test/ring: add contention stress test
authorKonstantin Ananyev <konstantin.ananyev@intel.com>
Mon, 20 Apr 2020 12:28:22 +0000 (13:28 +0100)
committerDavid Marchand <david.marchand@redhat.com>
Tue, 21 Apr 2020 09:34:09 +0000 (11:34 +0200)
Introduce stress test for ring enqueue/dequeue operations.
Performs the following pattern on each slave worker:
dequeue/read-write data from the dequeued objects/enqueue.
Serves as both functional and performance test of ring
enqueue/dequeue operations under high contention
(for both over committed and non-over committed scenarios).

Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Acked-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
app/test/Makefile
app/test/meson.build
app/test/test_ring_mpmc_stress.c [new file with mode: 0644]
app/test/test_ring_stress.c [new file with mode: 0644]
app/test/test_ring_stress.h [new file with mode: 0644]
app/test/test_ring_stress_impl.h [new file with mode: 0644]

index be53d33..a23a011 100644 (file)
@@ -77,7 +77,9 @@ SRCS-y += test_external_mem.c
 SRCS-y += test_rand_perf.c
 
 SRCS-y += test_ring.c
+SRCS-y += test_ring_mpmc_stress.c
 SRCS-y += test_ring_perf.c
+SRCS-y += test_ring_stress.c
 SRCS-y += test_pmd_perf.c
 
 ifeq ($(CONFIG_RTE_LIBRTE_TABLE),y)
index 04b59cf..8824f36 100644 (file)
@@ -100,7 +100,9 @@ test_sources = files('commands.c',
        'test_rib.c',
        'test_rib6.c',
        'test_ring.c',
+       'test_ring_mpmc_stress.c',
        'test_ring_perf.c',
+       'test_ring_stress.c',
        'test_rwlock.c',
        'test_sched.c',
        'test_service_cores.c',
diff --git a/app/test/test_ring_mpmc_stress.c b/app/test/test_ring_mpmc_stress.c
new file mode 100644 (file)
index 0000000..1524b02
--- /dev/null
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "test_ring_stress_impl.h"
+
+static inline uint32_t
+_st_ring_dequeue_bulk(struct rte_ring *r, void **obj, uint32_t n,
+       uint32_t *avail)
+{
+       return rte_ring_mc_dequeue_bulk(r, obj, n, avail);
+}
+
+static inline uint32_t
+_st_ring_enqueue_bulk(struct rte_ring *r, void * const *obj, uint32_t n,
+       uint32_t *free)
+{
+       return rte_ring_mp_enqueue_bulk(r, obj, n, free);
+}
+
+static int
+_st_ring_init(struct rte_ring *r, const char *name, uint32_t num)
+{
+       return rte_ring_init(r, name, num, 0);
+}
+
+const struct test test_ring_mpmc_stress = {
+       .name = "MP/MC",
+       .nb_case = RTE_DIM(tests),
+       .cases = tests,
+};
diff --git a/app/test/test_ring_stress.c b/app/test/test_ring_stress.c
new file mode 100644 (file)
index 0000000..60706f7
--- /dev/null
@@ -0,0 +1,48 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "test_ring_stress.h"
+
+static int
+run_test(const struct test *test)
+{
+       int32_t rc;
+       uint32_t i, k;
+
+       for (i = 0, k = 0; i != test->nb_case; i++) {
+
+               printf("TEST-CASE %s %s START\n",
+                       test->name, test->cases[i].name);
+
+               rc = test->cases[i].func(test->cases[i].wfunc);
+               k += (rc == 0);
+
+               if (rc != 0)
+                       printf("TEST-CASE %s %s FAILED\n",
+                               test->name, test->cases[i].name);
+               else
+                       printf("TEST-CASE %s %s OK\n",
