[RFC] random: use per lcore state

[Morten Brørup]

> From: Stephen Hemminger [mailto:stephen at networkplumber.org]
> Sent: Wednesday, 6 September 2023 19.20
> 
> Move the random number state into thread local storage.
> This has a several benefits.
>  - no false cache sharing from cpu prefetching
>  - fixes initialization of random state for non-DPDK threads
>  - fixes unsafe usage of random state by non-DPDK threads
> 
> The initialization of random number state is done by the
> lcore (lazy initialization).
> 
> Signed-off-by: Stephen Hemminger <stephen at networkplumber.org>
> ---
>  lib/eal/common/rte_random.c | 38 +++++++++++++++++++------------------
>  1 file changed, 20 insertions(+), 18 deletions(-)
> 
> diff --git a/lib/eal/common/rte_random.c b/lib/eal/common/rte_random.c
> index 53636331a27b..9657adf6ad3b 100644
> --- a/lib/eal/common/rte_random.c
> +++ b/lib/eal/common/rte_random.c
> @@ -19,13 +19,14 @@ struct rte_rand_state {
>  	uint64_t z3;
>  	uint64_t z4;
>  	uint64_t z5;
> -} __rte_cache_aligned;
> +	uint64_t seed;
> +};
> 
> -/* One instance each for every lcore id-equipped thread, and one
> - * additional instance to be shared by all others threads (i.e., all
> - * unregistered non-EAL threads).
> - */
> -static struct rte_rand_state rand_states[RTE_MAX_LCORE + 1];
> +/* Global random seed */
> +static uint64_t rte_rand_seed;
> +
> +/* Per lcore random state. */
> +static RTE_DEFINE_PER_LCORE(struct rte_rand_state, rte_rand_state);
> 
>  static uint32_t
>  __rte_rand_lcg32(uint32_t *seed)
> @@ -81,11 +82,7 @@ __rte_srand_lfsr258(uint64_t seed, struct
> rte_rand_state *state)
>  void
>  rte_srand(uint64_t seed)
>  {
> -	unsigned int lcore_id;
> -
> -	/* add lcore_id to seed to avoid having the same sequence */
> -	for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
> -		__rte_srand_lfsr258(seed + lcore_id,
> &rand_states[lcore_id]);
> +	__atomic_store_n(&rte_rand_seed, seed, __ATOMIC_RELAXED);
>  }
> 
>  static __rte_always_inline uint64_t
> @@ -119,15 +116,18 @@ __rte_rand_lfsr258(struct rte_rand_state *state)
>  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)) {

Please note that rte_rand_seed lives in a completely different cache line than RTE_PER_LCORE(rte_rand_state), so the comparison with rte_rand_seed requires reading one more cache line than the original implementation, which only uses the cache line holding rand_states[idx].

This is in the hot path.

If we could register a per-thread INIT function, the lazy initialization could be avoided, and only one cache line accessed.

> +		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);
> +	}
> 
> -	return &rand_states[idx];
> +	return rand_state;
>  }
> 
>  uint64_t
> @@ -227,7 +227,9 @@ RTE_INIT(rte_rand_init)
>  {
>  	uint64_t seed;
> 
> -	seed = __rte_random_initial_seed();
> +	do
> +		seed = __rte_random_initial_seed();
> +	while (seed == 0);
> 
>  	rte_srand(seed);
>  }
> --
> 2.39.2