Merge branch 'net-gve-header-split-support'

Ziwei Xiao says:

====================
gve: Add header split support

Currently, the ethtool's ringparam has added a new field
tcp-data-split for enabling and disabling header split. These three
patches will utilize that ethtool flag to support header split in GVE
driver.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/google/gve/gve.h

@@ -9,6 +9,7 @@
 
 #include <linux/dma-mapping.h>
 #include <linux/dmapool.h>
+#include <linux/ethtool_netlink.h>
 #include <linux/netdevice.h>
 #include <linux/pci.h>
 #include <linux/u64_stats_sync.h>
@@ -51,12 +52,16 @@
 
 #define GVE_DEFAULT_RX_BUFFER_SIZE 2048
 
+#define GVE_MAX_RX_BUFFER_SIZE 4096
+
 #define GVE_DEFAULT_RX_BUFFER_OFFSET 2048
 
 #define GVE_XDP_ACTIONS 5
 
 #define GVE_GQ_TX_MIN_PKT_DESC_BYTES 182
 
+#define GVE_DEFAULT_HEADER_BUFFER_SIZE 128
+
 #define DQO_QPL_DEFAULT_TX_PAGES 512
 #define DQO_QPL_DEFAULT_RX_PAGES 2048
 
@@ -150,6 +155,11 @@ struct gve_rx_compl_queue_dqo {
 	u32 mask; /* Mask for indices to the size of the ring */
 };
 
+struct gve_header_buf {
+	u8 *data;
+	dma_addr_t addr;
+};
+
 /* Stores state for tracking buffers posted to HW */
 struct gve_rx_buf_state_dqo {
 	/* The page posted to HW. */
@@ -252,19 +262,26 @@ struct gve_rx_ring {
 
 			/* track number of used buffers */
 			u16 used_buf_states_cnt;
+
+			/* Address info of the buffers for header-split */
+			struct gve_header_buf hdr_bufs;
 		} dqo;
 	};
 
 	u64 rbytes; /* free-running bytes received */
+	u64 rx_hsplit_bytes; /* free-running header bytes received */
 	u64 rpackets; /* free-running packets received */
 	u32 cnt; /* free-running total number of completed packets */
 	u32 fill_cnt; /* free-running total number of descs and buffs posted */
 	u32 mask; /* masks the cnt and fill_cnt to the size of the ring */
+	u64 rx_hsplit_pkt; /* free-running packets with headers split */
 	u64 rx_copybreak_pkt; /* free-running count of copybreak packets */
 	u64 rx_copied_pkt; /* free-running total number of copied packets */
 	u64 rx_skb_alloc_fail; /* free-running count of skb alloc fails */
 	u64 rx_buf_alloc_fail; /* free-running count of buffer alloc fails */
 	u64 rx_desc_err_dropped_pkt; /* free-running count of packets dropped by descriptor error */
+	/* free-running count of unsplit packets due to header buffer overflow or hdr_len is 0 */
+	u64 rx_hsplit_unsplit_pkt;
 	u64 rx_cont_packet_cnt; /* free-running multi-fragment packets received */
 	u64 rx_frag_flip_cnt; /* free-running count of rx segments where page_flip was used */
 	u64 rx_frag_copy_cnt; /* free-running count of rx segments copied */
@@ -664,6 +681,7 @@ struct gve_rx_alloc_rings_cfg {
 	struct gve_qpl_config *qpl_cfg;
 
 	u16 ring_size;
+	u16 packet_buffer_size;
 	bool raw_addressing;
 	bool enable_header_split;
 
@@ -778,13 +796,17 @@ struct gve_priv {
 	struct gve_ptype_lut *ptype_lut_dqo;
 
 	/* Must be a power of two. */
-	int data_buffer_size_dqo;
+	u16 data_buffer_size_dqo;
+	u16 max_rx_buffer_size; /* device limit */
 
 	enum gve_queue_format queue_format;
 
