1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2018 Intel Corporation
6 * This is a simple functional test for rte_smp_mb() implementation.
7 * I.E. make sure that LOAD and STORE operations that precede the
8 * rte_smp_mb() call are globally visible across the lcores
9 * before the the LOAD and STORE operations that follows it.
10 * The test uses simple implementation of Peterson's lock algorithm
11 * (https://en.wikipedia.org/wiki/Peterson%27s_algorithm)
12 * for two execution units to make sure that rte_smp_mb() prevents
13 * store-load reordering to happen.
14 * Also when executed on a single lcore could be used as a approxiamate
15 * estimation of number of cycles particular implementation of rte_smp_mb()
24 #include <rte_memory.h>
25 #include <rte_per_lcore.h>
26 #include <rte_launch.h>
27 #include <rte_atomic.h>
29 #include <rte_lcore.h>
30 #include <rte_pause.h>
31 #include <rte_random.h>
32 #include <rte_cycles.h>
34 #include <rte_debug.h>
39 #define ITER_MAX 0x1000000
48 volatile uint32_t flag[2];
49 volatile uint32_t victim;
50 enum plock_use_type utype;
54 * Lock plus protected by it two counters.
63 * Each active lcore shares plock_test struct with it's left and right
66 struct lcore_plock_test {
67 struct plock_test *pt[2]; /* shared, lock-protected data */
68 uint32_t sum[2]; /* local copy of the shared data */
69 uint32_t iter; /* number of iterations to perfom */
70 uint32_t lc; /* given lcore id */
74 store_load_barrier(uint32_t utype)
78 else if (utype == USE_SMP_MB)
85 * Peterson lock implementation.
88 plock_lock(struct plock *l, uint32_t self)
97 store_load_barrier(l->utype);
99 while (l->flag[other] == 1 && l->victim == self)
104 plock_unlock(struct plock *l, uint32_t self)
111 plock_reset(struct plock *l, enum plock_use_type utype)
113 memset(l, 0, sizeof(*l));
118 * grab the lock, update both counters, release the lock.
121 plock_add(struct plock_test *pt, uint32_t self, uint32_t n)
123 plock_lock(&pt->lock, self);
126 plock_unlock(&pt->lock, self);
130 plock_test1_lcore(void *data)
133 uint32_t i, lc, ln, n;
134 struct lcore_plock_test *lpt;
139 /* find lcore_plock_test struct for given lcore */
140 for (ln = rte_lcore_count(); ln != 0 && lpt->lc != lc; lpt++, ln--)
144 printf("%s(%u) error at init\n", __func__, lc);
148 n = rte_rand() % ADD_MAX;
149 tm = rte_get_timer_cycles();
152 * for each iteration:
153 * - update shared, locked protected data in a safe manner
154 * - update local copy of the shared data
156 for (i = 0; i != lpt->iter; i++) {
158 plock_add(lpt->pt[0], 0, n);
159 plock_add(lpt->pt[1], 1, n);
164 n = (n + 1) % ADD_MAX;
167 tm = rte_get_timer_cycles() - tm;
169 printf("%s(%u): %u iterations finished, in %" PRIu64
170 " cycles, %#Lf cycles/iteration, "
171 "local sum={%u, %u}\n",
172 __func__, lc, i, tm, (long double)tm / i,
173 lpt->sum[0], lpt->sum[1]);
178 * For N active lcores we allocate N+1 lcore_plock_test structures.
179 * Each active lcore shares one lcore_plock_test structure with its
180 * left lcore neighbor and one lcore_plock_test structure with its
181 * right lcore neighbor.
182 * During the test each lcore updates data in both shared structures and
183 * its local copies. Then at validation phase we check that our shared
184 * and local data are the same.
187 plock_test(uint32_t iter, enum plock_use_type utype)
192 struct plock_test *pt;
193 struct lcore_plock_test *lpt;
195 /* init phase, allocate and initialize shared data */
197 n = rte_lcore_count();
198 pt = calloc(n + 1, sizeof(*pt));
199 lpt = calloc(n, sizeof(*lpt));
200 sum = calloc(n + 1, sizeof(*sum));
202 printf("%s(iter=%u, utype=%u) started on %u lcores\n",
203 __func__, iter, utype, n);
205 if (pt == NULL || lpt == NULL) {
206 printf("%s: failed to allocate memory for %u lcores\n",
214 for (i = 0; i != n + 1; i++)
215 plock_reset(&pt[i].lock, utype);
218 RTE_LCORE_FOREACH(lc) {
222 lpt[i].pt[0] = pt + i;
223 lpt[i].pt[1] = pt + i + 1;
227 lpt[i - 1].pt[1] = pt;
229 for (i = 0; i != n; i++)
230 printf("lpt[%u]={lc=%u, pt={%p, %p},};\n",
231 i, lpt[i].lc, lpt[i].pt[0], lpt[i].pt[1]);
234 /* test phase - start and wait for completion on each active lcore */
236 rte_eal_mp_remote_launch(plock_test1_lcore, lpt, CALL_MASTER);
237 rte_eal_mp_wait_lcore();
239 /* validation phase - make sure that shared and local data match */
241 for (i = 0; i != n; i++) {
242 sum[i] += lpt[i].sum[0];
243 sum[i + 1] += lpt[i].sum[1];
249 for (i = 0; i != n; i++) {
250 printf("%s: sum[%u]=%u, pt[%u].val=%u, pt[%u].iter=%u;\n",
251 __func__, i, sum[i], i, pt[i].val, i, pt[i].iter);
253 /* race condition occurred, lock doesn't work properly */
254 if (sum[i] != pt[i].val || 2 * iter != pt[i].iter) {
255 printf("error: local and shared sums don't much\n");
264 printf("%s(utype=%u) returns %d\n", __func__, utype, rc);
271 int32_t i, ret, rc[USE_NUM];
273 for (i = 0; i != RTE_DIM(rc); i++)
274 rc[i] = plock_test(ITER_MAX, i);
277 for (i = 0; i != RTE_DIM(rc); i++) {
278 printf("%s for utype=%d %s\n",
279 __func__, i, rc[i] == 0 ? "passed" : "failed");
286 REGISTER_TEST_COMMAND(barrier_autotest, test_barrier);