* [PATCH 14/19] net/xsc: add ethdev Rx burst
@ 2024-09-06 12:14 WanRenyong
0 siblings, 0 replies; only message in thread
From: WanRenyong @ 2024-09-06 12:14 UTC (permalink / raw)
To: dev; +Cc: ferruh.yigit, nana, WanRenyong, Xiaoxiong Zhang
Implement xsc PMD receive function.
Signed-off-by: WanRenyong <wanry@yunsilicon.com>
Signed-off-by: Xiaoxiong Zhang <zhangxx@yunsilicon.com>
---
drivers/net/xsc/xsc_rxtx.c | 189 ++++++++++++++++++++++++++++++++++++-
drivers/net/xsc/xsc_rxtx.h | 9 ++
2 files changed, 197 insertions(+), 1 deletion(-)
diff --git a/drivers/net/xsc/xsc_rxtx.c b/drivers/net/xsc/xsc_rxtx.c
index 66b1511c6a..28360e62ff 100644
--- a/drivers/net/xsc/xsc_rxtx.c
+++ b/drivers/net/xsc/xsc_rxtx.c
@@ -7,11 +7,198 @@
#include "xsc_dev.h"
#include "xsc_ethdev.h"
#include "xsc_rxtx.h"
+#include "xsc_utils.h"
+#include "xsc_ctrl.h"
+
+#define XSC_CQE_OWNER_MASK 0x1
+#define XSC_CQE_OWNER_HW 0x2
+#define XSC_CQE_OWNER_SW 0x4
+#define XSC_CQE_OWNER_ERR 0x8
+
+#define XSC_MAX_RX_BURST_MBUFS 64
+
+static __rte_always_inline int
+check_cqe_own(volatile struct xsc_cqe *cqe, const uint16_t cqe_n,
+ const uint16_t ci)
+
+{
+ if (unlikely(((cqe->owner & XSC_CQE_OWNER_MASK) !=
+ ((ci >> cqe_n) & XSC_CQE_OWNER_MASK))))
+ return XSC_CQE_OWNER_HW;
+
+ rte_io_rmb();
+ if (cqe->msg_len <= 0 && cqe->is_error)
+ return XSC_CQE_OWNER_ERR;
+
+ return XSC_CQE_OWNER_SW;
+}
+
+static inline void
+xsc_cq_to_mbuf(struct xsc_rxq_data *rxq, struct rte_mbuf *pkt,
+ volatile struct xsc_cqe *cqe)
+{
+ uint32_t rss_hash_res = 0;
+ pkt->port = rxq->port_id;
+ if (rxq->rss_hash) {
+ rss_hash_res = rte_be_to_cpu_32(cqe->vni);
+ if (rss_hash_res) {
+ pkt->hash.rss = rss_hash_res;
+ pkt->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
+ }
+ }
+}
+
+static inline int
+xsc_rx_poll_len(struct xsc_rxq_data *rxq, volatile struct xsc_cqe *cqe)
+{
+ int len;
+
+ do {
+ len = 0;
+ int ret;
+
+ ret = check_cqe_own(cqe, rxq->cqe_n, rxq->cq_ci);
+ if (unlikely(ret != XSC_CQE_OWNER_SW)) {
+ if (unlikely(ret == XSC_CQE_OWNER_ERR)) {
+ /* TODO */
+ if (ret == XSC_CQE_OWNER_HW ||
+ ret == -1)
+ return 0;
+ } else {
+ return 0;
+ }
+ }
+
+ rxq->cq_ci += 1;
+ len = rte_le_to_cpu_32(cqe->msg_len);
+ return len;
+ } while (1);
+}
+
+static __rte_always_inline void
+xsc_pkt_info_sync(struct rte_mbuf *rep, struct rte_mbuf *seg)
+{
+ if (rep != NULL && seg != NULL) {
+ rep->data_len = seg->data_len;
+ rep->pkt_len = seg->pkt_len;
+ rep->data_off = seg->data_off;
+ rep->port = seg->port;
+ }
+}
uint16_t
xsc_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
- return 0;
+ struct xsc_rxq_data *rxq = dpdk_rxq;
+ const uint32_t wqe_m = rxq->wqe_m;
+ const uint32_t cqe_m = rxq->cqe_m;
+ const uint32_t sge_n = rxq->sge_n;
+ struct rte_mbuf *pkt = NULL;
+ struct rte_mbuf *seg = NULL;
+ volatile struct xsc_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_m];
+ uint32_t nb_pkts = 0;
+ uint32_t rq_ci = rxq->rq_ci;
+ int len = 0;
+ uint32_t cq_ci_two = 0;
