]> www.infradead.org Git - users/jedix/linux-maple.git/commitdiff
net: ethernet: ti: am65-cpsw: fix XDP_DROP, XDP_TX and XDP_REDIRECT
authorRoger Quadros <rogerq@kernel.org>
Thu, 29 Aug 2024 12:03:19 +0000 (15:03 +0300)
committerPaolo Abeni <pabeni@redhat.com>
Tue, 3 Sep 2024 08:43:46 +0000 (10:43 +0200)
The following XDP_DROP test from [1] stalls the interface after
250 packets.
~# xdb-bench drop -m native eth0
This is because new RX requests are never queued. Fix that.

The below XDP_TX test from [1] fails with a warning
[  499.947381] XDP_WARN: xdp_update_frame_from_buff(line:277): Driver BUG: missing reserved tailroom
~# xdb-bench tx -m native eth0
Fix that by using PAGE_SIZE during xdp_init_buf().

In XDP_REDIRECT case only 1 packet was processed in rx_poll.
Fix it to process up to budget packets.

Fix all XDP error cases to call trace_xdp_exception() and drop the packet
in am65_cpsw_run_xdp().

[1] xdp-tools suite https://github.com/xdp-project/xdp-tools

Fixes: 8acacc40f733 ("net: ethernet: ti: am65-cpsw: Add minimal XDP support")
Signed-off-by: Roger Quadros <rogerq@kernel.org>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Acked-by: Julien Panis <jpanis@baylibre.com>
Reviewed-by: MD Danish Anwar <danishanwar@ti.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
drivers/net/ethernet/ti/am65-cpsw-nuss.c

index 81d9f21086ec93288a8211437b4d29d9ca1e350d..9fd2ba26716cb86a585cb84500a45b42acc0c108 100644 (file)
 #define AM65_CPSW_CPPI_TX_PKT_TYPE 0x7
 
 /* XDP */
-#define AM65_CPSW_XDP_CONSUMED 2
-#define AM65_CPSW_XDP_REDIRECT 1
+#define AM65_CPSW_XDP_CONSUMED BIT(1)
+#define AM65_CPSW_XDP_REDIRECT BIT(0)
 #define AM65_CPSW_XDP_PASS     0
 
 /* Include headroom compatible with both skb and xdpf */
-#define AM65_CPSW_HEADROOM (max(NET_SKB_PAD, XDP_PACKET_HEADROOM) + NET_IP_ALIGN)
+#define AM65_CPSW_HEADROOM_NA (max(NET_SKB_PAD, XDP_PACKET_HEADROOM) + NET_IP_ALIGN)
+#define AM65_CPSW_HEADROOM ALIGN(AM65_CPSW_HEADROOM_NA, sizeof(long))
 
 static void am65_cpsw_port_set_sl_mac(struct am65_cpsw_port *slave,
                                      const u8 *dev_addr)
@@ -933,7 +934,7 @@ static int am65_cpsw_xdp_tx_frame(struct net_device *ndev,
        host_desc = k3_cppi_desc_pool_alloc(tx_chn->desc_pool);
        if (unlikely(!host_desc)) {
                ndev->stats.tx_dropped++;
-               return -ENOMEM;
+               return AM65_CPSW_XDP_CONSUMED;  /* drop */
        }
 
        am65_cpsw_nuss_set_buf_type(tx_chn, host_desc, buf_type);
@@ -942,7 +943,7 @@ static int am65_cpsw_xdp_tx_frame(struct net_device *ndev,
                                 pkt_len, DMA_TO_DEVICE);
        if (unlikely(dma_mapping_error(tx_chn->dma_dev, dma_buf))) {
                ndev->stats.tx_dropped++;
-               ret = -ENOMEM;
+               ret = AM65_CPSW_XDP_CONSUMED;   /* drop */
                goto pool_free;
        }
 
@@ -977,6 +978,7 @@ static int am65_cpsw_xdp_tx_frame(struct net_device *ndev,
                /* Inform BQL */
                netdev_tx_completed_queue(netif_txq, 1, pkt_len);
                ndev->stats.tx_errors++;
+               ret = AM65_CPSW_XDP_CONSUMED; /* drop */
                goto dma_unmap;
        }
 
@@ -1004,6 +1006,7 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common,
        struct bpf_prog *prog;
        struct page *page;
        u32 act;
+       int err;
 
        prog = READ_ONCE(port->xdp_prog);
        if (!prog)
@@ -1023,14 +1026,14 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common,
 
                xdpf = xdp_convert_buff_to_frame(xdp);
                if (unlikely(!xdpf))
-                       break;
+                       goto drop;
 
