[PATCH v2 1/4] net/ark: add PCIe IDS for newly supported devices

[PATCH v2 1/4] net/ark: add PCIe IDS for newly supported devices

[Ed Czeck]

Update documentation and release notes
update link speed capabilities for 200G and 400G

Signed-off-by: Ed Czeck <ed.czeck@atomicrules.com>
---
v2:
- included 2 more device id
- updated link speeds
---
 doc/guides/nics/ark.rst      |  7 +++++++
 drivers/net/ark/ark_ethdev.c | 10 +++++++++-
 2 files changed, 16 insertions(+), 1 deletion(-)

diff --git a/doc/guides/nics/ark.rst b/doc/guides/nics/ark.rst
index 6aabde2ed5..c8ed57289b 100644
--- a/doc/guides/nics/ark.rst
+++ b/doc/guides/nics/ark.rst
@@ -310,6 +310,12 @@ ARK PMD supports the following Arkville RTL PCIe instances including:
 * ``1d6c:1024`` - AR-TK242 [2x100GbE Packet Capture Device]
 * ``1d6c:1025`` - AR-TK242-FX2 [2x100GbE Gen5 Packet Capture Device]
 * ``1d6c:1026`` - AR-TK242-FX2 [1x200GbE Gen5 Packet Capture Device]
+* ``1d6c:102a`` - AR-TK242-FX2 [4x100GbE Gen5 Packet Capture Device]
+* ``1d6c:102b`` - AR-ARKV-FX1 [Arkville 128B DPDK Data Mover for Versal/CPM5]
+* ``1d6c:102c`` - AR-TK242-V80 [Gen5 PCAP Processor]
+* ``1d6c:102d`` - AR-TK242-FX2 [8x10GbE Gen5 Packet Capture-Replay Device]
+* ``1d6c:102e`` - AR-TK242-FX2 [8x25GbE Gen5 Packet Capture-Replay Device]
+* ``1d6c:102f`` - AR-TK242-FX2 [1x400GbE Gen5 Packet Capture-Replay Device]
 
 Arkville RTL Core Configurations
 --------------------------------
@@ -322,6 +328,7 @@ stream interfaces for both AMD/Xilinx and Intel FPGAs.
 
 * ARK-FX0 - 256-bit 32B datapath (PCIe Gen3, Gen4)
 * ARK-FX1 - 512-bit 64B datapath (PCIe Gen3, Gen4, Gen5)
+* ARKV-FX1 - 1024-bit 128B datapath (AMD PCIe Versal, Gen5 )
 * ARK-FX2 - 1024-bit 128B datapath (PCIe Gen5x16 Only)
 
 DPDK and Arkville Firmware Versioning
diff --git a/drivers/net/ark/ark_ethdev.c b/drivers/net/ark/ark_ethdev.c
index c029dc46b3..8d3c0a1b02 100644
--- a/drivers/net/ark/ark_ethdev.c
+++ b/drivers/net/ark/ark_ethdev.c
@@ -102,6 +102,12 @@ static const struct rte_pci_id pci_id_ark_map[] = {
 	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x1024)},
 	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x1025)},
 	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x1026)},
+	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x102a)},
+	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x102b)},
+	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x102c)},
+	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x102d)},
+	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x102e)},
+	{RTE_PCI_DEVICE(AR_VENDOR_ID, 0x102f)},
 	{.vendor_id = 0, /* sentinel */ },
 };
 
@@ -733,7 +739,9 @@ eth_ark_dev_info_get(struct rte_eth_dev *dev,
 				RTE_ETH_LINK_SPEED_25G |
 				RTE_ETH_LINK_SPEED_40G |
 				RTE_ETH_LINK_SPEED_50G |
-				RTE_ETH_LINK_SPEED_100G);
+				RTE_ETH_LINK_SPEED_100G |
+				RTE_ETH_LINK_SPEED_200G |
+				RTE_ETH_LINK_SPEED_400G);
 
 	dev_info->rx_offload_capa = RTE_ETH_RX_OFFLOAD_TIMESTAMP;
 
-- 
2.34.1

[PATCH v2 2/4] net/ark: remove double mbuf frees

[Ed Czeck]

