X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=app%2Ftest-crypto-perf%2Fcperf_test_latency.c;h=58b21abd9a42986fad05e5e6bc2718340663156a;hb=f6fadc3e6310;hp=657d0772a85ab67478102d6cf758abb320bb88ff;hpb=b2085cce60c309ba39e771b391a813ad7faea8ae;p=dpdk.git diff --git a/app/test-crypto-perf/cperf_test_latency.c b/app/test-crypto-perf/cperf_test_latency.c index 657d0772a8..58b21abd9a 100644 --- a/app/test-crypto-perf/cperf_test_latency.c +++ b/app/test-crypto-perf/cperf_test_latency.c @@ -39,26 +39,6 @@ #include "cperf_ops.h" -struct cperf_latency_results { - - uint64_t ops_failed; - - uint64_t enqd_tot; - uint64_t enqd_max; - uint64_t enqd_min; - - uint64_t deqd_tot; - uint64_t deqd_max; - uint64_t deqd_min; - - uint64_t cycles_tot; - uint64_t cycles_max; - uint64_t cycles_min; - - uint64_t burst_num; - uint64_t num; -}; - struct cperf_op_result { uint64_t tsc_start; uint64_t tsc_end; @@ -84,7 +64,10 @@ struct cperf_latency_ctx { const struct cperf_options *options; const struct cperf_test_vector *test_vector; struct cperf_op_result *res; - struct cperf_latency_results results; +}; + +struct priv_op_data { + struct cperf_op_result *result; }; #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b) @@ -96,8 +79,10 @@ cperf_latency_test_free(struct cperf_latency_ctx *ctx, uint32_t mbuf_nb) uint32_t i; if (ctx) { - if (ctx->sess) - rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess); + if (ctx->sess) { + rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess); + rte_cryptodev_sym_session_free(ctx->sess); + } if (ctx->mbufs_in) { for (i = 0; i < mbuf_nb; i++) @@ -136,8 +121,8 @@ cperf_mbuf_create(struct rte_mempool *mempool, const struct cperf_test_vector *test_vector) { struct rte_mbuf *mbuf; - uint32_t segment_sz = options->buffer_sz / segments_nb; - uint32_t last_sz = options->buffer_sz % segments_nb; + uint32_t segment_sz = options->max_buffer_size / segments_nb; + uint32_t last_sz = options->max_buffer_size % segments_nb; uint8_t *mbuf_data; uint8_t *test_data = (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ? @@ -184,14 +169,14 @@ cperf_mbuf_create(struct rte_mempool *mempool, if (options->op_type != CPERF_CIPHER_ONLY) { mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, - options->auth_digest_sz); + options->digest_sz); if (mbuf_data == NULL) goto error; } if (options->op_type == CPERF_AEAD) { uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf, - RTE_ALIGN_CEIL(options->auth_aad_sz, 16)); + RTE_ALIGN_CEIL(options->aead_aad_sz, 16)); if (aead == NULL) goto error; @@ -208,7 +193,8 @@ error: } void * -cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id, +cperf_latency_test_constructor(struct rte_mempool *sess_mp, + uint8_t dev_id, uint16_t qp_id, const struct cperf_options *options, const struct cperf_test_vector *test_vector, const struct cperf_op_fns *op_fns) @@ -228,7 +214,13 @@ cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id, ctx->options = options; ctx->test_vector = test_vector; - ctx->sess = op_fns->sess_create(dev_id, options, test_vector); + /* IV goes at the end of the crypto operation */ + uint16_t iv_offset = sizeof(struct rte_crypto_op) + + sizeof(struct rte_crypto_sym_op) + + sizeof(struct cperf_op_result *); + + ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, test_vector, + iv_offset); if (ctx->sess == NULL) goto err; @@ -239,18 +231,15 @@ cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id, options->pool_sz * options->segments_nb, 0, 0, RTE_PKTMBUF_HEADROOM + RTE_CACHE_LINE_ROUNDUP( - (options->buffer_sz / options->segments_nb) + - (options->buffer_sz % options->segments_nb) + - options->auth_digest_sz), + (options->max_buffer_size / options->segments_nb) + + (options->max_buffer_size % options->segments_nb) + + options->digest_sz), rte_socket_id()); if (ctx->pkt_mbuf_pool_in == NULL) goto err; /* Generate mbufs_in with plaintext populated for test */ - if (ctx->options->pool_sz % ctx->options->burst_sz) - goto err; - ctx->mbufs_in = rte_malloc(NULL, (sizeof(struct rte_mbuf *) * ctx->options->pool_sz), 0); @@ -273,8 +262,8 @@ cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id, pool_name, options->pool_sz, 0, 0, RTE_PKTMBUF_HEADROOM + RTE_CACHE_LINE_ROUNDUP( - options->buffer_sz + - options->auth_digest_sz), + options->max_buffer_size + + options->digest_sz), rte_socket_id()); if (ctx->pkt_mbuf_pool_out == NULL) @@ -300,9 +289,14 @@ cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id, snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d", dev_id); + uint16_t priv_size = sizeof(struct priv_op_data) + + test_vector->cipher_iv.length + + test_vector->auth_iv.length + + test_vector->aead_iv.length; ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name, - RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 0, 0, - rte_socket_id()); + RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, + 512, priv_size, rte_socket_id()); + if (ctx->crypto_op_pool == NULL) goto err; @@ -319,24 +313,32 @@ err: return NULL; } +static inline void +store_timestamp(struct rte_crypto_op *op, uint64_t timestamp) +{ + struct priv_op_data *priv_data; + + priv_data = (struct priv_op_data *) (op->sym + 1); + priv_data->result->status = op->status; + priv_data->result->tsc_end = timestamp; +} + int cperf_latency_test_runner(void *arg) { struct cperf_latency_ctx *ctx = arg; - struct cperf_op_result *pres; + uint16_t test_burst_size; + uint8_t burst_size_idx = 0; + + static int only_once; if (ctx == NULL) return 0; - struct rte_crypto_op *ops[ctx->options->burst_sz]; - struct rte_crypto_op *ops_processed[ctx->options->burst_sz]; - uint64_t ops_enqd = 0, ops_deqd = 0; - uint64_t m_idx = 0, b_idx = 0, i; - - uint64_t tsc_val, tsc_end, tsc_start; - uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0; - uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0; - uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0; + struct rte_crypto_op *ops[ctx->options->max_burst_size]; + struct rte_crypto_op *ops_processed[ctx->options->max_burst_size]; + uint64_t i; + struct priv_op_data *priv_data; uint32_t lcore = rte_lcore_id(); @@ -359,140 +361,220 @@ cperf_latency_test_runner(void *arg) for (i = 0; i < ctx->options->total_ops; i++) rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); - while (enqd_tot < ctx->options->total_ops) { - - uint16_t burst_size = ((enqd_tot + ctx->options->burst_sz) - <= ctx->options->total_ops) ? - ctx->options->burst_sz : - ctx->options->total_ops - - enqd_tot; - - /* Allocate crypto ops from pool */ - if (burst_size != rte_crypto_op_bulk_alloc( - ctx->crypto_op_pool, - RTE_CRYPTO_OP_TYPE_SYMMETRIC, - ops, burst_size)) - return -1; + /* Get first size from range or list */ + if (ctx->options->inc_burst_size != 0) + test_burst_size = ctx->options->min_burst_size; + else + test_burst_size = ctx->options->burst_size_list[0]; + + uint16_t iv_offset = sizeof(struct rte_crypto_op) + + sizeof(struct rte_crypto_sym_op) + + sizeof(struct cperf_op_result *); + + while (test_burst_size <= ctx->options->max_burst_size) { + uint64_t ops_enqd = 0, ops_deqd = 0; + uint64_t m_idx = 0, b_idx = 0; + + uint64_t tsc_val, tsc_end, tsc_start; + uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0; + uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0; + uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0; + + while (enqd_tot < ctx->options->total_ops) { + + uint16_t burst_size = ((enqd_tot + test_burst_size) + <= ctx->options->total_ops) ? + test_burst_size : + ctx->options->total_ops - + enqd_tot; + + /* Allocate crypto ops from pool */ + if (burst_size != rte_crypto_op_bulk_alloc( + ctx->crypto_op_pool, + RTE_CRYPTO_OP_TYPE_SYMMETRIC, + ops, burst_size)) { + RTE_LOG(ERR, USER1, + "Failed to allocate more crypto operations " + "from the the crypto operation pool.