mempool: fix slow allocation of large mempools
[dpdk.git] / app / test-crypto-perf / cperf_test_pmd_cyclecount.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2017 Intel Corporation
3  */
4
5 #include <stdbool.h>
6
7 #include <rte_crypto.h>
8 #include <rte_cryptodev.h>
9 #include <rte_cycles.h>
10 #include <rte_malloc.h>
11
12 #include "cperf_ops.h"
13 #include "cperf_test_pmd_cyclecount.h"
14 #include "cperf_test_common.h"
15
16 #define PRETTY_HDR_FMT "%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n"
17 #define PRETTY_LINE_FMT "%12u%12u%12u%12u%12u%12u%12u%12.0f%12.0f%12.0f\n"
18 #define CSV_HDR_FMT "%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n"
19 #define CSV_LINE_FMT "%10u;%10u;%u;%u;%u;%u;%u;%.3f;%.3f;%.3f\n"
20
21 struct cperf_pmd_cyclecount_ctx {
22         uint8_t dev_id;
23         uint16_t qp_id;
24         uint8_t lcore_id;
25
26         struct rte_mempool *pool;
27         struct rte_crypto_op **ops;
28         struct rte_crypto_op **ops_processed;
29
30         struct rte_cryptodev_sym_session *sess;
31
32         cperf_populate_ops_t populate_ops;
33
34         uint32_t src_buf_offset;
35         uint32_t dst_buf_offset;
36
37         const struct cperf_options *options;
38         const struct cperf_test_vector *test_vector;
39 };
40
41 struct pmd_cyclecount_state {
42         struct cperf_pmd_cyclecount_ctx *ctx;
43         const struct cperf_options *opts;
44         uint32_t lcore;
45         uint64_t delay;
46         int linearize;
47         uint32_t ops_enqd;
48         uint32_t ops_deqd;
49         uint32_t ops_enq_retries;
50         uint32_t ops_deq_retries;
51         double cycles_per_build;
52         double cycles_per_enq;
53         double cycles_per_deq;
54 };
55
56 static const uint16_t iv_offset =
57                 sizeof(struct rte_crypto_op) + sizeof(struct rte_crypto_sym_op);
58
59 static void
60 cperf_pmd_cyclecount_test_free(struct cperf_pmd_cyclecount_ctx *ctx)
61 {
62         if (ctx) {
63                 if (ctx->sess) {
64                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
65                         rte_cryptodev_sym_session_free(ctx->sess);
66                 }
67
68                 if (ctx->pool)
69                         rte_mempool_free(ctx->pool);
70
71                 if (ctx->ops)
72                         rte_free(ctx->ops);
73
74                 if (ctx->ops_processed)
75                         rte_free(ctx->ops_processed);
76
77                 rte_free(ctx);
78         }
79 }
80
81 void *
82 cperf_pmd_cyclecount_test_constructor(struct rte_mempool *sess_mp,
83                 struct rte_mempool *sess_priv_mp,
84                 uint8_t dev_id, uint16_t qp_id,
85                 const struct cperf_options *options,
86                 const struct cperf_test_vector *test_vector,
87                 const struct cperf_op_fns *op_fns)
88 {
89         struct cperf_pmd_cyclecount_ctx *ctx = NULL;
90
91         /* preallocate buffers for crypto ops as they can get quite big */
92         size_t alloc_sz = sizeof(struct rte_crypto_op *) *
93                         options->nb_descriptors;
94
95         ctx = rte_malloc(NULL, sizeof(struct cperf_pmd_cyclecount_ctx), 0);
96         if (ctx == NULL)
97                 goto err;
98
99         ctx->dev_id = dev_id;
100         ctx->qp_id = qp_id;
101
102         ctx->populate_ops = op_fns->populate_ops;
103         ctx->options = options;
104         ctx->test_vector = test_vector;
105
106         /* IV goes at the end of the crypto operation */
107         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
108                         sizeof(struct rte_crypto_sym_op);
109
110         ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
111                         test_vector, iv_offset);
112         if (ctx->sess == NULL)
113                 goto err;
114
115         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
116                         &ctx->src_buf_offset, &ctx->dst_buf_offset,
117                         &ctx->pool) < 0)
118                 goto err;
119
120         ctx->ops = rte_malloc("ops", alloc_sz, 0);
121         if (!ctx->ops)
122                 goto err;
123
124         ctx->ops_processed = rte_malloc("ops_processed", alloc_sz, 0);
125         if (!ctx->ops_processed)
126                 goto err;
127
128         return ctx;
129
130 err:
131         cperf_pmd_cyclecount_test_free(ctx);
132
133         return NULL;
134 }
135
136 /* benchmark alloc-build-free of ops */
137 static inline int
138 pmd_cyclecount_bench_ops(struct pmd_cyclecount_state *state, uint32_t cur_op,
139                 uint16_t test_burst_size)
140 {
141         uint32_t iter_ops_left = state->opts->total_ops - cur_op;
142         uint32_t iter_ops_needed =
143                         RTE_MIN(state->opts->nb_descriptors, iter_ops_left);
