eal: replace libc-based random generation with LFSR
authorMattias Rönnblom <mattias.ronnblom@ericsson.com>
Fri, 28 Jun 2019 09:01:20 +0000 (11:01 +0200)
committerThomas Monjalon <thomas@monjalon.net>
Fri, 28 Jun 2019 13:23:38 +0000 (15:23 +0200)
This commit replaces rte_rand()'s use of lrand48() with a DPDK-native
combined Linear Feedback Shift Register (LFSR) (also known as
Tausworthe) pseudo-random number generator.

This generator is faster and produces better-quality random numbers
than the linear congruential generator (LCG) of lib's lrand48(). The
implementation, as opposed to lrand48(), is multi-thread safe in
regards to concurrent rte_rand() calls from different lcore threads.
A LCG is still used, but only to seed the five per-lcore LFSR
sequences.

In addition, this patch also addresses the issue of the legacy
implementation only producing 62 bits of pseudo randomness, while the
API requires all 64 bits to be random.

This pseudo-random number generator is not cryptographically secure -
just like lrand48().

Bugzilla ID: 114
Bugzilla ID: 276

Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
13 files changed:
MAINTAINERS
app/test/Makefile
app/test/meson.build
app/test/test_rand_perf.c [new file with mode: 0644]
doc/guides/rel_notes/release_19_08.rst
lib/librte_eal/common/include/rte_random.h
lib/librte_eal/common/meson.build
lib/librte_eal/common/rte_random.c [new file with mode: 0644]
lib/librte_eal/freebsd/eal/Makefile
lib/librte_eal/freebsd/eal/eal.c
lib/librte_eal/linux/eal/Makefile
lib/librte_eal/linux/eal/eal.c
lib/librte_eal/rte_eal_version.map

index 0c3b489..bbec198 100644 (file)
@@ -227,6 +227,12 @@ M: Joyce Kong <joyce.kong@arm.com>
 F: lib/librte_eal/common/include/generic/rte_ticketlock.h
 F: app/test/test_ticketlock.c
 
+Pseudo-random Number Generation
+M: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
+F: lib/librte_eal/common/include/rte_random.h
+F: lib/librte_eal/common/rte_random.c
+F: app/test/test_rand_perf.c
+
 ARM v7
 M: Jan Viktorin <viktorin@rehivetech.com>
 M: Gavin Hu <gavin.hu@arm.com>
index 68d6b4f..be0f392 100644 (file)
@@ -73,6 +73,7 @@ SRCS-y += test_reciprocal_division.c
 SRCS-y += test_reciprocal_division_perf.c
 SRCS-y += test_fbarray.c
 SRCS-y += test_external_mem.c
+SRCS-y += test_rand_perf.c
 
 SRCS-y += test_ring.c
 SRCS-y += test_ring_perf.c
index f1db02f..562b93e 100644 (file)
@@ -90,6 +90,7 @@ test_sources = files('commands.c',
        'test_power_acpi_cpufreq.c',
        'test_power_kvm_vm.c',
        'test_prefetch.c',
+       'test_rand_perf.c',
        'test_rcu_qsbr.c',
        'test_rcu_qsbr_perf.c',
        'test_reciprocal_division.c',
@@ -256,6 +257,7 @@ perf_test_names = [
         'pmd_perf_autotest',
         'stack_perf_autotest',
         'stack_lf_perf_autotest',
+        'rand_perf_autotest',
 ]
 
 driver_test_names = [
diff --git a/app/test/test_rand_perf.c b/app/test/test_rand_perf.c
new file mode 100644 (file)
index 0000000..7717137
--- /dev/null
@@ -0,0 +1,75 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Ericsson AB
+ */
+
+#include <inttypes.h>
+#include <stdio.h>
+
+#include <rte_common.h>
+#include <rte_cycles.h>
+#include <rte_random.h>
+
+#include "test.h"
+
+static volatile uint64_t vsum;
+
+#define ITERATIONS (100000000)
+
+enum rand_type {
+       rand_type_64
+};
+
+static const char *
+rand_type_desc(enum rand_type rand_type)
+{
+       switch (rand_type) {
+       case rand_type_64:
+               return "Full 64-bit [rte_rand()]";
+       default:
+               return NULL;
+       }
+}
+
+static __rte_always_inline void
+test_rand_perf_type(enum rand_type rand_type)
+{
+       uint64_t start;
+       uint32_t i;
+       uint64_t end;
+       uint64_t sum = 0;
+       uint64_t op_latency;
+
+       start = rte_rdtsc();
+
+       for (i = 0; i < ITERATIONS; i++) {
+               switch (rand_type) {
+               case rand_type_64:
+                       sum += rte_rand();
+                       break;
+               }
+       }
+
+       end = rte_rdtsc();
+
+       /* to avoid an optimizing compiler removing the whole loop */
+       vsum = sum;
+
+       op_latency = (end - start) / ITERATIONS;
+
+       printf("%s: %"PRId64" TSC cycles/op\n", rand_type_desc(rand_type),
+              op_latency);
+}
+
+static int
+test_rand_perf(void)
+{
+       rte_srand(42);
+
+       printf("Pseudo-random number generation latencies:\n");
+
+       test_rand_perf_type(rand_type_64);
+
+       return 0;
+}
+
+REGISTER_TEST_COMMAND(rand_perf_autotest, test_rand_perf);
index 888ab19..5928f22 100644 (file)
@@ -54,6 +54,16 @@ New Features
      Also, make sure to start the actual text at the margin.
      =========================================================
 