From: John Miller <john.miller@atomicrules.com>

Fixes: 8b154b690266 ("net/ark: add Rx initial version")
Cc: stable@dpdk.org

Signed-off-by: John Miller <john.miller@atomicrules.com>
---
 drivers/net/ark/ark_ethdev_rx.c | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/drivers/net/ark/ark_ethdev_rx.c b/drivers/net/ark/ark_ethdev_rx.c
index 80e431f1ae..6077d98ae0 100644
--- a/drivers/net/ark/ark_ethdev_rx.c
+++ b/drivers/net/ark/ark_ethdev_rx.c
@@ -538,7 +538,6 @@ void
 eth_ark_dev_rx_queue_release(void *vqueue)
 {
 	struct ark_rx_queue *queue;
-	uint32_t i;
 
 	queue = (struct ark_rx_queue *)vqueue;
 	if (queue == 0)
@@ -551,9 +550,6 @@ eth_ark_dev_rx_queue_release(void *vqueue)
 	/* Need to clear out mbufs here, dropping packets along the way */
 	eth_ark_rx_queue_drain(queue);
 
-	for (i = 0; i < queue->queue_size; ++i)
-		rte_pktmbuf_free(queue->reserve_q[i]);
-
 	rte_free(queue->reserve_q);
 	rte_free(queue->paddress_q);
 	rte_free(queue);
-- 
2.34.1

[PATCH v2 3/4] net/ark: improve ring handling for segmented packets

[Ed Czeck]

handle case where complete packet is not yet available

Signed-off-by: Ed Czeck <ed.czeck@atomicrules.com>
---
 drivers/net/ark/ark_ethdev_rx.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ark/ark_ethdev_rx.c b/drivers/net/ark/ark_ethdev_rx.c
index 6077d98ae0..1b5c4b64a4 100644
--- a/drivers/net/ark/ark_ethdev_rx.c
+++ b/drivers/net/ark/ark_ethdev_rx.c
@@ -313,11 +313,15 @@ eth_ark_recv_pkts(void *rx_queue,
 			}
 		}
 
-		if (unlikely(meta->pkt_len > queue->dataroom))
-			cons_index = eth_ark_rx_jumbo
-				(queue, meta, mbuf, cons_index + 1);
-		else
+		if (unlikely(meta->pkt_len > queue->dataroom)) {
+			uint32_t tcons = eth_ark_rx_jumbo(queue, meta, mbuf, cons_index + 1);
+			if ((int32_t)(prod_index - tcons) >= 0)
+				cons_index = tcons; /* nominal condition */
+			else
+				break;
+		} else {
 			cons_index += 1;
+		}
 
 		rx_pkts[nb] = mbuf;
 		nb++;
-- 
2.34.1

[PATCH v2 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion

[Ed Czeck]

use 4-K page aligned buffers to reduce PCIe requests
reduce message spew
attempt to allocate smaller chunks of buffers during starvation

Signed-off-by: Ed Czeck <ed.czeck@atomicrules.com>
---
v2:
- reduced message to single line.
- Added comments on buffer alignment.  PCIe devices deal with page
size of 4096 bytes.  By aligning this buffer to a 4K boundary, we
will reduce the number of PCIe read requests.
---
 drivers/net/ark/ark_ethdev_rx.c | 34 ++++++++++++++++++++++-----------
 1 file changed, 23 insertions(+), 11 deletions(-)

diff --git a/drivers/net/ark/ark_ethdev_rx.c b/drivers/net/ark/ark_ethdev_rx.c
index 1b5c4b64a4..74f6d70d1e 100644
--- a/drivers/net/ark/ark_ethdev_rx.c
+++ b/drivers/net/ark/ark_ethdev_rx.c
@@ -42,6 +42,7 @@ struct __rte_cache_aligned ark_rx_queue {
 	rx_user_meta_hook_fn rx_user_meta_hook;
 	void *ext_user_data;
 
+	uint32_t starvation;
 	uint32_t dataroom;
 	uint32_t headroom;
 
@@ -57,8 +58,6 @@ struct __rte_cache_aligned ark_rx_queue {
 	/* The queue Index is used within the dpdk device structures */
 	uint16_t queue_index;
 
-	uint32_t unused;
-
 	/* next cache line - fields written by device */
 	alignas(RTE_CACHE_LINE_MIN_SIZE) RTE_MARKER cacheline1;
 