+                               test->name, test->cases[i].name);
+       }
+
+       return k;
+}
+
+static int
+test_ring_stress(void)
+{
+       uint32_t n, k;
+
+       n = 0;
+       k = 0;
+
+       n += test_ring_mpmc_stress.nb_case;
+       k += run_test(&test_ring_mpmc_stress);
+
+       printf("Number of tests:\t%u\nSuccess:\t%u\nFailed:\t%u\n",
+               n, k, n - k);
+       return (k != n);
+}
+
+REGISTER_TEST_COMMAND(ring_stress_autotest, test_ring_stress);
diff --git a/app/test/test_ring_stress.h b/app/test/test_ring_stress.h
new file mode 100644 (file)
index 0000000..60eac62
--- /dev/null
@@ -0,0 +1,35 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+
+#include <inttypes.h>
+#include <stddef.h>
+#include <stdalign.h>
+#include <string.h>
+#include <stdio.h>
+#include <unistd.h>
+
+#include <rte_ring.h>
+#include <rte_cycles.h>
+#include <rte_launch.h>
+#include <rte_pause.h>
+#include <rte_random.h>
+#include <rte_malloc.h>
+#include <rte_spinlock.h>
+
+#include "test.h"
+
+struct test_case {
+       const char *name;
+       int (*func)(int (*)(void *));
+       int (*wfunc)(void *arg);
+};
+
+struct test {
+       const char *name;
+       uint32_t nb_case;
+       const struct test_case *cases;
+};
+
+extern const struct test test_ring_mpmc_stress;
diff --git a/app/test/test_ring_stress_impl.h b/app/test/test_ring_stress_impl.h
new file mode 100644 (file)
index 0000000..222d62b
--- /dev/null
@@ -0,0 +1,396 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include "test_ring_stress.h"
+
+/**
+ * Stress test for ring enqueue/dequeue operations.
+ * Performs the following pattern on each slave worker:
+ * dequeue/read-write data from the dequeued objects/enqueue.
+ * Serves as both functional and performance test of ring
+ * enqueue/dequeue operations under high contention
+ * (for both over committed and non-over committed scenarios).
+ */
+
+#define RING_NAME      "RING_STRESS"
+#define BULK_NUM       32
+#define RING_SIZE      (2 * BULK_NUM * RTE_MAX_LCORE)
+
+enum {
+       WRK_CMD_STOP,
+       WRK_CMD_RUN,
+};
+
+static volatile uint32_t wrk_cmd __rte_cache_aligned;
+
+/* test run-time in seconds */
+static const uint32_t run_time = 60;
+static const uint32_t verbose;
+
+struct lcore_stat {
+       uint64_t nb_cycle;
+       struct {
+               uint64_t nb_call;
+               uint64_t nb_obj;
+               uint64_t nb_cycle;
+               uint64_t max_cycle;
+               uint64_t min_cycle;
+       } op;
+};
+
+struct lcore_arg {
+       struct rte_ring *rng;
+       struct lcore_stat stats;
+} __rte_cache_aligned;
+
+struct ring_elem {
+       uint32_t cnt[RTE_CACHE_LINE_SIZE / sizeof(uint32_t)];
+} __rte_cache_aligned;
+
+/*
+ * redefinable functions
+ */
+static uint32_t
+_st_ring_dequeue_bulk(struct rte_ring *r, void **obj, uint32_t n,
+       uint32_t *avail);
+
+static uint32_t
+_st_ring_enqueue_bulk(struct rte_ring *r, void * const *obj, uint32_t n,
+       uint32_t *free);
+
+static int
+_st_ring_init(struct rte_ring *r, const char *name, uint32_t num);
+
+
+static void
+lcore_stat_update(struct lcore_stat *ls, uint64_t call, uint64_t obj,
+       uint64_t tm, int32_t prcs)
+{
+       ls->op.nb_call += call;
+       ls->op.nb_obj += obj;
+       ls->op.nb_cycle += tm;
+       if (prcs) {
+               ls->op.max_cycle = RTE_MAX(ls->op.max_cycle, tm);
+               ls->op.min_cycle = RTE_MIN(ls->op.min_cycle, tm);
+       }
+}
+
+static void
+lcore_op_stat_aggr(struct lcore_stat *ms, const struct lcore_stat *ls)
+{
+
+       ms->op.nb_call += ls->op.nb_call;
+       ms->op.nb_obj += ls->op.nb_obj;
+       ms->op.nb_cycle += ls->op.nb_cycle;
+       ms->op.max_cycle = RTE_MAX(ms->op.max_cycle, ls->op.max_cycle);
+       ms->op.min_cycle = RTE_MIN(ms->op.min_cycle, ls->op.min_cycle);
+}
+
+static void
+lcore_stat_aggr(struct lcore_stat *ms, const struct lcore_stat *ls)
+{
+       ms->nb_cycle = RTE_MAX(ms->nb_cycle, ls->nb_cycle);
+       lcore_op_stat_aggr(ms, ls);
+}
+
+static void
+lcore_stat_dump(FILE *f, uint32_t lc, const struct lcore_stat *ls)
+{
+       long double st;
+
+       st = (long double)rte_get_timer_hz() / US_PER_S;
+
+       if (lc == UINT32_MAX)
+               fprintf(f, "%s(AGGREGATE)={\n", __func__);
+       else
+               fprintf(f, "%s(lcore=%u)={\n", __func__, lc);
+
+       fprintf(f, "\tnb_cycle=%" PRIu64 "(%.2Lf usec),\n",
+               ls->nb_cycle, (long double)ls->nb_cycle / st);
+
+       fprintf(f, "\tDEQ+ENQ={\n");
+
+       fprintf(f, "\t\tnb_call=%" PRIu64 ",\n", ls->op.nb_call);
+       fprintf(f, "\t\tnb_obj=%" PRIu64 ",\n", ls->op.nb_obj);
+       fprintf(f, "\t\tnb_cycle=%" PRIu64 ",\n", ls->op.nb_cycle);
+       fprintf(f, "\t\tobj/call(avg): %.2Lf\n",
+               (long double)ls->op.nb_obj / ls->op.nb_call);
+       fprintf(f, "\t\tcycles/obj(avg): %.2Lf\n",
+               (long double)ls->op.nb_cycle / ls->op.nb_obj);
+       fprintf(f, "\t\tcycles/call(avg): %.2Lf\n",
+               (long double)ls->op.nb_cycle / ls->op.nb_call);
+
+       /* if min/max cycles per call stats was collected */
+       if (ls->op.min_cycle != UINT64_MAX) {
+               fprintf(f, "\t\tmax cycles/call=%" PRIu64 "(%.2Lf usec),\n",
+                       ls->op.max_cycle,
+                       (long double)ls->op.max_cycle / st);
+               fprintf(f, "\t\tmin cycles/call=%" PRIu64 "(%.2Lf usec),\n",
+                       ls->op.min_cycle,
+                       (long double)ls->op.min_cycle / st);
+       }
+
+       fprintf(f, "\t},\n");
+       fprintf(f, "};\n");
+}
+
+static void
+fill_ring_elm(struct ring_elem *elm, uint32_t fill)
+{
+       uint32_t i;
+
+       for (i = 0; i != RTE_DIM(elm->cnt); i++)
+               elm->cnt[i] = fill;
+}
+
+static int32_t
+check_updt_elem(struct ring_elem *elm[], uint32_t num,
+       const struct ring_elem *check, const struct ring_elem *fill)
+{
+       uint32_t i;
+
+       static rte_spinlock_t dump_lock;
+
+       for (i = 0; i != num; i++) {
+               if (memcmp(check, elm[i], sizeof(*check)) != 0) {
+                       rte_spinlock_lock(&dump_lock);
+                       printf("%s(lc=%u, num=%u) failed at %u-th iter, "
+                               "offending object: %p\n",
+                               __func__, rte_lcore_id(), num, i, elm[i]);
+                       rte_memdump(stdout, "expected", check, sizeof(*check));
+                       rte_memdump(stdout, "result", elm[i], sizeof(elm[i]));
