[PATCH 24.11 0/3] backport dmaperf bugfix

[PATCH 24.11 0/3] backport dmaperf bugfix

[Chengwen Feng]

This patchset contain two dmaperf bugfix backport, and also include
one commit which will lead to compile error of vduse.h

Chengwen Feng (3):
  vhost: fix compile error when without vduse.h
  app/dma-perf: fix on-flight DMA when verifying data
  app/dma-perf: fix stopping device

 app/test-dma-perf/benchmark.c | 89 ++++++++++++++++++++++++-----------
 lib/vhost/vduse.h             |  2 +-
 2 files changed, 62 insertions(+), 29 deletions(-)

-- 
2.17.1

[PATCH 24.11 1/3] [PATCH 24.11] vhost: fix compile error when without vduse.h

[Chengwen Feng]

Fix compile error when don't set VHOST_HAS_VDUSE.

Fixes: 0ea4b51e84ce ("vhost: fix external buffer in VDUSE")

Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
---
 lib/vhost/vduse.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/lib/vhost/vduse.h b/lib/vhost/vduse.h
index 353f297d5a..e923aeac56 100644
--- a/lib/vhost/vduse.h
+++ b/lib/vhost/vduse.h
@@ -17,7 +17,7 @@ int vduse_device_destroy(const char *path);
 #else
 
 static inline int
-vduse_device_create(const char *path, bool compliant_ol_flags, bool extbuf, bool linearbuf);
+vduse_device_create(const char *path, bool compliant_ol_flags, bool extbuf, bool linearbuf)
 {
 	RTE_SET_USED(compliant_ol_flags);
 	RTE_SET_USED(extbuf);
-- 
2.17.1

Re: [PATCH 24.11 1/3] [PATCH 24.11] vhost: fix compile error when without vduse.h

[Kevin Traynor]

On 25/11/2025 01:51, Chengwen Feng wrote:
> Fix compile error when don't set VHOST_HAS_VDUSE.
> 
> Fixes: 0ea4b51e84ce ("vhost: fix external buffer in VDUSE")
> 
> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
> ---
>  lib/vhost/vduse.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/lib/vhost/vduse.h b/lib/vhost/vduse.h
> index 353f297d5a..e923aeac56 100644
> --- a/lib/vhost/vduse.h
> +++ b/lib/vhost/vduse.h
> @@ -17,7 +17,7 @@ int vduse_device_destroy(const char *path);
>  #else
>  
>  static inline int
> -vduse_device_create(const char *path, bool compliant_ol_flags, bool extbuf, bool linearbuf);
> +vduse_device_create(const char *path, bool compliant_ol_flags, bool extbuf, bool linearbuf)
>  {
>  	RTE_SET_USED(compliant_ol_flags);
>  	RTE_SET_USED(extbuf);

Ah, thanks for finding this. This was a mistake in rebasing for 24.11
that was not caught as the VHOST_HAS_VDUSE is set.

The commit referenced is not pushed yet, so I will fix it directly there
before push, rather than introducing the issue and then fixing it.

Thanks again for catching it,
Kevin.

Re: [PATCH 24.11 1/3] [PATCH 24.11] vhost: fix compile error when without vduse.h

[fengchengwen]

On 11/25/2025 6:30 PM, Kevin Traynor wrote:
> On 25/11/2025 01:51, Chengwen Feng wrote:
>> Fix compile error when don't set VHOST_HAS_VDUSE.
>>
>> Fixes: 0ea4b51e84ce ("vhost: fix external buffer in VDUSE")
>>
>> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
>> ---
>>  lib/vhost/vduse.h | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/lib/vhost/vduse.h b/lib/vhost/vduse.h
>> index 353f297d5a..e923aeac56 100644
>> --- a/lib/vhost/vduse.h
>> +++ b/lib/vhost/vduse.h
>> @@ -17,7 +17,7 @@ int vduse_device_destroy(const char *path);
>>  #else
>>  
>>  static inline int
>> -vduse_device_create(const char *path, bool compliant_ol_flags, bool extbuf, bool linearbuf);
>> +vduse_device_create(const char *path, bool compliant_ol_flags, bool extbuf, bool linearbuf)
>>  {
>>  	RTE_SET_USED(compliant_ol_flags);
>>  	RTE_SET_USED(extbuf);
> 
> Ah, thanks for finding this. This was a mistake in rebasing for 24.11
> that was not caught as the VHOST_HAS_VDUSE is set.
> 
> The commit referenced is not pushed yet, so I will fix it directly there
> before push, rather than introducing the issue and then fixing it.

