Re: [PATCH v1 1/2] random-bits: Factor out entropy generating function

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 04/04/2022 15:52, Noah Goldstein wrote:
> On Mon, Apr 4, 2022 at 1:38 PM Adhemerval Zanella
> <adhemerval.zanella@linaro.org> wrote:
>>
>>
>>
>> On 04/04/2022 15:23, Noah Goldstein wrote:
>>> On Mon, Apr 4, 2022 at 12:42 PM Adhemerval Zanella
>>> <adhemerval.zanella@linaro.org> wrote:
>>>>
>>>>
>>>>
>>>> On 01/04/2022 15:01, Cristian Rodríguez wrote:
>>>>> On Thu, Mar 31, 2022 at 8:05 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>>>>>
>>>>>> AFAIK our goal is entropy more so than security. For example
>>>>>> if this is used to generate jiffies to stagger threads its not a security
>>>>>> issue in any sense, it's just not ideal for performance.
>>>>>
>>>>> In any case this should be more than fast enough for the other use
>>>>> cases of random_bits() .. maybe one new random_bits_fast() function
>>>>> foe edge cases where even this is too slow?
>>>>
>>>> I think we are bike-shedding in the same issue OpenBSD guys stumbled
>>>> and which they have solved 10 years ago [1].  Essentially, we need to
>>>> come  up with a internal PRNG interface that can be used internally and
>>>> externally, instead of reinventing cleaver ways to use the timer as
>>>> entropy source.
>>>>
>>>> The issue is not really the cypher used, ideally it could be replace if
>>>> we find out that it does not fit.  The main issue is to glue together
>>>> all the requirements to have a concise internal interface, taking in
>>>> consideration the glibc constrains to work with multiple kernel version
>>>> and environments (where we can't assume we have access to a source or
>>>> reliable entropy like getrandom syscall).
>>>>
>>>> My plan to rehearse Florian arc4random proposal to have some simpler
>>>> to where we might improve upon (a simpler fork detection for kernels
>>>> without MADV_WIPEONFORK that just issue a atfork handle, maybe using
>>>> ChaCha20 as virtually all other systems do, no per-thread state).
>>>
>>> Why no per-thread / per-cpu? It seems otherwise there will need to be
>>> some explicit synchronization on the stream.
>>
>> Mainly because it simplifies a lot the *initial* implementation. I would
>> prefer to incremental add per-thread optimization than dump a large patch
>> so we can review in integrate the code more easier.
>>
>>>
>>>
>>> Regarding this patch, do we want to skip it and just wait on arc4random
>>> interface in kernel/glibc or should I go forward with it and some arch
>>> specific entropy functions in the mean-time?
>>
>> I don't have a strong opinion on this patch, it does improve x86_64
>> latency on random_bits although currently internal usage are far from
>> latency sensitive.  It is really a microoptimization without much real
>> work gain for current code.
> 
> That's fair. The motivation is so random_bits can be used for lightweight
> jitter i.e in cases like:
> 
> [v2] nptl: Add backoff mechanism to spinlock loop
> 
> The x86 backend with `rdtsc` is just because it's a simple improvement,
> other arch will hopefully be able to get off syscall if they don't have vdso
> gettime.
> But agree the improvement gains a minimal so if people don't feel its worth
> the added complexity we can wait on a strong arch-random interface.

I think for mutex optimization it would be better to just add a arch-specific
jitter code and use the backoff optimization iff the arch-specific code is
used. And maybe not tied to random_bits(), since I am not sure if an interface
like arc4random would be good to use in such scenario (since a possible state
reschedule might call getentropy with my add unexpected latency).