+ int read_cqe_num = 0;
+ int read_cqe_num_len = 0;
+ volatile struct xsc_cqe_u64 *cqe_u64 = NULL;
+ struct rte_mbuf *rep;
+
+ while (pkts_n) {
+ uint32_t idx = rq_ci & wqe_m;
+ volatile struct xsc_wqe_data_seg *wqe =
+ &((volatile struct xsc_wqe_data_seg *)rxq->wqes)[idx << sge_n];
+ seg = (*rxq->elts)[idx];
+ rte_prefetch0(cqe);
+ rte_prefetch0(wqe);
+
+ rep = rte_mbuf_raw_alloc(seg->pool);
+ if (unlikely(rep == NULL))
+ break;
+
+ if (!pkt) {
+ if (read_cqe_num) {
+ cqe = cqe + 1;
+ len = read_cqe_num_len;
+ read_cqe_num = 0;
+ } else if ((rxq->cq_ci % 2 == 0) && (pkts_n > 1)) {
+ cq_ci_two = (rxq->cq_ci & rxq->cqe_m) / 2;
+ cqe_u64 = &(*rxq->cqes_u64)[cq_ci_two];
+ cqe = (volatile struct xsc_cqe *)cqe_u64;
+ len = xsc_rx_poll_len(rxq, cqe);
+ if (len > 0) {
+ read_cqe_num_len = xsc_rx_poll_len(rxq, cqe + 1);
+ if (read_cqe_num_len > 0)
+ read_cqe_num = 1;
+ }
+ } else {
+ cqe = &(*rxq->cqes)[rxq->cq_ci & rxq->cqe_m];
+ len = xsc_rx_poll_len(rxq, cqe);
+ }
+
+ if (!len) {
+ rte_mbuf_raw_free(rep);
+ break;
+ }
+
+ if (len > rte_pktmbuf_data_len(seg)) {
+ rte_mbuf_raw_free(rep);
+ pkt = NULL;
+ ++rq_ci;
+ continue;
+ }
+
+ pkt = seg;
+ pkt->ol_flags &= RTE_MBUF_F_EXTERNAL;
+ xsc_cq_to_mbuf(rxq, pkt, cqe);
+
+ if (rxq->crc_present)
+ len -= RTE_ETHER_CRC_LEN;
+ rte_pktmbuf_pkt_len(pkt) = len;
+ }
+
+ xsc_pkt_info_sync(rep, seg);
+ (*rxq->elts)[idx] = rep;
+
+ /* Fill wqe */
+ wqe->va = rte_cpu_to_le_64(rte_pktmbuf_iova(rep));
+ rte_pktmbuf_data_len(seg) = len;
+
+ *(pkts++) = pkt;
+ pkt = NULL;
+ --pkts_n;
+ ++nb_pkts;
+ ++rq_ci;
+ }
+
+ if (unlikely(nb_pkts == 0 && rq_ci == rxq->rq_ci))
+ return 0;
+
+ rxq->rq_ci = rq_ci;
+ rxq->nb_rx_hold += nb_pkts;
+
+ if (rxq->nb_rx_hold >= rxq->rx_free_thresh) {
+ union xsc_cq_doorbell cq_db = {
+ .cq_data = 0
+ };
+ cq_db.next_cid = rxq->cq_ci;
+ cq_db.cq_num = rxq->cqn;
+
+ union xsc_recv_doorbell rq_db = {
+ .recv_data = 0
+ };
+ rq_db.next_pid = (rxq->rq_ci << sge_n);
+ rq_db.qp_num = rxq->qpn;
+
+ rte_io_wmb();
+ *rxq->cq_db = rte_cpu_to_le_32(cq_db.cq_data);
+ rte_io_wmb();
+ *rxq->rq_db = rte_cpu_to_le_32(rq_db.recv_data);
+ rxq->nb_rx_hold = 0;
+ }
+
+ return nb_pkts;
}
uint16_t
diff --git a/drivers/net/xsc/xsc_rxtx.h b/drivers/net/xsc/xsc_rxtx.h
index 0b2fded42b..454d5c0378 100644
--- a/drivers/net/xsc/xsc_rxtx.h
+++ b/drivers/net/xsc/xsc_rxtx.h
@@ -175,6 +175,15 @@ union xsc_recv_doorbell {
uint32_t recv_data;
};
+union xsc_send_doorbell {
+ struct {
+ uint32_t next_pid : 16;
+ uint32_t qp_num : 15;
+ uint32_t rsv : 1;
+ };
+ uint32_t send_data;
+};
+
uint16_t xsc_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
uint16_t xsc_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n);
--
2.25.1
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2024-09-06 12:15 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-09-06 12:14 [PATCH 14/19] net/xsc: add ethdev Rx burst WanRenyong
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).