Okay, I happened to compile dpdk on the old kernel which don't support vduse.

> 
> Thanks again for catching it,
> Kevin.
> 
> 

[PATCH 24.11 2/3] [PATCH 24.11] app/dma-perf: fix on-flight DMA when verifying data

[Chengwen Feng]

[ upstream commit d1b3b669674a17c58eabf3d631b21aaad7232403 ]

There maybe on-flight DMA when verify_data() because the DMA device
may still working when worker exit.

This commit add wait DMA complete stage before worker exit.

Fixes: 623dc9364dc6 ("app/dma-perf: introduce DMA performance test")
Cc: stable@dpdk.org

Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
---
 app/test-dma-perf/benchmark.c | 55 +++++++++++++++++++++++------------
 1 file changed, 36 insertions(+), 19 deletions(-)

diff --git a/app/test-dma-perf/benchmark.c b/app/test-dma-perf/benchmark.c
index c08c5c8dc6..fc583725c4 100644
--- a/app/test-dma-perf/benchmark.c
+++ b/app/test-dma-perf/benchmark.c
@@ -19,7 +19,6 @@
 #define MAX_DMA_CPL_NB 255
 
 #define TEST_WAIT_U_SECOND 10000
-#define POLL_MAX 1000
 
 #define CSV_LINE_DMA_FMT "Scenario %u,%u,%s,%u,%u,%u,%u,%.2lf,%" PRIu64 ",%.3lf,%.3lf\n"
 #define CSV_LINE_CPU_FMT "Scenario %u,%u,NA,NA,NA,%u,%u,%.2lf,%" PRIu64 ",%.3lf,%.3lf\n"
@@ -282,6 +281,40 @@ do_dma_submit_and_poll(uint16_t dev_id, uint64_t *async_cnt,
 	worker_info->total_cpl += nr_cpl;
 }
 
+static int
+do_dma_submit_and_wait_cpl(uint16_t dev_id, uint64_t async_cnt)
+{
+#define MAX_WAIT_MSEC	1000
+#define MAX_POLL	1000
+#define DEQ_SZ		64
+	enum rte_dma_vchan_status st;
+	uint32_t poll_cnt = 0;
+	uint32_t wait_ms = 0;
+	uint16_t nr_cpl;
+
+	rte_dma_submit(dev_id, 0);
+
+	if (rte_dma_vchan_status(dev_id, 0, &st) < 0) {
+		rte_delay_ms(MAX_WAIT_MSEC);
+		goto wait_cpl;
+	}
+
+	while (st == RTE_DMA_VCHAN_ACTIVE && wait_ms++ < MAX_WAIT_MSEC) {
+		rte_delay_ms(1);
+		rte_dma_vchan_status(dev_id, 0, &st);
+	}
+
+wait_cpl:
+	while ((async_cnt > 0) && (poll_cnt++ < MAX_POLL)) {
+		nr_cpl = rte_dma_completed(dev_id, 0, MAX_DMA_CPL_NB, NULL, NULL);
+		async_cnt -= nr_cpl;
+	}
+	if (async_cnt > 0)
+		PRINT_ERR("Error: wait DMA %u failed!\n", dev_id);
+
+	return async_cnt == 0 ? 0 : -1;
+}
+
 static inline int
 do_dma_plain_mem_copy(void *p)
 {
@@ -293,10 +326,8 @@ do_dma_plain_mem_copy(void *p)
 	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;
@@ -327,13 +358,7 @@ do_dma_plain_mem_copy(void *p)
 			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;
+	return do_dma_submit_and_wait_cpl(dev_id, async_cnt);
 }
 
 static inline int
@@ -349,8 +374,6 @@ do_dma_sg_mem_copy(void *p)
 	const uint16_t dev_id = para->dev_id;
 	uint32_t nr_buf = para->nr_buf;
 	uint64_t async_cnt = 0;
-	uint32_t poll_cnt = 0;
-	uint16_t nr_cpl;
 	uint32_t i, j;
 	int ret;
 
@@ -386,13 +409,7 @@ do_dma_sg_mem_copy(void *p)
 			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;
+	return do_dma_submit_and_wait_cpl(dev_id, async_cnt);
 }
 
 static inline int
-- 
2.17.1

[PATCH 24.11 3/3] [PATCH 24.11] app/dma-perf: fix stopping device

[Chengwen Feng]

[ upstream commit cca4c3bc3feb16226480ea9b3c1dc5e7f0116fee ]

The benchmark do stop-dmadev at last, it has two problem:
1. it still invoke stop-dmadev when something wrong before
   config-dmadev.
2. the dmadev may working but the memory are freed.