@@ -187,10 +186,11 @@ eth_ark_dev_rx_queue_setup(struct rte_eth_dev *dev,
 				   nb_desc * sizeof(struct rte_mbuf *),
 				   512,
 				   socket_id);
+	/* Align buffer to PCIe's page size of 4K to reduce upstream read requests from FPGA */
 	queue->paddress_q =
 		rte_zmalloc_socket("Ark_rx_queue paddr",
 				   nb_desc * sizeof(rte_iova_t),
-				   512,
+				   4096,
 				   socket_id);
 
 	if (queue->reserve_q == 0 || queue->paddress_q == 0) {
@@ -265,6 +265,9 @@ eth_ark_recv_pkts(void *rx_queue,
 		return 0;
 	if (unlikely(nb_pkts == 0))
 		return 0;
+	if (unlikely(queue->starvation))
+		eth_ark_rx_seed_mbufs(queue);
+
 	prod_index = queue->prod_index;
 	cons_index = queue->cons_index;
 	if (prod_index == cons_index)
@@ -453,7 +456,7 @@ eth_ark_rx_stop_queue(struct rte_eth_dev *dev, uint16_t queue_id)
 static inline int
 eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
 {
-	uint32_t limit = (queue->cons_index & ~(ARK_RX_MPU_CHUNK - 1)) +
+	uint32_t limit = RTE_ALIGN_FLOOR(queue->cons_index, ARK_RX_MPU_CHUNK) +
 		queue->queue_size;
 	uint32_t seed_index = queue->seed_index;
 
@@ -461,23 +464,32 @@ eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
 	uint32_t seed_m = queue->seed_index & queue->queue_mask;
 
 	uint32_t nb = limit - seed_index;
+	int status;
 
 	/* Handle wrap around -- remainder is filled on the next call */
 	if (unlikely(seed_m + nb > queue->queue_size))
 		nb = queue->queue_size - seed_m;
 
 	struct rte_mbuf **mbufs = &queue->reserve_q[seed_m];
-	int status = rte_pktmbuf_alloc_bulk(queue->mb_pool, mbufs, nb);
+	do {
+		status = rte_pktmbuf_alloc_bulk(queue->mb_pool, mbufs, nb);
+		if (status == 0)
+			break;
+		/* Try again with a smaller request, keeping aligned with chunk size */
+		nb = RTE_ALIGN_FLOOR(nb / 2, ARK_RX_MPU_CHUNK);
+	} while (nb >= ARK_RX_MPU_CHUNK);
 
 	if (unlikely(status != 0)) {
-		ARK_PMD_LOG(NOTICE,
-			    "Could not allocate %u mbufs from pool"
-			    " for RX queue %u;"
-			    " %u free buffers remaining in queue\n",
-			    nb, queue->queue_index,
-			    queue->seed_index - queue->cons_index);
+		if (queue->starvation == 0) {
+			ARK_PMD_LOG(NOTICE,
+				    "Could not allocate %u mbufs from pool for RX queue %u; %u free buffers remaining\n",
+				    ARK_RX_MPU_CHUNK, queue->queue_index,
+				    queue->seed_index - queue->cons_index);
+			queue->starvation = 1;
+		}
 		return -1;
 	}
+	queue->starvation = 0;
 
 	if (ARK_DEBUG_CORE) {		/* DEBUG */
 		while (count != nb) {
-- 
2.34.1

[PATCH v2 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion

[Ed Czeck]

use 4-K page aligned buffers to reduce PCIe requests
reduce message spew
attempt to allocate smaller chunks of buffers during starvation

Signed-off-by: Ed Czeck <ed.czeck@atomicrules.com>
---
v2:
- reduced message to single line.
- Added comments on buffer alignment.  PCIe devices deal with page
size of 4096 bytes.  By aligning this buffer to a 4K boundary, we
will reduce the number of PCIe read requests.
---
 drivers/net/ark/ark_ethdev_rx.c | 34 ++++++++++++++++++++++-----------
 1 file changed, 23 insertions(+), 11 deletions(-)

diff --git a/drivers/net/ark/ark_ethdev_rx.c b/drivers/net/ark/ark_ethdev_rx.c
index 1b5c4b64a4..74f6d70d1e 100644
--- a/drivers/net/ark/ark_ethdev_rx.c
+++ b/drivers/net/ark/ark_ethdev_rx.c
@@ -42,6 +42,7 @@ struct __rte_cache_aligned ark_rx_queue {
 	rx_user_meta_hook_fn rx_user_meta_hook;
 	void *ext_user_data;
 
+	uint32_t starvation;
 	uint32_t dataroom;
 	uint32_t headroom;
 
@@ -57,8 +58,6 @@ struct __rte_cache_aligned ark_rx_queue {
 	/* The queue Index is used within the dpdk device structures */
 	uint16_t queue_index;
 