                __netif_tx_lock(netif_txq, cpu);
-               ret = am65_cpsw_xdp_tx_frame(ndev, tx_chn, xdpf,
+               err = am65_cpsw_xdp_tx_frame(ndev, tx_chn, xdpf,
                                             AM65_CPSW_TX_BUF_TYPE_XDP_TX);
                __netif_tx_unlock(netif_txq);
-               if (ret)
-                       break;
+               if (err)
+                       goto drop;
 
                ndev->stats.rx_bytes += *len;
                ndev->stats.rx_packets++;
@@ -1038,7 +1041,7 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common,
                goto out;
        case XDP_REDIRECT:
                if (unlikely(xdp_do_redirect(ndev, xdp, prog)))
-                       break;
+                       goto drop;
 
                ndev->stats.rx_bytes += *len;
                ndev->stats.rx_packets++;
@@ -1048,6 +1051,7 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common,
                bpf_warn_invalid_xdp_action(ndev, prog, act);
                fallthrough;
        case XDP_ABORTED:
+drop:
                trace_xdp_exception(ndev, prog, act);
                fallthrough;
        case XDP_DROP:
@@ -1056,7 +1060,6 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common,
 
        page = virt_to_head_page(xdp->data);
        am65_cpsw_put_page(rx_chn, page, true, desc_idx);
-
 out:
        return ret;
 }
@@ -1095,7 +1098,7 @@ static void am65_cpsw_nuss_rx_csum(struct sk_buff *skb, u32 csum_info)
 }
 
 static int am65_cpsw_nuss_rx_packets(struct am65_cpsw_common *common,
-                                    u32 flow_idx, int cpu)
+                                    u32 flow_idx, int cpu, int *xdp_state)
 {
        struct am65_cpsw_rx_chn *rx_chn = &common->rx_chns;
        u32 buf_dma_len, pkt_len, port_id = 0, csum_info;
@@ -1114,6 +1117,7 @@ static int am65_cpsw_nuss_rx_packets(struct am65_cpsw_common *common,
        void **swdata;
        u32 *psdata;
 
+       *xdp_state = AM65_CPSW_XDP_PASS;
        ret = k3_udma_glue_pop_rx_chn(rx_chn->rx_chn, flow_idx, &desc_dma);
        if (ret) {
                if (ret != -ENODATA)
@@ -1161,15 +1165,13 @@ static int am65_cpsw_nuss_rx_packets(struct am65_cpsw_common *common,
        }
 
        if (port->xdp_prog) {
-               xdp_init_buff(&xdp, AM65_CPSW_MAX_PACKET_SIZE, &port->xdp_rxq);
-
-               xdp_prepare_buff(&xdp, page_addr, skb_headroom(skb),
+               xdp_init_buff(&xdp, PAGE_SIZE, &port->xdp_rxq);
+               xdp_prepare_buff(&xdp, page_addr, AM65_CPSW_HEADROOM,
                                 pkt_len, false);
-
-               ret = am65_cpsw_run_xdp(common, port, &xdp, desc_idx,
-                                       cpu, &pkt_len);
-               if (ret != AM65_CPSW_XDP_PASS)
-                       return ret;
+               *xdp_state = am65_cpsw_run_xdp(common, port, &xdp, desc_idx,
+                                              cpu, &pkt_len);
+               if (*xdp_state != AM65_CPSW_XDP_PASS)
+                       goto allocate;
 
                /* Compute additional headroom to be reserved */
                headroom = (xdp.data - xdp.data_hard_start) - skb_headroom(skb);
@@ -1193,9 +1195,13 @@ static int am65_cpsw_nuss_rx_packets(struct am65_cpsw_common *common,
        stats->rx_bytes += pkt_len;
        u64_stats_update_end(&stats->syncp);
 
+allocate:
        new_page = page_pool_dev_alloc_pages(rx_chn->page_pool);
-       if (unlikely(!new_page))
+       if (unlikely(!new_page)) {
+               dev_err(dev, "page alloc failed\n");
                return -ENOMEM;
+       }
+
        rx_chn->pages[desc_idx] = new_page;
 
        if (netif_dormant(ndev)) {
@@ -1229,8 +1235,9 @@ static int am65_cpsw_nuss_rx_poll(struct napi_struct *napi_rx, int budget)
        struct am65_cpsw_common *common = am65_cpsw_napi_to_common(napi_rx);
        int flow = AM65_CPSW_MAX_RX_FLOWS;
        int cpu = smp_processor_id();
-       bool xdp_redirect = false;
+       int xdp_state_or = 0;
        int cur_budget, ret;
+       int xdp_state;
        int num_rx = 0;
 
        /* process every flow */
@@ -1238,12 +1245,11 @@ static int am65_cpsw_nuss_rx_poll(struct napi_struct *napi_rx, int budget)
                cur_budget = budget - num_rx;
 
                while (cur_budget--) {
-                       ret = am65_cpsw_nuss_rx_packets(common, flow, cpu);
-                       if (ret) {
-                               if (ret == AM65_CPSW_XDP_REDIRECT)
-                                       xdp_redirect = true;
+                       ret = am65_cpsw_nuss_rx_packets(common, flow, cpu,
+                                                       &xdp_state);
+                       xdp_state_or |= xdp_state;
+                       if (ret)
                                break;
-                       }
                        num_rx++;
                }
 
@@ -1251,7 +1257,7 @@ static int am65_cpsw_nuss_rx_poll(struct napi_struct *napi_rx, int budget)
                        break;
        }
 
-       if (xdp_redirect)
+       if (xdp_state_or & AM65_CPSW_XDP_REDIRECT)
                xdp_do_flush();
 
        dev_dbg(common->dev, "%s num_rx:%d %d\n", __func__, num_rx, budget);