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