 	/* Interrupt coalescing settings */
 	u32 tx_coalesce_usecs;
 	u32 rx_coalesce_usecs;
+
+	u16 header_buf_size; /* device configured, header-split supported if non-zero */
+	bool header_split_enabled; /* True if the header split is enabled by the user */
 };
 
 enum gve_service_task_flags_bit {
@@ -1122,6 +1144,9 @@ void gve_rx_free_rings_gqi(struct gve_priv *priv,
 			   struct gve_rx_alloc_rings_cfg *cfg);
 void gve_rx_start_ring_gqi(struct gve_priv *priv, int idx);
 void gve_rx_stop_ring_gqi(struct gve_priv *priv, int idx);
+u16 gve_get_pkt_buf_size(const struct gve_priv *priv, bool enable_hplit);
+bool gve_header_split_supported(const struct gve_priv *priv);
+int gve_set_hsplit_config(struct gve_priv *priv, u8 tcp_data_split);
 /* Reset */
 void gve_schedule_reset(struct gve_priv *priv);
 int gve_reset(struct gve_priv *priv, bool attempt_teardown);

drivers/net/ethernet/google/gve/gve_adminq.c

@@ -40,7 +40,8 @@ void gve_parse_device_option(struct gve_priv *priv,
 			     struct gve_device_option_gqi_qpl **dev_op_gqi_qpl,
 			     struct gve_device_option_dqo_rda **dev_op_dqo_rda,
 			     struct gve_device_option_jumbo_frames **dev_op_jumbo_frames,
-			     struct gve_device_option_dqo_qpl **dev_op_dqo_qpl)
+			     struct gve_device_option_dqo_qpl **dev_op_dqo_qpl,
+			     struct gve_device_option_buffer_sizes **dev_op_buffer_sizes)
 {
 	u32 req_feat_mask = be32_to_cpu(option->required_features_mask);
 	u16 option_length = be16_to_cpu(option->option_length);
@@ -147,6 +148,23 @@ void gve_parse_device_option(struct gve_priv *priv,
 		}
 		*dev_op_jumbo_frames = (void *)(option + 1);
 		break;
+	case GVE_DEV_OPT_ID_BUFFER_SIZES:
+		if (option_length < sizeof(**dev_op_buffer_sizes) ||
+		    req_feat_mask != GVE_DEV_OPT_REQ_FEAT_MASK_BUFFER_SIZES) {
+			dev_warn(&priv->pdev->dev, GVE_DEVICE_OPTION_ERROR_FMT,
+				 "Buffer Sizes",
+				 (int)sizeof(**dev_op_buffer_sizes),
+				 GVE_DEV_OPT_REQ_FEAT_MASK_BUFFER_SIZES,
+				 option_length, req_feat_mask);
+			break;
+		}
+
+		if (option_length > sizeof(**dev_op_buffer_sizes))
+			dev_warn(&priv->pdev->dev,
+				 GVE_DEVICE_OPTION_TOO_BIG_FMT,
+				 "Buffer Sizes");
+		*dev_op_buffer_sizes = (void *)(option + 1);
+		break;
 	default:
 		/* If we don't recognize the option just continue
 		 * without doing anything.
@@ -164,7 +182,8 @@ gve_process_device_options(struct gve_priv *priv,
 			   struct gve_device_option_gqi_qpl **dev_op_gqi_qpl,
 			   struct gve_device_option_dqo_rda **dev_op_dqo_rda,
 			   struct gve_device_option_jumbo_frames **dev_op_jumbo_frames,
-			   struct gve_device_option_dqo_qpl **dev_op_dqo_qpl)
+			   struct gve_device_option_dqo_qpl **dev_op_dqo_qpl,
+			   struct gve_device_option_buffer_sizes **dev_op_buffer_sizes)
 {
 	const int num_options = be16_to_cpu(descriptor->num_device_options);
 	struct gve_device_option *dev_opt;
@@ -185,7 +204,7 @@ gve_process_device_options(struct gve_priv *priv,
 		gve_parse_device_option(priv, descriptor, dev_opt,
 					dev_op_gqi_rda, dev_op_gqi_qpl,
 					dev_op_dqo_rda, dev_op_jumbo_frames,
-					dev_op_dqo_qpl);
+					dev_op_dqo_qpl, dev_op_buffer_sizes);
 		dev_opt = next_opt;
 	}
 
