
tests have run, because there was no system entropy. As soon as I
start typing, the tests start running again.
[side note: yes, I'm aware that FreeBSD 4.X (my system) has a
poor /dev/random implementation, and I'm aware that FreeBSD 5.X no
longer suffers this problem. But I've also seen the problem
happen on flaky RedHat boxes as well.]
So perhaps svn_repos_create() shouldn't be calling apr_uuid_get() at
all? Perhaps svn should have its own pseudorandom-nonblocking uuid
generator? Perhaps APR could grow one?