+* **Updated the EAL Pseudo-random Number Generator.**
+
+  The lrand48()-based rte_rand() function is replaced with a
+  DPDK-native combined Linear Feedback Shift Register (LFSR)
+  pseudo-random number generator (PRNG).
+
+  This new PRNG implementation is multi-thread safe, provides
+  higher-quality pseudo-random numbers (including full 64 bit
+  support) and improved performance.
+
 * **Updated the bnxt PMD.**
 
   Updated the bnxt PMD. The major enhancements include:
index b2ca1c2..66dfe8a 100644 (file)
@@ -16,7 +16,6 @@ extern "C" {
 #endif
 
 #include <stdint.h>
-#include <stdlib.h>
 
 /**
  * Seed the pseudo-random generator.
@@ -25,34 +24,28 @@ extern "C" {
  * value. It may need to be re-seeded by the user with a real random
  * value.
  *
+ * This function is not multi-thread safe in regards to other
+ * rte_srand() calls, nor is it in relation to concurrent rte_rand()
+ * calls.
+ *
  * @param seedval
  *   The value of the seed.
  */
-static inline void
-rte_srand(uint64_t seedval)
-{
-       srand48((long)seedval);
-}
+void
+rte_srand(uint64_t seedval);
 
 /**
  * Get a pseudo-random value.
  *
- * This function generates pseudo-random numbers using the linear
- * congruential algorithm and 48-bit integer arithmetic, called twice
- * to generate a 64-bit value.
+ * The generator is not cryptographically secure.
+ *
+ * If called from lcore threads, this function is thread-safe.
  *
  * @return
  *   A pseudo-random value between 0 and (1<<64)-1.
  */
-static inline uint64_t
-rte_rand(void)
-{
-       uint64_t val;
-       val = (uint64_t)lrand48();
-       val <<= 32;
-       val += (uint64_t)lrand48();
-       return val;
-}
+uint64_t
+rte_rand(void);
 
