[net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Suman Ghosh]

This patch implements below changes,

1. To avoid concurrency with normal traffic uses
   XDP queues.
2. Since there are chances that XDP and AF_XDP can
   fall under same queue uses separate flags to handle
   dma buffers.

Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 .../marvell/octeontx2/nic/otx2_common.c       |  4 ++
 .../marvell/octeontx2/nic/otx2_common.h       |  6 +++
 .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  2 +-
 .../marvell/octeontx2/nic/otx2_txrx.c         | 49 +++++++++++++++----
 .../marvell/octeontx2/nic/otx2_txrx.h         |  2 +
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 43 +++++++++++++++-
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  3 ++
 7 files changed, 96 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 6f7c11d59942..84cd029a85aa 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -1037,6 +1037,10 @@ int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
 
 	sq->stats.bytes = 0;
 	sq->stats.pkts = 0;
+	/* Attach XSK_BUFF_POOL to XDP queue */
+	if (qidx > pfvf->hw.xdp_queues)
+		otx2_attach_xsk_buff(pfvf, sq, (qidx - pfvf->hw.xdp_queues));
+
 
 	chan_offset = qidx % pfvf->hw.tx_chan_cnt;
 	err = pfvf->hw_ops->sq_aq_init(pfvf, qidx, chan_offset, sqb_aura);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index 19e9e2e72233..1e88422825be 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -129,6 +129,12 @@ enum otx2_errcodes_re {
 	ERRCODE_IL4_CSUM = 0x22,
 };
 
+enum otx2_xdp_action {
+	OTX2_XDP_TX	  = BIT(0),
+	OTX2_XDP_REDIRECT = BIT(1),
+	OTX2_AF_XDP_FRAME = BIT(2),
+};
+
 struct otx2_dev_stats {
 	u64 rx_bytes;
 	u64 rx_frames;
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
index 188ab6b6fb16..47b05a9c3db5 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
@@ -2693,7 +2693,7 @@ static int otx2_xdp_xmit_tx(struct otx2_nic *pf, struct xdp_frame *xdpf,
 		return -ENOMEM;
 
 	err = otx2_xdp_sq_append_pkt(pf, xdpf, dma_addr, xdpf->len,
-				     qidx, XDP_REDIRECT);
+				     qidx, OTX2_XDP_REDIRECT);
 	if (!err) {
 		otx2_dma_unmap_page(pf, dma_addr, xdpf->len, DMA_TO_DEVICE);
 		xdp_return_frame(xdpf);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
index 9a6c1f1a3ee0..6bc5ce5a9f61 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
@@ -20,6 +20,7 @@
 #include "otx2_txrx.h"
 #include "otx2_ptp.h"
 #include "cn10k.h"
+#include "otx2_xsk.h"
 
 #define CQE_ADDR(CQ, idx) ((CQ)->cqe_base + ((CQ)->cqe_size * (idx)))
 #define PTP_PORT	        0x13F
@@ -103,13 +104,19 @@ static unsigned int frag_num(unsigned int i)
 
 static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf,
 				     struct otx2_snd_queue *sq,
-				     struct nix_cqe_tx_s *cqe)
+				     struct nix_cqe_tx_s *cqe,
+				     int *xsk_frames)
 {
 	struct nix_send_comp_s *snd_comp = &cqe->comp;
 	struct sg_list *sg;
 
 	sg = &sq->sg[snd_comp->sqe_id];
-	if (sg->flags & XDP_REDIRECT)
+	if (sg->flags & OTX2_AF_XDP_FRAME) {
+		(*xsk_frames)++;
+		return;
+	}
+
+	if (sg->flags & OTX2_XDP_REDIRECT)
 		otx2_dma_unmap_page(pfvf, sg->dma_addr[0], sg->size[0], DMA_TO_DEVICE);
 	xdp_return_frame((struct xdp_frame *)sg->skb);
 	sg->skb = (u64)NULL;
@@ -434,6 +441,18 @@ int otx2_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
 	return cnt - cq->pool_ptrs;
 }
 
+static void otx2_zc_submit_pkts(struct otx2_nic *pfvf, struct xsk_buff_pool *xsk_pool,
+				int *xsk_frames, int qidx, int budget)
+{
+	if (*xsk_frames)
+		xsk_tx_completed(xsk_pool, *xsk_frames);
+
+	if (xsk_uses_need_wakeup(xsk_pool))
+		xsk_set_tx_need_wakeup(xsk_pool);
+
+	otx2_zc_napi_handler(pfvf, xsk_pool, qidx, budget);
+}
+
 static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 				struct otx2_cq_queue *cq, int budget)
 {
@@ -442,16 +461,22 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 	struct nix_cqe_tx_s *cqe;
 	struct net_device *ndev;
 	int processed_cqe = 0;
+	int xsk_frames = 0;
+
+	qidx = cq->cq_idx - pfvf->hw.rx_queues;
+	sq = &pfvf->qset.sq[qidx];
 
 	if (cq->pend_cqe >= budget)
 		goto process_cqe;
 
-	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
+	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) {
+		if (sq->xsk_pool)
+			otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames,
+					    qidx, budget);
 		return 0;
+	}
 
 process_cqe:
-	qidx = cq->cq_idx - pfvf->hw.rx_queues;
-	sq = &pfvf->qset.sq[qidx];
 
 	while (likely(processed_cqe < budget) && cq->pend_cqe) {
 		cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq);
@@ -461,10 +486,8 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 			break;
 		}
 
-		qidx = cq->cq_idx - pfvf->hw.rx_queues;
-
 		if (cq->cq_type == CQ_XDP)
-			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe);
+			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe, &xsk_frames);
 		else
 			otx2_snd_pkt_handler(pfvf, cq, &pfvf->qset.sq[qidx],
 					     cqe, budget, &tx_pkts, &tx_bytes);
@@ -505,6 +528,10 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 		    netif_carrier_ok(ndev))
 			netif_tx_wake_queue(txq);
 	}
+
+	if (sq->xsk_pool)
+		otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames, qidx, budget);
+
 	return 0;
 }
 
@@ -1499,8 +1526,10 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 		qidx += pfvf->hw.tx_queues;
 		cq->pool_ptrs++;
 		xdpf = xdp_convert_buff_to_frame(&xdp);
-		return otx2_xdp_sq_append_pkt(pfvf, xdpf, cqe->sg.seg_addr,
-					      cqe->sg.seg_size, qidx, XDP_TX);
+		return otx2_xdp_sq_append_pkt(pfvf, xdpf,
+					      cqe->sg.seg_addr,
+					      cqe->sg.seg_size,
+					      qidx, OTX2_XDP_TX);
 	case XDP_REDIRECT:
 		cq->pool_ptrs++;
 		if (xsk_buff) {
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
index 8f346fbc8221..acf259d72008 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
@@ -106,6 +106,8 @@ struct otx2_snd_queue {
 	/* SQE ring and CPT response queue for Inline IPSEC */
 	struct qmem		*sqe_ring;
 	struct qmem		*cpt_resp;
+	/* Buffer pool for af_xdp zero-copy */
+	struct xsk_buff_pool    *xsk_pool;
 } ____cacheline_aligned_in_smp;
 
 enum cq_type {
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
index 3bdee1603fac..ce10caea8511 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
@@ -140,11 +140,14 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qidx)
 {
 	struct net_device *netdev = pf->netdev;
 	struct xsk_buff_pool *pool;
+	struct otx2_snd_queue *sq;
 
 	pool = xsk_get_pool_from_qid(netdev, qidx);
 	if (!pool)
 		return -EINVAL;
 
+	sq = &pf->qset.sq[qidx + pf->hw.tx_queues];
+	sq->xsk_pool = NULL;
 	otx2_clean_up_rq(pf, qidx);
 	clear_bit(qidx, pf->af_xdp_zc_qidx);
 	xsk_pool_dma_unmap(pool, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
@@ -171,7 +174,7 @@ int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 	if (pf->flags & OTX2_FLAG_INTF_DOWN)
 		return -ENETDOWN;
 
-	if (queue_id >= pf->hw.rx_queues)
+	if (queue_id >= pf->hw.rx_queues || queue_id >= pf->hw.tx_queues)
 		return -EINVAL;
 
 	cq_poll = &qset->napi[queue_id];
@@ -179,8 +182,44 @@ int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 		return -EINVAL;
 
 	/* Trigger interrupt */
-	if (!napi_if_scheduled_mark_missed(&cq_poll->napi))
+	if (!napi_if_scheduled_mark_missed(&cq_poll->napi)) {
 		otx2_write64(pf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), BIT_ULL(0));
+		otx2_write64(pf, NIX_LF_CINTX_INT_W1S(cq_poll->cint_idx), BIT_ULL(0));
+	}
 
 	return 0;
 }
+
+void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx)
+{
+	if (test_bit(qidx, pfvf->af_xdp_zc_qidx))
+		sq->xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, qidx);
+}
+
+void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
+			  int queue, int budget)
+{
+	struct xdp_desc *xdp_desc = pool->tx_descs;
+	int err, i, work_done = 0, batch;
+
+	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
+	batch = xsk_tx_peek_release_desc_batch(pool, budget);
+	if (!batch)
+		return;
+
+	for (i = 0; i < batch; i++) {
+		dma_addr_t dma_addr;
+
+		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
+		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr, xdp_desc[i].len,
+					     queue, OTX2_AF_XDP_FRAME);
+		if (!err) {
+			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer packet err%d\n", err);
+			break;
+		}
+		work_done++;
+	}
+
+	if (work_done)
+		xsk_tx_release(pool);
+}
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
index 022b3433edbb..8047fafee8fe 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
@@ -17,5 +17,8 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qid);
 int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
 			    dma_addr_t *dma, int idx);
 int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
+void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
+			  int queue, int budget);
+void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx);
 
 #endif /* OTX2_XSK_H */
-- 
2.25.1

[net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Suman Ghosh]

This patch implements below changes,

1. To avoid concurrency with normal traffic uses
   XDP queues.
2. Since there are chances that XDP and AF_XDP can
   fall under same queue uses separate flags to handle
   dma buffers.

Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 .../marvell/octeontx2/nic/otx2_common.c       |  4 ++
 .../marvell/octeontx2/nic/otx2_common.h       |  6 +++
 .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  2 +-
 .../marvell/octeontx2/nic/otx2_txrx.c         | 49 +++++++++++++++----
 .../marvell/octeontx2/nic/otx2_txrx.h         |  2 +
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 43 +++++++++++++++-
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  3 ++
 7 files changed, 96 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 6f7c11d59942..84cd029a85aa 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -1037,6 +1037,10 @@ int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
 
 	sq->stats.bytes = 0;
 	sq->stats.pkts = 0;
+	/* Attach XSK_BUFF_POOL to XDP queue */
+	if (qidx > pfvf->hw.xdp_queues)
+		otx2_attach_xsk_buff(pfvf, sq, (qidx - pfvf->hw.xdp_queues));
+
 
 	chan_offset = qidx % pfvf->hw.tx_chan_cnt;
 	err = pfvf->hw_ops->sq_aq_init(pfvf, qidx, chan_offset, sqb_aura);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index 19e9e2e72233..1e88422825be 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -129,6 +129,12 @@ enum otx2_errcodes_re {
 	ERRCODE_IL4_CSUM = 0x22,
 };
 
+enum otx2_xdp_action {
+	OTX2_XDP_TX	  = BIT(0),
+	OTX2_XDP_REDIRECT = BIT(1),
+	OTX2_AF_XDP_FRAME = BIT(2),
+};
+
 struct otx2_dev_stats {
 	u64 rx_bytes;
 	u64 rx_frames;
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
index 188ab6b6fb16..47b05a9c3db5 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
@@ -2693,7 +2693,7 @@ static int otx2_xdp_xmit_tx(struct otx2_nic *pf, struct xdp_frame *xdpf,
 		return -ENOMEM;
 
 	err = otx2_xdp_sq_append_pkt(pf, xdpf, dma_addr, xdpf->len,
-				     qidx, XDP_REDIRECT);
+				     qidx, OTX2_XDP_REDIRECT);
 	if (!err) {
 		otx2_dma_unmap_page(pf, dma_addr, xdpf->len, DMA_TO_DEVICE);
 		xdp_return_frame(xdpf);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
index 9a6c1f1a3ee0..6bc5ce5a9f61 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
@@ -20,6 +20,7 @@
 #include "otx2_txrx.h"
 #include "otx2_ptp.h"
 #include "cn10k.h"
+#include "otx2_xsk.h"
 
 #define CQE_ADDR(CQ, idx) ((CQ)->cqe_base + ((CQ)->cqe_size * (idx)))
 #define PTP_PORT	        0x13F
@@ -103,13 +104,19 @@ static unsigned int frag_num(unsigned int i)
 
 static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf,
 				     struct otx2_snd_queue *sq,
-				     struct nix_cqe_tx_s *cqe)
+				     struct nix_cqe_tx_s *cqe,
+				     int *xsk_frames)
 {
 	struct nix_send_comp_s *snd_comp = &cqe->comp;
 	struct sg_list *sg;
 
 	sg = &sq->sg[snd_comp->sqe_id];
-	if (sg->flags & XDP_REDIRECT)
+	if (sg->flags & OTX2_AF_XDP_FRAME) {
+		(*xsk_frames)++;
+		return;
+	}
+
+	if (sg->flags & OTX2_XDP_REDIRECT)
 		otx2_dma_unmap_page(pfvf, sg->dma_addr[0], sg->size[0], DMA_TO_DEVICE);
 	xdp_return_frame((struct xdp_frame *)sg->skb);
 	sg->skb = (u64)NULL;
@@ -434,6 +441,18 @@ int otx2_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
 	return cnt - cq->pool_ptrs;
 }
 
+static void otx2_zc_submit_pkts(struct otx2_nic *pfvf, struct xsk_buff_pool *xsk_pool,
+				int *xsk_frames, int qidx, int budget)
+{
+	if (*xsk_frames)
+		xsk_tx_completed(xsk_pool, *xsk_frames);
+
+	if (xsk_uses_need_wakeup(xsk_pool))
+		xsk_set_tx_need_wakeup(xsk_pool);
+
+	otx2_zc_napi_handler(pfvf, xsk_pool, qidx, budget);
+}
+
 static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 				struct otx2_cq_queue *cq, int budget)
 {
@@ -442,16 +461,22 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 	struct nix_cqe_tx_s *cqe;
 	struct net_device *ndev;
 	int processed_cqe = 0;
+	int xsk_frames = 0;
+
+	qidx = cq->cq_idx - pfvf->hw.rx_queues;
+	sq = &pfvf->qset.sq[qidx];
 
 	if (cq->pend_cqe >= budget)
 		goto process_cqe;
 
-	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
+	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) {
+		if (sq->xsk_pool)
+			otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames,
+					    qidx, budget);
 		return 0;
+	}
 
 process_cqe:
-	qidx = cq->cq_idx - pfvf->hw.rx_queues;
-	sq = &pfvf->qset.sq[qidx];
 
 	while (likely(processed_cqe < budget) && cq->pend_cqe) {
 		cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq);
@@ -461,10 +486,8 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 			break;
 		}
 