@@ -640,6 +659,9 @@ static int gve_adminq_create_rx_queue(struct gve_priv *priv, u32 queue_index)
 			cpu_to_be16(rx_buff_ring_entries);
 		cmd.create_rx_queue.enable_rsc =
 			!!(priv->dev->features & NETIF_F_LRO);
+		if (priv->header_split_enabled)
+			cmd.create_rx_queue.header_buffer_size =
+				cpu_to_be16(priv->header_buf_size);
 	}
 
 	return gve_adminq_issue_cmd(priv, &cmd);
@@ -755,7 +777,9 @@ static void gve_enable_supported_features(struct gve_priv *priv,
 					  const struct gve_device_option_jumbo_frames
 					  *dev_op_jumbo_frames,
 					  const struct gve_device_option_dqo_qpl
-					  *dev_op_dqo_qpl)
+					  *dev_op_dqo_qpl,
+					  const struct gve_device_option_buffer_sizes
+					  *dev_op_buffer_sizes)
 {
 	/* Before control reaches this point, the page-size-capped max MTU from
 	 * the gve_device_descriptor field has already been stored in
@@ -779,10 +803,22 @@ static void gve_enable_supported_features(struct gve_priv *priv,
 		if (priv->rx_pages_per_qpl == 0)
 			priv->rx_pages_per_qpl = DQO_QPL_DEFAULT_RX_PAGES;
 	}
+
+	if (dev_op_buffer_sizes &&
+	    (supported_features_mask & GVE_SUP_BUFFER_SIZES_MASK)) {
+		priv->max_rx_buffer_size =
+			be16_to_cpu(dev_op_buffer_sizes->packet_buffer_size);
+		priv->header_buf_size =
+			be16_to_cpu(dev_op_buffer_sizes->header_buffer_size);
+		dev_info(&priv->pdev->dev,
+			 "BUFFER SIZES device option enabled with max_rx_buffer_size of %u, header_buf_size of %u.\n",
+			 priv->max_rx_buffer_size, priv->header_buf_size);
+	}
 }
 
 int gve_adminq_describe_device(struct gve_priv *priv)
 {
+	struct gve_device_option_buffer_sizes *dev_op_buffer_sizes = NULL;
 	struct gve_device_option_jumbo_frames *dev_op_jumbo_frames = NULL;
 	struct gve_device_option_gqi_rda *dev_op_gqi_rda = NULL;
 	struct gve_device_option_gqi_qpl *dev_op_gqi_qpl = NULL;
@@ -816,7 +852,8 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 	err = gve_process_device_options(priv, descriptor, &dev_op_gqi_rda,
 					 &dev_op_gqi_qpl, &dev_op_dqo_rda,
 					 &dev_op_jumbo_frames,
-					 &dev_op_dqo_qpl);
+					 &dev_op_dqo_qpl,
+					 &dev_op_buffer_sizes);
 	if (err)
 		goto free_device_descriptor;
 
@@ -885,7 +922,8 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 	priv->default_num_queues = be16_to_cpu(descriptor->default_num_queues);
 
 	gve_enable_supported_features(priv, supported_features_mask,
-				      dev_op_jumbo_frames, dev_op_dqo_qpl);
+				      dev_op_jumbo_frames, dev_op_dqo_qpl,
+				      dev_op_buffer_sizes);
 
 free_device_descriptor:
 	dma_pool_free(priv->adminq_pool, descriptor, descriptor_bus);

drivers/net/ethernet/google/gve/gve_adminq.h

@@ -125,6 +125,15 @@ struct gve_device_option_jumbo_frames {
 
 static_assert(sizeof(struct gve_device_option_jumbo_frames) == 8);
 
+struct gve_device_option_buffer_sizes {
+	/* GVE_SUP_BUFFER_SIZES_MASK bit should be set */
+	__be32 supported_features_mask;
+	__be16 packet_buffer_size;
+	__be16 header_buffer_size;
+};
+
+static_assert(sizeof(struct gve_device_option_buffer_sizes) == 8);
+
 /* Terminology:
  *
  * RDA - Raw DMA Addressing - Buffers associated with SKBs are directly DMA
@@ -140,6 +149,7 @@ enum gve_dev_opt_id {
 	GVE_DEV_OPT_ID_DQO_RDA = 0x4,
 	GVE_DEV_OPT_ID_DQO_QPL = 0x7,
 	GVE_DEV_OPT_ID_JUMBO_FRAMES = 0x8,
+	GVE_DEV_OPT_ID_BUFFER_SIZES = 0xa,
 };
 
 enum gve_dev_opt_req_feat_mask {
@@ -149,10 +159,12 @@ enum gve_dev_opt_req_feat_mask {
 	GVE_DEV_OPT_REQ_FEAT_MASK_DQO_RDA = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_JUMBO_FRAMES = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_DQO_QPL = 0x0,
+	GVE_DEV_OPT_REQ_FEAT_MASK_BUFFER_SIZES = 0x0,
 };
 
 enum gve_sup_feature_mask {
 	GVE_SUP_JUMBO_FRAMES_MASK = 1 << 2,
+	GVE_SUP_BUFFER_SIZES_MASK = 1 << 4,
 };
 