 #ifdef __cplusplus
 }
index 0670e41..bafd232 100644 (file)
@@ -35,6 +35,7 @@ common_sources = files(
        'rte_keepalive.c',
        'rte_malloc.c',
        'rte_option.c',
+       'rte_random.c',
        'rte_reciprocal.c',
        'rte_service.c'
 )
diff --git a/lib/librte_eal/common/rte_random.c b/lib/librte_eal/common/rte_random.c
new file mode 100644 (file)
index 0000000..4d3cf52
--- /dev/null
@@ -0,0 +1,139 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Ericsson AB
+ */
+
+#include <stdlib.h>
+
+#include <rte_branch_prediction.h>
+#include <rte_cycles.h>
+#include <rte_eal.h>
+#include <rte_lcore.h>
+#include <rte_memory.h>
+#include <rte_random.h>
+
+struct rte_rand_state {
+       uint64_t z1;
+       uint64_t z2;
+       uint64_t z3;
+       uint64_t z4;
+       uint64_t z5;
+} __rte_cache_aligned;
+
+static struct rte_rand_state rand_states[RTE_MAX_LCORE];
+
+static uint32_t
+__rte_rand_lcg32(uint32_t *seed)
+{
+       *seed = 1103515245U * *seed + 12345U;
+
+       return *seed;
+}
+
+static uint64_t
+__rte_rand_lcg64(uint32_t *seed)
+{
+       uint64_t low;
+       uint64_t high;
+
+       /* A 64-bit LCG would have been much cleaner, but good
+        * multiplier/increments for such seem hard to come by.
+        */
+
+       low = __rte_rand_lcg32(seed);
+       high = __rte_rand_lcg32(seed);
+
+       return low | (high << 32);
+}
+
+static uint64_t
+__rte_rand_lfsr258_gen_seed(uint32_t *seed, uint64_t min_value)
+{
+       uint64_t res;
+
+       res = __rte_rand_lcg64(seed);
+
+       if (res < min_value)
+               res += min_value;
+
+       return res;
+}
+
+static void
+__rte_srand_lfsr258(uint64_t seed, struct rte_rand_state *state)
+{
+       uint32_t lcg_seed;
+
+       lcg_seed = (uint32_t)(seed ^ (seed >> 32));
+
+       state->z1 = __rte_rand_lfsr258_gen_seed(&lcg_seed, 2UL);
+       state->z2 = __rte_rand_lfsr258_gen_seed(&lcg_seed, 512UL);
+       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);
+}
+
+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]);
+}
+
+static __rte_always_inline uint64_t
+__rte_rand_lfsr258_comp(uint64_t z, uint64_t a, uint64_t b, uint64_t c,
+                       uint64_t d)
+{
+       return ((z & c) << d) ^ (((z << a) ^ z) >> b);
+}
+
+/* Based on L’Ecuyer, P.: Tables of maximally equidistributed combined
+ * LFSR generators.
+ */
+
+static __rte_always_inline uint64_t
+__rte_rand_lfsr258(struct rte_rand_state *state)
+{
+       state->z1 = __rte_rand_lfsr258_comp(state->z1, 1UL, 53UL,
+                                           18446744073709551614UL, 10UL);
+       state->z2 = __rte_rand_lfsr258_comp(state->z2, 24UL, 50UL,
+                                           18446744073709551104UL, 5UL);
+       state->z3 = __rte_rand_lfsr258_comp(state->z3, 3UL, 23UL,
+                                           18446744073709547520UL, 29UL);
+       state->z4 = __rte_rand_lfsr258_comp(state->z4, 5UL, 24UL,
+                                           18446744073709420544UL, 23UL);
+       state->z5 = __rte_rand_lfsr258_comp(state->z5, 3UL, 33UL,
+                                           18446744073701163008UL, 8UL);
+
+       return state->z1 ^ state->z2 ^ state->z3 ^ state->z4 ^ state->z5;
+}
+
+static __rte_always_inline
+struct rte_rand_state *__rte_rand_get_state(void)
+{
+       unsigned int lcore_id;
+
+       lcore_id = rte_lcore_id();
+
+       if (unlikely(lcore_id == LCORE_ID_ANY))
+               lcore_id = rte_get_master_lcore();
+
+       return &rand_states[lcore_id];
+}
+
+uint64_t
+rte_rand(void)
+{
+       struct rte_rand_state *state;
+
+       state = __rte_rand_get_state();
+
+       return __rte_rand_lfsr258(state);
+}
+
+RTE_INIT(rte_rand_init)
+{
+       rte_srand(rte_get_timer_cycles());
+}
index 19854ee..ca616c4 100644 (file)
@@ -69,6 +69,7 @@ SRCS-$(CONFIG_RTE_EXEC_ENV_FREEBSD) += malloc_mp.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_FREEBSD) += rte_keepalive.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_FREEBSD) += rte_option.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_FREEBSD) += rte_service.c
+SRCS-$(CONFIG_RTE_EXEC_ENV_FREEBSD) += rte_random.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_FREEBSD) += rte_reciprocal.c
 
 # from arch dir
index 1364d55..2c77a86 100644 (file)
@@ -758,8 +758,6 @@ rte_eal_init(int argc, char **argv)
 #endif
        }
 
-       rte_srand(rte_rdtsc());
-
        /* in secondary processes, memory init may allocate additional fbarrays
         * not present in primary processes, so to avoid any potential issues,
         * initialize memzones first.
index 6e52611..729795a 100644 (file)
@@ -77,6 +77,7 @@ SRCS-$(CONFIG_RTE_EXEC_ENV_LINUX) += malloc_mp.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUX) += rte_keepalive.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUX) += rte_option.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUX) += rte_service.c
+SRCS-$(CONFIG_RTE_EXEC_ENV_LINUX) += rte_random.c
 SRCS-$(CONFIG_RTE_EXEC_ENV_LINUX) += rte_reciprocal.c
 
 # from arch dir
index dfbbedd..aa0137e 100644 (file)
@@ -1122,8 +1122,6 @@ rte_eal_init(int argc, char **argv)
 #endif
        }
 
-       rte_srand(rte_rdtsc());
-
        if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0) {
                rte_eal_init_alert("Cannot init logging.");
                rte_errno = ENOMEM;
index 824edf0..20c1a90 100644 (file)
@@ -292,6 +292,8 @@ DPDK_19.08 {
 
        rte_lcore_index;
        rte_lcore_to_socket_id;
+       rte_rand;
+       rte_srand;
 
 } DPDK_19.05;