+                       rte_spinlock_unlock(&dump_lock);
+                       return -EINVAL;
+               }
+               memcpy(elm[i], fill, sizeof(*elm[i]));
+       }
+
+       return 0;
+}
+
+static int
+check_ring_op(uint32_t exp, uint32_t res, uint32_t lc,
+       const char *fname, const char *opname)
+{
+       if (exp != res) {
+               printf("%s(lc=%u) failure: %s expected: %u, returned %u\n",
+                       fname, lc, opname, exp, res);
+               return -ENOSPC;
+       }
+       return 0;
+}
+
+static int
+test_worker(void *arg, const char *fname, int32_t prcs)
+{
+       int32_t rc;
+       uint32_t lc, n, num;
+       uint64_t cl, tm0, tm1;
+       struct lcore_arg *la;
+       struct ring_elem def_elm, loc_elm;
+       struct ring_elem *obj[2 * BULK_NUM];
+
+       la = arg;
+       lc = rte_lcore_id();
+
+       fill_ring_elm(&def_elm, UINT32_MAX);
+       fill_ring_elm(&loc_elm, lc);
+
+       while (wrk_cmd != WRK_CMD_RUN) {
+               rte_smp_rmb();
+               rte_pause();
+       }
+
+       cl = rte_rdtsc_precise();
+
+       do {
+               /* num in interval [7/8, 11/8] of BULK_NUM */
+               num = 7 * BULK_NUM / 8 + rte_rand() % (BULK_NUM / 2);
+
+               /* reset all pointer values */
+               memset(obj, 0, sizeof(obj));
+
+               /* dequeue num elems */
+               tm0 = (prcs != 0) ? rte_rdtsc_precise() : 0;
+               n = _st_ring_dequeue_bulk(la->rng, (void **)obj, num, NULL);
+               tm0 = (prcs != 0) ? rte_rdtsc_precise() - tm0 : 0;
+
+               /* check return value and objects */
+               rc = check_ring_op(num, n, lc, fname,
+                       RTE_STR(_st_ring_dequeue_bulk));
+               if (rc == 0)
+                       rc = check_updt_elem(obj, num, &def_elm, &loc_elm);
+               if (rc != 0)
+                       break;
+
+               /* enqueue num elems */
+               rte_compiler_barrier();
+               rc = check_updt_elem(obj, num, &loc_elm, &def_elm);
+               if (rc != 0)
+                       break;
+
+               tm1 = (prcs != 0) ? rte_rdtsc_precise() : 0;
+               n = _st_ring_enqueue_bulk(la->rng, (void **)obj, num, NULL);
+               tm1 = (prcs != 0) ? rte_rdtsc_precise() - tm1 : 0;
+
+               /* check return value */
+               rc = check_ring_op(num, n, lc, fname,
+                       RTE_STR(_st_ring_enqueue_bulk));
+               if (rc != 0)
+                       break;
+
+               lcore_stat_update(&la->stats, 1, num, tm0 + tm1, prcs);
+
+       } while (wrk_cmd == WRK_CMD_RUN);
+
+       cl = rte_rdtsc_precise() - cl;
+       if (prcs == 0)
+               lcore_stat_update(&la->stats, 0, 0, cl, 0);
+       la->stats.nb_cycle = cl;
+       return rc;
+}
+static int
+test_worker_prcs(void *arg)
+{
+       return test_worker(arg, __func__, 1);
+}
+
+static int
+test_worker_avg(void *arg)
+{
+       return test_worker(arg, __func__, 0);
+}
+
+static void
+mt1_fini(struct rte_ring *rng, void *data)
+{
+       rte_free(rng);
+       rte_free(data);
+}
+
+static int
+mt1_init(struct rte_ring **rng, void **data, uint32_t num)
+{
+       int32_t rc;
+       size_t sz;
