From 5e24db550bd6f484d2c7687ee488708260e1f84a Mon Sep 17 00:00:00 2001 From: Roger Quadros Date: Thu, 29 Aug 2024 15:03:19 +0300 Subject: [PATCH 1/3] net: ethernet: ti: am65-cpsw: fix XDP_DROP, XDP_TX and XDP_REDIRECT 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 Reviewed-by: Jacob Keller Acked-by: Julien Panis Reviewed-by: MD Danish Anwar Signed-off-by: Paolo Abeni --- drivers/net/ethernet/ti/am65-cpsw-nuss.c | 62 +++++++++++++----------- 1 file changed, 34 insertions(+), 28 deletions(-) diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c index 81d9f21086ec..9fd2ba26716c 100644 --- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c +++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c @@ -156,12 +156,13 @@ #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); From 0a50c35277f96481a5a6ed5faf347f282040c57d Mon Sep 17 00:00:00 2001 From: Roger Quadros Date: Thu, 29 Aug 2024 15:03:20 +0300 Subject: [PATCH 2/3] net: ethernet: ti: am65-cpsw: Fix NULL dereference on XDP_TX If number of TX queues are set to 1 we get a NULL pointer dereference during XDP_TX. ~# ethtool -L eth0 tx 1 ~# ./xdp-trafficgen udp -A -a eth0 -t 2 Transmitting on eth0 (ifindex 2) [ 241.135257] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000030 Fix this by using actual TX queues instead of max TX queues when picking the TX channel in am65_cpsw_ndo_xdp_xmit(). Fixes: 8acacc40f733 ("net: ethernet: ti: am65-cpsw: Add minimal XDP support") Signed-off-by: Roger Quadros Reviewed-by: Jacob Keller Acked-by: Julien Panis Reviewed-by: MD Danish Anwar Signed-off-by: Paolo Abeni --- drivers/net/ethernet/ti/am65-cpsw-nuss.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c index 9fd2ba26716c..03577a008df2 100644 --- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c +++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c @@ -1924,12 +1924,13 @@ static int am65_cpsw_ndo_bpf(struct net_device *ndev, struct netdev_bpf *bpf) static int am65_cpsw_ndo_xdp_xmit(struct net_device *ndev, int n, struct xdp_frame **frames, u32 flags) { + struct am65_cpsw_common *common = am65_ndev_to_common(ndev); struct am65_cpsw_tx_chn *tx_chn; struct netdev_queue *netif_txq; int cpu = smp_processor_id(); int i, nxmit = 0; - tx_chn = &am65_ndev_to_common(ndev)->tx_chns[cpu % AM65_CPSW_MAX_TX_QUEUES]; + tx_chn = &common->tx_chns[cpu % common->tx_ch_num]; netif_txq = netdev_get_tx_queue(ndev, tx_chn->id); __netif_tx_lock(netif_txq, cpu); From 624d3291484f9cada10660f820db926c0bce7741 Mon Sep 17 00:00:00 2001 From: Roger Quadros Date: Thu, 29 Aug 2024 15:03:21 +0300 Subject: [PATCH 3/3] net: ethernet: ti: am65-cpsw: Fix RX statistics for XDP_TX and XDP_REDIRECT We are not using ndev->stats for rx_packets and rx_bytes anymore. Instead, we use per CPU stats which are collated in am65_cpsw_nuss_ndo_get_stats(). Fix RX statistics for XDP_TX and XDP_REDIRECT cases. Fixes: 8acacc40f733 ("net: ethernet: ti: am65-cpsw: Add minimal XDP support") Signed-off-by: Roger Quadros Reviewed-by: Jacob Keller Acked-by: Julien Panis Reviewed-by: MD Danish Anwar Signed-off-by: Paolo Abeni --- drivers/net/ethernet/ti/am65-cpsw-nuss.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-) diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c index 03577a008df2..b06b8872b4eb 100644 --- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c +++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c @@ -998,7 +998,9 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common, int desc_idx, int cpu, int *len) { struct am65_cpsw_rx_chn *rx_chn = &common->rx_chns; + struct am65_cpsw_ndev_priv *ndev_priv; struct net_device *ndev = port->ndev; + struct am65_cpsw_ndev_stats *stats; int ret = AM65_CPSW_XDP_CONSUMED; struct am65_cpsw_tx_chn *tx_chn; struct netdev_queue *netif_txq; @@ -1016,6 +1018,9 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common, /* XDP prog might have changed packet data and boundaries */ *len = xdp->data_end - xdp->data; + ndev_priv = netdev_priv(ndev); + stats = this_cpu_ptr(ndev_priv->stats); + switch (act) { case XDP_PASS: ret = AM65_CPSW_XDP_PASS; @@ -1035,16 +1040,20 @@ static int am65_cpsw_run_xdp(struct am65_cpsw_common *common, if (err) goto drop; - ndev->stats.rx_bytes += *len; - ndev->stats.rx_packets++; + u64_stats_update_begin(&stats->syncp); + stats->rx_bytes += *len; + stats->rx_packets++; + u64_stats_update_end(&stats->syncp); ret = AM65_CPSW_XDP_CONSUMED; goto out; case XDP_REDIRECT: if (unlikely(xdp_do_redirect(ndev, xdp, prog))) goto drop; - ndev->stats.rx_bytes += *len; - ndev->stats.rx_packets++; + u64_stats_update_begin(&stats->syncp); + stats->rx_bytes += *len; + stats->rx_packets++; + u64_stats_update_end(&stats->syncp); ret = AM65_CPSW_XDP_REDIRECT; goto out; default: