[EXT] [PATCH v8] app/dma-perf: introduce dma-perf application

[Anoob Joseph]

Hi Cheng,

Thanks for the new version. Please see inline.

Thanks,
Anoob

> -----Original Message-----
> From: Cheng Jiang <cheng1.jiang at intel.com>
> Sent: Tuesday, June 20, 2023 12:24 PM
> To: thomas at monjalon.net; bruce.richardson at intel.com;
> mb at smartsharesystems.com; chenbo.xia at intel.com; Amit Prakash Shukla
> <amitprakashs at marvell.com>; Anoob Joseph <anoobj at marvell.com>;
> huangdengdui at huawei.com; kevin.laatz at intel.com;
> fengchengwen at huawei.com; Jerin Jacob Kollanukkaran
> <jerinj at marvell.com>
> Cc: dev at dpdk.org; jiayu.hu at intel.com; xuan.ding at intel.com;
> wenwux.ma at intel.com; yuanx.wang at intel.com; xingguang.he at intel.com;
> weix.ling at intel.com; Cheng Jiang <cheng1.jiang at intel.com>
> Subject: [EXT] [PATCH v8] app/dma-perf: introduce dma-perf application
> 
> External Email
> 
> ----------------------------------------------------------------------
> There are many high-performance DMA devices supported in DPDK now,
> and these DMA devices can also be integrated into other modules of DPDK as
> accelerators, such as Vhost. Before integrating DMA into applications,
> developers need to know the performance of these DMA devices in various
> scenarios and the performance of CPUs in the same scenario, such as
> different buffer lengths. Only in this way can we know the target
> performance of the application accelerated by using them. This patch
> introduces a high-performance testing tool, which supports comparing the
> performance of CPU and DMA in different scenarios automatically with a pre-
> set config file. Memory Copy performance test are supported for now.
> 
> Signed-off-by: Cheng Jiang <cheng1.jiang at intel.com>
> Signed-off-by: Jiayu Hu <jiayu.hu at intel.com>
> Signed-off-by: Yuan Wang <yuanx.wang at intel.com>
> Acked-by: Morten Brørup <mb at smartsharesystems.com>
> Acked-by: Chenbo Xia <chenbo.xia at intel.com>
> ---
> v8:
>   fixed string copy issue in parse_lcore();
>   improved some data display format;
>   added doc in doc/guides/tools;
>   updated release notes;
> 
> v7:
>   fixed some strcpy issues;
>   removed cache setup in calling rte_pktmbuf_pool_create();
>   fixed some typos;
>   added some memory free and null set operations;
>   improved result calculation;
> v6:
>   improved code based on Anoob's comments;
>   fixed some code structure issues;
> v5:
>   fixed some LONG_LINE warnings;
> v4:
>   fixed inaccuracy of the memory footprint display;
> v3:
>   fixed some typos;
> v2:
>   added lcore/dmadev designation;
>   added error case process;
>   removed worker_threads parameter from config.ini;
>   improved the logs;
>   improved config file;
> 
>  app/meson.build                        |   1 +
>  app/test-dma-perf/benchmark.c          | 498 +++++++++++++++++++++
>  app/test-dma-perf/config.ini           |  61 +++
>  app/test-dma-perf/main.c               | 594 +++++++++++++++++++++++++
>  app/test-dma-perf/main.h               |  69 +++
>  app/test-dma-perf/meson.build          |  17 +
>  doc/guides/rel_notes/release_23_07.rst |   6 +
>  doc/guides/tools/dmaperf.rst           | 103 +++++
>  doc/guides/tools/index.rst             |   1 +
>  9 files changed, 1350 insertions(+)
>  create mode 100644 app/test-dma-perf/benchmark.c  create mode 100644
> app/test-dma-perf/config.ini  create mode 100644 app/test-dma-
> perf/main.c  create mode 100644 app/test-dma-perf/main.h  create mode
> 100644 app/test-dma-perf/meson.build  create mode 100644
> doc/guides/tools/dmaperf.rst
> 

<snip>

> +/* Configuration of device. */
> +static void
> +configure_dmadev_queue(uint32_t dev_id, uint32_t ring_size) {
> +	uint16_t vchan = 0;
> +	struct rte_dma_info info;
> +	struct rte_dma_conf dev_config = { .nb_vchans = 1 };
> +	struct rte_dma_vchan_conf qconf = {
> +		.direction = RTE_DMA_DIR_MEM_TO_MEM,
> +		.nb_desc = ring_size
> +	};
> +
> +	if (rte_dma_configure(dev_id, &dev_config) != 0)
> +		rte_exit(EXIT_FAILURE, "Error with dma configure.\n");
> +
> +	if (rte_dma_vchan_setup(dev_id, vchan, &qconf) != 0)
> +		rte_exit(EXIT_FAILURE, "Error with queue configuration.\n");
> +
> +	rte_dma_info_get(dev_id, &info);

[Anoob] This API can return errors. Better to add handling.

> +	if (info.nb_vchans != 1)
> +		rte_exit(EXIT_FAILURE, "Error, no configured queues
> reported on device id. %u\n",
> +				dev_id);
> +
> +	if (rte_dma_start(dev_id) != 0)
> +		rte_exit(EXIT_FAILURE, "Error with dma start.\n"); }
> +
> +static int
> +config_dmadevs(struct test_configure *cfg) {
> +	uint32_t ring_size = cfg->ring_size.cur;
> +	struct lcore_dma_map_t *ldm = &cfg->lcore_dma_map;
> +	uint32_t nb_workers = ldm->cnt;
> +	uint32_t i;
> +	int dev_id;
> +	uint16_t nb_dmadevs = 0;
> +	char *dma_name;
> +
> +	for (i = 0; i < ldm->cnt; i++) {
> +		dma_name = ldm->dma_names[i];
> +		dev_id = rte_dma_get_dev_id_by_name(dma_name);
> +		if (dev_id == -1) {

[Anoob] Can you check the above API definition? I think it returns not just -1 in case of errors.

> +			fprintf(stderr, "Error: Fail to find DMA %s.\n",
> dma_name);
> +			goto end;
> +		}
> +
> +		ldm->dma_ids[i] = dev_id;
> +		configure_dmadev_queue(dev_id, ring_size);
> +		++nb_dmadevs;
> +	}
> +
> +end:
> +	if (nb_dmadevs < nb_workers) {
> +		printf("Not enough dmadevs (%u) for all workers (%u).\n",
> nb_dmadevs, nb_workers);
> +		return -1;
> +	}
> +
> +	printf("Number of used dmadevs: %u.\n", nb_dmadevs);
> +
> +	return 0;
> +}
> +
> +static inline void
> +do_dma_submit_and_poll(uint16_t dev_id, uint64_t *async_cnt,
> +			volatile struct worker_info *worker_info) {
> +	int ret;
> +	uint16_t nr_cpl;
> +
> +	ret = rte_dma_submit(dev_id, 0);
> +	if (ret < 0) {
> +		rte_dma_stop(dev_id);
> +		rte_dma_close(dev_id);
> +		rte_exit(EXIT_FAILURE, "Error with dma submit.\n");
> +	}
> +
> +	nr_cpl = rte_dma_completed(dev_id, 0, MAX_DMA_CPL_NB, NULL,
> NULL);
> +	*async_cnt -= nr_cpl;
> +	worker_info->total_cpl += nr_cpl;
> +}
> +
> +static inline int
> +do_dma_mem_copy(void *p)

[Anoob] Just curious, why not pass struct lcore_params *para itself? Is it because the pointer is volatile? If yes, then we can take an AI to split the struct into volatile and non-volatile parts.

> +{
> +	const uint16_t *para_idx = (uint16_t *)p;
> +	volatile struct lcore_params *para = lcores_p[*para_idx].v_ptr;
> +	volatile struct worker_info *worker_info = &(para->worker_info);
> +	const uint16_t dev_id = para->dev_id;
> +	const uint32_t nr_buf = para->nr_buf;
> +	const uint16_t kick_batch = para->kick_batch;
> +	const uint32_t buf_size = para->buf_size;
> +	struct rte_mbuf **srcs = para->srcs;
> +	struct rte_mbuf **dsts = para->dsts;
> +	uint16_t nr_cpl;
> +	uint64_t async_cnt = 0;
> +	uint32_t i;
> +	uint32_t poll_cnt = 0;
> +	int ret;
> +
> +	worker_info->stop_flag = false;
> +	worker_info->ready_flag = true;
> +
> +	while (!worker_info->start_flag)
> +		;
> +
> +	while (1) {
> +		for (i = 0; i < nr_buf; i++) {
> +dma_copy:
> +			ret = rte_dma_copy(dev_id, 0,
> rte_pktmbuf_iova(srcs[i]),
> +				rte_pktmbuf_iova(dsts[i]), buf_size, 0);

[Anoob] Do we need to use ' rte_mbuf_data_iova' here instead of 'rte_pktmbuf_iova'? 

> +			if (unlikely(ret < 0)) {
> +				if (ret == -ENOSPC) {
> +					do_dma_submit_and_poll(dev_id,
> &async_cnt, worker_info);
> +					goto dma_copy;
> +				} else {
> +					/* Error exit */
> +					rte_dma_stop(dev_id);

[Anoob] Missing rte_dma_close() here. Also, may be introduce a static void function so that rte_exit call etc won't be part of the fastpath loop.

