mempool: fix slow allocation of large mempools
[dpdk.git] / app / test-compress-perf / comp_perf_test_verify.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2018 Intel Corporation
3  */
4
5 #include <rte_malloc.h>
6 #include <rte_eal.h>
7 #include <rte_log.h>
8 #include <rte_compressdev.h>
9
10 #include "comp_perf_test_verify.h"
11 #include "comp_perf_test_common.h"
12
13 void
14 cperf_verify_test_destructor(void *arg)
15 {
16         if (arg) {
17                 comp_perf_free_memory(
18                                 ((struct cperf_verify_ctx *)arg)->options,
19                                 &((struct cperf_verify_ctx *)arg)->mem);
20                 rte_free(arg);
21         }
22 }
23
24 void *
25 cperf_verify_test_constructor(uint8_t dev_id, uint16_t qp_id,
26                 struct comp_test_data *options)
27 {
28         struct cperf_verify_ctx *ctx = NULL;
29
30         ctx = rte_malloc(NULL, sizeof(struct cperf_verify_ctx), 0);
31
32         if (ctx == NULL)
33                 return NULL;
34
35         ctx->mem.dev_id = dev_id;
36         ctx->mem.qp_id = qp_id;
37         ctx->options = options;
38
39         if (!comp_perf_allocate_memory(ctx->options, &ctx->mem) &&
40                         !prepare_bufs(ctx->options, &ctx->mem))
41                 return ctx;
42
43         cperf_verify_test_destructor(ctx);
44         return NULL;
45 }
46
47 static int
48 main_loop(struct cperf_verify_ctx *ctx, enum rte_comp_xform_type type)
49 {
50         struct comp_test_data *test_data = ctx->options;
51         uint8_t *output_data_ptr;
52         size_t *output_data_sz;
53         struct cperf_mem_resources *mem = &ctx->mem;
54
55         uint8_t dev_id = mem->dev_id;
56         uint32_t i, iter, num_iter;
57         struct rte_comp_op **ops, **deq_ops;
58         void *priv_xform = NULL;
59         struct rte_comp_xform xform;
60         size_t output_size = 0;
61         struct rte_mbuf **input_bufs, **output_bufs;
62         int res = 0;
63         int allocated = 0;
64         uint32_t out_seg_sz;
65
66         if (test_data == NULL || !test_data->burst_sz) {
67                 RTE_LOG(ERR, USER1,
68                         "Unknown burst size\n");
69                 return -1;
70         }
71
72         ops = rte_zmalloc_socket(NULL,
73                 2 * mem->total_bufs * sizeof(struct rte_comp_op *),
74                 0, rte_socket_id());
75
76         if (ops == NULL) {
77                 RTE_LOG(ERR, USER1,
78                         "Can't allocate memory for ops strucures\n");
79                 return -1;
80         }
81
82         deq_ops = &ops[mem->total_bufs];
83
84         if (type == RTE_COMP_COMPRESS) {
85                 xform = (struct rte_comp_xform) {
86                         .type = RTE_COMP_COMPRESS,
87                         .compress = {
88                                 .algo = RTE_COMP_ALGO_DEFLATE,
89                                 .deflate.huffman = test_data->huffman_enc,
90                                 .level = test_data->level,
91                                 .window_size = test_data->window_sz,
92                                 .chksum = RTE_COMP_CHECKSUM_NONE,
93                                 .hash_algo = RTE_COMP_HASH_ALGO_NONE
94                         }
95                 };
96                 output_data_ptr = ctx->mem.compressed_data;
97                 output_data_sz = &ctx->comp_data_sz;
98                 input_bufs = mem->decomp_bufs;
99                 output_bufs = mem->comp_bufs;
100                 out_seg_sz = test_data->out_seg_sz;
101         } else {
102                 xform = (struct rte_comp_xform) {
103                         .type = RTE_COMP_DECOMPRESS,
104                         .decompress = {
105                                 .algo = RTE_COMP_ALGO_DEFLATE,
106                                 .chksum = RTE_COMP_CHECKSUM_NONE,
107                                 .window_size = test_data->window_sz,
108                                 .hash_algo = RTE_COMP_HASH_ALGO_NONE
109                         }
110                 };
111                 output_data_ptr = ctx->mem.decompressed_data;
112                 output_data_sz = &ctx->decomp_data_sz;
113                 input_bufs = mem->comp_bufs;
114                 output_bufs = mem->decomp_bufs;
115                 out_seg_sz = test_data->seg_sz;
116         }
117
118         /* Create private xform */
119         if (rte_compressdev_private_xform_create(dev_id, &xform,
120                         &priv_xform) < 0) {
121                 RTE_LOG(ERR, USER1, "Private xform could not be created\n");
122                 res = -1;
123                 goto end;
124         }
125
126         num_iter = 1;
127
128         for (iter = 0; iter < num_iter; iter++) {
129                 uint32_t total_ops = mem->total_bufs;
130                 uint32_t remaining_ops = mem->total_bufs;
131                 uint32_t total_deq_ops = 0;
132                 uint32_t total_enq_ops = 0;
133                 uint16_t ops_unused = 0;
134                 uint16_t num_enq = 0;
135                 uint16_t num_deq = 0;
136
137                 output_size = 0;
138
139                 while (remaining_ops > 0) {
