Re: [RFC] random: use per lcore state

["Mattias Rönnblom" <hofors@lysator.liu.se>]

On 2023-09-06 21:55, Stephen Hemminger wrote:
> On Wed,  6 Sep 2023 10:20:13 -0700
> Stephen Hemminger <stephen@networkplumber.org> wrote:
> 
>>   static __rte_always_inline
>>   struct rte_rand_state *__rte_rand_get_state(void)
>>   {
>> -	unsigned int idx;
>> +	struct rte_rand_state *rand_state = &RTE_PER_LCORE(rte_rand_state);
>> +	uint64_t seed;
>>   
>> -	idx = rte_lcore_id();
>> +	seed = __atomic_load_n(&rte_rand_seed, __ATOMIC_RELAXED);
>> +	if (unlikely(seed != rand_state->seed)) {
>> +		rand_state->seed = seed;
>>   
>> -	/* last instance reserved for unregistered non-EAL threads */
>> -	if (unlikely(idx == LCORE_ID_ANY))
>> -		idx = RTE_MAX_LCORE;
>> +		seed += rte_thread_self().opaque_id;
>> +		__rte_srand_lfsr258(seed, rand_state);
>> +	}
> 
> Not sure about this.
> It would change the semantics of rte_srand so that if passed the same
> value across multiple runs, it would still generate different values because
> thread_id is not the same.  Using rte_lcore() instead would cause repeatablity
> but then there would still be a bug if two non-EAL threads used random.

If the non-EAL threads register themselves (and thus acquires a lcore 
id), they will get this kind of repeatability, provided the order of 
registration is the same. A "big if", no doubt.

> Both threads would get the same sequence of numbers, but that is true
> with current code.