-		qidx = cq->cq_idx - pfvf->hw.rx_queues;
-
 		if (cq->cq_type == CQ_XDP)
-			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe);
+			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe, &xsk_frames);
 		else
 			otx2_snd_pkt_handler(pfvf, cq, &pfvf->qset.sq[qidx],
 					     cqe, budget, &tx_pkts, &tx_bytes);
@@ -505,6 +528,10 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 		    netif_carrier_ok(ndev))
 			netif_tx_wake_queue(txq);
 	}
+
+	if (sq->xsk_pool)
+		otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames, qidx, budget);
+
 	return 0;
 }
 
@@ -1499,8 +1526,10 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 		qidx += pfvf->hw.tx_queues;
 		cq->pool_ptrs++;
 		xdpf = xdp_convert_buff_to_frame(&xdp);
-		return otx2_xdp_sq_append_pkt(pfvf, xdpf, cqe->sg.seg_addr,
-					      cqe->sg.seg_size, qidx, XDP_TX);
+		return otx2_xdp_sq_append_pkt(pfvf, xdpf,
+					      cqe->sg.seg_addr,
+					      cqe->sg.seg_size,
+					      qidx, OTX2_XDP_TX);
 	case XDP_REDIRECT:
 		cq->pool_ptrs++;
 		if (xsk_buff) {
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
index 8f346fbc8221..acf259d72008 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
@@ -106,6 +106,8 @@ struct otx2_snd_queue {
 	/* SQE ring and CPT response queue for Inline IPSEC */
 	struct qmem		*sqe_ring;
 	struct qmem		*cpt_resp;
+	/* Buffer pool for af_xdp zero-copy */
+	struct xsk_buff_pool    *xsk_pool;
 } ____cacheline_aligned_in_smp;
 
 enum cq_type {
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
index 3bdee1603fac..ce10caea8511 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
@@ -140,11 +140,14 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qidx)
 {
 	struct net_device *netdev = pf->netdev;
 	struct xsk_buff_pool *pool;
+	struct otx2_snd_queue *sq;
 
 	pool = xsk_get_pool_from_qid(netdev, qidx);
 	if (!pool)
 		return -EINVAL;
 
+	sq = &pf->qset.sq[qidx + pf->hw.tx_queues];
+	sq->xsk_pool = NULL;
 	otx2_clean_up_rq(pf, qidx);
 	clear_bit(qidx, pf->af_xdp_zc_qidx);
 	xsk_pool_dma_unmap(pool, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
@@ -171,7 +174,7 @@ int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 	if (pf->flags & OTX2_FLAG_INTF_DOWN)
 		return -ENETDOWN;
 
-	if (queue_id >= pf->hw.rx_queues)
+	if (queue_id >= pf->hw.rx_queues || queue_id >= pf->hw.tx_queues)
 		return -EINVAL;
 
 	cq_poll = &qset->napi[queue_id];
@@ -179,8 +182,44 @@ int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 		return -EINVAL;
 
 	/* Trigger interrupt */
-	if (!napi_if_scheduled_mark_missed(&cq_poll->napi))
+	if (!napi_if_scheduled_mark_missed(&cq_poll->napi)) {
 		otx2_write64(pf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), BIT_ULL(0));
+		otx2_write64(pf, NIX_LF_CINTX_INT_W1S(cq_poll->cint_idx), BIT_ULL(0));
+	}
 
 	return 0;
 }
+
+void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx)
+{
+	if (test_bit(qidx, pfvf->af_xdp_zc_qidx))
+		sq->xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, qidx);
+}
+
+void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
+			  int queue, int budget)
+{
+	struct xdp_desc *xdp_desc = pool->tx_descs;
+	int err, i, work_done = 0, batch;
+
+	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
+	batch = xsk_tx_peek_release_desc_batch(pool, budget);
+	if (!batch)
+		return;
+
+	for (i = 0; i < batch; i++) {
+		dma_addr_t dma_addr;
+
+		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
+		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr, xdp_desc[i].len,
+					     queue, OTX2_AF_XDP_FRAME);
+		if (!err) {
+			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer packet err%d\n", err);
+			break;
+		}
+		work_done++;
+	}
+
+	if (work_done)
+		xsk_tx_release(pool);
+}
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
index 022b3433edbb..8047fafee8fe 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
@@ -17,5 +17,8 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qid);
 int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
 			    dma_addr_t *dma, int idx);
 int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
+void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
+			  int queue, int budget);
+void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx);
 
 #endif /* OTX2_XSK_H */
-- 
2.25.1

[net-next PATCH v6 0/6] Add af_xdp support for cn10k

[Suman Ghosh]

This patchset includes changes to support AF_XDP for cn10k chipsets. Both
non-zero copy and zero copy will be supported after these changes. Also,
the RSS will be reconfigured once a particular receive queue is
added/removed to/from AF_XDP support.

Patch #1: octeontx2-pf: use xdp_return_frame() to free xdp buffers

Patch #2: octeontx2-pf: Add AF_XDP non-zero copy support

Patch #3: octeontx2-pf: AF_XDP zero copy receive support

Patch #4: octeontx2-pf: Reconfigure RSS table after enabling AF_XDP
zerocopy on rx queue

Patch #5: octeontx2-pf: Prepare for AF_XDP transmit

Patch #6: octeontx2-pf: AF_XDP zero copy transmit support

Geetha sowjanya (1):
  octeontx2-pf: use xdp_return_frame() to free xdp buffers

Hariprasad Kelam (2):
  octeontx2-pf: Prepare for AF_XDP
  octeontx2-pf: AF_XDP zero copy transmit support

Suman Ghosh (3):
  octeontx2-pf: Add AF_XDP non-zero copy support
  octeontx2-pf: AF_XDP zero copy receive support
  octeontx2-pf: Reconfigure RSS table after enabling AF_XDP zerocopy on
    rx queue

v6 changes:
- Updated patch #1,#3,#5 and #6 to address review comments
  from Simon for some code re-arrangement

v5 changes:
- Updated patch #1 to use xdp_return_frame 
- Updated patch #6 to use xdp_return_frame

v4 changes:
- Addressed minor comments from Paolo regarding adding fixes tag in patch#2
  and removed one unnecessary NULL check from patch#3

v3 changes:
- Rearrenged patch ordering to fix individual patch compilation issue
- Fixed un-initialized variable declaration and reverse x-mas tree issue
  pointed by Simon

v2 changes:
- Addressed minor review comments from Simon regrading smatch warnings

 .../ethernet/marvell/octeontx2/nic/Makefile   |   2 +-
 .../ethernet/marvell/octeontx2/nic/cn10k.c    |   7 +-
 .../marvell/octeontx2/nic/otx2_common.c       | 122 +++++++---
 .../marvell/octeontx2/nic/otx2_common.h       |  17 +-
 .../marvell/octeontx2/nic/otx2_ethtool.c      |   6 +-
 .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  32 +--
 .../marvell/octeontx2/nic/otx2_txrx.c         | 188 +++++++++++----
 .../marvell/octeontx2/nic/otx2_txrx.h         |   9 +
 .../ethernet/marvell/octeontx2/nic/otx2_vf.c  |  12 +-
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 225 ++++++++++++++++++
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  24 ++
 .../ethernet/marvell/octeontx2/nic/qos_sq.c   |   2 +-
 12 files changed, 554 insertions(+), 92 deletions(-)
 create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
 create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h

-- 
2.25.1

[net-next PATCH v6 1/6] octeontx2-pf: use xdp_return_frame() to free xdp buffers

[Suman Ghosh]

xdp_return_frames() will help to free the xdp frames and their
associated pages back to page pool.

Signed-off-by: Geetha sowjanya <gakula@marvell.com>
Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 .../marvell/octeontx2/nic/otx2_common.h       |  4 +-
 .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  7 ++-
 .../marvell/octeontx2/nic/otx2_txrx.c         | 53 +++++++++++--------
 .../marvell/octeontx2/nic/otx2_txrx.h         |  1 +
 4 files changed, 38 insertions(+), 27 deletions(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index 65814e3dc93f..d5fbccb289df 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -21,6 +21,7 @@
 #include <linux/time64.h>
 #include <linux/dim.h>
 #include <uapi/linux/if_macsec.h>
+#include <net/page_pool/helpers.h>
 
 #include <mbox.h>
 #include <npc.h>
@@ -1094,7 +1095,8 @@ int otx2_del_macfilter(struct net_device *netdev, const u8 *mac);
 int otx2_add_macfilter(struct net_device *netdev, const u8 *mac);
 int otx2_enable_rxvlan(struct otx2_nic *pf, bool enable);
 int otx2_install_rxvlan_offload_flow(struct otx2_nic *pfvf);
-bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, u64 iova, int len, u16 qidx);
+bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, struct xdp_frame *xdpf,
+			    u64 iova, int len, u16 qidx, u16 flags);
 u16 otx2_get_max_mtu(struct otx2_nic *pfvf);
 int otx2_handle_ntuple_tc_features(struct net_device *netdev,
 				   netdev_features_t features);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
index e1dde93e8af8..4347a3c95350 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
@@ -2691,7 +2691,6 @@ static int otx2_get_vf_config(struct net_device *netdev, int vf,
 static int otx2_xdp_xmit_tx(struct otx2_nic *pf, struct xdp_frame *xdpf,
 			    int qidx)
 {
-	struct page *page;
 	u64 dma_addr;
 	int err = 0;
 
@@ -2701,11 +2700,11 @@ static int otx2_xdp_xmit_tx(struct otx2_nic *pf, struct xdp_frame *xdpf,
 	if (dma_mapping_error(pf->dev, dma_addr))
 		return -ENOMEM;
 
-	err = otx2_xdp_sq_append_pkt(pf, dma_addr, xdpf->len, qidx);
+	err = otx2_xdp_sq_append_pkt(pf, xdpf, dma_addr, xdpf->len,
+				     qidx, XDP_REDIRECT);
 	if (!err) {
 		otx2_dma_unmap_page(pf, dma_addr, xdpf->len, DMA_TO_DEVICE);
-		page = virt_to_page(xdpf->data);
-		put_page(page);
+		xdp_return_frame(xdpf);
 		return -ENOMEM;
 	}
 	return 0;
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
index 224cef938927..4a7275043103 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
@@ -96,20 +96,16 @@ static unsigned int frag_num(unsigned int i)
 
 static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf,
 				     struct otx2_snd_queue *sq,
-				 struct nix_cqe_tx_s *cqe)
+				     struct nix_cqe_tx_s *cqe)
 {
 	struct nix_send_comp_s *snd_comp = &cqe->comp;
 	struct sg_list *sg;
-	struct page *page;
-	u64 pa;
 
 	sg = &sq->sg[snd_comp->sqe_id];
-
-	pa = otx2_iova_to_phys(pfvf->iommu_domain, sg->dma_addr[0]);
-	otx2_dma_unmap_page(pfvf, sg->dma_addr[0],
-			    sg->size[0], DMA_TO_DEVICE);
-	page = virt_to_page(phys_to_virt(pa));
-	put_page(page);
+	if (sg->flags & XDP_REDIRECT)
+		otx2_dma_unmap_page(pfvf, sg->dma_addr[0], sg->size[0], DMA_TO_DEVICE);
+	xdp_return_frame((struct xdp_frame *)sg->skb);
+	sg->skb = (u64)NULL;
 }
 
 static void otx2_snd_pkt_handler(struct otx2_nic *pfvf,
@@ -1359,8 +1355,9 @@ void otx2_free_pending_sqe(struct otx2_nic *pfvf)
 	}
 }
 
-static void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq, u64 dma_addr,
-				int len, int *offset)
+static void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq,
+				struct xdp_frame *xdpf,
+				u64 dma_addr, int len, int *offset, u16 flags)
 {
 	struct nix_sqe_sg_s *sg = NULL;
 	u64 *iova = NULL;
@@ -1377,9 +1374,12 @@ static void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq, u64 dma_addr,
 	sq->sg[sq->head].dma_addr[0] = dma_addr;
 	sq->sg[sq->head].size[0] = len;
 	sq->sg[sq->head].num_segs = 1;
+	sq->sg[sq->head].flags = flags;
+	sq->sg[sq->head].skb = (u64)xdpf;
 }
 
-bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, u64 iova, int len, u16 qidx)
+bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, struct xdp_frame *xdpf,
+			    u64 iova, int len, u16 qidx, u16 flags)
 {
 	struct nix_sqe_hdr_s *sqe_hdr;
 	struct otx2_snd_queue *sq;
@@ -1405,7 +1405,7 @@ bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, u64 iova, int len, u16 qidx)
 
 	offset = sizeof(*sqe_hdr);
 
-	otx2_xdp_sqe_add_sg(sq, iova, len, &offset);
+	otx2_xdp_sqe_add_sg(sq, xdpf, iova, len, &offset, flags);
 	sqe_hdr->sizem1 = (offset / 16) - 1;
 	pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx);
 
@@ -1419,6 +1419,8 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 				     bool *need_xdp_flush)
 {
 	unsigned char *hard_start;
+	struct otx2_pool *pool;
+	struct xdp_frame *xdpf;
 	int qidx = cq->cq_idx;
 	struct xdp_buff xdp;
 	struct page *page;
@@ -1426,6 +1428,7 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 	u32 act;
 	int err;
 
+	pool = &pfvf->qset.pool[qidx];
 	iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM;
 	pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
 	page = virt_to_page(phys_to_virt(pa));
@@ -1444,19 +1447,21 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 	case XDP_TX:
 		qidx += pfvf->hw.tx_queues;
 		cq->pool_ptrs++;
-		return otx2_xdp_sq_append_pkt(pfvf, iova,
-					      cqe->sg.seg_size, qidx);
+		xdpf = xdp_convert_buff_to_frame(&xdp);
+		return otx2_xdp_sq_append_pkt(pfvf, xdpf, cqe->sg.seg_addr,
+					      cqe->sg.seg_size, qidx, XDP_TX);
 	case XDP_REDIRECT:
 		cq->pool_ptrs++;
 		err = xdp_do_redirect(pfvf->netdev, &xdp, prog);
-
-		otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
-				    DMA_FROM_DEVICE);
 		if (!err) {
 			*need_xdp_flush = true;
 			return true;
 		}
-		put_page(page);
+
+		otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
+				    DMA_FROM_DEVICE);
+		xdpf = xdp_convert_buff_to_frame(&xdp);
+		xdp_return_frame(xdpf);
 		break;
 	default:
 		bpf_warn_invalid_xdp_action(pfvf->netdev, prog, act);
@@ -1465,10 +1470,14 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 		trace_xdp_exception(pfvf->netdev, prog, act);
 		break;
 	case XDP_DROP:
-		otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
-				    DMA_FROM_DEVICE);
-		put_page(page);
 		cq->pool_ptrs++;
+		if (page->pp) {
+			page_pool_recycle_direct(pool->page_pool, page);
+		} else {
+			otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
+					    DMA_FROM_DEVICE);
+			put_page(page);
+		}
 		return true;
 	}
 	return false;
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
index d23810963fdb..92e1e84cad75 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
@@ -76,6 +76,7 @@ struct otx2_rcv_queue {
 
 struct sg_list {
 	u16	num_segs;
+	u16	flags;
 	u64	skb;
 	u64	size[OTX2_MAX_FRAGS_IN_SQE];
 	u64	dma_addr[OTX2_MAX_FRAGS_IN_SQE];
-- 
2.25.1

[net-next PATCH v6 2/6] octeontx2-pf: Add AF_XDP non-zero copy support

[Suman Ghosh]

Set xdp rx ring memory type as MEM_TYPE_PAGE_POOL for
af-xdp to work. This is needed since xdp_return_frame
internally will use page pools.

