net: ti: icssg-prueth: Add XDP support
authorRoger Quadros <rogerq@kernel.org>
Wed, 5 Mar 2025 10:14:22 +0000 (15:44 +0530)
committerPaolo Abeni <pabeni@redhat.com>
Tue, 11 Mar 2025 10:09:56 +0000 (11:09 +0100)
Add native XDP support. We do not support zero copy yet.

Signed-off-by: Roger Quadros <rogerq@kernel.org>
Signed-off-by: MD Danish Anwar <danishanwar@ti.com>
Signed-off-by: Meghana Malladi <m-malladi@ti.com>
Link: https://patch.msgid.link/20250305101422.1908370-4-m-malladi@ti.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
drivers/net/ethernet/ti/icssg/icssg_common.c
drivers/net/ethernet/ti/icssg/icssg_prueth.c
drivers/net/ethernet/ti/icssg/icssg_prueth.h

index fee1204db36722f1aab0c42fc231edd5e03a5564..df5da7a98abf08c1278656a27db2d6283ed30f92 100644 (file)
@@ -98,11 +98,20 @@ void prueth_xmit_free(struct prueth_tx_chn *tx_chn,
 {
        struct cppi5_host_desc_t *first_desc, *next_desc;
        dma_addr_t buf_dma, next_desc_dma;
+       struct prueth_swdata *swdata;
+       struct page *page;
        u32 buf_dma_len;
 
        first_desc = desc;
        next_desc = first_desc;
 
+       swdata = cppi5_hdesc_get_swdata(desc);
+       if (swdata->type == PRUETH_SWDATA_PAGE) {
+               page = swdata->data.page;
+               page_pool_recycle_direct(page->pp, swdata->data.page);
+               goto free_desc;
+       }
+
        cppi5_hdesc_get_obuf(first_desc, &buf_dma, &buf_dma_len);
        k3_udma_glue_tx_cppi5_to_dma_addr(tx_chn->tx_chn, &buf_dma);
 
@@ -126,6 +135,7 @@ void prueth_xmit_free(struct prueth_tx_chn *tx_chn,
                k3_cppi_desc_pool_free(tx_chn->desc_pool, next_desc);
        }
 
+free_desc:
        k3_cppi_desc_pool_free(tx_chn->desc_pool, first_desc);
 }
 EXPORT_SYMBOL_GPL(prueth_xmit_free);
@@ -139,6 +149,7 @@ int emac_tx_complete_packets(struct prueth_emac *emac, int chn,
        struct prueth_swdata *swdata;
        struct prueth_tx_chn *tx_chn;
        unsigned int total_bytes = 0;
+       struct xdp_frame *xdpf;
        struct sk_buff *skb;
        dma_addr_t desc_dma;
        int res, num_tx = 0;
@@ -161,16 +172,28 @@ int emac_tx_complete_packets(struct prueth_emac *emac, int chn,
                desc_tx = k3_cppi_desc_pool_dma2virt(tx_chn->desc_pool,
                                                     desc_dma);
                swdata = cppi5_hdesc_get_swdata(desc_tx);
-               prueth_xmit_free(tx_chn, desc_tx);
-               if (swdata->type != PRUETH_SWDATA_SKB)
+
+               switch (swdata->type) {
+               case PRUETH_SWDATA_SKB:
+                       skb = swdata->data.skb;
+                       dev_sw_netstats_tx_add(skb->dev, 1, skb->len);
+                       total_bytes += skb->len;
+                       napi_consume_skb(skb, budget);
+                       break;
+               case PRUETH_SWDATA_XDPF:
+                       xdpf = swdata->data.xdpf;
+                       dev_sw_netstats_tx_add(ndev, 1, xdpf->len);
+                       total_bytes += xdpf->len;
+                       xdp_return_frame(xdpf);
+                       break;
+               default:
+                       netdev_err(ndev, "tx_complete: invalid swdata type %d\n", swdata->type);
+                       prueth_xmit_free(tx_chn, desc_tx);
+                       ndev->stats.tx_dropped++;
                        continue;
+               }
 
