RE: [PATCH] eal: have unregistered non-EAL threads use dedicated PRNG

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Mattias Rönnblom [mailto:mattias.ronnblom@ericsson.com]
> Sent: Monday, 5 December 2022 11.04
> 
> Prior to this change, unregistered non-EAL threads shared a PRNG
> instance with the main lcore. The main lcore may well be used for fast
> path processing, potentially making rte_rand() calls in the
> process. It should not need to synchronize with control threads.
> 
> With this change, all unregistered non-EAL threads share one dedicated
> PRNG instance.
> 
> The API documentation is updated to use the proper terminology when
> referring to threads equipped with an lcore id.
> 
> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> ---
>  lib/eal/common/rte_random.c  | 17 +++++++++++------
>  lib/eal/include/rte_random.h | 10 +++++++---
>  2 files changed, 18 insertions(+), 9 deletions(-)
> 
> diff --git a/lib/eal/common/rte_random.c b/lib/eal/common/rte_random.c
> index 166b0d8921..565f2401ce 100644
> --- a/lib/eal/common/rte_random.c
> +++ b/lib/eal/common/rte_random.c
> @@ -20,7 +20,11 @@ struct rte_rand_state {
>  	uint64_t z5;
>  } __rte_cache_aligned;
> 
> -static struct rte_rand_state rand_states[RTE_MAX_LCORE];
> +/* 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];
> 
>  static uint32_t
>  __rte_rand_lcg32(uint32_t *seed)
> @@ -114,14 +118,15 @@ __rte_rand_lfsr258(struct rte_rand_state *state)
>  static __rte_always_inline
>  struct rte_rand_state *__rte_rand_get_state(void)
>  {
> -	unsigned int lcore_id;
> +	unsigned int idx;
> 
> -	lcore_id = rte_lcore_id();
> +	idx = rte_lcore_id();
> 
> -	if (unlikely(lcore_id == LCORE_ID_ANY))
> -		lcore_id = rte_get_main_lcore();
> +	/* last instance reserved for unregistered non-EAL threads */
> +	if (unlikely(idx == LCORE_ID_ANY))
> +		idx = RTE_MAX_LCORE;
> 
> -	return &rand_states[lcore_id];
> +	return &rand_states[idx];
>  }
> 
>  uint64_t
> diff --git a/lib/eal/include/rte_random.h
> b/lib/eal/include/rte_random.h
> index d90e4d2192..2edf5d210b 100644
> --- a/lib/eal/include/rte_random.h
> +++ b/lib/eal/include/rte_random.h
> @@ -41,7 +41,8 @@ rte_srand(uint64_t seedval);
>   *
>   * The generator is not cryptographically secure.
>   *
> - * If called from lcore threads, this function is thread-safe.
> + * If called from EAL threads or registered non-EAL threads, this
> function
> + * is thread-safe.
>   *
>   * @return
>   *   A pseudo-random value between 0 and (1<<64)-1.
> @@ -55,7 +56,8 @@ rte_rand(void);
>   * This function returns an uniformly distributed (unbiased) random
>   * number less than a user-specified maximum value.
>   *
> - * If called from lcore threads, this function is thread-safe.
> + * If called from EAL threads or registered non-EAL threads, this
> function
> + * is thread-safe.
>   *
>   * @param upper_bound
>   *   The upper bound of the generated number.
> @@ -75,7 +77,9 @@ rte_rand_max(uint64_t upper_bound);
>   * number uniformly distributed over the interval [0.0, 1.0).
>   *
>   * The generator is not cryptographically secure.
> - * If called from lcore threads, this function is thread-safe.
> + *
> + * If called from EAL threads or registered non-EAL threads, this
> function
> + * is thread-safe.
>   *
>   * @return
>   *   A pseudo-random value between 0 and 1.0.
> --
> 2.34.1
> 

A nice improvement.

Acked-by: Morten Brørup <mb@smartsharesystems.com>


Here's some serious feature creep...

Instead of using "static struct rte_rand_state rand_states[RTE_MAX_LCORE + 1];", we could use thread local storage ("__tread rte_rand_state rand_state;") to keep the state per O/S thread (independent of lcore_id etc.), making it completely thread safe.

But then, how do we seed the state?

Currently, we use the RTE_INIT() constructor attribute to seed the array of rand_states; but there is no thread constructor attribute. So here comes the feature creep:

It would be very useful with RTE_THREAD_INIT()/_FINI constructor/destructor macros, so libraries and applications could define functions to be called by thread_func_wrapper() before/after calling tread_func.

Using arrays like some_variable[RTE_MAX_LCORE (+ 1)] is common practice in DPDK, but only really required for variables that are not private to the thread, i.e. variables that other threads need access to.

Per-thread constructors/destructors is a generic feature suggestion, so please don't hold back this rte_random patch!

-Morten