Fixes: 623dc9364dc6 ("app/dma-perf: introduce DMA performance test")
Cc: stable@dpdk.org

Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
---
 app/test-dma-perf/benchmark.c | 34 +++++++++++++++++++++++++---------
 1 file changed, 25 insertions(+), 9 deletions(-)

diff --git a/app/test-dma-perf/benchmark.c b/app/test-dma-perf/benchmark.c
index fc583725c4..537cc7618c 100644
--- a/app/test-dma-perf/benchmark.c
+++ b/app/test-dma-perf/benchmark.c
@@ -257,6 +257,25 @@ config_dmadevs(struct test_configure *cfg)
 	return 0;
 }
 
+static void
+stop_dmadev(struct test_configure *cfg, bool *stopped)
+{
+	struct lcore_dma_map_t *lcore_dma_map;
+	uint32_t i;
+
+	if (*stopped)
+		return;
+
+	if (cfg->is_dma) {
+		for (i = 0; i < cfg->num_worker; i++) {
+			lcore_dma_map = &cfg->dma_config[i].lcore_dma_map;
+			printf("Stopping dmadev %d\n", lcore_dma_map->dma_id);
+			rte_dma_stop(lcore_dma_map->dma_id);
+		}
+	}
+	*stopped = true;
+}
+
 static void
 error_exit(int dev_id)
 {
@@ -709,6 +728,7 @@ mem_copy_benchmark(struct test_configure *cfg)
 	float memory = 0;
 	uint32_t avg_cycles = 0;
 	uint32_t avg_cycles_total;
+	bool dev_stopped = false;
 	float mops, mops_total;
 	float bandwidth, bandwidth_total;
 	uint32_t nr_sgsrc = 0, nr_sgdst = 0;
@@ -770,7 +790,7 @@ mem_copy_benchmark(struct test_configure *cfg)
 		    vchan_dev->tdir == RTE_DMA_DIR_MEM_TO_DEV) {
 			if (attach_ext_buffer(vchan_dev, lcores[i], cfg->is_sg,
 					      (nr_sgsrc/nb_workers), (nr_sgdst/nb_workers)) < 0)
-				goto out;
+				goto stop_dmadev;
 		}
 
 		rte_eal_remote_launch(get_work_function(cfg), (void *)(lcores[i]), lcore_id);
@@ -805,6 +825,8 @@ mem_copy_benchmark(struct test_configure *cfg)
 
 	rte_eal_mp_wait_lcore();
 
+	stop_dmadev(cfg, &dev_stopped);
+
 	for (k = 0; k < nb_workers; k++) {
 		struct rte_mbuf **src_buf = NULL, **dst_buf = NULL;
 		uint32_t nr_buf_pt = nr_buf / nb_workers;
@@ -889,6 +911,8 @@ mem_copy_benchmark(struct test_configure *cfg)
 			cfg->scenario_id, nr_buf, memory * nb_workers,
 			(avg_cycles_total * (float) 1.0) / nb_workers, bandwidth_total, mops_total);
 
+stop_dmadev:
+	stop_dmadev(cfg, &dev_stopped);
 out:
 
 	for (k = 0; k < nb_workers; k++) {
@@ -944,13 +968,5 @@ mem_copy_benchmark(struct test_configure *cfg)
 		lcores[i] = NULL;
 	}
 
-	if (cfg->is_dma) {
-		for (i = 0; i < nb_workers; i++) {
-			lcore_dma_map = &cfg->dma_config[i].lcore_dma_map;
-			printf("Stopping dmadev %d\n", lcore_dma_map->dma_id);
-			rte_dma_stop(lcore_dma_map->dma_id);
-		}
-	}
-
 	return ret;
 }
-- 
2.17.1

Re: [PATCH 24.11 0/3] backport dmaperf bugfix

[Kevin Traynor]

On 25/11/2025 01:51, Chengwen Feng wrote:
> This patchset contain two dmaperf bugfix backport, and also include
> one commit which will lead to compile error of vduse.h
> 
> Chengwen Feng (3):
>   vhost: fix compile error when without vduse.h
>   app/dma-perf: fix on-flight DMA when verifying data
>   app/dma-perf: fix stopping device
> 
>  app/test-dma-perf/benchmark.c | 89 ++++++++++++++++++++++++-----------
>  lib/vhost/vduse.h             |  2 +-
>  2 files changed, 62 insertions(+), 29 deletions(-)
> 

Thanks Chengwen. I applied the dma-perf patches to 24.11 branch.
Kevin.