Fixes: 06059a1a9a4a ("octeontx2-pf: Add XDP support to netdev PF")
Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 2b49bfec7869..161cf33ef89e 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -1047,6 +1047,7 @@ static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
 	int err, pool_id, non_xdp_queues;
 	struct nix_aq_enq_req *aq;
 	struct otx2_cq_queue *cq;
+	struct otx2_pool *pool;
 
 	cq = &qset->cq[qidx];
 	cq->cq_idx = qidx;
@@ -1055,8 +1056,13 @@ static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
 		cq->cq_type = CQ_RX;
 		cq->cint_idx = qidx;
 		cq->cqe_cnt = qset->rqe_cnt;
-		if (pfvf->xdp_prog)
+		if (pfvf->xdp_prog) {
+			pool = &qset->pool[qidx];
 			xdp_rxq_info_reg(&cq->xdp_rxq, pfvf->netdev, qidx, 0);
+			xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
+						   MEM_TYPE_PAGE_POOL,
+						   pool->page_pool);
+		}
 	} else if (qidx < non_xdp_queues) {
 		cq->cq_type = CQ_TX;
 		cq->cint_idx = qidx - pfvf->hw.rx_queues;
-- 
2.25.1

[net-next PATCH v6 3/6] octeontx2-pf: AF_XDP zero copy receive support

[Suman Ghosh]

This patch adds support to AF_XDP zero copy for CN10K.
This patch specifically adds receive side support. In this approach once
a xdp program with zero copy support on a specific rx queue is enabled,
then that receive quse is disabled/detached from the existing kernel
queue and re-assigned to the umem memory.

Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 .../ethernet/marvell/octeontx2/nic/Makefile   |   2 +-
 .../ethernet/marvell/octeontx2/nic/cn10k.c    |   7 +-
 .../marvell/octeontx2/nic/otx2_common.c       | 114 ++++++++---
 .../marvell/octeontx2/nic/otx2_common.h       |   6 +-
 .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  25 ++-
 .../marvell/octeontx2/nic/otx2_txrx.c         |  73 +++++--
 .../marvell/octeontx2/nic/otx2_txrx.h         |   6 +
 .../ethernet/marvell/octeontx2/nic/otx2_vf.c  |  12 +-
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 182 ++++++++++++++++++
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  21 ++
 .../ethernet/marvell/octeontx2/nic/qos_sq.c   |   2 +-
 11 files changed, 389 insertions(+), 61 deletions(-)
 create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
 create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/Makefile b/drivers/net/ethernet/marvell/octeontx2/nic/Makefile
index cb6513ab35e7..69e0778f9ac1 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/Makefile
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/Makefile
@@ -9,7 +9,7 @@ obj-$(CONFIG_RVU_ESWITCH) += rvu_rep.o
 
 rvu_nicpf-y := otx2_pf.o otx2_common.o otx2_txrx.o otx2_ethtool.o \
                otx2_flows.o otx2_tc.o cn10k.o otx2_dmac_flt.o \
-               otx2_devlink.o qos_sq.o qos.o
+               otx2_devlink.o qos_sq.o qos.o otx2_xsk.o
 rvu_nicvf-y := otx2_vf.o
 rvu_rep-y := rep.o
 
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c b/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
index a15cc86635d6..c3b6e0f60a79 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
@@ -112,9 +112,12 @@ int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
 	struct otx2_nic *pfvf = dev;
 	int cnt = cq->pool_ptrs;
 	u64 ptrs[NPA_MAX_BURST];
+	struct otx2_pool *pool;
 	dma_addr_t bufptr;
 	int num_ptrs = 1;
 
+	pool = &pfvf->qset.pool[cq->cq_idx];
+
 	/* Refill pool with new buffers */
 	while (cq->pool_ptrs) {
 		if (otx2_alloc_buffer(pfvf, cq, &bufptr)) {
@@ -124,7 +127,9 @@ int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
 			break;
 		}
 		cq->pool_ptrs--;
-		ptrs[num_ptrs] = (u64)bufptr + OTX2_HEAD_ROOM;
+		ptrs[num_ptrs] = pool->xsk_pool ?
+				 (u64)bufptr : (u64)bufptr + OTX2_HEAD_ROOM;
+
 		num_ptrs++;
 		if (num_ptrs == NPA_MAX_BURST || cq->pool_ptrs == 0) {
 			__cn10k_aura_freeptr(pfvf, cq->cq_idx, ptrs,
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 161cf33ef89e..92b0dba07853 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -17,6 +17,7 @@
 #include "otx2_common.h"
 #include "otx2_struct.h"
 #include "cn10k.h"
+#include "otx2_xsk.h"
 
 static bool otx2_is_pfc_enabled(struct otx2_nic *pfvf)
 {
@@ -549,10 +550,13 @@ static int otx2_alloc_pool_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
 }
 
 static int __otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
-			     dma_addr_t *dma)
+			     dma_addr_t *dma, int qidx, int idx)
 {
 	u8 *buf;
 
+	if (pool->xsk_pool)
+		return otx2_xsk_pool_alloc_buf(pfvf, pool, dma, idx);
+
 	if (pool->page_pool)
 		return otx2_alloc_pool_buf(pfvf, pool, dma);
 
@@ -571,12 +575,12 @@ static int __otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
 }
 
 int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
-		    dma_addr_t *dma)
+		    dma_addr_t *dma, int qidx, int idx)
 {
 	int ret;
 
 	local_bh_disable();
-	ret = __otx2_alloc_rbuf(pfvf, pool, dma);
+	ret = __otx2_alloc_rbuf(pfvf, pool, dma, qidx, idx);
 	local_bh_enable();
 	return ret;
 }
@@ -584,7 +588,8 @@ int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
 int otx2_alloc_buffer(struct otx2_nic *pfvf, struct otx2_cq_queue *cq,
 		      dma_addr_t *dma)
 {
-	if (unlikely(__otx2_alloc_rbuf(pfvf, cq->rbpool, dma)))
+	if (unlikely(__otx2_alloc_rbuf(pfvf, cq->rbpool, dma,
+				       cq->cq_idx, cq->pool_ptrs - 1)))
 		return -ENOMEM;
 	return 0;
 }
@@ -884,7 +889,7 @@ void otx2_sqb_flush(struct otx2_nic *pfvf)
 #define RQ_PASS_LVL_AURA (255 - ((95 * 256) / 100)) /* RED when 95% is full */
 #define RQ_DROP_LVL_AURA (255 - ((99 * 256) / 100)) /* Drop when 99% is full */
 
-static int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura)
+int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura)
 {
 	struct otx2_qset *qset = &pfvf->qset;
 	struct nix_aq_enq_req *aq;
@@ -1041,7 +1046,7 @@ int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
 
 }
 
-static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
+int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
 {
 	struct otx2_qset *qset = &pfvf->qset;
 	int err, pool_id, non_xdp_queues;
@@ -1057,11 +1062,18 @@ static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
 		cq->cint_idx = qidx;
 		cq->cqe_cnt = qset->rqe_cnt;
 		if (pfvf->xdp_prog) {
-			pool = &qset->pool[qidx];
 			xdp_rxq_info_reg(&cq->xdp_rxq, pfvf->netdev, qidx, 0);
-			xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
-						   MEM_TYPE_PAGE_POOL,
-						   pool->page_pool);
+			pool = &qset->pool[qidx];
+			if (pool->xsk_pool) {
+				xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
+							   MEM_TYPE_XSK_BUFF_POOL,
+							   NULL);
+				xsk_pool_set_rxq_info(pool->xsk_pool, &cq->xdp_rxq);
+			} else if (pool->page_pool) {
+				xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
+							   MEM_TYPE_PAGE_POOL,
+							   pool->page_pool);
+			}
 		}
 	} else if (qidx < non_xdp_queues) {
 		cq->cq_type = CQ_TX;
@@ -1281,9 +1293,10 @@ void otx2_free_bufs(struct otx2_nic *pfvf, struct otx2_pool *pool,
 
 	pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
 	page = virt_to_head_page(phys_to_virt(pa));
-
 	if (pool->page_pool) {
 		page_pool_put_full_page(pool->page_pool, page, true);
+	} else if (pool->xsk_pool) {
+		/* Note: No way of identifying xdp_buff */
 	} else {
 		dma_unmap_page_attrs(pfvf->dev, iova, size,
 				     DMA_FROM_DEVICE,
@@ -1298,6 +1311,7 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
 	int pool_id, pool_start = 0, pool_end = 0, size = 0;
 	struct otx2_pool *pool;
 	u64 iova;
+	int idx;
 
 	if (type == AURA_NIX_SQ) {
 		pool_start = otx2_get_pool_idx(pfvf, type, 0);
@@ -1312,16 +1326,21 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
 
 	/* Free SQB and RQB pointers from the aura pool */
 	for (pool_id = pool_start; pool_id < pool_end; pool_id++) {
-		iova = otx2_aura_allocptr(pfvf, pool_id);
 		pool = &pfvf->qset.pool[pool_id];
+		iova = otx2_aura_allocptr(pfvf, pool_id);
 		while (iova) {
 			if (type == AURA_NIX_RQ)
 				iova -= OTX2_HEAD_ROOM;
-
 			otx2_free_bufs(pfvf, pool, iova, size);
-
 			iova = otx2_aura_allocptr(pfvf, pool_id);
 		}
+
+		for (idx = 0 ; idx < pool->xdp_cnt; idx++) {
+			if (!pool->xdp[idx])
+				continue;
+
+			xsk_buff_free(pool->xdp[idx]);
+		}
 	}
 }
 
@@ -1338,7 +1357,8 @@ void otx2_aura_pool_free(struct otx2_nic *pfvf)
 		qmem_free(pfvf->dev, pool->stack);
 		qmem_free(pfvf->dev, pool->fc_addr);
 		page_pool_destroy(pool->page_pool);
-		pool->page_pool = NULL;
+		devm_kfree(pfvf->dev, pool->xdp);
+		pool->xsk_pool = NULL;
 	}
 	devm_kfree(pfvf->dev, pfvf->qset.pool);
 	pfvf->qset.pool = NULL;
@@ -1425,6 +1445,7 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
 		   int stack_pages, int numptrs, int buf_size, int type)
 {
 	struct page_pool_params pp_params = { 0 };
+	struct xsk_buff_pool *xsk_pool;
 	struct npa_aq_enq_req *aq;
 	struct otx2_pool *pool;
 	int err;
@@ -1468,21 +1489,35 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
 	aq->ctype = NPA_AQ_CTYPE_POOL;
 	aq->op = NPA_AQ_INSTOP_INIT;
 
-	if (type != AURA_NIX_RQ) {
-		pool->page_pool = NULL;
+	if (type != AURA_NIX_RQ)
+		return 0;
+
+	if (!test_bit(pool_id, pfvf->af_xdp_zc_qidx)) {
+		pp_params.order = get_order(buf_size);
+		pp_params.flags = PP_FLAG_DMA_MAP;
+		pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
+		pp_params.nid = NUMA_NO_NODE;
+		pp_params.dev = pfvf->dev;
+		pp_params.dma_dir = DMA_FROM_DEVICE;
+		pool->page_pool = page_pool_create(&pp_params);
+		if (IS_ERR(pool->page_pool)) {
+			netdev_err(pfvf->netdev, "Creation of page pool failed\n");
+			return PTR_ERR(pool->page_pool);
+		}
 		return 0;
 	}
 
-	pp_params.order = get_order(buf_size);
-	pp_params.flags = PP_FLAG_DMA_MAP;
-	pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
-	pp_params.nid = NUMA_NO_NODE;
-	pp_params.dev = pfvf->dev;
-	pp_params.dma_dir = DMA_FROM_DEVICE;
-	pool->page_pool = page_pool_create(&pp_params);
-	if (IS_ERR(pool->page_pool)) {
-		netdev_err(pfvf->netdev, "Creation of page pool failed\n");
-		return PTR_ERR(pool->page_pool);
+	/* Set XSK pool to support AF_XDP zero-copy */
+	xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, pool_id);
+	if (xsk_pool) {
+		pool->xsk_pool = xsk_pool;
+		pool->xdp_cnt = numptrs;
+		pool->xdp = devm_kcalloc(pfvf->dev,
+					 numptrs, sizeof(struct xdp_buff *), GFP_KERNEL);
+		if (IS_ERR(pool->xdp)) {
+			netdev_err(pfvf->netdev, "Creation of xsk pool failed\n");
+			return PTR_ERR(pool->xdp);
+		}
 	}
 
 	return 0;
@@ -1543,9 +1578,18 @@ int otx2_sq_aura_pool_init(struct otx2_nic *pfvf)
 		}
 
 		for (ptr = 0; ptr < num_sqbs; ptr++) {
-			err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
-			if (err)
+			err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
+			if (err) {
+				if (pool->xsk_pool) {
+					ptr--;
+					while (ptr >= 0) {
+						xsk_buff_free(pool->xdp[ptr]);
+						ptr--;
+					}
+				}
 				goto err_mem;
+			}
+
 			pfvf->hw_ops->aura_freeptr(pfvf, pool_id, bufptr);
 			sq->sqb_ptrs[sq->sqb_count++] = (u64)bufptr;
 		}
@@ -1595,11 +1639,19 @@ int otx2_rq_aura_pool_init(struct otx2_nic *pfvf)
 	/* Allocate pointers and free them to aura/pool */
 	for (pool_id = 0; pool_id < hw->rqpool_cnt; pool_id++) {
 		pool = &pfvf->qset.pool[pool_id];
+
 		for (ptr = 0; ptr < num_ptrs; ptr++) {
-			err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
-			if (err)
+			err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
+			if (err) {
+				if (pool->xsk_pool) {
+					while (ptr)
+						xsk_buff_free(pool->xdp[--ptr]);
+				}
 				return -ENOMEM;
+			}
+
 			pfvf->hw_ops->aura_freeptr(pfvf, pool_id,
+						   pool->xsk_pool ? bufptr :
 						   bufptr + OTX2_HEAD_ROOM);
 		}
 	}
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index d5fbccb289df..60508971b62f 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -532,6 +532,8 @@ struct otx2_nic {
 
 	/* Inline ipsec */
 	struct cn10k_ipsec	ipsec;
+	/* af_xdp zero-copy */
+	unsigned long		*af_xdp_zc_qidx;
 };
 
 static inline bool is_otx2_lbkvf(struct pci_dev *pdev)
@@ -1003,7 +1005,7 @@ void otx2_txschq_free_one(struct otx2_nic *pfvf, u16 lvl, u16 schq);
 void otx2_free_pending_sqe(struct otx2_nic *pfvf);
 void otx2_sqb_flush(struct otx2_nic *pfvf);
 int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
-		    dma_addr_t *dma);
+		    dma_addr_t *dma, int qidx, int idx);
 int otx2_rxtx_enable(struct otx2_nic *pfvf, bool enable);
 void otx2_ctx_disable(struct mbox *mbox, int type, bool npa);
 int otx2_nix_config_bp(struct otx2_nic *pfvf, bool enable);
@@ -1033,6 +1035,8 @@ void otx2_pfaf_mbox_destroy(struct otx2_nic *pf);
 void otx2_disable_mbox_intr(struct otx2_nic *pf);
 void otx2_disable_napi(struct otx2_nic *pf);
 irqreturn_t otx2_cq_intr_handler(int irq, void *cq_irq);
+int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura);
+int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx);
 