 #define GVE_DEV_OPT_LEN_GQI_RAW_ADDRESSING 0x0
@@ -165,6 +177,7 @@ enum gve_driver_capbility {
 	gve_driver_capability_dqo_qpl = 2, /* reserved for future use */
 	gve_driver_capability_dqo_rda = 3,
 	gve_driver_capability_alt_miss_compl = 4,
+	gve_driver_capability_flexible_buffer_size = 5,
 };
 
 #define GVE_CAP1(a) BIT((int)a)
@@ -176,7 +189,8 @@ enum gve_driver_capbility {
 	(GVE_CAP1(gve_driver_capability_gqi_qpl) | \
 	 GVE_CAP1(gve_driver_capability_gqi_rda) | \
 	 GVE_CAP1(gve_driver_capability_dqo_rda) | \
-	 GVE_CAP1(gve_driver_capability_alt_miss_compl))
+	 GVE_CAP1(gve_driver_capability_alt_miss_compl) | \
+	 GVE_CAP1(gve_driver_capability_flexible_buffer_size))
 
 #define GVE_DRIVER_CAPABILITY_FLAGS2 0x0
 #define GVE_DRIVER_CAPABILITY_FLAGS3 0x0
@@ -260,7 +274,9 @@ struct gve_adminq_create_rx_queue {
 	__be16 packet_buffer_size;
 	__be16 rx_buff_ring_size;
 	u8 enable_rsc;
-	u8 padding[5];
+	u8 padding1;
+	__be16 header_buffer_size;
+	u8 padding2[2];
 };
 
 static_assert(sizeof(struct gve_adminq_create_rx_queue) == 56);

drivers/net/ethernet/google/gve/gve_ethtool.c

@@ -4,7 +4,6 @@
  * Copyright (C) 2015-2021 Google, Inc.
  */
 
-#include <linux/ethtool.h>
 #include <linux/rtnetlink.h>
 #include "gve.h"
 #include "gve_adminq.h"
@@ -40,17 +39,18 @@ static u32 gve_get_msglevel(struct net_device *netdev)
  * as declared in enum xdp_action inside file uapi/linux/bpf.h .
  */
 static const char gve_gstrings_main_stats[][ETH_GSTRING_LEN] = {
-	"rx_packets", "tx_packets", "rx_bytes", "tx_bytes",
-	"rx_dropped", "tx_dropped", "tx_timeouts",
+	"rx_packets", "rx_hsplit_pkt", "tx_packets", "rx_bytes",
+	"tx_bytes", "rx_dropped", "tx_dropped", "tx_timeouts",
 	"rx_skb_alloc_fail", "rx_buf_alloc_fail", "rx_desc_err_dropped_pkt",
+	"rx_hsplit_unsplit_pkt",
 	"interface_up_cnt", "interface_down_cnt", "reset_cnt",
 	"page_alloc_fail", "dma_mapping_error", "stats_report_trigger_cnt",
 };
 