-               skb = swdata->data.skb;
-               ndev = skb->dev;
-               ndev->stats.tx_packets++;
-               ndev->stats.tx_bytes += skb->len;
-               total_bytes += skb->len;
-               napi_consume_skb(skb, budget);
+               prueth_xmit_free(tx_chn, desc_tx);
                num_tx++;
        }
 
@@ -529,7 +552,153 @@ void emac_rx_timestamp(struct prueth_emac *emac,
        ssh->hwtstamp = ns_to_ktime(ns);
 }
 
-static int emac_rx_packet(struct prueth_emac *emac, u32 flow_id)
+/**
+ * emac_xmit_xdp_frame - transmits an XDP frame
+ * @emac: emac device
+ * @xdpf: data to transmit
+ * @page: page from page pool if already DMA mapped
+ * @q_idx: queue id
+ *
+ * Return: XDP state
+ */
+u32 emac_xmit_xdp_frame(struct prueth_emac *emac,
+                       struct xdp_frame *xdpf,
+                       struct page *page,
+                       unsigned int q_idx)
+{
+       struct cppi5_host_desc_t *first_desc;
+       struct net_device *ndev = emac->ndev;
+       struct prueth_tx_chn *tx_chn;
+       dma_addr_t desc_dma, buf_dma;
+       struct prueth_swdata *swdata;
+       u32 *epib;
+       int ret;
+
+       if (q_idx >= PRUETH_MAX_TX_QUEUES) {
+               netdev_err(ndev, "xdp tx: invalid q_id %d\n", q_idx);
+               return ICSSG_XDP_CONSUMED;      /* drop */
+       }
+
+       tx_chn = &emac->tx_chns[q_idx];
+
+       first_desc = k3_cppi_desc_pool_alloc(tx_chn->desc_pool);
+       if (!first_desc) {
+               netdev_dbg(ndev, "xdp tx: failed to allocate descriptor\n");
+               goto drop_free_descs;   /* drop */
+       }
+
+       if (page) { /* already DMA mapped by page_pool */
+               buf_dma = page_pool_get_dma_addr(page);
+               buf_dma += xdpf->headroom + sizeof(struct xdp_frame);
+       } else { /* Map the linear buffer */
+               buf_dma = dma_map_single(tx_chn->dma_dev, xdpf->data, xdpf->len, DMA_TO_DEVICE);
+               if (dma_mapping_error(tx_chn->dma_dev, buf_dma)) {
+                       netdev_err(ndev, "xdp tx: failed to map data buffer\n");
+                       goto drop_free_descs;   /* drop */
+               }
+       }
+
+       cppi5_hdesc_init(first_desc, CPPI5_INFO0_HDESC_EPIB_PRESENT,
+                        PRUETH_NAV_PS_DATA_SIZE);
+       cppi5_hdesc_set_pkttype(first_desc, 0);
+       epib = first_desc->epib;
+       epib[0] = 0;
+       epib[1] = 0;
+
+       /* set dst tag to indicate internal qid at the firmware which is at
+        * bit8..bit15. bit0..bit7 indicates port num for directed
+        * packets in case of switch mode operation
+        */
+       cppi5_desc_set_tags_ids(&first_desc->hdr, 0, (emac->port_id | (q_idx << 8)));
+       k3_udma_glue_tx_dma_to_cppi5_addr(tx_chn->tx_chn, &buf_dma);
+       cppi5_hdesc_attach_buf(first_desc, buf_dma, xdpf->len, buf_dma, xdpf->len);
+       swdata = cppi5_hdesc_get_swdata(first_desc);
+       if (page) {
+               swdata->type = PRUETH_SWDATA_PAGE;
+               swdata->data.page = page;
+       } else {
+               swdata->type = PRUETH_SWDATA_XDPF;