140                         uint16_t num_ops = RTE_MIN(remaining_ops,
141                                                    test_data->burst_sz);
142                         uint16_t ops_needed = num_ops - ops_unused;
143
144                         /*
145                          * Move the unused operations from the previous
146                          * enqueue_burst call to the front, to maintain order
147                          */
148                         if ((ops_unused > 0) && (num_enq > 0)) {
149                                 size_t nb_b_to_mov =
150                                       ops_unused * sizeof(struct rte_comp_op *);
151
152                                 memmove(ops, &ops[num_enq], nb_b_to_mov);
153                         }
154
155                         /* Allocate compression operations */
156                         if (ops_needed && !rte_comp_op_bulk_alloc(
157                                                 mem->op_pool,
158                                                 &ops[ops_unused],
159                                                 ops_needed)) {
160                                 RTE_LOG(ERR, USER1,
161                                       "Could not allocate enough operations\n");
162                                 res = -1;
163                                 goto end;
164                         }
165                         allocated += ops_needed;
166
167                         for (i = 0; i < ops_needed; i++) {
168                                 /*
169                                  * Calculate next buffer to attach to operation
170                                  */
171                                 uint32_t buf_id = total_enq_ops + i +
172                                                 ops_unused;
173                                 uint16_t op_id = ops_unused + i;
174                                 /* Reset all data in output buffers */
175                                 struct rte_mbuf *m = output_bufs[buf_id];
176
177                                 m->pkt_len = out_seg_sz * m->nb_segs;
178                                 while (m) {
179                                         m->data_len = m->buf_len - m->data_off;
180                                         m = m->next;
181                                 }
182                                 ops[op_id]->m_src = input_bufs[buf_id];
183                                 ops[op_id]->m_dst = output_bufs[buf_id];
184                                 ops[op_id]->src.offset = 0;
185                                 ops[op_id]->src.length =
186                                         rte_pktmbuf_pkt_len(input_bufs[buf_id]);
187                                 ops[op_id]->dst.offset = 0;
188                                 ops[op_id]->flush_flag = RTE_COMP_FLUSH_FINAL;
189                                 ops[op_id]->input_chksum = buf_id;
190                                 ops[op_id]->private_xform = priv_xform;
191                         }
192
193                         if (unlikely(test_data->perf_comp_force_stop))
194                                 goto end;
195
196                         num_enq = rte_compressdev_enqueue_burst(dev_id,
197                                                                 mem->qp_id, ops,
198                                                                 num_ops);
199                         if (num_enq == 0) {
200                                 struct rte_compressdev_stats stats;
201
202                                 rte_compressdev_stats_get(dev_id, &stats);
203                                 if (stats.enqueue_err_count) {
204                                         res = -1;
205                                         goto end;
206                                 }
207                         }
208
209                         ops_unused = num_ops - num_enq;
210                         remaining_ops -= num_enq;
211                         total_enq_ops += num_enq;
212
213                         num_deq = rte_compressdev_dequeue_burst(dev_id,
214                                                            mem->qp_id,
215                                                            deq_ops,
216                                                            test_data->burst_sz);
217                         total_deq_ops += num_deq;
218
219                         for (i = 0; i < num_deq; i++) {
220                                 struct rte_comp_op *op = deq_ops[i];
221
222                                 if (op->status ==
223                                   RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED ||
224                                   op->status ==
225                                   RTE_COMP_OP_STATUS_OUT_OF_SPACE_RECOVERABLE) {
226                                         RTE_LOG(ERR, USER1,
227 "Out of space error occurred due to uncompressible input data expanding to larger than destination buffer. Increase the EXPANSE_RATIO constant to use this data.\n");
228                                         res = -1;
229                                         goto end;
230                                 } else if (op->status !=
231                                                 RTE_COMP_OP_STATUS_SUCCESS) {
