[PATCH] eal: add cache guard to per-lcore PRNG state

[Stephen Hemminger]

On Mon, 4 Sep 2023 13:57:19 +0200
Mattias Rönnblom <hofors at lysator.liu.se> wrote:

> On 2023-09-04 11:26, Morten Brørup wrote:
> > The per-lcore random state is frequently updated by their individual
> > lcores, so add a cache guard to prevent CPU cache thrashing.
> >   
> 
> "to prevent false sharing in case the CPU employs a next-N-lines (or 
> similar) hardware prefetcher"
> 
> In my world, cache trashing and cache line contention are two different 
> things.
> 
> Other than that,
> Acked-by: Mattias Rönnblom <mattias.ronnblom at ericsson.com>

Could the per-lcore state be thread local?

Something like this:

From 3df5e28a7e5589d05e1eade62a0979e84697853d Mon Sep 17 00:00:00 2001
From: Stephen Hemminger <stephen at networkplumber.org>
Date: Wed, 6 Sep 2023 09:22:42 -0700
Subject: [PATCH] random: use per lcore state

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 | 35 +++++++++++++++++------------------
 1 file changed, 17 insertions(+), 18 deletions(-)

diff --git a/lib/eal/common/rte_random.c b/lib/eal/common/rte_random.c
index 53636331a27b..62f36038ac52 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)
@@ -76,16 +77,14 @@ __rte_srand_lfsr258(uint64_t seed, struct rte_rand_state *state)
 	state->z3 = __rte_rand_lfsr258_gen_seed(&lcg_seed, 4096UL);
 	state->z4 = __rte_rand_lfsr258_gen_seed(&lcg_seed, 131072UL);
 	state->z5 = __rte_rand_lfsr258_gen_seed(&lcg_seed, 8388608UL);
+
+	state->seed = seed;
 }
 
 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 +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 idx;
-
-	idx = rte_lcore_id();
+	struct rte_rand_state *rand_state = &RTE_PER_LCORE(rte_rand_state);
+	uint64_t seed;
 
-	/* last instance reserved for unregistered non-EAL threads */
-	if (unlikely(idx == LCORE_ID_ANY))
-		idx = RTE_MAX_LCORE;
+	/* did seed change */
+	seed = __atomic_load_n(&rte_rand_seed, __ATOMIC_RELAXED);
+	if (unlikely(seed != rand_state->seed))
+		__rte_srand_lfsr258(seed, rand_state);
 
-	return &rand_states[idx];
+	return rand_state;
 }
 
 uint64_t
-- 
2.39.2