+               swdata->data.xdpf = xdpf;
+       }
+
+       cppi5_hdesc_set_pktlen(first_desc, xdpf->len);
+       desc_dma = k3_cppi_desc_pool_virt2dma(tx_chn->desc_pool, first_desc);
+
+       ret = k3_udma_glue_push_tx_chn(tx_chn->tx_chn, first_desc, desc_dma);
+       if (ret) {
+               netdev_err(ndev, "xdp tx: push failed: %d\n", ret);
+               goto drop_free_descs;
+       }
+
+       return ICSSG_XDP_TX;
+
+drop_free_descs:
+       prueth_xmit_free(tx_chn, first_desc);
+       return ICSSG_XDP_CONSUMED;
+}
+EXPORT_SYMBOL_GPL(emac_xmit_xdp_frame);
+
+/**
+ * emac_run_xdp - run an XDP program
+ * @emac: emac device
+ * @xdp: XDP buffer containing the frame
+ * @page: page with RX data if already DMA mapped
+ * @len: Rx descriptor packet length
+ *
+ * Return: XDP state
+ */
+static u32 emac_run_xdp(struct prueth_emac *emac, struct xdp_buff *xdp,
+                       struct page *page, u32 *len)
+{
+       struct net_device *ndev = emac->ndev;
+       struct bpf_prog *xdp_prog;
+       struct xdp_frame *xdpf;
+       u32 pkt_len = *len;
+       u32 act, result;
+       int q_idx, err;
+
+       xdp_prog = READ_ONCE(emac->xdp_prog);
+       act = bpf_prog_run_xdp(xdp_prog, xdp);
+       switch (act) {
+       case XDP_PASS:
+               return ICSSG_XDP_PASS;
+       case XDP_TX:
+               /* Send packet to TX ring for immediate transmission */
+               xdpf = xdp_convert_buff_to_frame(xdp);
+               if (unlikely(!xdpf)) {
+                       ndev->stats.tx_dropped++;
+                       goto drop;
+               }
+
+               q_idx = smp_processor_id() % emac->tx_ch_num;
+               result = emac_xmit_xdp_frame(emac, xdpf, page, q_idx);
+               if (result == ICSSG_XDP_CONSUMED)
+                       goto drop;
+
+               dev_sw_netstats_rx_add(ndev, xdpf->len);
+               return result;
+       case XDP_REDIRECT:
+               err = xdp_do_redirect(emac->ndev, xdp, xdp_prog);
+               if (err)
+                       goto drop;
+
+               dev_sw_netstats_rx_add(ndev, pkt_len);
+               return ICSSG_XDP_REDIR;
+       default:
+               bpf_warn_invalid_xdp_action(emac->ndev, xdp_prog, act);
+               fallthrough;
+       case XDP_ABORTED:
+drop:
+               trace_xdp_exception(emac->ndev, xdp_prog, act);
+               fallthrough; /* handle aborts by dropping packet */
+       case XDP_DROP:
+               ndev->stats.rx_dropped++;
+               page_pool_recycle_direct(emac->rx_chns.pg_pool, page);
+               return ICSSG_XDP_CONSUMED;
+       }
+}
+
+static int emac_rx_packet(struct prueth_emac *emac, u32 flow_id, u32 *xdp_state)
 {
        struct prueth_rx_chn *rx_chn = &emac->rx_chns;
        u32 buf_dma_len, pkt_len, port_id = 0;
@@ -540,10 +709,12 @@ static int emac_rx_packet(struct prueth_emac *emac, u32 flow_id)
        struct page *page, *new_page;
        struct page_pool *pool;
        struct sk_buff *skb;
+       struct xdp_buff xdp;
        u32 *psdata;
        void *pa;
        int ret;
 