+       uint32_t i, nr;
+       struct rte_ring *r;
+       struct ring_elem *elm;
+       void *p;
+
+       *rng = NULL;
+       *data = NULL;
+
+       sz = num * sizeof(*elm);
+       elm = rte_zmalloc(NULL, sz, __alignof__(*elm));
+       if (elm == NULL) {
+               printf("%s: alloc(%zu) for %u elems data failed",
+                       __func__, sz, num);
+               return -ENOMEM;
+       }
+
+       *data = elm;
+
+       /* alloc ring */
+       nr = 2 * num;
+       sz = rte_ring_get_memsize(nr);
+       r = rte_zmalloc(NULL, sz, __alignof__(*r));
+       if (r == NULL) {
+               printf("%s: alloc(%zu) for FIFO with %u elems failed",
+                       __func__, sz, nr);
+               return -ENOMEM;
+       }
+
+       *rng = r;
+
+       rc = _st_ring_init(r, RING_NAME, nr);
+       if (rc != 0) {
+               printf("%s: _st_ring_init(%p, %u) failed, error: %d(%s)\n",
+                       __func__, r, nr, rc, strerror(-rc));
+               return rc;
+       }
+
+       for (i = 0; i != num; i++) {
+               fill_ring_elm(elm + i, UINT32_MAX);
+               p = elm + i;
+               if (_st_ring_enqueue_bulk(r, &p, 1, NULL) != 1)
+                       break;
+       }
+
+       if (i != num) {
+               printf("%s: _st_ring_enqueue(%p, %u) returned %u\n",
+                       __func__, r, num, i);
+               return -ENOSPC;
+       }
+
+       return 0;
+}
+
+static int
+test_mt1(int (*test)(void *))
+{
+       int32_t rc;
+       uint32_t lc, mc;
+       struct rte_ring *r;
+       void *data;
+       struct lcore_arg arg[RTE_MAX_LCORE];
+
+       static const struct lcore_stat init_stat = {
+               .op.min_cycle = UINT64_MAX,
+       };
+
+       rc = mt1_init(&r, &data, RING_SIZE);
+       if (rc != 0) {
+               mt1_fini(r, data);
+               return rc;
+       }
+
+       memset(arg, 0, sizeof(arg));
+
+       /* launch on all slaves */
+       RTE_LCORE_FOREACH_SLAVE(lc) {
+               arg[lc].rng = r;
+               arg[lc].stats = init_stat;
+               rte_eal_remote_launch(test, &arg[lc], lc);
+       }
+
+       /* signal worker to start test */
+       wrk_cmd = WRK_CMD_RUN;
+       rte_smp_wmb();
+
+       usleep(run_time * US_PER_S);
+
+       /* signal worker to start test */
+       wrk_cmd = WRK_CMD_STOP;
+       rte_smp_wmb();
+
+       /* wait for slaves and collect stats. */
+       mc = rte_lcore_id();
+       arg[mc].stats = init_stat;
+
+       rc = 0;
+       RTE_LCORE_FOREACH_SLAVE(lc) {
+               rc |= rte_eal_wait_lcore(lc);
+               lcore_stat_aggr(&arg[mc].stats, &arg[lc].stats);
+               if (verbose != 0)
+                       lcore_stat_dump(stdout, lc, &arg[lc].stats);
+       }
+
+       lcore_stat_dump(stdout, UINT32_MAX, &arg[mc].stats);
+       mt1_fini(r, data);
+       return rc;
+}
+
+static const struct test_case tests[] = {
+       {
+               .name = "MT-WRK_ENQ_DEQ-MST_NONE-PRCS",
+               .func = test_mt1,
+               .wfunc = test_worker_prcs,
+       },
+       {
+               .name = "MT-WRK_ENQ_DEQ-MST_NONE-AVG",
+               .func = test_mt1,
+               .wfunc = test_worker_avg,
+       },
+};