net/ark: add PCIe IDS for newly supported devices
PATCH v2 1/4 net/ark: add PCIe IDS for newly supported devices
Update documentation and release notes
update link speed capabilities for 200G and 400G
Signed-off-by Ed Czeck <[email protected]>
---
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(-)
doc/guides/nics/ark.rst +7 -0
@@ -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
drivers/net/ark/ark_ethdev.c +9 -1
@@ -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 ;
Signature · 1 line
2.34.1
PATCH v2 2/4 net/ark: remove double mbuf frees
From: John Miller <[email protected]>
Fixes 8b154b690266 ("net/ark: add Rx initial version")
Signed-off-by John Miller <[email protected]>
--- drivers/net/ark/ark_ethdev_rx.c | 4 ---- 1 file changed, 4 deletions(-)
drivers/net/ark/ark_ethdev_rx.c +0 -4
@@ -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);
Signature · 1 line
2.34.1
PATCH v2 3/4 net/ark: improve ring handling for segmented packets
In reply to Ed Czeckhandle case where complete packet is not yet available
Signed-off-by Ed Czeck <[email protected]>
--- drivers/net/ark/ark_ethdev_rx.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-)
drivers/net/ark/ark_ethdev_rx.c +8 -4
@@ -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++;
Signature · 1 line
2.34.1
PATCH v2 4/4 net/ark: improve Rx queue recovery after mbuf exhaustion
In reply to Ed Czeckuse 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 <[email protected]>
---
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(-)
drivers/net/ark/ark_ethdev_rx.c +23 -11
@@ -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) {
Signature · 1 line
2.34.1
Re: [PATCH v2 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion
Quoted 11 lines · On Wed, 10 Sep 2025 14:57:20 -0400 Ed Czeck <[email protected]> wrote:
use 4-K page aligned buffers to reduce PCIe requests reduce message spew
attempt to allocate smaller chunks of buffers during starvationSigned-off-by Ed Czeck <[email protected]>---
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.