144         uint32_t cur_iter_op;
145         uint32_t imix_idx = 0;
146
147         for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
148                         cur_iter_op += test_burst_size) {
149                 uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
150                                 test_burst_size);
151                 struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
152
153                 /* Allocate objects containing crypto operations and mbufs */
154                 if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
155                                         burst_size) != 0) {
156                         RTE_LOG(ERR, USER1,
157                                         "Failed to allocate more crypto operations "
158                                         "from the crypto operation pool.\n"
159                                         "Consider increasing the pool size "
160                                         "with --pool-sz\n");
161                                 return -1;
162                 }
163
164                 /* Setup crypto op, attach mbuf etc */
165                 (state->ctx->populate_ops)(ops,
166                                 state->ctx->src_buf_offset,
167                                 state->ctx->dst_buf_offset,
168                                 burst_size,
169                                 state->ctx->sess, state->opts,
170                                 state->ctx->test_vector, iv_offset,
171                                 &imix_idx);
172
173 #ifdef CPERF_LINEARIZATION_ENABLE
174                 /* Check if source mbufs require coalescing */
175                 if (state->linearize) {
176                         uint8_t i;
177                         for (i = 0; i < burst_size; i++) {
178                                 struct rte_mbuf *src = ops[i]->sym->m_src;
179                                 rte_pktmbuf_linearize(src);
180                         }
181                 }
182 #endif /* CPERF_LINEARIZATION_ENABLE */
183                 rte_mempool_put_bulk(state->ctx->pool, (void **)ops,
184                                 burst_size);
185         }
186
187         return 0;
188 }
189
190 /* allocate and build ops (no free) */
191 static int
192 pmd_cyclecount_build_ops(struct pmd_cyclecount_state *state,
193                 uint32_t iter_ops_needed, uint16_t test_burst_size)
194 {
195         uint32_t cur_iter_op;
196         uint32_t imix_idx = 0;
197
198         for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
199                         cur_iter_op += test_burst_size) {
200                 uint32_t burst_size = RTE_MIN(
201                                 iter_ops_needed - cur_iter_op, test_burst_size);
202                 struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
203
204                 /* Allocate objects containing crypto operations and mbufs */
205                 if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
206                                         burst_size) != 0) {
207                         RTE_LOG(ERR, USER1,
208                                         "Failed to allocate more crypto operations "
209                                         "from the crypto operation pool.\n"
210                                         "Consider increasing the pool size "
211                                         "with --pool-sz\n");
212                                 return -1;
213                 }
214
215                 /* Setup crypto op, attach mbuf etc */
216                 (state->ctx->populate_ops)(ops,
217                                 state->ctx->src_buf_offset,
218                                 state->ctx->dst_buf_offset,
219                                 burst_size,
220                                 state->ctx->sess, state->opts,
221                                 state->ctx->test_vector, iv_offset,
222                                 &imix_idx);
223         }
224         return 0;
225 }
226
227 /* benchmark enqueue, returns number of ops enqueued */
228 static uint32_t
229 pmd_cyclecount_bench_enq(struct pmd_cyclecount_state *state,
230                 uint32_t iter_ops_needed, uint16_t test_burst_size)
231 {
232         /* Enqueue full descriptor ring of ops on crypto device */
233         uint32_t cur_iter_op = 0;
234         while (cur_iter_op < iter_ops_needed) {
235                 uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
236                                 test_burst_size);
237                 struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
238                 uint32_t burst_enqd;
239
240                 burst_enqd = rte_cryptodev_enqueue_burst(state->ctx->dev_id,
241                                 state->ctx->qp_id, ops, burst_size);
242
243                 /* if we couldn't enqueue anything, the queue is full */
