xref: /dpdk/app/test/test_barrier.c (revision 68a03efeed657e6e05f281479b33b51102797e15)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2018 Intel Corporation
3  */
4 
5  /*
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()
16   * will take.
17   */
18 
19 #include <stdio.h>
20 #include <string.h>
21 #include <stdint.h>
22 #include <inttypes.h>
23 
24 #include <rte_memory.h>
25 #include <rte_per_lcore.h>
26 #include <rte_launch.h>
27 #include <rte_atomic.h>
28 #include <rte_eal.h>
29 #include <rte_lcore.h>
30 #include <rte_pause.h>
31 #include <rte_random.h>
32 #include <rte_cycles.h>
33 #include <rte_vect.h>
34 #include <rte_debug.h>
35 
36 #include "test.h"
37 
38 #define ADD_MAX		8
39 #define ITER_MAX	0x1000000
40 
41 enum plock_use_type {
42 	USE_MB,
43 	USE_SMP_MB,
44 	USE_NUM
45 };
46 
47 struct plock {
48 	volatile uint32_t flag[2];
49 	volatile uint32_t victim;
50 	enum plock_use_type utype;
51 };
52 
53 /*
54  * Lock plus protected by it two counters.
55  */
56 struct plock_test {
57 	struct plock lock;
58 	uint64_t val;
59 	uint64_t iter;
60 };
61 
62 /*
63  * Each active lcore shares plock_test struct with it's left and right
64  * neighbours.
65  */
66 struct lcore_plock_test {
67 	struct plock_test *pt[2]; /* shared, lock-protected data */
68 	uint64_t sum[2];          /* local copy of the shared data */
69 	uint64_t iter;            /* number of iterations to perfom */
70 	uint32_t lc;              /* given lcore id */
71 };
72 
73 static inline void
74 store_load_barrier(uint32_t utype)
75 {
76 	if (utype == USE_MB)
77 		rte_mb();
78 	else if (utype == USE_SMP_MB)
79 		rte_smp_mb();
80 	else
81 		RTE_VERIFY(0);
82 }
83 
84 /*
85  * Peterson lock implementation.
86  */
87 static void
88 plock_lock(struct plock *l, uint32_t self)
89 {
90 	uint32_t other;
91 
92 	other = self ^ 1;
93 
94 	l->flag[self] = 1;
95 	rte_smp_wmb();
96 	l->victim = self;
97 
98 	store_load_barrier(l->utype);
99 
100 	while (l->flag[other] == 1 && l->victim == self)
101 		rte_pause();
102 	rte_smp_rmb();
103 }
104 
105 static void
106 plock_unlock(struct plock *l, uint32_t self)
107 {
108 	rte_smp_wmb();
109 	l->flag[self] = 0;
110 }
111 
112 static void
113 plock_reset(struct plock *l, enum plock_use_type utype)
114 {
115 	memset(l, 0, sizeof(*l));
116 	l->utype = utype;
117 }
118 
119 /*
120  * grab the lock, update both counters, release the lock.
121  */
122 static void
123 plock_add(struct plock_test *pt, uint32_t self, uint32_t n)
124 {
125 	plock_lock(&pt->lock, self);
126 	pt->iter++;
127 	pt->val += n;
128 	plock_unlock(&pt->lock, self);
129 }
130 
131 static int
132 plock_test1_lcore(void *data)
133 {
134 	uint64_t tm;
135 	uint32_t lc, ln;
136 	uint64_t i, n;
137 	struct lcore_plock_test *lpt;
138 
139 	lpt = data;
140 	lc = rte_lcore_id();
141 
142 	/* find lcore_plock_test struct for given lcore */
143 	for (ln = rte_lcore_count(); ln != 0 && lpt->lc != lc; lpt++, ln--)
144 		;
145 
146 	if (ln == 0) {
147 		printf("%s(%u) error at init\n", __func__, lc);
148 		return -1;
149 	}
150 
151 	n = rte_rand() % ADD_MAX;
152 	tm = rte_get_timer_cycles();
153 
154 	/*
155 	 * for each iteration:
156 	 * - update shared, locked protected data in a safe manner
157 	 * - update local copy of the shared data
158 	 */
159 	for (i = 0; i != lpt->iter; i++) {
160 
161 		plock_add(lpt->pt[0], 0, n);
162 		plock_add(lpt->pt[1], 1, n);
163 
164 		lpt->sum[0] += n;
165 		lpt->sum[1] += n;
166 
167 		n = (n + 1) % ADD_MAX;
168 	}
169 
170 	tm = rte_get_timer_cycles() - tm;
171 
172 	printf("%s(%u): %" PRIu64 " iterations finished, in %" PRIu64
173 		" cycles, %#Lf cycles/iteration, "
174 		"local sum={%" PRIu64 ", %" PRIu64 "}\n",
175 		__func__, lc, i, tm, (long double)tm / i,
176 		lpt->sum[0], lpt->sum[1]);
177 	return 0;
178 }
179 
180 /*
181  * For N active lcores we allocate N+1 lcore_plock_test structures.
