Re: [PATCH] Stochastic Fair Blue queue discipline

[Juliusz Chroboczek <Juliusz.Chroboczek@pps.jussieu.fr>]

> (read Documentation/CodingStyle)

Will do, thanks for the pointer.

>> +static unsigned sfb_hash(struct sk_buff *skb, int hash, int filter,
>> +                         struct sfb_sched_data *q)

> I think we have multiple qdiscs now doing very similar such hashes.
> Any chance this could be factored out into a common library function 
> first?

I'll be glad to do so.  Where shall I put it?

> Getting high res time can be actually somewhat expensive on different
> platforms. I would recommend to only use it if it's actually needed in
> the inelastic flow case below and use jiffies to check for rehashing.

Good point.

> /dev/random is a precious resource (some systems have very little
> true entropy) and getting that much data from it regularly is not
> a good idea because you take it away from other users who really
> need it

In the default configuration, I'm getting 32 bytes every 10 minutes,
which seemed reasonable.  This code was copied from sch_sfq (function
sfq_perturbation), which does the very same thing.

> But in any case it is likely better you only get the
> get_random_bytes() entropy once and then use that to init some kind
> of RNG (either a secure one if it's for security or some random fast
> one if it's not) and then use that output for rehashing.

Could do, but in that case the RNG should be shared between sfq and
sfb.  Shall I simply use net_random (which calls random32) ?

                                        Juliusz