244                 if (!burst_enqd) {
245                         /* don't try to dequeue anything we didn't enqueue */
246                         return cur_iter_op;
247                 }
248
249                 if (burst_enqd < burst_size)
250                         state->ops_enq_retries++;
251                 state->ops_enqd += burst_enqd;
252                 cur_iter_op += burst_enqd;
253         }
254         return iter_ops_needed;
255 }
256
257 /* benchmark dequeue */
258 static void
259 pmd_cyclecount_bench_deq(struct pmd_cyclecount_state *state,
260                 uint32_t iter_ops_needed, uint16_t test_burst_size)
261 {
262         /* Dequeue full descriptor ring of ops on crypto device */
263         uint32_t cur_iter_op = 0;
264         while (cur_iter_op < iter_ops_needed) {
265                 uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
266                                 test_burst_size);
267                 struct rte_crypto_op **ops_processed =
268                                 &state->ctx->ops[cur_iter_op];
269                 uint32_t burst_deqd;
270
271                 burst_deqd = rte_cryptodev_dequeue_burst(state->ctx->dev_id,
272                                 state->ctx->qp_id, ops_processed, burst_size);
273
274                 if (burst_deqd < burst_size)
275                         state->ops_deq_retries++;
276                 state->ops_deqd += burst_deqd;
277                 cur_iter_op += burst_deqd;
278         }
279 }
280
281 /* run benchmark per burst size */
282 static inline int
283 pmd_cyclecount_bench_burst_sz(
284                 struct pmd_cyclecount_state *state, uint16_t test_burst_size)
285 {
286         uint64_t tsc_start;
287         uint64_t tsc_end;
288         uint64_t tsc_op;
289         uint64_t tsc_enq;
290         uint64_t tsc_deq;
291         uint32_t cur_op;
292
293         /* reset all counters */
294         tsc_enq = 0;
295         tsc_deq = 0;
296         state->ops_enqd = 0;
297         state->ops_enq_retries = 0;
298         state->ops_deqd = 0;
299         state->ops_deq_retries = 0;
300
301         /*
302          * Benchmark crypto op alloc-build-free separately.
303          */
304         tsc_start = rte_rdtsc_precise();
305
306         for (cur_op = 0; cur_op < state->opts->total_ops;
307                         cur_op += state->opts->nb_descriptors) {
308                 if (unlikely(pmd_cyclecount_bench_ops(
309                                 state, cur_op, test_burst_size)))
310                         return -1;
311         }
312
313         tsc_end = rte_rdtsc_precise();
314         tsc_op = tsc_end - tsc_start;
315
316
317         /*
318          * Hardware acceleration cyclecount benchmarking loop.
319          *
320          * We're benchmarking raw enq/deq performance by filling up the device
321          * queue, so we never get any failed enqs unless the driver won't accept
322          * the exact number of descriptors we requested, or the driver won't
323          * wrap around the end of the TX ring. However, since we're only
324          * dequeueing once we've filled up the queue, we have to benchmark it
325          * piecemeal and then average out the results.
326          */
327         cur_op = 0;
328         while (cur_op < state->opts->total_ops) {
329                 uint32_t iter_ops_left = state->opts->total_ops - cur_op;
330                 uint32_t iter_ops_needed = RTE_MIN(
331                                 state->opts->nb_descriptors, iter_ops_left);
332                 uint32_t iter_ops_allocd = iter_ops_needed;
333
334                 /* allocate and build ops */
335                 if (unlikely(pmd_cyclecount_build_ops(state, iter_ops_needed,
336                                 test_burst_size)))
337                         return -1;
338
339                 tsc_start = rte_rdtsc_precise();
340
341                 /* fill up TX ring */
342                 iter_ops_needed = pmd_cyclecount_bench_enq(state,
343                                 iter_ops_needed, test_burst_size);
344
345                 tsc_end = rte_rdtsc_precise();
346
347                 tsc_enq += tsc_end - tsc_start;
348
349                 /* allow for HW to catch up */
350                 if (state->delay)
351                         rte_delay_us_block(state->delay);
352
353                 tsc_start = rte_rdtsc_precise();
354
355                 /* drain RX ring */
356                 pmd_cyclecount_bench_deq(state, iter_ops_needed,
357                                 test_burst_size);
358
359                 tsc_end = rte_rdtsc_precise();
360
361                 tsc_deq += tsc_end - tsc_start;
362
363                 cur_op += iter_ops_needed;
364
365                 /*
366                  * we may not have processed all ops that we allocated, so
367                  * free everything we've allocated.