+       *xdp_state = 0;
        pool = rx_chn->pg_pool;
        ret = k3_udma_glue_pop_rx_chn(rx_chn->rx_chn, flow_id, &desc_dma);
        if (ret) {
@@ -584,9 +755,21 @@ static int emac_rx_packet(struct prueth_emac *emac, u32 flow_id)
                goto requeue;
        }
 
-       /* prepare skb and send to n/w stack */
        pa = page_address(page);
-       skb = napi_build_skb(pa, PAGE_SIZE);
+       if (emac->xdp_prog) {
+               xdp_init_buff(&xdp, PAGE_SIZE, &rx_chn->xdp_rxq);
+               xdp_prepare_buff(&xdp, pa, PRUETH_HEADROOM, pkt_len, false);
+
+               *xdp_state = emac_run_xdp(emac, &xdp, page, &pkt_len);
+               if (*xdp_state == ICSSG_XDP_PASS)
+                       skb = xdp_build_skb_from_buff(&xdp);
+               else
+                       goto requeue;
+       } else {
+               /* prepare skb and send to n/w stack */
+               skb = napi_build_skb(pa, PAGE_SIZE);
+       }
+
        if (!skb) {
                ndev->stats.rx_dropped++;
                page_pool_recycle_direct(pool, page);
@@ -849,13 +1032,23 @@ static void prueth_tx_cleanup(void *data, dma_addr_t desc_dma)
        struct prueth_tx_chn *tx_chn = data;
        struct cppi5_host_desc_t *desc_tx;
        struct prueth_swdata *swdata;
+       struct xdp_frame *xdpf;
        struct sk_buff *skb;
 
        desc_tx = k3_cppi_desc_pool_dma2virt(tx_chn->desc_pool, desc_dma);
        swdata = cppi5_hdesc_get_swdata(desc_tx);
-       if (swdata->type == PRUETH_SWDATA_SKB) {
+
+       switch (swdata->type) {
+       case PRUETH_SWDATA_SKB:
                skb = swdata->data.skb;
                dev_kfree_skb_any(skb);
+               break;
+       case PRUETH_SWDATA_XDPF:
+               xdpf = swdata->data.xdpf;
+               xdp_return_frame(xdpf);
+               break;
+       default:
+               break;
        }
 
        prueth_xmit_free(tx_chn, desc_tx);
@@ -892,15 +1085,18 @@ int icssg_napi_rx_poll(struct napi_struct *napi_rx, int budget)
                PRUETH_RX_FLOW_DATA_SR1 : PRUETH_RX_FLOW_DATA;
        int flow = emac->is_sr1 ?
                PRUETH_MAX_RX_FLOWS_SR1 : PRUETH_MAX_RX_FLOWS;
+       int xdp_state_or = 0;
        int num_rx = 0;
        int cur_budget;
+       u32 xdp_state;
        int ret;
 
        while (flow--) {
                cur_budget = budget - num_rx;
 
                while (cur_budget--) {
-                       ret = emac_rx_packet(emac, flow);
+                       ret = emac_rx_packet(emac, flow, &xdp_state);
+                       xdp_state_or |= xdp_state;
                        if (ret)
                                break;
                        num_rx++;
@@ -910,6 +1106,9 @@ int icssg_napi_rx_poll(struct napi_struct *napi_rx, int budget)
                        break;
        }
 
+       if (xdp_state_or & ICSSG_XDP_REDIR)
+               xdp_do_flush();
+
        if (num_rx < budget && napi_complete_done(napi_rx, num_rx)) {
                if (unlikely(emac->rx_pace_timeout_ns)) {
                        hrtimer_start(&emac->rx_hrtimer,
index 3ff8c322f9d93dfa54dbe228c2f59695b11fb756..c4a995630eca9c7e8ddcecb9f232c5decdac942f 100644 (file)
@@ -559,6 +559,33 @@ const struct icss_iep_clockops prueth_iep_clockops = {
        .perout_enable = prueth_perout_enable,
 };
 
+static int prueth_create_xdp_rxqs(struct prueth_emac *emac)
+{
+       struct xdp_rxq_info *rxq = &emac->rx_chns.xdp_rxq;
+       struct page_pool *pool = emac->rx_chns.pg_pool;
+       int ret;
+
+       ret = xdp_rxq_info_reg(rxq, emac->ndev, 0, emac->napi_rx.napi_id);
+       if (ret)
+               return ret;
+
+       ret = xdp_rxq_info_reg_mem_model(rxq, MEM_TYPE_PAGE_POOL, pool);
+       if (ret)
+               xdp_rxq_info_unreg(rxq);
+
+       return ret;
+}
+
+static void prueth_destroy_xdp_rxqs(struct prueth_emac *emac)
+{
+       struct xdp_rxq_info *rxq = &emac->rx_chns.xdp_rxq;
+
+       if (!xdp_rxq_info_is_reg(rxq))
+               return;
+
+       xdp_rxq_info_unreg(rxq);
+}
+
 static int icssg_prueth_add_mcast(struct net_device *ndev, const u8 *addr)
 {
        struct net_device *real_dev;
@@ -780,10 +807,14 @@ static int emac_ndo_open(struct net_device *ndev)
        if (ret)
                goto free_tx_ts_irq;
 