\n" + "Consider increasing the pool size " + "with --pool-sz\n"); + return -1; + } - /* Setup crypto op, attach mbuf etc */ - (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx], - &ctx->mbufs_out[m_idx], - burst_size, ctx->sess, ctx->options, - ctx->test_vector); + /* Setup crypto op, attach mbuf etc */ + (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx], + &ctx->mbufs_out[m_idx], + burst_size, ctx->sess, ctx->options, + ctx->test_vector, iv_offset); - tsc_start = rte_rdtsc_precise(); + tsc_start = rte_rdtsc_precise(); #ifdef CPERF_LINEARIZATION_ENABLE - if (linearize) { - /* PMD doesn't support scatter-gather and source buffer - * is segmented. - * We need to linearize it before enqueuing. - */ - for (i = 0; i < burst_size; i++) - rte_pktmbuf_linearize(ops[i]->sym->m_src); - } + if (linearize) { + /* PMD doesn't support scatter-gather and source buffer + * is segmented. + * We need to linearize it before enqueuing. + */ + for (i = 0; i < burst_size; i++) + rte_pktmbuf_linearize(ops[i]->sym->m_src); + } #endif /* CPERF_LINEARIZATION_ENABLE */ - /* Enqueue burst of ops on crypto device */ - ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, - ops, burst_size); + /* Enqueue burst of ops on crypto device */ + ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, + ops, burst_size); + + /* Dequeue processed burst of ops from crypto device */ + ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, + ops_processed, test_burst_size); + + tsc_end = rte_rdtsc_precise(); + + /* Free memory for not enqueued operations */ + if (ops_enqd != burst_size) + rte_mempool_put_bulk(ctx->crypto_op_pool, + (void **)&ops[ops_enqd], + burst_size - ops_enqd); + + for (i = 0; i < ops_enqd; i++) { + ctx->res[tsc_idx].tsc_start = tsc_start; + /* + * Private data structure starts after the end of the + * rte_crypto_sym_op structure. + */ + priv_data = (struct priv_op_data *) (ops[i]->sym + 1); + priv_data->result = (void *)&ctx->res[tsc_idx]; + tsc_idx++; + } - /* Dequeue processed burst of ops from crypto device */ - ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, - ops_processed, ctx->options->burst_sz); + if (likely(ops_deqd)) { + /* + * free crypto ops so they can be reused. We don't free + * the mbufs here as we don't want to reuse them as + * the crypto operation will change the data and cause + * failures. + */ + for (i = 0; i < ops_deqd; i++) + store_timestamp(ops_processed[i], tsc_end); + + rte_mempool_put_bulk(ctx->crypto_op_pool, + (void **)ops_processed, ops_deqd); + + deqd_tot += ops_deqd; + deqd_max = max(ops_deqd, deqd_max); + deqd_min = min(ops_deqd, deqd_min); + } - tsc_end = rte_rdtsc_precise(); + enqd_tot += ops_enqd; + enqd_max = max(ops_enqd, enqd_max); + enqd_min = min(ops_enqd, enqd_min); - for (i = 0; i < ops_enqd; i++) { - ctx->res[tsc_idx].tsc_start = tsc_start; - ops[i]->opaque_data = (void *)&ctx->res[tsc_idx]; - tsc_idx++; + m_idx += ops_enqd; + m_idx = m_idx + test_burst_size > ctx->options->pool_sz ? + 0 : m_idx; + b_idx++; } - /* Free memory for not enqueued operations */ - for (i = ops_enqd; i < burst_size; i++) - rte_crypto_op_free(ops[i]); - - if (likely(ops_deqd)) { - /* - * free crypto ops so they can be reused. We don't free - * the mbufs here as we don't want to reuse them as - * the crypto operation will change the data and cause - * failures. - */ - for (i = 0; i < ops_deqd; i++) { - pres = (struct cperf_op_result *) - (ops_processed[i]->opaque_data); - pres->status = ops_processed[i]->status; - pres->tsc_end = tsc_end; - - rte_crypto_op_free(ops_processed[i]); + /* Dequeue any operations still in the crypto device */ + while (deqd_tot < ctx->options->total_ops) { + /* Sending 0 length burst to flush sw crypto device */ + rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); + + /* dequeue burst */ + ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, + ops_processed, test_burst_size); + + tsc_end = rte_rdtsc_precise(); + + if (ops_deqd != 0) { + for (i = 0; i < ops_deqd; i++) + store_timestamp(ops_processed[i], tsc_end); + + rte_mempool_put_bulk(ctx->crypto_op_pool, + (void **)ops_processed, ops_deqd); + + deqd_tot += ops_deqd; + deqd_max = max(ops_deqd, deqd_max); + deqd_min = min(ops_deqd, deqd_min); } + } - deqd_tot += ops_deqd; - deqd_max = max(ops_deqd, deqd_max); - deqd_min = min(ops_deqd, deqd_min); + for (i = 0; i < tsc_idx; i++) { + tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start; + tsc_max = max(tsc_val, tsc_max); + tsc_min = min(tsc_val, tsc_min); + tsc_tot += tsc_val; } - enqd_tot += ops_enqd; - enqd_max = max(ops_enqd, enqd_max); - enqd_min = min(ops_enqd, enqd_min); + double time_tot, time_avg, time_max, time_min; - m_idx += ops_enqd; - m_idx = m_idx + ctx->options->burst_sz > ctx->options->pool_sz ? - 0 : m_idx; - b_idx++; - } + const uint64_t tunit = 1000000; /* us */ + const uint64_t tsc_hz = rte_get_tsc_hz(); - /* Dequeue any operations still in the crypto device */ - while (deqd_tot < ctx->options->total_ops) { - /* Sending 0 length burst to flush sw crypto device */ - rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); + uint64_t enqd_avg = enqd_tot / b_idx; + uint64_t deqd_avg = deqd_tot / b_idx; + uint64_t tsc_avg = tsc_tot / tsc_idx; + + time_tot = tunit*(double)(tsc_tot) / tsc_hz; + time_avg = tunit*(double)(tsc_avg) / tsc_hz; + time_max = tunit*(double)(tsc_max) / tsc_hz; + time_min = tunit*(double)(tsc_min) / tsc_hz; - /* dequeue burst */ - ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, - ops_processed, ctx->options->burst_sz); + if (ctx->options->csv) { + if (!only_once) + printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, " + "Packet Size, cycles, time (us)"); - tsc_end = rte_rdtsc_precise(); + for (i = 0; i < ctx->options->total_ops; i++) { - if (ops_deqd != 0) { - for (i = 0; i < ops_deqd; i++) { - pres = (struct cperf_op_result *) - (ops_processed[i]->opaque_data); - pres->status = ops_processed[i]->status; - pres->tsc_end = tsc_end; + printf("\n%u;%u;%u;%"PRIu64";%"PRIu64";%.3f", + ctx->lcore_id, ctx->options->test_buffer_size, + test_burst_size, i + 1, + ctx->res[i].tsc_end - ctx->res[i].tsc_start, + tunit * (double) (ctx->res[i].tsc_end + - ctx->res[i].tsc_start) + / tsc_hz); - rte_crypto_op_free(ops_processed[i]); } + only_once = 1; + } else { + printf("\n# Device %d on lcore %u\n", ctx->dev_id, + ctx->lcore_id); + printf("\n# total operations: %u", ctx->options->total_ops); + printf("\n# Buffer size: %u", ctx->options->test_buffer_size); + printf("\n# Burst size: %u", test_burst_size); + printf("\n# Number of bursts: %"PRIu64, + b_idx); + + printf("\n#"); + printf("\n# \t Total\t Average\t " + "Maximum\t