-	uint32_t unused;
-
 	/* next cache line - fields written by device */
 	alignas(RTE_CACHE_LINE_MIN_SIZE) RTE_MARKER cacheline1;
 
@@ -187,10 +186,11 @@ eth_ark_dev_rx_queue_setup(struct rte_eth_dev *dev,
 				   nb_desc * sizeof(struct rte_mbuf *),
 				   512,
 				   socket_id);
+	/* Align buffer to PCIe's page size of 4K to reduce upstream read requests from FPGA */
 	queue->paddress_q =
 		rte_zmalloc_socket("Ark_rx_queue paddr",
 				   nb_desc * sizeof(rte_iova_t),
-				   512,
+				   4096,
 				   socket_id);
 
 	if (queue->reserve_q == 0 || queue->paddress_q == 0) {
@@ -265,6 +265,9 @@ eth_ark_recv_pkts(void *rx_queue,
 		return 0;
 	if (unlikely(nb_pkts == 0))
 		return 0;
+	if (unlikely(queue->starvation))
+		eth_ark_rx_seed_mbufs(queue);
+
 	prod_index = queue->prod_index;
 	cons_index = queue->cons_index;
 	if (prod_index == cons_index)
@@ -453,7 +456,7 @@ eth_ark_rx_stop_queue(struct rte_eth_dev *dev, uint16_t queue_id)
 static inline int
 eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
 {
-	uint32_t limit = (queue->cons_index & ~(ARK_RX_MPU_CHUNK - 1)) +
+	uint32_t limit = RTE_ALIGN_FLOOR(queue->cons_index, ARK_RX_MPU_CHUNK) +
 		queue->queue_size;
 	uint32_t seed_index = queue->seed_index;
 
@@ -461,23 +464,32 @@ eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
 	uint32_t seed_m = queue->seed_index & queue->queue_mask;
 
 	uint32_t nb = limit - seed_index;
+	int status;
 
 	/* Handle wrap around -- remainder is filled on the next call */
 	if (unlikely(seed_m + nb > queue->queue_size))
 		nb = queue->queue_size - seed_m;
 
 	struct rte_mbuf **mbufs = &queue->reserve_q[seed_m];
-	int status = rte_pktmbuf_alloc_bulk(queue->mb_pool, mbufs, nb);
+	do {
+		status = rte_pktmbuf_alloc_bulk(queue->mb_pool, mbufs, nb);
+		if (status == 0)
+			break;
+		/* Try again with a smaller request, keeping aligned with chunk size */
+		nb = RTE_ALIGN_FLOOR(nb / 2, ARK_RX_MPU_CHUNK);
+	} while (nb >= ARK_RX_MPU_CHUNK);
 
 	if (unlikely(status != 0)) {
-		ARK_PMD_LOG(NOTICE,
-			    "Could not allocate %u mbufs from pool"
-			    " for RX queue %u;"
-			    " %u free buffers remaining in queue\n",
-			    nb, queue->queue_index,
-			    queue->seed_index - queue->cons_index);
+		if (queue->starvation == 0) {
+			ARK_PMD_LOG(NOTICE,
+				    "Could not allocate %u mbufs from pool for RX queue %u; %u free buffers remaining\n",
+				    ARK_RX_MPU_CHUNK, queue->queue_index,
+				    queue->seed_index - queue->cons_index);
+			queue->starvation = 1;
+		}
 		return -1;
 	}
+	queue->starvation = 0;
 
 	if (ARK_DEBUG_CORE) {		/* DEBUG */
 		while (count != nb) {
-- 
2.34.1

Re: [PATCH v2 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion

[Stephen Hemminger]

On Wed, 10 Sep 2025 14:57:20 -0400
Ed Czeck <ed.czeck@atomicrules.com> wrote:

> use 4-K page aligned buffers to reduce PCIe requests
> reduce message spew
> attempt to allocate smaller chunks of buffers during starvation
> 
> Signed-off-by: Ed Czeck <ed.czeck@atomicrules.com>
> ---
> v2:
> - reduced message to single line.
> - Added comments on buffer alignment.  PCIe devices deal with page
> size of 4096 bytes.  By aligning this buffer to a 4K boundary, we
> will reduce the number of PCIe read requests.

It would be clearer if there was a #define for the magic 4096 size.
Is the IOMMU page size? The PCI devices themselves don't usually
have page visibility.

Alignment in this case is really a workaround for "don't cross
the boundary". It looks like the rte_malloc library doesn't have
better way to express that.