 /* RSS configuration APIs*/
 int otx2_rss_init(struct otx2_nic *pfvf);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
index 4347a3c95350..50a42cd5d50a 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
@@ -27,6 +27,7 @@
 #include "qos.h"
 #include <rvu_trace.h>
 #include "cn10k_ipsec.h"
+#include "otx2_xsk.h"
 
 #define DRV_NAME	"rvu_nicpf"
 #define DRV_STRING	"Marvell RVU NIC Physical Function Driver"
@@ -1662,9 +1663,7 @@ void otx2_free_hw_resources(struct otx2_nic *pf)
 	struct nix_lf_free_req *free_req;
 	struct mbox *mbox = &pf->mbox;
 	struct otx2_cq_queue *cq;
-	struct otx2_pool *pool;
 	struct msg_req *req;
-	int pool_id;
 	int qidx;
 
 	/* Ensure all SQE are processed */
@@ -1705,13 +1704,6 @@ void otx2_free_hw_resources(struct otx2_nic *pf)
 	/* Free RQ buffer pointers*/
 	otx2_free_aura_ptr(pf, AURA_NIX_RQ);
 
-	for (qidx = 0; qidx < pf->hw.rx_queues; qidx++) {
-		pool_id = otx2_get_pool_idx(pf, AURA_NIX_RQ, qidx);
-		pool = &pf->qset.pool[pool_id];
-		page_pool_destroy(pool->page_pool);
-		pool->page_pool = NULL;
-	}
-
 	otx2_free_cq_res(pf);
 
 	/* Free all ingress bandwidth profiles allocated */
@@ -2788,6 +2780,8 @@ static int otx2_xdp(struct net_device *netdev, struct netdev_bpf *xdp)
 	switch (xdp->command) {
 	case XDP_SETUP_PROG:
 		return otx2_xdp_setup(pf, xdp->prog);
+	case XDP_SETUP_XSK_POOL:
+		return otx2_xsk_pool_setup(pf, xdp->xsk.pool, xdp->xsk.queue_id);
 	default:
 		return -EINVAL;
 	}
@@ -2865,6 +2859,7 @@ static const struct net_device_ops otx2_netdev_ops = {
 	.ndo_set_vf_vlan	= otx2_set_vf_vlan,
 	.ndo_get_vf_config	= otx2_get_vf_config,
 	.ndo_bpf		= otx2_xdp,
+	.ndo_xsk_wakeup		= otx2_xsk_wakeup,
 	.ndo_xdp_xmit           = otx2_xdp_xmit,
 	.ndo_setup_tc		= otx2_setup_tc,
 	.ndo_set_vf_trust	= otx2_ndo_set_vf_trust,
@@ -3203,16 +3198,26 @@ static int otx2_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	/* Enable link notifications */
 	otx2_cgx_config_linkevents(pf, true);
 
+	pf->af_xdp_zc_qidx = bitmap_zalloc(qcount, GFP_KERNEL);
+	if (!pf->af_xdp_zc_qidx) {
+		err = -ENOMEM;
+		goto err_sriov_cleannup;
+	}
+
 #ifdef CONFIG_DCB
 	err = otx2_dcbnl_set_ops(netdev);
 	if (err)
-		goto err_pf_sriov_init;
+		goto err_free_zc_bmap;
 #endif
 
 	otx2_qos_init(pf, qos_txqs);
 
 	return 0;
 
+err_free_zc_bmap:
+	bitmap_free(pf->af_xdp_zc_qidx);
+err_sriov_cleannup:
+	otx2_sriov_vfcfg_cleanup(pf);
 err_pf_sriov_init:
 	otx2_shutdown_tc(pf);
 err_mcam_flow_del:
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
index 4a7275043103..00b6903ba250 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
@@ -12,6 +12,7 @@
 #include <linux/bpf_trace.h>
 #include <net/ip6_checksum.h>
 #include <net/xfrm.h>
+#include <net/xdp.h>
 
 #include "otx2_reg.h"
 #include "otx2_common.h"
@@ -523,9 +524,10 @@ static void otx2_adjust_adaptive_coalese(struct otx2_nic *pfvf, struct otx2_cq_p
 int otx2_napi_handler(struct napi_struct *napi, int budget)
 {
 	struct otx2_cq_queue *rx_cq = NULL;
+	struct otx2_cq_queue *cq = NULL;
+	struct otx2_pool *pool = NULL;
 	struct otx2_cq_poll *cq_poll;
 	int workdone = 0, cq_idx, i;
-	struct otx2_cq_queue *cq;
 	struct otx2_qset *qset;
 	struct otx2_nic *pfvf;
 	int filled_cnt = -1;
@@ -550,6 +552,7 @@ int otx2_napi_handler(struct napi_struct *napi, int budget)
 
 	if (rx_cq && rx_cq->pool_ptrs)
 		filled_cnt = pfvf->hw_ops->refill_pool_ptrs(pfvf, rx_cq);
+
 	/* Clear the IRQ */
 	otx2_write64(pfvf, NIX_LF_CINTX_INT(cq_poll->cint_idx), BIT_ULL(0));
 
@@ -562,20 +565,31 @@ int otx2_napi_handler(struct napi_struct *napi, int budget)
 		if (pfvf->flags & OTX2_FLAG_ADPTV_INT_COAL_ENABLED)
 			otx2_adjust_adaptive_coalese(pfvf, cq_poll);
 
+		if (likely(cq))
+			pool = &pfvf->qset.pool[cq->cq_idx];
+
 		if (unlikely(!filled_cnt)) {
 			struct refill_work *work;
 			struct delayed_work *dwork;
 
-			work = &pfvf->refill_wrk[cq->cq_idx];
-			dwork = &work->pool_refill_work;
-			/* Schedule a task if no other task is running */
-			if (!cq->refill_task_sched) {
-				work->napi = napi;
-				cq->refill_task_sched = true;
-				schedule_delayed_work(dwork,
-						      msecs_to_jiffies(100));
+			if (likely(cq)) {
+				work = &pfvf->refill_wrk[cq->cq_idx];
+				dwork = &work->pool_refill_work;
+				/* Schedule a task if no other task is running */
+				if (!cq->refill_task_sched) {
+					work->napi = napi;
+					cq->refill_task_sched = true;
+					schedule_delayed_work(dwork,
+							      msecs_to_jiffies(100));
+				}
+				/* Call wake-up for not able to fill buffers */
+				if (pool->xsk_pool)
+					xsk_set_rx_need_wakeup(pool->xsk_pool);
 			}
 		} else {
+			/* Clear wake-up, since buffers are filled successfully */
+			if (pool && pool->xsk_pool)
+				xsk_clear_rx_need_wakeup(pool->xsk_pool);
 			/* Re-enable interrupts */
 			otx2_write64(pfvf,
 				     NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx),
@@ -1226,15 +1240,19 @@ void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq, int q
 	u16 pool_id;
 	u64 iova;
 
-	if (pfvf->xdp_prog)
+	pool_id = otx2_get_pool_idx(pfvf, AURA_NIX_RQ, qidx);
+	pool = &pfvf->qset.pool[pool_id];
+
+	if (pfvf->xdp_prog) {
+		if (pool->page_pool)
+			xdp_rxq_info_unreg_mem_model(&cq->xdp_rxq);
+
 		xdp_rxq_info_unreg(&cq->xdp_rxq);
+	}
 
 	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
 		return;
 
-	pool_id = otx2_get_pool_idx(pfvf, AURA_NIX_RQ, qidx);
-	pool = &pfvf->qset.pool[pool_id];
-
 	while (cq->pend_cqe) {
 		cqe = (struct nix_cqe_rx_s *)otx2_get_next_cqe(cq);
 		processed_cqe++;
@@ -1418,17 +1436,28 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 				     struct otx2_cq_queue *cq,
 				     bool *need_xdp_flush)
 {
+	struct xdp_buff xdp, *xsk_buff = NULL;
 	unsigned char *hard_start;
 	struct otx2_pool *pool;
 	struct xdp_frame *xdpf;
 	int qidx = cq->cq_idx;
-	struct xdp_buff xdp;
 	struct page *page;
 	u64 iova, pa;
 	u32 act;
 	int err;
 
 	pool = &pfvf->qset.pool[qidx];
+
+	if (pool->xsk_pool) {
+		xsk_buff = pool->xdp[--cq->rbpool->xdp_top];
+		if (!xsk_buff)
+			return false;
+
+		xsk_buff->data_end = xsk_buff->data + cqe->sg.seg_size;
+		act = bpf_prog_run_xdp(prog, xsk_buff);
+		goto handle_xdp_verdict;
+	}
+
 	iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM;
 	pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
 	page = virt_to_page(phys_to_virt(pa));
@@ -1441,6 +1470,7 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 
 	act = bpf_prog_run_xdp(prog, &xdp);
 
+handle_xdp_verdict:
 	switch (act) {
 	case XDP_PASS:
 		break;
@@ -1452,6 +1482,15 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 					      cqe->sg.seg_size, qidx, XDP_TX);
 	case XDP_REDIRECT:
 		cq->pool_ptrs++;
+		if (xsk_buff) {
+			err = xdp_do_redirect(pfvf->netdev, xsk_buff, prog);
+			if (!err) {
+				*need_xdp_flush = true;
+				return true;
+			}
+			return false;
+		}
+
 		err = xdp_do_redirect(pfvf->netdev, &xdp, prog);
 		if (!err) {
 			*need_xdp_flush = true;
@@ -1467,11 +1506,15 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 		bpf_warn_invalid_xdp_action(pfvf->netdev, prog, act);
 		break;
 	case XDP_ABORTED:
+		if (xsk_buff)
+			xsk_buff_free(xsk_buff);
 		trace_xdp_exception(pfvf->netdev, prog, act);
 		break;
 	case XDP_DROP:
 		cq->pool_ptrs++;
-		if (page->pp) {
+		if (xsk_buff) {
+			xsk_buff_free(xsk_buff);
+		} else if (page->pp) {
 			page_pool_recycle_direct(pool->page_pool, page);
 		} else {
 			otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
index 92e1e84cad75..8f346fbc8221 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
@@ -12,6 +12,7 @@
 #include <linux/iommu.h>
 #include <linux/if_vlan.h>
 #include <net/xdp.h>
+#include <net/xdp_sock_drv.h>
 
 #define LBK_CHAN_BASE	0x000
 #define SDP_CHAN_BASE	0x700
@@ -128,7 +129,11 @@ struct otx2_pool {
 	struct qmem		*stack;
 	struct qmem		*fc_addr;
 	struct page_pool	*page_pool;
+	struct xsk_buff_pool	*xsk_pool;
+	struct xdp_buff		**xdp;
+	u16			xdp_cnt;
 	u16			rbsize;
+	u16			xdp_top;
 };
 
 struct otx2_cq_queue {
@@ -145,6 +150,7 @@ struct otx2_cq_queue {
 	void			*cqe_base;
 	struct qmem		*cqe;
 	struct otx2_pool	*rbpool;
+	bool			xsk_zc_en;
 	struct xdp_rxq_info xdp_rxq;
 } ____cacheline_aligned_in_smp;
 
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_vf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_vf.c
index e926c6ce96cf..63ddd262d122 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_vf.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_vf.c
@@ -722,15 +722,25 @@ static int otx2vf_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	if (err)
 		goto err_shutdown_tc;
 
+	vf->af_xdp_zc_qidx = bitmap_zalloc(qcount, GFP_KERNEL);
+	if (!vf->af_xdp_zc_qidx) {
+		err = -ENOMEM;
+		goto err_unreg_devlink;
+	}
+
 #ifdef CONFIG_DCB
 	err = otx2_dcbnl_set_ops(netdev);
 	if (err)
-		goto err_shutdown_tc;
+		goto err_free_zc_bmap;
 #endif
 	otx2_qos_init(vf, qos_txqs);
 
 	return 0;
 