-       ret = k3_udma_glue_enable_rx_chn(emac->rx_chns.rx_chn);
+       ret = prueth_create_xdp_rxqs(emac);
        if (ret)
                goto reset_rx_chn;
 
+       ret = k3_udma_glue_enable_rx_chn(emac->rx_chns.rx_chn);
+       if (ret)
+               goto destroy_xdp_rxqs;
+
        for (i = 0; i < emac->tx_ch_num; i++) {
                ret = k3_udma_glue_enable_tx_chn(emac->tx_chns[i].tx_chn);
                if (ret)
@@ -809,6 +840,8 @@ reset_tx_chan:
         * any SKB for completion. So set false to free_skb
         */
        prueth_reset_tx_chan(emac, i, false);
+destroy_xdp_rxqs:
+       prueth_destroy_xdp_rxqs(emac);
 reset_rx_chn:
        prueth_reset_rx_chan(&emac->rx_chns, max_rx_flows, false);
 free_tx_ts_irq:
@@ -879,7 +912,7 @@ static int emac_ndo_stop(struct net_device *ndev)
        k3_udma_glue_tdown_rx_chn(emac->rx_chns.rx_chn, true);
 
        prueth_reset_rx_chan(&emac->rx_chns, max_rx_flows, true);
-
+       prueth_destroy_xdp_rxqs(emac);
        napi_disable(&emac->napi_rx);
        hrtimer_cancel(&emac->rx_hrtimer);
 
@@ -1024,6 +1057,93 @@ static int emac_ndo_vlan_rx_del_vid(struct net_device *ndev,
        return 0;
 }
 
+/**
+ * emac_xdp_xmit - Implements ndo_xdp_xmit
+ * @dev: netdev
+ * @n: number of frames
+ * @frames: array of XDP buffer pointers
+ * @flags: XDP extra info
+ *
+ * Return: number of frames successfully sent. Failed frames
+ * will be free'ed by XDP core.
+ *
+ * For error cases, a negative errno code is returned and no-frames
+ * are transmitted (caller must handle freeing frames).
+ **/
+static int emac_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames,
+                        u32 flags)
+{
+       struct prueth_emac *emac = netdev_priv(dev);
+       struct net_device *ndev = emac->ndev;
+       struct xdp_frame *xdpf;
+       unsigned int q_idx;
+       int nxmit = 0;
+       u32 err;
+       int i;
+
+       q_idx = smp_processor_id() % emac->tx_ch_num;
+
+       if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
+               return -EINVAL;
+
+       for (i = 0; i < n; i++) {
+               xdpf = frames[i];
+               err = emac_xmit_xdp_frame(emac, xdpf, NULL, q_idx);
+               if (err != ICSSG_XDP_TX) {
+                       ndev->stats.tx_dropped++;
+                       break;
+               }
+               nxmit++;
+       }
+
+       return nxmit;
+}
+
+/**
+ * emac_xdp_setup - add/remove an XDP program
+ * @emac: emac device
+ * @bpf: XDP program
+ *
+ * Return: Always 0 (Success)
+ **/
+static int emac_xdp_setup(struct prueth_emac *emac, struct netdev_bpf *bpf)
+{
+       struct bpf_prog *prog = bpf->prog;
+       xdp_features_t val;
+
+       val = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
+             NETDEV_XDP_ACT_NDO_XMIT;
+       xdp_set_features_flag(emac->ndev, val);
+
+       if (!emac->xdpi.prog && !prog)
+               return 0;
+
+       WRITE_ONCE(emac->xdp_prog, prog);
+
+       xdp_attachment_setup(&emac->xdpi, bpf);
+
+       return 0;
+}
+
+/**
+ * emac_ndo_bpf - implements ndo_bpf for icssg_prueth
+ * @ndev: network adapter device
+ * @bpf: XDP program
+ *
+ * Return: 0 on success, error code on failure.
+ **/
+static int emac_ndo_bpf(struct net_device *ndev, struct netdev_bpf *bpf)
+{
+       struct prueth_emac *emac = netdev_priv(ndev);
+
+       switch (bpf->command) {
+       case XDP_SETUP_PROG:
+               return emac_xdp_setup(emac, bpf);
+       default:
+               return -EINVAL;
+       }
+}
+
 static const struct net_device_ops emac_netdev_ops = {
        .ndo_open = emac_ndo_open,
        .ndo_stop = emac_ndo_stop,
@@ -1038,6 +1158,8 @@ static const struct net_device_ops emac_netdev_ops = {
        .ndo_fix_features = emac_ndo_fix_features,
        .ndo_vlan_rx_add_vid = emac_ndo_vlan_rx_add_vid,
        .ndo_vlan_rx_kill_vid = emac_ndo_vlan_rx_del_vid,
+       .ndo_bpf = emac_ndo_bpf,
+       .ndo_xdp_xmit = emac_xdp_xmit,
 };
 