May be something like below and you can call it from here and "do_dma_submit_and_poll".

static void
error_exit(int dev_id)
{
	/* Error exit */
	rte_dma_stop(dev_id);
	rte_dma_close(dev_id);
	rte_exit(EXIT_FAILURE, "DMA enqueue failed\n");
}

> +					rte_exit(EXIT_FAILURE, "DMA
> enqueue failed\n");
> +				}
> +			}
> +			async_cnt++;
> +
> +			if ((async_cnt % kick_batch) == 0)
> +				do_dma_submit_and_poll(dev_id,
> &async_cnt, worker_info);
> +		}
> +
> +		if (worker_info->stop_flag)
> +			break;
> +	}
> +
> +	rte_dma_submit(dev_id, 0);
> +	while ((async_cnt > 0) && (poll_cnt++ < POLL_MAX)) {
> +		nr_cpl = rte_dma_completed(dev_id, 0,
> MAX_DMA_CPL_NB, NULL, NULL);
> +		async_cnt -= nr_cpl;
> +	}
> +
> +	return 0;
> +}
> +

<snip>

> +static int
> +setup_memory_env(struct test_configure *cfg, struct rte_mbuf ***srcs,
> +			struct rte_mbuf ***dsts)
> +{
> +	unsigned int buf_size = cfg->buf_size.cur;
> +	unsigned int nr_sockets;
> +	uint32_t nr_buf = cfg->nr_buf;
> +
> +	nr_sockets = rte_socket_count();
> +	if (cfg->src_numa_node >= nr_sockets ||
> +		cfg->dst_numa_node >= nr_sockets) {
> +		printf("Error: Source or destination numa exceeds the acture
> numa nodes.\n");
> +		return -1;
> +	}
> +
> +	src_pool = rte_pktmbuf_pool_create("Benchmark_DMA_SRC",
> +			nr_buf,
> +			0,
> +			0,
> +			buf_size + RTE_PKTMBUF_HEADROOM,
> +			cfg->src_numa_node);
> +	if (src_pool == NULL) {
> +		PRINT_ERR("Error with source mempool creation.\n");
> +		return -1;
> +	}
> +
> +	dst_pool = rte_pktmbuf_pool_create("Benchmark_DMA_DST",
> +			nr_buf,
> +			0,
> +			0,
> +			buf_size + RTE_PKTMBUF_HEADROOM,
> +			cfg->dst_numa_node);
> +	if (dst_pool == NULL) {
> +		PRINT_ERR("Error with destination mempool creation.\n");
> +		return -1;
> +	}
> +
> +	*srcs = rte_malloc(NULL, nr_buf * sizeof(struct rte_mbuf *), 0);
> +	if (*srcs == NULL) {
> +		printf("Error: srcs malloc failed.\n");
> +		return -1;
> +	}
> +
> +	*dsts = rte_malloc(NULL, nr_buf * sizeof(struct rte_mbuf *), 0);
> +	if (*dsts == NULL) {
> +		printf("Error: dsts malloc failed.\n");
> +		return -1;
> +	}
> +
> +	if (rte_mempool_get_bulk(src_pool, (void **)*srcs, nr_buf) != 0) {

[Anoob] Might be better to use 'rte_pktmbuf_alloc_bulk' since we use ' rte_mbuf_data_iova' in the datapath and it is desirable to initialize it properly as an mbuf. Same comment below as well.

<snip>

> +
> +	for (i = 0; i < nb_workers; i++) {
> +		calc_result(buf_size, nr_buf, nb_workers, test_secs,
> +			lcores_p[i].v_ptr->worker_info.test_cpl,
> +			&memory, &avg_cycles, &bandwidth, &mops);
> +		output_result(cfg->scenario_id, lcores_p[i].v_ptr->lcore_id,
> +					lcores_p[i].v_ptr->dma_name,
> avg_cycles, buf_size,
> +					nr_buf / nb_workers, memory,
> bandwidth, mops, is_dma);
> +	}

[Anoob] Can you also print total_bandwidth & total_mops? It can be a simple aggregation in the above loop. Would help when we are dealing with larger number of queues but single hardware block.

> +
> +out:
> +	/* free mbufs used in the test */
> +	if (srcs)

[Anoob] DPDK coding guidelines recommend a usage like below,
	(if (srcs != NULL)

> +		rte_pktmbuf_free_bulk(srcs, nr_buf);
> +	if (dsts)
> +		rte_pktmbuf_free_bulk(dsts, nr_buf);
> +
> +	/* free the points for the mbufs */
> +	rte_free(srcs);
> +	srcs = NULL;
> +	rte_free(dsts);
> +	dsts = NULL;
> +
> +	if (src_pool) {
> +		rte_mempool_free(src_pool);
> +		src_pool = NULL;
> +	}
> +	if (dst_pool) {
> +		rte_mempool_free(dst_pool);
> +		src_pool = NULL;

[Anoob] Should be dst_pool, right?

<snip>

> diff --git a/app/test-dma-perf/main.h b/app/test-dma-perf/main.h new file
> mode 100644 index 0000000000..215ac42673
> --- /dev/null
> +++ b/app/test-dma-perf/main.h
> @@ -0,0 +1,69 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2023 Intel Corporation
> + */
> +
> +#ifndef _MAIN_H_
> +#define _MAIN_H_
> +
> +
> +#include <rte_common.h>
> +#include <rte_cycles.h>
> +#include <rte_dev.h>
> +#include <rte_dmadev.h>

[Anoob] Is the above include (rte_dmadev.h) required?

> +
> +#ifndef __maybe_unused
> +#define __maybe_unused	__rte_unused
> +#endif

[Anoob] Can you try to avoid this and use rte_unused or RTE_SET_USED in the cache_flush_buf() function?