+err_free_zc_bmap:
+	bitmap_free(vf->af_xdp_zc_qidx);
+err_unreg_devlink:
+	otx2_unregister_dl(vf);
 err_shutdown_tc:
 	otx2_shutdown_tc(vf);
 err_unreg_netdev:
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
new file mode 100644
index 000000000000..894c1e0aea6f
--- /dev/null
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Marvell RVU Ethernet driver
+ *
+ * Copyright (C) 2024 Marvell.
+ *
+ */
+
+#include <linux/bpf_trace.h>
+#include <linux/stringify.h>
+#include <net/xdp_sock_drv.h>
+#include <net/xdp.h>
+
+#include "otx2_common.h"
+#include "otx2_xsk.h"
+
+int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
+			    dma_addr_t *dma, int idx)
+{
+	struct xdp_buff *xdp;
+	int delta;
+
+	xdp = xsk_buff_alloc(pool->xsk_pool);
+	if (!xdp)
+		return -ENOMEM;
+
+	pool->xdp[pool->xdp_top++] = xdp;
+	*dma = OTX2_DATA_ALIGN(xsk_buff_xdp_get_dma(xdp));
+	/* Adjust xdp->data for unaligned addresses */
+	delta = *dma - xsk_buff_xdp_get_dma(xdp);
+	xdp->data += delta;
+
+	return 0;
+}
+
+static int otx2_xsk_ctx_disable(struct otx2_nic *pfvf, u16 qidx, int aura_id)
+{
+	struct nix_cn10k_aq_enq_req *cn10k_rq_aq;
+	struct npa_aq_enq_req *aura_aq;
+	struct npa_aq_enq_req *pool_aq;
+	struct nix_aq_enq_req *rq_aq;
+
+	if (test_bit(CN10K_LMTST, &pfvf->hw.cap_flag)) {
+		cn10k_rq_aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
+		if (!cn10k_rq_aq)
+			return -ENOMEM;
+		cn10k_rq_aq->qidx = qidx;
+		cn10k_rq_aq->rq.ena = 0;
+		cn10k_rq_aq->rq_mask.ena = 1;
+		cn10k_rq_aq->ctype = NIX_AQ_CTYPE_RQ;
+		cn10k_rq_aq->op = NIX_AQ_INSTOP_WRITE;
+	} else {
+		rq_aq = otx2_mbox_alloc_msg_nix_aq_enq(&pfvf->mbox);
+		if (!rq_aq)
+			return -ENOMEM;
+		rq_aq->qidx = qidx;
+		rq_aq->sq.ena = 0;
+		rq_aq->sq_mask.ena = 1;
+		rq_aq->ctype = NIX_AQ_CTYPE_RQ;
+		rq_aq->op = NIX_AQ_INSTOP_WRITE;
+	}
+
+	aura_aq = otx2_mbox_alloc_msg_npa_aq_enq(&pfvf->mbox);
+	if (!aura_aq)
+		goto fail;
+
+	aura_aq->aura_id = aura_id;
+	aura_aq->aura.ena = 0;
+	aura_aq->aura_mask.ena = 1;
+	aura_aq->ctype = NPA_AQ_CTYPE_AURA;
+	aura_aq->op = NPA_AQ_INSTOP_WRITE;
+
+	pool_aq = otx2_mbox_alloc_msg_npa_aq_enq(&pfvf->mbox);
+	if (!pool_aq)
+		goto fail;
+
+	pool_aq->aura_id = aura_id;
+	pool_aq->pool.ena = 0;
+	pool_aq->pool_mask.ena = 1;
+
+	pool_aq->ctype = NPA_AQ_CTYPE_POOL;
+	pool_aq->op = NPA_AQ_INSTOP_WRITE;
+
+	return otx2_sync_mbox_msg(&pfvf->mbox);
+
+fail:
+	otx2_mbox_reset(&pfvf->mbox.mbox, 0);
+	return -ENOMEM;
+}
+
+static void otx2_clean_up_rq(struct otx2_nic *pfvf, int qidx)
+{
+	struct otx2_qset *qset = &pfvf->qset;
+	struct otx2_cq_queue *cq;
+	struct otx2_pool *pool;
+	u64 iova;
+
+	/* If the DOWN flag is set SQs are already freed */
+	if (pfvf->flags & OTX2_FLAG_INTF_DOWN)
+		return;
+
+	cq = &qset->cq[qidx];
+	if (cq)
+		otx2_cleanup_rx_cqes(pfvf, cq, qidx);
+
+	pool = &pfvf->qset.pool[qidx];
+	iova = otx2_aura_allocptr(pfvf, qidx);
+	while (iova) {
+		iova -= OTX2_HEAD_ROOM;
+		otx2_free_bufs(pfvf, pool, iova, pfvf->rbsize);
+		iova = otx2_aura_allocptr(pfvf, qidx);
+	}
+
+	mutex_lock(&pfvf->mbox.lock);
+	otx2_xsk_ctx_disable(pfvf, qidx, qidx);
+	mutex_unlock(&pfvf->mbox.lock);
+}
+
+int otx2_xsk_pool_enable(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qidx)
+{
+	u16 rx_queues = pf->hw.rx_queues;
+	u16 tx_queues = pf->hw.tx_queues;
+	int err;
+
+	if (qidx >= rx_queues || qidx >= tx_queues)
+		return -EINVAL;
+
+	err = xsk_pool_dma_map(pool, pf->dev, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
+	if (err)
+		return err;
+
+	set_bit(qidx, pf->af_xdp_zc_qidx);
+	otx2_clean_up_rq(pf, qidx);
+	/* Kick start the NAPI context so that receiving will start */
+	return otx2_xsk_wakeup(pf->netdev, qidx, XDP_WAKEUP_RX);
+}
+
+int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qidx)
+{
+	struct net_device *netdev = pf->netdev;
+	struct xsk_buff_pool *pool;
+
+	pool = xsk_get_pool_from_qid(netdev, qidx);
+	if (!pool)
+		return -EINVAL;
+
+	otx2_clean_up_rq(pf, qidx);
+	clear_bit(qidx, pf->af_xdp_zc_qidx);
+	xsk_pool_dma_unmap(pool, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
+
+	return 0;
+}
+
+int otx2_xsk_pool_setup(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qidx)
+{
+	if (pool)
+		return otx2_xsk_pool_enable(pf, pool, qidx);
+
+	return otx2_xsk_pool_disable(pf, qidx);
+}
+
+int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
+{
+	struct otx2_nic *pf = netdev_priv(dev);
+	struct otx2_cq_poll *cq_poll = NULL;
+	struct otx2_qset *qset = &pf->qset;
+
+	if (pf->flags & OTX2_FLAG_INTF_DOWN)
+		return -ENETDOWN;
+
+	if (queue_id >= pf->hw.rx_queues)
+		return -EINVAL;
+
+	cq_poll = &qset->napi[queue_id];
+	if (!cq_poll)
+		return -EINVAL;
+
+	/* Trigger interrupt */
+	if (!napi_if_scheduled_mark_missed(&cq_poll->napi))
+		otx2_write64(pf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), BIT_ULL(0));
+
+	return 0;
+}
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
new file mode 100644
index 000000000000..022b3433edbb
--- /dev/null
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
@@ -0,0 +1,21 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Marvell RVU PF/VF Netdev Devlink
+ *
+ * Copyright (C) 2024 Marvell.
+ *
+ */
+
+#ifndef	OTX2_XSK_H
+#define	OTX2_XSK_H
+
+struct otx2_nic;
+struct xsk_buff_pool;
+
+int otx2_xsk_pool_setup(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qid);
+int otx2_xsk_pool_enable(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qid);
+int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qid);
+int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
+			    dma_addr_t *dma, int idx);
+int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
+
+#endif /* OTX2_XSK_H */
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/qos_sq.c b/drivers/net/ethernet/marvell/octeontx2/nic/qos_sq.c
index 9d887bfc3108..c5dbae0e513b 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/qos_sq.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/qos_sq.c
@@ -82,7 +82,7 @@ static int otx2_qos_sq_aura_pool_init(struct otx2_nic *pfvf, int qidx)
 	}
 
 	for (ptr = 0; ptr < num_sqbs; ptr++) {
-		err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
+		err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
 		if (err)
 			goto sqb_free;
 		pfvf->hw_ops->aura_freeptr(pfvf, pool_id, bufptr);
-- 
2.25.1

[net-next PATCH v6 4/6] octeontx2-pf: Reconfigure RSS table after enabling AF_XDP zerocopy on rx queue

[Suman Ghosh]

RSS table needs to be reconfigured once a rx queue is enabled or
disabled for AF_XDP zerocopy support. After enabling UMEM on a rx queue,
that queue should not be part of RSS queue selection algorithm.
Similarly the queue should be considered again after UMEM is disabled.

Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c  | 4 ++++
 drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c | 6 +++++-
 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c     | 4 ++++
 3 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 92b0dba07853..6f7c11d59942 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -331,6 +331,10 @@ int otx2_set_rss_table(struct otx2_nic *pfvf, int ctx_id)
 	rss_ctx = rss->rss_ctx[ctx_id];
 	/* Get memory to put this msg */
 	for (idx = 0; idx < rss->rss_size; idx++) {
+		/* Ignore the queue if AF_XDP zero copy is enabled */
+		if (test_bit(rss_ctx->ind_tbl[idx], pfvf->af_xdp_zc_qidx))
+			continue;
+
 		aq = otx2_mbox_alloc_msg_nix_aq_enq(mbox);
 		if (!aq) {
 			/* The shared memory buffer can be full.
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c
index 2d53dc77ef1e..010385b29988 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c
@@ -910,8 +910,12 @@ static int otx2_get_rxfh(struct net_device *dev,
 		return -ENOENT;
 
 	if (indir) {
-		for (idx = 0; idx < rss->rss_size; idx++)
+		for (idx = 0; idx < rss->rss_size; idx++) {
+			/* Ignore if the rx queue is AF_XDP zero copy enabled */
+			if (test_bit(rss_ctx->ind_tbl[idx], pfvf->af_xdp_zc_qidx))
+				continue;
 			indir[idx] = rss_ctx->ind_tbl[idx];
+		}
 	}
 	if (rxfh->key)
 		memcpy(rxfh->key, rss->key, sizeof(rss->key));
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
index 894c1e0aea6f..3bdee1603fac 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
@@ -130,6 +130,8 @@ int otx2_xsk_pool_enable(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qi
 
 	set_bit(qidx, pf->af_xdp_zc_qidx);
 	otx2_clean_up_rq(pf, qidx);
+	/* Reconfigure RSS table as 'qidx' cannot be part of RSS now */
+	otx2_set_rss_table(pf, DEFAULT_RSS_CONTEXT_GROUP);
 	/* Kick start the NAPI context so that receiving will start */
 	return otx2_xsk_wakeup(pf->netdev, qidx, XDP_WAKEUP_RX);
 }
@@ -146,6 +148,8 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qidx)
 	otx2_clean_up_rq(pf, qidx);
 	clear_bit(qidx, pf->af_xdp_zc_qidx);
 	xsk_pool_dma_unmap(pool, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
+	/* Reconfigure RSS table as 'qidx' now need to be part of RSS now */
+	otx2_set_rss_table(pf, DEFAULT_RSS_CONTEXT_GROUP);
 
 	return 0;
 }
-- 
2.25.1

[net-next PATCH v6 5/6] octeontx2-pf: Prepare for AF_XDP

[Suman Ghosh]

Implement necessary APIs required for AF_XDP transmit.

Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 .../marvell/octeontx2/nic/otx2_common.h       |  1 +
 .../marvell/octeontx2/nic/otx2_txrx.c         | 25 +++++++++++++++++--
 2 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index 60508971b62f..19e9e2e72233 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -1181,4 +1181,5 @@ static inline int mcam_entry_cmp(const void *a, const void *b)
 dma_addr_t otx2_dma_map_skb_frag(struct otx2_nic *pfvf,
 				 struct sk_buff *skb, int seg, int *len);
 void otx2_dma_unmap_skb_frags(struct otx2_nic *pfvf, struct sg_list *sg);
+int otx2_read_free_sqe(struct otx2_nic *pfvf, u16 qidx);
 #endif /* OTX2_COMMON_H */
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
index 00b6903ba250..9a6c1f1a3ee0 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
@@ -30,6 +30,12 @@
 
 DEFINE_STATIC_KEY_FALSE(cn10k_ipsec_sa_enabled);
 
+static int otx2_get_free_sqe(struct otx2_snd_queue *sq)
+{
+	return (sq->cons_head - sq->head - 1 + sq->sqe_cnt)
+		& (sq->sqe_cnt - 1);
+}
+
 static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 				     struct bpf_prog *prog,
 				     struct nix_cqe_rx_s *cqe,
@@ -1157,7 +1163,7 @@ bool otx2_sq_append_skb(void *dev, struct netdev_queue *txq,
 	/* Check if there is enough room between producer
 	 * and consumer index.
 	 */
-	free_desc = (sq->cons_head - sq->head - 1 + sq->sqe_cnt) & (sq->sqe_cnt - 1);
+	free_desc = otx2_get_free_sqe(sq);
 	if (free_desc < sq->sqe_thresh)
 		return false;
 
@@ -1396,6 +1402,21 @@ static void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq,
 	sq->sg[sq->head].skb = (u64)xdpf;
 }
 
+int otx2_read_free_sqe(struct otx2_nic *pfvf, u16 qidx)
+{
+	struct otx2_snd_queue *sq;
+	int free_sqe;
+
+	sq = &pfvf->qset.sq[qidx];
+	free_sqe = otx2_get_free_sqe(sq);
+	if (free_sqe < sq->sqe_thresh) {
+		netdev_warn(pfvf->netdev, "No free sqe for Send queue%d\n", qidx);
+		return 0;
+	}
+
+	return free_sqe - sq->sqe_thresh;
+}
+
 bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, struct xdp_frame *xdpf,
 			    u64 iova, int len, u16 qidx, u16 flags)
 {
@@ -1404,7 +1425,7 @@ bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, struct xdp_frame *xdpf,
 	int offset, free_sqe;
 
 	sq = &pfvf->qset.sq[qidx];
-	free_sqe = (sq->num_sqbs - *sq->aura_fc_addr) * sq->sqe_per_sqb;
+	free_sqe = otx2_get_free_sqe(sq);
 	if (free_sqe < sq->sqe_thresh)
 		return false;
 
-- 
2.25.1

[net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Suman Ghosh]

This patch implements below changes,

1. To avoid concurrency with normal traffic uses
   XDP queues.
2. Since there are chances that XDP and AF_XDP can
   fall under same queue uses separate flags to handle
   dma buffers.

Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
Signed-off-by: Suman Ghosh <sumang@marvell.com>
---
 .../marvell/octeontx2/nic/otx2_common.c       |  4 ++
 .../marvell/octeontx2/nic/otx2_common.h       |  6 +++
 .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  2 +-
 .../marvell/octeontx2/nic/otx2_txrx.c         | 49 +++++++++++++++----
 .../marvell/octeontx2/nic/otx2_txrx.h         |  2 +
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 43 +++++++++++++++-
 .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  3 ++
 7 files changed, 96 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 6f7c11d59942..84cd029a85aa 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -1037,6 +1037,10 @@ int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
 
 	sq->stats.bytes = 0;
 	sq->stats.pkts = 0;
+	/* Attach XSK_BUFF_POOL to XDP queue */
+	if (qidx > pfvf->hw.xdp_queues)
+		otx2_attach_xsk_buff(pfvf, sq, (qidx - pfvf->hw.xdp_queues));
+
 
 	chan_offset = qidx % pfvf->hw.tx_chan_cnt;
 	err = pfvf->hw_ops->sq_aq_init(pfvf, qidx, chan_offset, sqb_aura);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index 19e9e2e72233..1e88422825be 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -129,6 +129,12 @@ enum otx2_errcodes_re {
 	ERRCODE_IL4_CSUM = 0x22,
 };
 
+enum otx2_xdp_action {
+	OTX2_XDP_TX	  = BIT(0),
+	OTX2_XDP_REDIRECT = BIT(1),
+	OTX2_AF_XDP_FRAME = BIT(2),
+};
+
 struct otx2_dev_stats {
 	u64 rx_bytes;
 	u64 rx_frames;
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
index 50a42cd5d50a..c7c562f0f5e5 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
@@ -2693,7 +2693,7 @@ static int otx2_xdp_xmit_tx(struct otx2_nic *pf, struct xdp_frame *xdpf,
 		return -ENOMEM;
 
 	err = otx2_xdp_sq_append_pkt(pf, xdpf, dma_addr, xdpf->len,
-				     qidx, XDP_REDIRECT);
+				     qidx, OTX2_XDP_REDIRECT);
 	if (!err) {
 		otx2_dma_unmap_page(pf, dma_addr, xdpf->len, DMA_TO_DEVICE);
 		xdp_return_frame(xdpf);
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
index 9a6c1f1a3ee0..6bc5ce5a9f61 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
@@ -20,6 +20,7 @@
 #include "otx2_txrx.h"
 #include "otx2_ptp.h"
 #include "cn10k.h"
+#include "otx2_xsk.h"
 
 #define CQE_ADDR(CQ, idx) ((CQ)->cqe_base + ((CQ)->cqe_size * (idx)))
 #define PTP_PORT	        0x13F
@@ -103,13 +104,19 @@ static unsigned int frag_num(unsigned int i)
 
 static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf,
 				     struct otx2_snd_queue *sq,
-				     struct nix_cqe_tx_s *cqe)
+				     struct nix_cqe_tx_s *cqe,
+				     int *xsk_frames)
 {
 	struct nix_send_comp_s *snd_comp = &cqe->comp;
 	struct sg_list *sg;
 
 	sg = &sq->sg[snd_comp->sqe_id];
-	if (sg->flags & XDP_REDIRECT)
+	if (sg->flags & OTX2_AF_XDP_FRAME) {
+		(*xsk_frames)++;
+		return;
+	}
+
+	if (sg->flags & OTX2_XDP_REDIRECT)
 		otx2_dma_unmap_page(pfvf, sg->dma_addr[0], sg->size[0], DMA_TO_DEVICE);
 	xdp_return_frame((struct xdp_frame *)sg->skb);
 	sg->skb = (u64)NULL;
@@ -434,6 +441,18 @@ int otx2_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
 	return cnt - cq->pool_ptrs;
 }
 