Minimum"); + printf("\n# enqueued\t%12"PRIu64"\t%10"PRIu64"\t" + "%10"PRIu64"\t%10"PRIu64, enqd_tot, + enqd_avg, enqd_max, enqd_min); + printf("\n# dequeued\t%12"PRIu64"\t%10"PRIu64"\t" + "%10"PRIu64"\t%10"PRIu64, deqd_tot, + deqd_avg, deqd_max, deqd_min); + printf("\n# cycles\t%12"PRIu64"\t%10"PRIu64"\t" + "%10"PRIu64"\t%10"PRIu64, tsc_tot, + tsc_avg, tsc_max, tsc_min); + printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f", + time_tot, time_avg, time_max, time_min); + printf("\n\n"); - deqd_tot += ops_deqd; - deqd_max = max(ops_deqd, deqd_max); - deqd_min = min(ops_deqd, deqd_min); } - } - for (i = 0; i < tsc_idx; i++) { - tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start; - tsc_max = max(tsc_val, tsc_max); - tsc_min = min(tsc_val, tsc_min); - tsc_tot += tsc_val; + /* Get next size from range or list */ + if (ctx->options->inc_burst_size != 0) + test_burst_size += ctx->options->inc_burst_size; + else { + if (++burst_size_idx == ctx->options->burst_size_count) + break; + test_burst_size = + ctx->options->burst_size_list[burst_size_idx]; + } } - ctx->results.enqd_tot = enqd_tot; - ctx->results.enqd_max = enqd_max; - ctx->results.enqd_min = enqd_min; - - ctx->results.deqd_tot = deqd_tot; - ctx->results.deqd_max = deqd_max; - ctx->results.deqd_min = deqd_min; - - ctx->results.cycles_tot = tsc_tot; - ctx->results.cycles_max = tsc_max; - ctx->results.cycles_min = tsc_min; - - ctx->results.burst_num = b_idx; - ctx->results.num = tsc_idx; - return 0; } @@ -500,74 +582,11 @@ void cperf_latency_test_destructor(void *arg) { struct cperf_latency_ctx *ctx = arg; - uint64_t i; + if (ctx == NULL) return; - static int only_once; - uint64_t etot, eavg, emax, emin; - uint64_t dtot, davg, dmax, dmin; - uint64_t ctot, cavg, cmax, cmin; - double ttot, tavg, tmax, tmin; - - const uint64_t tunit = 1000000; /* us */ - const uint64_t tsc_hz = rte_get_tsc_hz(); - - etot = ctx->results.enqd_tot; - eavg = ctx->results.enqd_tot / ctx->results.burst_num; - emax = ctx->results.enqd_max; - emin = ctx->results.enqd_min; - - dtot = ctx->results.deqd_tot; - davg = ctx->results.deqd_tot / ctx->results.burst_num; - dmax = ctx->results.deqd_max; - dmin = ctx->results.deqd_min; - - ctot = ctx->results.cycles_tot; - cavg = ctx->results.cycles_tot / ctx->results.num; - cmax = ctx->results.cycles_max; - cmin = ctx->results.cycles_min; - - ttot = tunit*(double)(ctot) / tsc_hz; - tavg = tunit*(double)(cavg) / tsc_hz; - tmax = tunit*(double)(cmax) / tsc_hz; - tmin = tunit*(double)(cmin) / tsc_hz; - - if (ctx->options->csv) { - if (!only_once) - printf("\n# lcore, Pakt Seq #, Packet Size, cycles," - " time (us)"); - - for (i = 0; i < ctx->options->total_ops; i++) { - - printf("\n%u;%"PRIu64";%"PRIu64";%.3f", - ctx->lcore_id, i + 1, - ctx->res[i].tsc_end - ctx->res[i].tsc_start, - tunit * (double) (ctx->res[i].tsc_end - - ctx->res[i].tsc_start) - / tsc_hz); - } - only_once = 1; - } else { - printf("\n# Device %d on lcore %u\n", ctx->dev_id, - ctx->lcore_id); - printf("\n# total operations: %u", ctx->options->total_ops); - printf("\n# burst number: %"PRIu64, - ctx->results.burst_num); - printf("\n#"); - printf("\n# \t Total\t Average\t Maximum\t " - " Minimum"); - printf("\n# enqueued\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t" - "%10"PRIu64, etot, eavg, emax, emin); - printf("\n# dequeued\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t" - "%10"PRIu64, dtot, davg, dmax, dmin); - printf("\n# cycles\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t" - "%10"PRIu64, ctot, cavg, cmax, cmin); - printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f", ttot, - tavg, tmax, tmin); - printf("\n\n"); + rte_cryptodev_stop(ctx->dev_id); - } cperf_latency_test_free(ctx, ctx->options->pool_sz); - }