182  * Each active lcore shares one lcore_plock_test structure with its
183  * left lcore neighbor and one lcore_plock_test structure with its
184  * right lcore neighbor.
185  * During the test each lcore updates data in both shared structures and
186  * its local copies. Then at validation phase we check that our shared
187  * and local data are the same.
188  */
189 static int
190 plock_test(uint64_t iter, enum plock_use_type utype)
191 {
192 	int32_t rc;
193 	uint32_t i, lc, n;
194 	uint64_t *sum;
195 	struct plock_test *pt;
196 	struct lcore_plock_test *lpt;
197 
198 	/* init phase, allocate and initialize shared data */
199 
200 	n = rte_lcore_count();
201 	pt = calloc(n + 1, sizeof(*pt));
202 	lpt = calloc(n, sizeof(*lpt));
203 	sum = calloc(n + 1, sizeof(*sum));
204 
205 	printf("%s(iter=%" PRIu64 ", utype=%u) started on %u lcores\n",
206 		__func__, iter, utype, n);
207 
208 	if (pt == NULL || lpt == NULL || sum == NULL) {
209 		printf("%s: failed to allocate memory for %u lcores\n",
210 			__func__, n);
211 		free(pt);
212 		free(lpt);
213 		free(sum);
214 		return -ENOMEM;
215 	}
216 
217 	for (i = 0; i != n + 1; i++)
218 		plock_reset(&pt[i].lock, utype);
219 
220 	i = 0;
221 	RTE_LCORE_FOREACH(lc) {
222 
223 		lpt[i].lc = lc;
224 		lpt[i].iter = iter;
225 		lpt[i].pt[0] = pt + i;
226 		lpt[i].pt[1] = pt + i + 1;
227 		i++;
228 	}
229 
230 	lpt[i - 1].pt[1] = pt;
231 
232 	for (i = 0; i != n; i++)
233 		printf("lpt[%u]={lc=%u, pt={%p, %p},};\n",
234 			i, lpt[i].lc, lpt[i].pt[0], lpt[i].pt[1]);
235 
236 
237 	/* test phase - start and wait for completion on each active lcore */
238 
239 	rte_eal_mp_remote_launch(plock_test1_lcore, lpt, CALL_MAIN);
240 	rte_eal_mp_wait_lcore();
241 
242 	/* validation phase - make sure that shared and local data match */
243 
244 	for (i = 0; i != n; i++) {
245 		sum[i] += lpt[i].sum[0];
246 		sum[i + 1] += lpt[i].sum[1];
247 	}
248 
249 	sum[0] += sum[i];
250 
251 	rc = 0;
252 	for (i = 0; i != n; i++) {
253 		printf("%s: sum[%u]=%" PRIu64 ", pt[%u].val=%" PRIu64 ", pt[%u].iter=%" PRIu64 ";\n",
254 			__func__, i, sum[i], i, pt[i].val, i, pt[i].iter);
255 
256 		/* race condition occurred, lock doesn't work properly */
257 		if (sum[i] != pt[i].val || 2 * iter != pt[i].iter) {
258 			printf("error: local and shared sums don't match\n");
259 			rc = -1;
260 		}
261 	}
262 
263 	free(pt);
264 	free(lpt);
265 	free(sum);
266 
267 	printf("%s(utype=%u) returns %d\n", __func__, utype, rc);
268 	return rc;
269 }
270 
271 static int
272 test_barrier(void)
273 {
274 	int32_t i, ret, rc[USE_NUM];
275 
276 	for (i = 0; i != RTE_DIM(rc); i++)
277 		rc[i] = plock_test(ITER_MAX, i);
278 
279 	ret = 0;
280 	for (i = 0; i != RTE_DIM(rc); i++) {
281 		printf("%s for utype=%d %s\n",
282 			__func__, i, rc[i] == 0 ? "passed" : "failed");
283 		ret |= rc[i];
284 	}
285 
286 	return ret;
287 }
288 
289 REGISTER_TEST_COMMAND(barrier_autotest, test_barrier);
290