+static void otx2_zc_submit_pkts(struct otx2_nic *pfvf, struct xsk_buff_pool *xsk_pool,
+				int *xsk_frames, int qidx, int budget)
+{
+	if (*xsk_frames)
+		xsk_tx_completed(xsk_pool, *xsk_frames);
+
+	if (xsk_uses_need_wakeup(xsk_pool))
+		xsk_set_tx_need_wakeup(xsk_pool);
+
+	otx2_zc_napi_handler(pfvf, xsk_pool, qidx, budget);
+}
+
 static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 				struct otx2_cq_queue *cq, int budget)
 {
@@ -442,16 +461,22 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 	struct nix_cqe_tx_s *cqe;
 	struct net_device *ndev;
 	int processed_cqe = 0;
+	int xsk_frames = 0;
+
+	qidx = cq->cq_idx - pfvf->hw.rx_queues;
+	sq = &pfvf->qset.sq[qidx];
 
 	if (cq->pend_cqe >= budget)
 		goto process_cqe;
 
-	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
+	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) {
+		if (sq->xsk_pool)
+			otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames,
+					    qidx, budget);
 		return 0;
+	}
 
 process_cqe:
-	qidx = cq->cq_idx - pfvf->hw.rx_queues;
-	sq = &pfvf->qset.sq[qidx];
 
 	while (likely(processed_cqe < budget) && cq->pend_cqe) {
 		cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq);
@@ -461,10 +486,8 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 			break;
 		}
 
-		qidx = cq->cq_idx - pfvf->hw.rx_queues;
-
 		if (cq->cq_type == CQ_XDP)
-			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe);
+			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe, &xsk_frames);
 		else
 			otx2_snd_pkt_handler(pfvf, cq, &pfvf->qset.sq[qidx],
 					     cqe, budget, &tx_pkts, &tx_bytes);
@@ -505,6 +528,10 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
 		    netif_carrier_ok(ndev))
 			netif_tx_wake_queue(txq);
 	}
+
+	if (sq->xsk_pool)
+		otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames, qidx, budget);
+
 	return 0;
 }
 
@@ -1499,8 +1526,10 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
 		qidx += pfvf->hw.tx_queues;
 		cq->pool_ptrs++;
 		xdpf = xdp_convert_buff_to_frame(&xdp);
-		return otx2_xdp_sq_append_pkt(pfvf, xdpf, cqe->sg.seg_addr,
-					      cqe->sg.seg_size, qidx, XDP_TX);
+		return otx2_xdp_sq_append_pkt(pfvf, xdpf,
+					      cqe->sg.seg_addr,
+					      cqe->sg.seg_size,
+					      qidx, OTX2_XDP_TX);
 	case XDP_REDIRECT:
 		cq->pool_ptrs++;
 		if (xsk_buff) {
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
index 8f346fbc8221..acf259d72008 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h
@@ -106,6 +106,8 @@ struct otx2_snd_queue {
 	/* SQE ring and CPT response queue for Inline IPSEC */
 	struct qmem		*sqe_ring;
 	struct qmem		*cpt_resp;
+	/* Buffer pool for af_xdp zero-copy */
+	struct xsk_buff_pool    *xsk_pool;
 } ____cacheline_aligned_in_smp;
 
 enum cq_type {
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
index 3bdee1603fac..ce10caea8511 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
@@ -140,11 +140,14 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qidx)
 {
 	struct net_device *netdev = pf->netdev;
 	struct xsk_buff_pool *pool;
+	struct otx2_snd_queue *sq;
 
 	pool = xsk_get_pool_from_qid(netdev, qidx);
 	if (!pool)
 		return -EINVAL;
 
+	sq = &pf->qset.sq[qidx + pf->hw.tx_queues];
+	sq->xsk_pool = NULL;
 	otx2_clean_up_rq(pf, qidx);
 	clear_bit(qidx, pf->af_xdp_zc_qidx);
 	xsk_pool_dma_unmap(pool, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
@@ -171,7 +174,7 @@ int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 	if (pf->flags & OTX2_FLAG_INTF_DOWN)
 		return -ENETDOWN;
 
-	if (queue_id >= pf->hw.rx_queues)
+	if (queue_id >= pf->hw.rx_queues || queue_id >= pf->hw.tx_queues)
 		return -EINVAL;
 
 	cq_poll = &qset->napi[queue_id];
@@ -179,8 +182,44 @@ int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 		return -EINVAL;
 
 	/* Trigger interrupt */
-	if (!napi_if_scheduled_mark_missed(&cq_poll->napi))
+	if (!napi_if_scheduled_mark_missed(&cq_poll->napi)) {
 		otx2_write64(pf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), BIT_ULL(0));
+		otx2_write64(pf, NIX_LF_CINTX_INT_W1S(cq_poll->cint_idx), BIT_ULL(0));
+	}
 
 	return 0;
 }
+
+void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx)
+{
+	if (test_bit(qidx, pfvf->af_xdp_zc_qidx))
+		sq->xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, qidx);
+}
+
+void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
+			  int queue, int budget)
+{
+	struct xdp_desc *xdp_desc = pool->tx_descs;
+	int err, i, work_done = 0, batch;
+
+	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
+	batch = xsk_tx_peek_release_desc_batch(pool, budget);
+	if (!batch)
+		return;
+
+	for (i = 0; i < batch; i++) {
+		dma_addr_t dma_addr;
+
+		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
+		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr, xdp_desc[i].len,
+					     queue, OTX2_AF_XDP_FRAME);
+		if (!err) {
+			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer packet err%d\n", err);
+			break;
+		}
+		work_done++;
+	}
+
+	if (work_done)
+		xsk_tx_release(pool);
+}
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
index 022b3433edbb..8047fafee8fe 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
@@ -17,5 +17,8 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qid);
 int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
 			    dma_addr_t *dma, int idx);
 int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
+void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
+			  int queue, int budget);
+void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx);
 
 #endif /* OTX2_XSK_H */
-- 
2.25.1

Re: [net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Paolo Abeni]

On 2/13/25 6:31 AM, Suman Ghosh wrote:
> +void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
> +			  int queue, int budget)
> +{
> +	struct xdp_desc *xdp_desc = pool->tx_descs;
> +	int err, i, work_done = 0, batch;
> +
> +	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
> +	batch = xsk_tx_peek_release_desc_batch(pool, budget);
> +	if (!batch)
> +		return;
> +
> +	for (i = 0; i < batch; i++) {
> +		dma_addr_t dma_addr;
> +
> +		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
> +		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr, xdp_desc[i].len,
> +					     queue, OTX2_AF_XDP_FRAME);
> +		if (!err) {
> +			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer packet err%d\n", err);

Here `err` is always 0, dumping it's value is quite confusing.

The root cause is that otx2_xdp_sq_append_pkt() returns a success
boolean value, the variable holding it should possibly be renamed
accordingly.

Since this is the only nit I could find, I think we are better without a
repost, but please follow-up on this chunk soon.

Thanks,

Paolo

Re: [net-next PATCH v6 0/6] Add af_xdp support for cn10k

[patchwork-bot+netdevbpf]

Hello:

This series was applied to netdev/net-next.git (main)
by Paolo Abeni <pabeni@redhat.com>:

On Thu, 13 Feb 2025 11:01:35 +0530 you wrote:
> This patchset includes changes to support AF_XDP for cn10k chipsets. Both
> non-zero copy and zero copy will be supported after these changes. Also,
> the RSS will be reconfigured once a particular receive queue is
> added/removed to/from AF_XDP support.
> 
> Patch #1: octeontx2-pf: use xdp_return_frame() to free xdp buffers
> 
> [...]

Here is the summary with links:
  - [net-next,v6,1/6] octeontx2-pf: use xdp_return_frame() to free xdp buffers
    https://git.kernel.org/netdev/net-next/c/94c80f748873
  - [net-next,v6,2/6] octeontx2-pf: Add AF_XDP non-zero copy support
    https://git.kernel.org/netdev/net-next/c/b4164de5041b
  - [net-next,v6,3/6] octeontx2-pf: AF_XDP zero copy receive support
    https://git.kernel.org/netdev/net-next/c/efabce290151
  - [net-next,v6,4/6] octeontx2-pf: Reconfigure RSS table after enabling AF_XDP zerocopy on rx queue
    https://git.kernel.org/netdev/net-next/c/25b07c1a8694
  - [net-next,v6,5/6] octeontx2-pf: Prepare for AF_XDP
    https://git.kernel.org/netdev/net-next/c/c5c2398eb88b
  - [net-next,v6,6/6] octeontx2-pf: AF_XDP zero copy transmit support
    https://git.kernel.org/netdev/net-next/c/53616af09b5a

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html

Re: [net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Maciej Fijalkowski]

On Thu, Feb 13, 2025 at 11:01:41AM +0530, Suman Ghosh wrote:
> This patch implements below changes,
> 
> 1. To avoid concurrency with normal traffic uses
>    XDP queues.
> 2. Since there are chances that XDP and AF_XDP can
>    fall under same queue uses separate flags to handle
>    dma buffers.
> 
> Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
> Signed-off-by: Suman Ghosh <sumang@marvell.com>
> ---

[...]

> +void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
> +			  int queue, int budget)
> +{
> +	struct xdp_desc *xdp_desc = pool->tx_descs;
> +	int err, i, work_done = 0, batch;
> +
> +	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
> +	batch = xsk_tx_peek_release_desc_batch(pool, budget);
> +	if (!batch)
> +		return;
> +
> +	for (i = 0; i < batch; i++) {
> +		dma_addr_t dma_addr;
> +
> +		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
> +		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr, xdp_desc[i].len,
> +					     queue, OTX2_AF_XDP_FRAME);
> +		if (!err) {
> +			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer packet err%d\n", err);
> +			break;
> +		}
> +		work_done++;
> +	}
> +
> +	if (work_done)
> +		xsk_tx_release(pool);

this is broken actually. the batch api you're using above is doing tx
release internally for you.

Sorry for not catching this earlier but i was never CCed in this series.

> +}
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
> index 022b3433edbb..8047fafee8fe 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
> @@ -17,5 +17,8 @@ int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qid);
>  int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
>  			    dma_addr_t *dma, int idx);
>  int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
> +void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool *pool,
> +			  int queue, int budget);
> +void otx2_attach_xsk_buff(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, int qidx);
>  
>  #endif /* OTX2_XSK_H */
> -- 
> 2.25.1
> 
> 

Re: [net-next PATCH v6 3/6] octeontx2-pf: AF_XDP zero copy receive support

[Maciej Fijalkowski]