232                                         RTE_LOG(ERR, USER1,
233                                                 "Some operations were not successful\n");
234                                         goto end;
235                                 }
236
237                                 const void *read_data_addr =
238                                                 rte_pktmbuf_read(op->m_dst, 0,
239                                                 op->produced, output_data_ptr);
240                                 if (read_data_addr == NULL) {
241                                         RTE_LOG(ERR, USER1,
242                                                 "Could not copy buffer in destination\n");
243                                         res = -1;
244                                         goto end;
245                                 }
246
247                                 if (read_data_addr != output_data_ptr)
248                                         rte_memcpy(output_data_ptr,
249                                                    rte_pktmbuf_mtod(op->m_dst,
250                                                                     uint8_t *),
251                                                    op->produced);
252                                 output_data_ptr += op->produced;
253                                 output_size += op->produced;
254
255                         }
256
257
258                         if (iter == num_iter - 1) {
259                                 for (i = 0; i < num_deq; i++) {
260                                         struct rte_comp_op *op = deq_ops[i];
261                                         struct rte_mbuf *m = op->m_dst;
262
263                                         m->pkt_len = op->produced;
264                                         uint32_t remaining_data = op->produced;
265                                         uint16_t data_to_append;
266
267                                         while (remaining_data > 0) {
268                                                 data_to_append =
269                                                         RTE_MIN(remaining_data,
270                                                         out_seg_sz);
271                                                 m->data_len = data_to_append;
272                                                 remaining_data -=
273                                                                 data_to_append;
274                                                 m = m->next;
275                                         }
276                                 }
277                         }
278                         rte_mempool_put_bulk(mem->op_pool,
279                                              (void **)deq_ops, num_deq);
280                         allocated -= num_deq;
281                 }
282
283                 /* Dequeue the last operations */
284                 while (total_deq_ops < total_ops) {
285                         if (unlikely(test_data->perf_comp_force_stop))
286                                 goto end;
287
288                         num_deq = rte_compressdev_dequeue_burst(dev_id,
289                                                         mem->qp_id,
290                                                         deq_ops,
291                                                         test_data->burst_sz);
292                         if (num_deq == 0) {
293                                 struct rte_compressdev_stats stats;
294
295                                 rte_compressdev_stats_get(dev_id, &stats);
296                                 if (stats.dequeue_err_count) {
297                                         res = -1;
298                                         goto end;
299                                 }
300                         }
301
302                         total_deq_ops += num_deq;
303
304                         for (i = 0; i < num_deq; i++) {
305                                 struct rte_comp_op *op = deq_ops[i];
306
307                                 if (op->status ==
308                                   RTE_COMP_OP_STATUS_OUT_OF_SPACE_TERMINATED ||
309                                   op->status ==
310                                   RTE_COMP_OP_STATUS_OUT_OF_SPACE_RECOVERABLE) {
311                                         RTE_LOG(ERR, USER1,
312 "Out of space error occurred due to uncompressible input data expanding to larger than destination buffer. Increase the EXPANSE_RATIO constant to use this data.\n");
313                                         res = -1;
314                                         goto end;
315                                 } else if (op->status !=
316                                                 RTE_COMP_OP_STATUS_SUCCESS) {
317                                         RTE_LOG(ERR, USER1,
318                                                 "Some operations were not successful\n");
319                                         goto end;
320                                 }
321                                 const void *read_data_addr =
322                                                 rte_pktmbuf_read(op->m_dst,
323                                                                  op->dst.offset,