 static int prueth_netdev_init(struct prueth *prueth,
@@ -1066,6 +1188,8 @@ static int prueth_netdev_init(struct prueth *prueth,
        emac->prueth = prueth;
        emac->ndev = ndev;
        emac->port_id = port;
+       emac->xdp_prog = NULL;
+       emac->ndev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS;
        emac->cmd_wq = create_singlethread_workqueue("icssg_cmd_wq");
        if (!emac->cmd_wq) {
                ret = -ENOMEM;
index 3bbabd0071299963ba6c1aa6a21a7b6369a6489b..1dd76e2adfcd0898df2cecf2131593599c6373df 100644 (file)
@@ -8,6 +8,8 @@
 #ifndef __NET_TI_ICSSG_PRUETH_H
 #define __NET_TI_ICSSG_PRUETH_H
 
+#include <linux/bpf.h>
+#include <linux/bpf_trace.h>
 #include <linux/etherdevice.h>
 #include <linux/genalloc.h>
 #include <linux/if_vlan.h>
@@ -134,6 +136,7 @@ struct prueth_rx_chn {
        unsigned int irq[ICSSG_MAX_RFLOWS];     /* separate irq per flow */
        char name[32];
        struct page_pool *pg_pool;
+       struct xdp_rxq_info xdp_rxq;
 };
 
 enum prueth_swdata_type {
@@ -141,6 +144,7 @@ enum prueth_swdata_type {
        PRUETH_SWDATA_SKB,
        PRUETH_SWDATA_PAGE,
        PRUETH_SWDATA_CMD,
+       PRUETH_SWDATA_XDPF,
 };
 
 struct prueth_swdata {
@@ -149,6 +153,7 @@ struct prueth_swdata {
                struct sk_buff *skb;
                struct page *page;
                u32 cmd;
+               struct xdp_frame *xdpf;
        } data;
 };
 
@@ -159,6 +164,12 @@ struct prueth_swdata {
 
 #define PRUETH_MAX_TX_TS_REQUESTS      50 /* Max simultaneous TX_TS requests */
 
+/* XDP BPF state */
+#define ICSSG_XDP_PASS           0
+#define ICSSG_XDP_CONSUMED       BIT(0)
+#define ICSSG_XDP_TX             BIT(1)
+#define ICSSG_XDP_REDIR          BIT(2)
+
 /* Minimum coalesce time in usecs for both Tx and Rx */
 #define ICSSG_MIN_COALESCE_USECS 20
 
@@ -227,6 +238,8 @@ struct prueth_emac {
        unsigned long rx_pace_timeout_ns;
 
        struct netdev_hw_addr_list vlan_mcast_list[MAX_VLAN_ID];
+       struct bpf_prog *xdp_prog;
+       struct xdp_attachment_info xdpi;
 };
 
 /* The buf includes headroom compatible with both skb and xdpf */
@@ -465,5 +478,9 @@ void prueth_put_cores(struct prueth *prueth, int slice);
 
 /* Revision specific helper */
 u64 icssg_ts_to_ns(u32 hi_sw, u32 hi, u32 lo, u32 cycle_time_ns);
+u32 emac_xmit_xdp_frame(struct prueth_emac *emac,
+                       struct xdp_frame *xdpf,
+                       struct page *page,
+                       unsigned int q_idx);
 
 #endif /* __NET_TI_ICSSG_PRUETH_H */