On Thu, Feb 13, 2025 at 11:01:38AM +0530, Suman Ghosh wrote:
> This patch adds support to AF_XDP zero copy for CN10K.
> This patch specifically adds receive side support. In this approach once
> a xdp program with zero copy support on a specific rx queue is enabled,
> then that receive quse is disabled/detached from the existing kernel
> queue and re-assigned to the umem memory.
> 
> Signed-off-by: Suman Ghosh <sumang@marvell.com>
> ---
>  .../ethernet/marvell/octeontx2/nic/Makefile   |   2 +-
>  .../ethernet/marvell/octeontx2/nic/cn10k.c    |   7 +-
>  .../marvell/octeontx2/nic/otx2_common.c       | 114 ++++++++---
>  .../marvell/octeontx2/nic/otx2_common.h       |   6 +-
>  .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  25 ++-
>  .../marvell/octeontx2/nic/otx2_txrx.c         |  73 +++++--
>  .../marvell/octeontx2/nic/otx2_txrx.h         |   6 +
>  .../ethernet/marvell/octeontx2/nic/otx2_vf.c  |  12 +-
>  .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 182 ++++++++++++++++++
>  .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  21 ++
>  .../ethernet/marvell/octeontx2/nic/qos_sq.c   |   2 +-
>  11 files changed, 389 insertions(+), 61 deletions(-)
>  create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
>  create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
> 
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/Makefile b/drivers/net/ethernet/marvell/octeontx2/nic/Makefile
> index cb6513ab35e7..69e0778f9ac1 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/Makefile
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/Makefile
> @@ -9,7 +9,7 @@ obj-$(CONFIG_RVU_ESWITCH) += rvu_rep.o
>  
>  rvu_nicpf-y := otx2_pf.o otx2_common.o otx2_txrx.o otx2_ethtool.o \
>                 otx2_flows.o otx2_tc.o cn10k.o otx2_dmac_flt.o \
> -               otx2_devlink.o qos_sq.o qos.o
> +               otx2_devlink.o qos_sq.o qos.o otx2_xsk.o
>  rvu_nicvf-y := otx2_vf.o
>  rvu_rep-y := rep.o
>  
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c b/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
> index a15cc86635d6..c3b6e0f60a79 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
> @@ -112,9 +112,12 @@ int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
>  	struct otx2_nic *pfvf = dev;
>  	int cnt = cq->pool_ptrs;
>  	u64 ptrs[NPA_MAX_BURST];
> +	struct otx2_pool *pool;
>  	dma_addr_t bufptr;
>  	int num_ptrs = 1;
>  
> +	pool = &pfvf->qset.pool[cq->cq_idx];
> +
>  	/* Refill pool with new buffers */
>  	while (cq->pool_ptrs) {
>  		if (otx2_alloc_buffer(pfvf, cq, &bufptr)) {
> @@ -124,7 +127,9 @@ int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
>  			break;
>  		}
>  		cq->pool_ptrs--;
> -		ptrs[num_ptrs] = (u64)bufptr + OTX2_HEAD_ROOM;
> +		ptrs[num_ptrs] = pool->xsk_pool ?
> +				 (u64)bufptr : (u64)bufptr + OTX2_HEAD_ROOM;
> +
>  		num_ptrs++;
>  		if (num_ptrs == NPA_MAX_BURST || cq->pool_ptrs == 0) {
>  			__cn10k_aura_freeptr(pfvf, cq->cq_idx, ptrs,
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
> index 161cf33ef89e..92b0dba07853 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
> @@ -17,6 +17,7 @@
>  #include "otx2_common.h"
>  #include "otx2_struct.h"
>  #include "cn10k.h"
> +#include "otx2_xsk.h"
>  
>  static bool otx2_is_pfc_enabled(struct otx2_nic *pfvf)
>  {
> @@ -549,10 +550,13 @@ static int otx2_alloc_pool_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
>  }
>  
>  static int __otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
> -			     dma_addr_t *dma)
> +			     dma_addr_t *dma, int qidx, int idx)
>  {
>  	u8 *buf;
>  
> +	if (pool->xsk_pool)
> +		return otx2_xsk_pool_alloc_buf(pfvf, pool, dma, idx);
> +
>  	if (pool->page_pool)
>  		return otx2_alloc_pool_buf(pfvf, pool, dma);
>  
> @@ -571,12 +575,12 @@ static int __otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
>  }
>  
>  int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
> -		    dma_addr_t *dma)
> +		    dma_addr_t *dma, int qidx, int idx)
>  {
>  	int ret;
>  
>  	local_bh_disable();
> -	ret = __otx2_alloc_rbuf(pfvf, pool, dma);
> +	ret = __otx2_alloc_rbuf(pfvf, pool, dma, qidx, idx);
>  	local_bh_enable();
>  	return ret;
>  }
> @@ -584,7 +588,8 @@ int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
>  int otx2_alloc_buffer(struct otx2_nic *pfvf, struct otx2_cq_queue *cq,
>  		      dma_addr_t *dma)
>  {
> -	if (unlikely(__otx2_alloc_rbuf(pfvf, cq->rbpool, dma)))
> +	if (unlikely(__otx2_alloc_rbuf(pfvf, cq->rbpool, dma,
> +				       cq->cq_idx, cq->pool_ptrs - 1)))
>  		return -ENOMEM;
>  	return 0;
>  }
> @@ -884,7 +889,7 @@ void otx2_sqb_flush(struct otx2_nic *pfvf)
>  #define RQ_PASS_LVL_AURA (255 - ((95 * 256) / 100)) /* RED when 95% is full */
>  #define RQ_DROP_LVL_AURA (255 - ((99 * 256) / 100)) /* Drop when 99% is full */
>  
> -static int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura)
> +int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura)
>  {
>  	struct otx2_qset *qset = &pfvf->qset;
>  	struct nix_aq_enq_req *aq;
> @@ -1041,7 +1046,7 @@ int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
>  
>  }
>  
> -static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
> +int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
>  {
>  	struct otx2_qset *qset = &pfvf->qset;
>  	int err, pool_id, non_xdp_queues;
> @@ -1057,11 +1062,18 @@ static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
>  		cq->cint_idx = qidx;
>  		cq->cqe_cnt = qset->rqe_cnt;
>  		if (pfvf->xdp_prog) {
> -			pool = &qset->pool[qidx];
>  			xdp_rxq_info_reg(&cq->xdp_rxq, pfvf->netdev, qidx, 0);
> -			xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
> -						   MEM_TYPE_PAGE_POOL,
> -						   pool->page_pool);
> +			pool = &qset->pool[qidx];
> +			if (pool->xsk_pool) {
> +				xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
> +							   MEM_TYPE_XSK_BUFF_POOL,
> +							   NULL);
> +				xsk_pool_set_rxq_info(pool->xsk_pool, &cq->xdp_rxq);
> +			} else if (pool->page_pool) {
> +				xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
> +							   MEM_TYPE_PAGE_POOL,
> +							   pool->page_pool);
> +			}
>  		}
>  	} else if (qidx < non_xdp_queues) {
>  		cq->cq_type = CQ_TX;
> @@ -1281,9 +1293,10 @@ void otx2_free_bufs(struct otx2_nic *pfvf, struct otx2_pool *pool,
>  
>  	pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
>  	page = virt_to_head_page(phys_to_virt(pa));
> -
>  	if (pool->page_pool) {
>  		page_pool_put_full_page(pool->page_pool, page, true);
> +	} else if (pool->xsk_pool) {
> +		/* Note: No way of identifying xdp_buff */
>  	} else {
>  		dma_unmap_page_attrs(pfvf->dev, iova, size,
>  				     DMA_FROM_DEVICE,
> @@ -1298,6 +1311,7 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
>  	int pool_id, pool_start = 0, pool_end = 0, size = 0;
>  	struct otx2_pool *pool;
>  	u64 iova;
> +	int idx;
>  
>  	if (type == AURA_NIX_SQ) {
>  		pool_start = otx2_get_pool_idx(pfvf, type, 0);
> @@ -1312,16 +1326,21 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
>  
>  	/* Free SQB and RQB pointers from the aura pool */
>  	for (pool_id = pool_start; pool_id < pool_end; pool_id++) {
> -		iova = otx2_aura_allocptr(pfvf, pool_id);
>  		pool = &pfvf->qset.pool[pool_id];
> +		iova = otx2_aura_allocptr(pfvf, pool_id);
>  		while (iova) {
>  			if (type == AURA_NIX_RQ)
>  				iova -= OTX2_HEAD_ROOM;
> -
>  			otx2_free_bufs(pfvf, pool, iova, size);
> -
>  			iova = otx2_aura_allocptr(pfvf, pool_id);
>  		}
> +
> +		for (idx = 0 ; idx < pool->xdp_cnt; idx++) {
> +			if (!pool->xdp[idx])
> +				continue;
> +
> +			xsk_buff_free(pool->xdp[idx]);
> +		}
>  	}
>  }
>  
> @@ -1338,7 +1357,8 @@ void otx2_aura_pool_free(struct otx2_nic *pfvf)
>  		qmem_free(pfvf->dev, pool->stack);
>  		qmem_free(pfvf->dev, pool->fc_addr);
>  		page_pool_destroy(pool->page_pool);
> -		pool->page_pool = NULL;
> +		devm_kfree(pfvf->dev, pool->xdp);
> +		pool->xsk_pool = NULL;
>  	}
>  	devm_kfree(pfvf->dev, pfvf->qset.pool);
>  	pfvf->qset.pool = NULL;
> @@ -1425,6 +1445,7 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
>  		   int stack_pages, int numptrs, int buf_size, int type)
>  {
>  	struct page_pool_params pp_params = { 0 };
> +	struct xsk_buff_pool *xsk_pool;
>  	struct npa_aq_enq_req *aq;
>  	struct otx2_pool *pool;
>  	int err;
> @@ -1468,21 +1489,35 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
>  	aq->ctype = NPA_AQ_CTYPE_POOL;
>  	aq->op = NPA_AQ_INSTOP_INIT;
>  
> -	if (type != AURA_NIX_RQ) {
> -		pool->page_pool = NULL;
> +	if (type != AURA_NIX_RQ)
> +		return 0;
> +
> +	if (!test_bit(pool_id, pfvf->af_xdp_zc_qidx)) {
> +		pp_params.order = get_order(buf_size);
> +		pp_params.flags = PP_FLAG_DMA_MAP;
> +		pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
> +		pp_params.nid = NUMA_NO_NODE;
> +		pp_params.dev = pfvf->dev;
> +		pp_params.dma_dir = DMA_FROM_DEVICE;
> +		pool->page_pool = page_pool_create(&pp_params);
> +		if (IS_ERR(pool->page_pool)) {
> +			netdev_err(pfvf->netdev, "Creation of page pool failed\n");
> +			return PTR_ERR(pool->page_pool);
> +		}
>  		return 0;
>  	}
>  
> -	pp_params.order = get_order(buf_size);
> -	pp_params.flags = PP_FLAG_DMA_MAP;
> -	pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
> -	pp_params.nid = NUMA_NO_NODE;
> -	pp_params.dev = pfvf->dev;
> -	pp_params.dma_dir = DMA_FROM_DEVICE;
> -	pool->page_pool = page_pool_create(&pp_params);
> -	if (IS_ERR(pool->page_pool)) {
> -		netdev_err(pfvf->netdev, "Creation of page pool failed\n");
> -		return PTR_ERR(pool->page_pool);
> +	/* Set XSK pool to support AF_XDP zero-copy */
> +	xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, pool_id);
> +	if (xsk_pool) {
> +		pool->xsk_pool = xsk_pool;
> +		pool->xdp_cnt = numptrs;
> +		pool->xdp = devm_kcalloc(pfvf->dev,
> +					 numptrs, sizeof(struct xdp_buff *), GFP_KERNEL);

What is the rationale behind having a buffer pool within your driver while
you have this very same thing within xsk_buff_pool?

You're doubling your work. Just use the xsk_buff_alloc_batch() and have a
simpler ZC implementation in your driver.

> +		if (IS_ERR(pool->xdp)) {
> +			netdev_err(pfvf->netdev, "Creation of xsk pool failed\n");
> +			return PTR_ERR(pool->xdp);
> +		}
>  	}
>  
>  	return 0;
> @@ -1543,9 +1578,18 @@ int otx2_sq_aura_pool_init(struct otx2_nic *pfvf)
>  		}
>  
>  		for (ptr = 0; ptr < num_sqbs; ptr++) {
> -			err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
> -			if (err)
> +			err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
> +			if (err) {
> +				if (pool->xsk_pool) {
> +					ptr--;
> +					while (ptr >= 0) {
> +						xsk_buff_free(pool->xdp[ptr]);
> +						ptr--;
> +					}
> +				}
>  				goto err_mem;
> +			}
> +
>  			pfvf->hw_ops->aura_freeptr(pfvf, pool_id, bufptr);
>  			sq->sqb_ptrs[sq->sqb_count++] = (u64)bufptr;
>  		}
> @@ -1595,11 +1639,19 @@ int otx2_rq_aura_pool_init(struct otx2_nic *pfvf)
>  	/* Allocate pointers and free them to aura/pool */
>  	for (pool_id = 0; pool_id < hw->rqpool_cnt; pool_id++) {
>  		pool = &pfvf->qset.pool[pool_id];
> +
>  		for (ptr = 0; ptr < num_ptrs; ptr++) {
> -			err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
> -			if (err)
> +			err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
> +			if (err) {
> +				if (pool->xsk_pool) {
> +					while (ptr)
> +						xsk_buff_free(pool->xdp[--ptr]);
> +				}
>  				return -ENOMEM;
> +			}
> +
>  			pfvf->hw_ops->aura_freeptr(pfvf, pool_id,
> +						   pool->xsk_pool ? bufptr :
>  						   bufptr + OTX2_HEAD_ROOM);
>  		}
>  	}
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
> index d5fbccb289df..60508971b62f 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
> @@ -532,6 +532,8 @@ struct otx2_nic {
>  
>  	/* Inline ipsec */
>  	struct cn10k_ipsec	ipsec;
> +	/* af_xdp zero-copy */
> +	unsigned long		*af_xdp_zc_qidx;
>  };
>  
>  static inline bool is_otx2_lbkvf(struct pci_dev *pdev)
> @@ -1003,7 +1005,7 @@ void otx2_txschq_free_one(struct otx2_nic *pfvf, u16 lvl, u16 schq);
>  void otx2_free_pending_sqe(struct otx2_nic *pfvf);
>  void otx2_sqb_flush(struct otx2_nic *pfvf);
>  int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
> -		    dma_addr_t *dma);
> +		    dma_addr_t *dma, int qidx, int idx);
>  int otx2_rxtx_enable(struct otx2_nic *pfvf, bool enable);
>  void otx2_ctx_disable(struct mbox *mbox, int type, bool npa);
>  int otx2_nix_config_bp(struct otx2_nic *pfvf, bool enable);
> @@ -1033,6 +1035,8 @@ void otx2_pfaf_mbox_destroy(struct otx2_nic *pf);
>  void otx2_disable_mbox_intr(struct otx2_nic *pf);
>  void otx2_disable_napi(struct otx2_nic *pf);
>  irqreturn_t otx2_cq_intr_handler(int irq, void *cq_irq);
> +int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura);
> +int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx);
>  
>  /* RSS configuration APIs*/
>  int otx2_rss_init(struct otx2_nic *pfvf);
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
> index 4347a3c95350..50a42cd5d50a 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
> @@ -27,6 +27,7 @@
>  #include "qos.h"
>  #include <rvu_trace.h>
>  #include "cn10k_ipsec.h"
> +#include "otx2_xsk.h"
>  
>  #define DRV_NAME	"rvu_nicpf"
>  #define DRV_STRING	"Marvell RVU NIC Physical Function Driver"
> @@ -1662,9 +1663,7 @@ void otx2_free_hw_resources(struct otx2_nic *pf)
>  	struct nix_lf_free_req *free_req;
>  	struct mbox *mbox = &pf->mbox;
>  	struct otx2_cq_queue *cq;
> -	struct otx2_pool *pool;
>  	struct msg_req *req;
> -	int pool_id;
>  	int qidx;
>  
>  	/* Ensure all SQE are processed */
> @@ -1705,13 +1704,6 @@ void otx2_free_hw_resources(struct otx2_nic *pf)
>  	/* Free RQ buffer pointers*/
>  	otx2_free_aura_ptr(pf, AURA_NIX_RQ);
>  
> -	for (qidx = 0; qidx < pf->hw.rx_queues; qidx++) {
> -		pool_id = otx2_get_pool_idx(pf, AURA_NIX_RQ, qidx);
> -		pool = &pf->qset.pool[pool_id];
> -		page_pool_destroy(pool->page_pool);
> -		pool->page_pool = NULL;
> -	}
> -
>  	otx2_free_cq_res(pf);
>  
>  	/* Free all ingress bandwidth profiles allocated */
> @@ -2788,6 +2780,8 @@ static int otx2_xdp(struct net_device *netdev, struct netdev_bpf *xdp)
>  	switch (xdp->command) {
>  	case XDP_SETUP_PROG:
>  		return otx2_xdp_setup(pf, xdp->prog);
> +	case XDP_SETUP_XSK_POOL:
> +		return otx2_xsk_pool_setup(pf, xdp->xsk.pool, xdp->xsk.queue_id);
>  	default:
>  		return -EINVAL;
>  	}
> @@ -2865,6 +2859,7 @@ static const struct net_device_ops otx2_netdev_ops = {
>  	.ndo_set_vf_vlan	= otx2_set_vf_vlan,
>  	.ndo_get_vf_config	= otx2_get_vf_config,
>  	.ndo_bpf		= otx2_xdp,
> +	.ndo_xsk_wakeup		= otx2_xsk_wakeup,
>  	.ndo_xdp_xmit           = otx2_xdp_xmit,
>  	.ndo_setup_tc		= otx2_setup_tc,
>  	.ndo_set_vf_trust	= otx2_ndo_set_vf_trust,
> @@ -3203,16 +3198,26 @@ static int otx2_probe(struct pci_dev *pdev, const struct pci_device_id *id)
>  	/* Enable link notifications */
>  	otx2_cgx_config_linkevents(pf, true);
>  
> +	pf->af_xdp_zc_qidx = bitmap_zalloc(qcount, GFP_KERNEL);

if this is taken from ice drivers then be aware we got rid of bitmap
tracking zc enabled queues. see adbf5a42341f ("ice: remove af_xdp_zc_qps
bitmap").

in case you would still have a need for that after going through
referenced commit, please provide us some justification why.

> +	if (!pf->af_xdp_zc_qidx) {
> +		err = -ENOMEM;
> +		goto err_sriov_cleannup;
> +	}
> +
>  #ifdef CONFIG_DCB
>  	err = otx2_dcbnl_set_ops(netdev);
>  	if (err)
> -		goto err_pf_sriov_init;
> +		goto err_free_zc_bmap;
>  #endif
>  
>  	otx2_qos_init(pf, qos_txqs);
>  
>  	return 0;
>  
> +err_free_zc_bmap:
> +	bitmap_free(pf->af_xdp_zc_qidx);
> +err_sriov_cleannup:
> +	otx2_sriov_vfcfg_cleanup(pf);
>  err_pf_sriov_init:
>  	otx2_shutdown_tc(pf);
>  err_mcam_flow_del:

[...]