368                  */
369                 rte_mempool_put_bulk(state->ctx->pool,
370                                 (void **)state->ctx->ops, iter_ops_allocd);
371         }
372
373         state->cycles_per_build = (double)tsc_op / state->opts->total_ops;
374         state->cycles_per_enq = (double)tsc_enq / state->ops_enqd;
375         state->cycles_per_deq = (double)tsc_deq / state->ops_deqd;
376
377         return 0;
378 }
379
380 int
381 cperf_pmd_cyclecount_test_runner(void *test_ctx)
382 {
383         struct pmd_cyclecount_state state = {0};
384         const struct cperf_options *opts;
385         uint16_t test_burst_size;
386         uint8_t burst_size_idx = 0;
387
388         state.ctx = test_ctx;
389         opts = state.ctx->options;
390         state.opts = opts;
391         state.lcore = rte_lcore_id();
392         state.linearize = 0;
393
394         static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
395         static bool warmup = true;
396
397         /*
398          * We need a small delay to allow for hardware to process all the crypto
399          * operations. We can't automatically figure out what the delay should
400          * be, so we leave it up to the user (by default it's 0).
401          */
402         state.delay = 1000 * opts->pmdcc_delay;
403
404 #ifdef CPERF_LINEARIZATION_ENABLE
405         struct rte_cryptodev_info dev_info;
406
407         /* Check if source mbufs require coalescing */
408         if (opts->segments_sz < ctx->options->max_buffer_size) {
409                 rte_cryptodev_info_get(state.ctx->dev_id, &dev_info);
410                 if ((dev_info.feature_flags &
411                                     RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) ==
412                                 0) {
413                         state.linearize = 1;
414                 }
415         }
416 #endif /* CPERF_LINEARIZATION_ENABLE */
417
418         state.ctx->lcore_id = state.lcore;
419
420         /* Get first size from range or list */
421         if (opts->inc_burst_size != 0)
422                 test_burst_size = opts->min_burst_size;
423         else
424                 test_burst_size = opts->burst_size_list[0];
425
426         while (test_burst_size <= opts->max_burst_size) {
427                 /* do a benchmark run */
428                 if (pmd_cyclecount_bench_burst_sz(&state, test_burst_size))
429                         return -1;
430
431                 /*
432                  * First run is always a warm up run.
433                  */
434                 if (warmup) {
435                         warmup = false;
436                         continue;
437                 }
438
439                 if (!opts->csv) {
440                         if (rte_atomic16_test_and_set(&display_once))
441                                 printf(PRETTY_HDR_FMT, "lcore id", "Buf Size",
442                                                 "Burst Size", "Enqueued",
443                                                 "Dequeued", "Enq Retries",
444                                                 "Deq Retries", "Cycles/Op",
445                                                 "Cycles/Enq", "Cycles/Deq");
446
447                         printf(PRETTY_LINE_FMT, state.ctx->lcore_id,
448                                         opts->test_buffer_size, test_burst_size,
449                                         state.ops_enqd, state.ops_deqd,
450                                         state.ops_enq_retries,
451                                         state.ops_deq_retries,
452                                         state.cycles_per_build,
453                                         state.cycles_per_enq,
454                                         state.cycles_per_deq);
455                 } else {
456                         if (rte_atomic16_test_and_set(&display_once))
457                                 printf(CSV_HDR_FMT, "# lcore id", "Buf Size",
458                                                 "Burst Size", "Enqueued",
459                                                 "Dequeued", "Enq Retries",
460                                                 "Deq Retries", "Cycles/Op",
461                                                 "Cycles/Enq", "Cycles/Deq");
462
463                         printf(CSV_LINE_FMT, state.ctx->lcore_id,
464                                         opts->test_buffer_size, test_burst_size,
465                                         state.ops_enqd, state.ops_deqd,
466                                         state.ops_enq_retries,
467                                         state.ops_deq_retries,
468                                         state.cycles_per_build,
469                                         state.cycles_per_enq,
470                                         state.cycles_per_deq);
471                 }
472
473                 /* Get next size from range or list */
474                 if (opts->inc_burst_size != 0)
475                         test_burst_size += opts->inc_burst_size;
476                 else {
477                         if (++burst_size_idx == opts->burst_size_count)
478                                 break;
479                         test_burst_size = opts->burst_size_list[burst_size_idx];
480                 }
481         }
482
483         return 0;
484 }
485
486 void
487 cperf_pmd_cyclecount_test_destructor(void *arg)
488 {
489         struct cperf_pmd_cyclecount_ctx *ctx = arg;
490
491         if (ctx == NULL)
492                 return;
493
494         cperf_pmd_cyclecount_test_free(ctx);
495 }