324                                                 op->produced, output_data_ptr);
325                                 if (read_data_addr == NULL) {
326                                         RTE_LOG(ERR, USER1,
327                                                 "Could not copy buffer in destination\n");
328                                         res = -1;
329                                         goto end;
330                                 }
331
332                                 if (read_data_addr != output_data_ptr)
333                                         rte_memcpy(output_data_ptr,
334                                                    rte_pktmbuf_mtod(
335                                                         op->m_dst, uint8_t *),
336                                                    op->produced);
337                                 output_data_ptr += op->produced;
338                                 output_size += op->produced;
339
340                         }
341
342                         if (iter == num_iter - 1) {
343                                 for (i = 0; i < num_deq; i++) {
344                                         struct rte_comp_op *op = deq_ops[i];
345                                         struct rte_mbuf *m = op->m_dst;
346
347                                         m->pkt_len = op->produced;
348                                         uint32_t remaining_data = op->produced;
349                                         uint16_t data_to_append;
350
351                                         while (remaining_data > 0) {
352                                                 data_to_append =
353                                                 RTE_MIN(remaining_data,
354                                                         out_seg_sz);
355                                                 m->data_len = data_to_append;
356                                                 remaining_data -=
357                                                                 data_to_append;
358                                                 m = m->next;
359                                         }
360                                 }
361                         }
362                         rte_mempool_put_bulk(mem->op_pool,
363                                              (void **)deq_ops, num_deq);
364                         allocated -= num_deq;
365                 }
366         }
367
368         if (output_data_sz)
369                 *output_data_sz = output_size;
370 end:
371         rte_mempool_put_bulk(mem->op_pool, (void **)ops, allocated);
372         rte_compressdev_private_xform_free(dev_id, priv_xform);
373         rte_free(ops);
374
375         if (test_data->perf_comp_force_stop) {
376                 RTE_LOG(ERR, USER1,
377                       "lcore: %d Perf. test has been aborted by user\n",
378                         mem->lcore_id);
379                 res = -1;
380         }
381
382         return res;
383 }
384
385 int
386 cperf_verify_test_runner(void *test_ctx)
387 {
388         struct cperf_verify_ctx *ctx = test_ctx;
389         struct comp_test_data *test_data = ctx->options;
390         int ret = EXIT_SUCCESS;
391         static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
392         uint32_t lcore = rte_lcore_id();
393
394         ctx->mem.lcore_id = lcore;
395
396         test_data->ratio = 0;
397
398         if (main_loop(ctx, RTE_COMP_COMPRESS) < 0) {
399                 ret = EXIT_FAILURE;
400                 goto end;
401         }
402
403         if (main_loop(ctx, RTE_COMP_DECOMPRESS) < 0) {
404                 ret = EXIT_FAILURE;
405                 goto end;
406         }
407
408         if (ctx->decomp_data_sz != test_data->input_data_sz) {
409                 RTE_LOG(ERR, USER1,
410            "Decompressed data length not equal to input data length\n");
411                 RTE_LOG(ERR, USER1,
412                         "Decompressed size = %zu, expected = %zu\n",
413                         ctx->decomp_data_sz, test_data->input_data_sz);
414                 ret = EXIT_FAILURE;
415                 goto end;
416         } else {
417                 if (memcmp(ctx->mem.decompressed_data,
418                                 test_data->input_data,
419                                 test_data->input_data_sz) != 0) {
420                         RTE_LOG(ERR, USER1,
421                     "Decompressed data is not the same as file data\n");
422                         ret = EXIT_FAILURE;
423                         goto end;
424                 }
425         }
426
427         ctx->ratio = (double) ctx->comp_data_sz /
428                         test_data->input_data_sz * 100;
429
430         if (!ctx->silent) {
431                 if (rte_atomic16_test_and_set(&display_once)) {
432                         printf("%12s%6s%12s%17s\n",
433                             "lcore id", "Level", "Comp size", "Comp ratio [%]");
434                 }
435                 printf("%12u%6u%12zu%17.2f\n",
436                        ctx->mem.lcore_id,
437                        test_data->level, ctx->comp_data_sz, ctx->ratio);
438         }
439
440 end:
441         return ret;
442 }