RE: [EXTERNAL] Re: [net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Suman Ghosh]

>> +void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool
>*pool,
>> +			  int queue, int budget)
>> +{
>> +	struct xdp_desc *xdp_desc = pool->tx_descs;
>> +	int err, i, work_done = 0, batch;
>> +
>> +	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
>> +	batch = xsk_tx_peek_release_desc_batch(pool, budget);
>> +	if (!batch)
>> +		return;
>> +
>> +	for (i = 0; i < batch; i++) {
>> +		dma_addr_t dma_addr;
>> +
>> +		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
>> +		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr,
>xdp_desc[i].len,
>> +					     queue, OTX2_AF_XDP_FRAME);
>> +		if (!err) {
>> +			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer
>packet
>> +err%d\n", err);
>
>Here `err` is always 0, dumping it's value is quite confusing.
>
>The root cause is that otx2_xdp_sq_append_pkt() returns a success
>boolean value, the variable holding it should possibly be renamed
>accordingly.
>
>Since this is the only nit I could find, I think we are better without a
>repost, but please follow-up on this chunk soon.
>
>Thanks,
>
>Paolo
[Suman] Sure, I will push a follow-up patch

RE: [EXTERNAL] Re: [net-next PATCH v6 6/6] octeontx2-pf: AF_XDP zero copy transmit support

[Suman Ghosh]

>> +void otx2_zc_napi_handler(struct otx2_nic *pfvf, struct xsk_buff_pool
>*pool,
>> +			  int queue, int budget)
>> +{
>> +	struct xdp_desc *xdp_desc = pool->tx_descs;
>> +	int err, i, work_done = 0, batch;
>> +
>> +	budget = min(budget, otx2_read_free_sqe(pfvf, queue));
>> +	batch = xsk_tx_peek_release_desc_batch(pool, budget);
>> +	if (!batch)
>> +		return;
>> +
>> +	for (i = 0; i < batch; i++) {
>> +		dma_addr_t dma_addr;
>> +
>> +		dma_addr = xsk_buff_raw_get_dma(pool, xdp_desc[i].addr);
>> +		err = otx2_xdp_sq_append_pkt(pfvf, NULL, dma_addr,
>xdp_desc[i].len,
>> +					     queue, OTX2_AF_XDP_FRAME);
>> +		if (!err) {
>> +			netdev_err(pfvf->netdev, "AF_XDP: Unable to transfer
>packet err%d\n", err);
>> +			break;
>> +		}
>> +		work_done++;
>> +	}
>> +
>> +	if (work_done)
>> +		xsk_tx_release(pool);
>
>this is broken actually. the batch api you're using above is doing tx
>release internally for you.
>
>Sorry for not catching this earlier but i was never CCed in this series.
[Suman] Thanks for the comment, I will push a follow-up patch.

RE: [EXTERNAL] Re: [net-next PATCH v6 3/6] octeontx2-pf: AF_XDP zero copy receive support

[Suman Ghosh]

>>
>> -	pp_params.order = get_order(buf_size);
>> -	pp_params.flags = PP_FLAG_DMA_MAP;
>> -	pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
>> -	pp_params.nid = NUMA_NO_NODE;
>> -	pp_params.dev = pfvf->dev;
>> -	pp_params.dma_dir = DMA_FROM_DEVICE;
>> -	pool->page_pool = page_pool_create(&pp_params);
>> -	if (IS_ERR(pool->page_pool)) {
>> -		netdev_err(pfvf->netdev, "Creation of page pool failed\n");
>> -		return PTR_ERR(pool->page_pool);
>> +	/* Set XSK pool to support AF_XDP zero-copy */
>> +	xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, pool_id);
>> +	if (xsk_pool) {
>> +		pool->xsk_pool = xsk_pool;
>> +		pool->xdp_cnt = numptrs;
>> +		pool->xdp = devm_kcalloc(pfvf->dev,
>> +					 numptrs, sizeof(struct xdp_buff *),
>GFP_KERNEL);
>
>What is the rationale behind having a buffer pool within your driver
>while you have this very same thing within xsk_buff_pool?
>
>You're doubling your work. Just use the xsk_buff_alloc_batch() and have
>a simpler ZC implementation in your driver.
[Suman] This series is the initial change, we will use the batch API in following update patches.
>
>> +		if (IS_ERR(pool->xdp)) {
>> +			netdev_err(pfvf->netdev, "Creation of xsk pool
>failed\n");
>> +			return PTR_ERR(pool->xdp);
>> +		}
>>  	}
>>
>>  	return 0;
>> @@ -1543,9 +1578,18 @@ int otx2_sq_aura_pool_init(struct otx2_nic
>*pfvf)
>>  		}
>>
>>  		for (ptr = 0; ptr < num_sqbs; ptr++) {
>> -			err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
>> -			if (err)
>> +			err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
>> +			if (err) {
>> +				if (pool->xsk_pool) {
>> +					ptr--;
>> +					while (ptr >= 0) {
>> +						xsk_buff_free(pool->xdp[ptr]);
>> +						ptr--;
>> +					}
>> +				}
>>  				goto err_mem;
>> +			}
>> +
>>  			pfvf->hw_ops->aura_freeptr(pfvf, pool_id, bufptr);
>>  			sq->sqb_ptrs[sq->sqb_count++] = (u64)bufptr;
>>  		}
>> @@ -1595,11 +1639,19 @@ int otx2_rq_aura_pool_init(struct otx2_nic
>*pfvf)
>>  	/* Allocate pointers and free them to aura/pool */
>>  	for (pool_id = 0; pool_id < hw->rqpool_cnt; pool_id++) {
>>  		pool = &pfvf->qset.pool[pool_id];
>> +
>>  		for (ptr = 0; ptr < num_ptrs; ptr++) {
>> -			err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
>> -			if (err)
>> +			err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
>> +			if (err) {
>> +				if (pool->xsk_pool) {
>> +					while (ptr)
>> +						xsk_buff_free(pool->xdp[--ptr]);
>> +				}
>>  				return -ENOMEM;
>> +			}
>> +
>>  			pfvf->hw_ops->aura_freeptr(pfvf, pool_id,
>> +						   pool->xsk_pool ? bufptr :
>>  						   bufptr + OTX2_HEAD_ROOM);
>>  		}
>>  	}
>> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
>> b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
>> index d5fbccb289df..60508971b62f 100644
>> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
>> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
>> @@ -532,6 +532,8 @@ struct otx2_nic {
>>
>>  	/* Inline ipsec */
>>  	struct cn10k_ipsec	ipsec;
>> +	/* af_xdp zero-copy */
>> +	unsigned long		*af_xdp_zc_qidx;
>>  };
>>
>>  static inline bool is_otx2_lbkvf(struct pci_dev *pdev) @@ -1003,7
>> +1005,7 @@ void otx2_txschq_free_one(struct otx2_nic *pfvf, u16 lvl,
>> u16 schq);  void otx2_free_pending_sqe(struct otx2_nic *pfvf);  void
>> otx2_sqb_flush(struct otx2_nic *pfvf);  int otx2_alloc_rbuf(struct
>> otx2_nic *pfvf, struct otx2_pool *pool,
>> -		    dma_addr_t *dma);
>> +		    dma_addr_t *dma, int qidx, int idx);
>>  int otx2_rxtx_enable(struct otx2_nic *pfvf, bool enable);  void
>> otx2_ctx_disable(struct mbox *mbox, int type, bool npa);  int
>> otx2_nix_config_bp(struct otx2_nic *pfvf, bool enable); @@ -1033,6
>> +1035,8 @@ void otx2_pfaf_mbox_destroy(struct otx2_nic *pf);  void
>> otx2_disable_mbox_intr(struct otx2_nic *pf);  void
>> otx2_disable_napi(struct otx2_nic *pf);  irqreturn_t
>> otx2_cq_intr_handler(int irq, void *cq_irq);
>> +int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura); int
>> +otx2_cq_init(struct otx2_nic *pfvf, u16 qidx);
>>
>>  /* RSS configuration APIs*/
>>  int otx2_rss_init(struct otx2_nic *pfvf); diff --git
>> a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
>> b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
>> index 4347a3c95350..50a42cd5d50a 100644
>> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
>> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
>> @@ -27,6 +27,7 @@
>>  #include "qos.h"
>>  #include <rvu_trace.h>
>>  #include "cn10k_ipsec.h"
>> +#include "otx2_xsk.h"
>>
>>  #define DRV_NAME	"rvu_nicpf"
>>  #define DRV_STRING	"Marvell RVU NIC Physical Function Driver"
>> @@ -1662,9 +1663,7 @@ void otx2_free_hw_resources(struct otx2_nic *pf)
>>  	struct nix_lf_free_req *free_req;
>>  	struct mbox *mbox = &pf->mbox;
>>  	struct otx2_cq_queue *cq;
>> -	struct otx2_pool *pool;
>>  	struct msg_req *req;
>> -	int pool_id;
>>  	int qidx;
>>
>>  	/* Ensure all SQE are processed */
>> @@ -1705,13 +1704,6 @@ void otx2_free_hw_resources(struct otx2_nic
>*pf)
>>  	/* Free RQ buffer pointers*/
>>  	otx2_free_aura_ptr(pf, AURA_NIX_RQ);
>>
>> -	for (qidx = 0; qidx < pf->hw.rx_queues; qidx++) {
>> -		pool_id = otx2_get_pool_idx(pf, AURA_NIX_RQ, qidx);
>> -		pool = &pf->qset.pool[pool_id];
>> -		page_pool_destroy(pool->page_pool);
>> -		pool->page_pool = NULL;
>> -	}
>> -
>>  	otx2_free_cq_res(pf);
>>
>>  	/* Free all ingress bandwidth profiles allocated */ @@ -2788,6
>> +2780,8 @@ static int otx2_xdp(struct net_device *netdev, struct
>netdev_bpf *xdp)
>>  	switch (xdp->command) {
>>  	case XDP_SETUP_PROG:
>>  		return otx2_xdp_setup(pf, xdp->prog);
>> +	case XDP_SETUP_XSK_POOL:
>> +		return otx2_xsk_pool_setup(pf, xdp->xsk.pool, xdp-
>>xsk.queue_id);
>>  	default:
>>  		return -EINVAL;
>>  	}
>> @@ -2865,6 +2859,7 @@ static const struct net_device_ops
>otx2_netdev_ops = {
>>  	.ndo_set_vf_vlan	= otx2_set_vf_vlan,
>>  	.ndo_get_vf_config	= otx2_get_vf_config,
>>  	.ndo_bpf		= otx2_xdp,
>> +	.ndo_xsk_wakeup		= otx2_xsk_wakeup,
>>  	.ndo_xdp_xmit           = otx2_xdp_xmit,
>>  	.ndo_setup_tc		= otx2_setup_tc,
>>  	.ndo_set_vf_trust	= otx2_ndo_set_vf_trust,
>> @@ -3203,16 +3198,26 @@ static int otx2_probe(struct pci_dev *pdev,
>const struct pci_device_id *id)
>>  	/* Enable link notifications */
>>  	otx2_cgx_config_linkevents(pf, true);
>>
>> +	pf->af_xdp_zc_qidx = bitmap_zalloc(qcount, GFP_KERNEL);
>
>if this is taken from ice drivers then be aware we got rid of bitmap
>tracking zc enabled queues. see adbf5a42341f ("ice: remove af_xdp_zc_qps
>bitmap").
>
>in case you would still have a need for that after going through
>referenced commit, please provide us some justification why.
[Suman] Thanks for pointing it out. I will check if it can be avoided.

Re: [net-next PATCH v6 3/6] octeontx2-pf: AF_XDP zero copy receive support

[Kees Bakker]

Op 13-02-2025 om 06:31 schreef Suman Ghosh:
> This patch adds support to AF_XDP zero copy for CN10K.
> This patch specifically adds receive side support. In this approach once
> a xdp program with zero copy support on a specific rx queue is enabled,
> then that receive quse is disabled/detached from the existing kernel
> queue and re-assigned to the umem memory.
>
> Signed-off-by: Suman Ghosh <sumang@marvell.com>
> ---
>   .../ethernet/marvell/octeontx2/nic/Makefile   |   2 +-
>   .../ethernet/marvell/octeontx2/nic/cn10k.c    |   7 +-
>   .../marvell/octeontx2/nic/otx2_common.c       | 114 ++++++++---
>   .../marvell/octeontx2/nic/otx2_common.h       |   6 +-
>   .../ethernet/marvell/octeontx2/nic/otx2_pf.c  |  25 ++-
>   .../marvell/octeontx2/nic/otx2_txrx.c         |  73 +++++--
>   .../marvell/octeontx2/nic/otx2_txrx.h         |   6 +
>   .../ethernet/marvell/octeontx2/nic/otx2_vf.c  |  12 +-
>   .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 182 ++++++++++++++++++
>   .../ethernet/marvell/octeontx2/nic/otx2_xsk.h |  21 ++
>   .../ethernet/marvell/octeontx2/nic/qos_sq.c   |   2 +-
>   11 files changed, 389 insertions(+), 61 deletions(-)
>   create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
>   create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h
>
>
> [...]
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
> new file mode 100644
> index 000000000000..894c1e0aea6f
> --- /dev/null
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c
>
> [...]
> +static void otx2_clean_up_rq(struct otx2_nic *pfvf, int qidx)
> +{
> +	struct otx2_qset *qset = &pfvf->qset;
> +	struct otx2_cq_queue *cq;
> +	struct otx2_pool *pool;
> +	u64 iova;
> +
> +	/* If the DOWN flag is set SQs are already freed */
> +	if (pfvf->flags & OTX2_FLAG_INTF_DOWN)
> +		return;
> +
> +	cq = &qset->cq[qidx];
> +	if (cq)
> +		otx2_cleanup_rx_cqes(pfvf, cq, qidx);
The if check makes no sense, cq is always != NULL
> +
> [...]
> +int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
> +{
> +	struct otx2_nic *pf = netdev_priv(dev);
> +	struct otx2_cq_poll *cq_poll = NULL;
> +	struct otx2_qset *qset = &pf->qset;
> +
> +	if (pf->flags & OTX2_FLAG_INTF_DOWN)
> +		return -ENETDOWN;
> +
> +	if (queue_id >= pf->hw.rx_queues)
> +		return -EINVAL;
> +
> +	cq_poll = &qset->napi[queue_id];
> +	if (!cq_poll)
> +		return -EINVAL;
cq_poll can never be NULL.
> [...]
>

Re: [net-next PATCH v6 1/6] octeontx2-pf: use xdp_return_frame() to free xdp buffers

[Yunsheng Lin]

On 2025/2/13 13:31, Suman Ghosh wrote:

> -		put_page(page);
>  		cq->pool_ptrs++;
> +		if (page->pp) {

It seems the above changing caused the below error for the DMA API misuse
patchset：
https://lore.kernel.org/oe-kbuild-all/202502280250.Bp3jD6ZE-lkp@intel.com/

And it seems this patch uses 'page->pp' being NULL or not to decide if a
page is page_pool owned or not, I am not sure if the buddy page allocator
will always ensure that memory that 'page->pp' points to will always be
zero, even if it is for now, It seems the driver should not use that to
decide if a page is page_pool owned or not.

The PP_SIGNATURE magic macro seems to be correct way to decide if the page
is page_pool owned or not when driver doesn't have its own way to decide
if a page is page_pool owned or not, see:
https://elixir.bootlin.com/linux/v6.14-rc1/source/net/core/skbuff.c#L924

> +			page_pool_recycle_direct(pool->page_pool, page);
> +		} else {
> +			otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
> +					    DMA_FROM_DEVICE);
> +			put_page(page);
> +		}
>  		return true;
>  	}
>  	return false;