 static const char gve_gstrings_rx_stats[][ETH_GSTRING_LEN] = {
-	"rx_posted_desc[%u]", "rx_completed_desc[%u]", "rx_consumed_desc[%u]", "rx_bytes[%u]",
-	"rx_cont_packet_cnt[%u]", "rx_frag_flip_cnt[%u]", "rx_frag_copy_cnt[%u]",
-	"rx_frag_alloc_cnt[%u]",
+	"rx_posted_desc[%u]", "rx_completed_desc[%u]", "rx_consumed_desc[%u]",
+	"rx_bytes[%u]", "rx_hsplit_bytes[%u]", "rx_cont_packet_cnt[%u]",
+	"rx_frag_flip_cnt[%u]", "rx_frag_copy_cnt[%u]", "rx_frag_alloc_cnt[%u]",
 	"rx_dropped_pkt[%u]", "rx_copybreak_pkt[%u]", "rx_copied_pkt[%u]",
 	"rx_queue_drop_cnt[%u]", "rx_no_buffers_posted[%u]",
 	"rx_drops_packet_over_mru[%u]", "rx_drops_invalid_checksum[%u]",
@@ -154,11 +154,13 @@ static void
 gve_get_ethtool_stats(struct net_device *netdev,
 		      struct ethtool_stats *stats, u64 *data)
 {
-	u64 tmp_rx_pkts, tmp_rx_bytes, tmp_rx_skb_alloc_fail,
-		tmp_rx_buf_alloc_fail, tmp_rx_desc_err_dropped_pkt,
+	u64 tmp_rx_pkts, tmp_rx_hsplit_pkt, tmp_rx_bytes, tmp_rx_hsplit_bytes,
+		tmp_rx_skb_alloc_fail, tmp_rx_buf_alloc_fail,
+		tmp_rx_desc_err_dropped_pkt, tmp_rx_hsplit_unsplit_pkt,
 		tmp_tx_pkts, tmp_tx_bytes;
-	u64 rx_buf_alloc_fail, rx_desc_err_dropped_pkt, rx_pkts,
-		rx_skb_alloc_fail, rx_bytes, tx_pkts, tx_bytes, tx_dropped;
+	u64 rx_buf_alloc_fail, rx_desc_err_dropped_pkt, rx_hsplit_unsplit_pkt,
+		rx_pkts, rx_hsplit_pkt, rx_skb_alloc_fail, rx_bytes, tx_pkts, tx_bytes,
+		tx_dropped;
 	int stats_idx, base_stats_idx, max_stats_idx;
 	struct stats *report_stats;
 	int *rx_qid_to_stats_idx;
@@ -185,8 +187,10 @@ gve_get_ethtool_stats(struct net_device *netdev,
 		kfree(rx_qid_to_stats_idx);
 		return;
 	}
-	for (rx_pkts = 0, rx_bytes = 0, rx_skb_alloc_fail = 0,
-	     rx_buf_alloc_fail = 0, rx_desc_err_dropped_pkt = 0, ring = 0;
+	for (rx_pkts = 0, rx_bytes = 0, rx_hsplit_pkt = 0,
+	     rx_skb_alloc_fail = 0, rx_buf_alloc_fail = 0,
+	     rx_desc_err_dropped_pkt = 0, rx_hsplit_unsplit_pkt = 0,
+	     ring = 0;
 	     ring < priv->rx_cfg.num_queues; ring++) {
 		if (priv->rx) {
 			do {
@@ -195,18 +199,23 @@ gve_get_ethtool_stats(struct net_device *netdev,
 				start =
 				  u64_stats_fetch_begin(&priv->rx[ring].statss);
 				tmp_rx_pkts = rx->rpackets;
+				tmp_rx_hsplit_pkt = rx->rx_hsplit_pkt;
 				tmp_rx_bytes = rx->rbytes;
 				tmp_rx_skb_alloc_fail = rx->rx_skb_alloc_fail;
 				tmp_rx_buf_alloc_fail = rx->rx_buf_alloc_fail;
 				tmp_rx_desc_err_dropped_pkt =
 					rx->rx_desc_err_dropped_pkt;
+				tmp_rx_hsplit_unsplit_pkt =
+					rx->rx_hsplit_unsplit_pkt;
 			} while (u64_stats_fetch_retry(&priv->rx[ring].statss,
 						       start));
 			rx_pkts += tmp_rx_pkts;
+			rx_hsplit_pkt += tmp_rx_hsplit_pkt;
 			rx_bytes += tmp_rx_bytes;
 			rx_skb_alloc_fail += tmp_rx_skb_alloc_fail;
 			rx_buf_alloc_fail += tmp_rx_buf_alloc_fail;
 			rx_desc_err_dropped_pkt += tmp_rx_desc_err_dropped_pkt;
+			rx_hsplit_unsplit_pkt += tmp_rx_hsplit_unsplit_pkt;
 		}
 	}
 	for (tx_pkts = 0, tx_bytes = 0, tx_dropped = 0, ring = 0;
@@ -227,6 +236,7 @@ gve_get_ethtool_stats(struct net_device *netdev,
 
 	i = 0;
 	data[i++] = rx_pkts;
+	data[i++] = rx_hsplit_pkt;
 	data[i++] = tx_pkts;
 	data[i++] = rx_bytes;
 	data[i++] = tx_bytes;
@@ -238,6 +248,7 @@ gve_get_ethtool_stats(struct net_device *netdev,
 	data[i++] = rx_skb_alloc_fail;
 	data[i++] = rx_buf_alloc_fail;
 	data[i++] = rx_desc_err_dropped_pkt;
+	data[i++] = rx_hsplit_unsplit_pkt;
 	data[i++] = priv->interface_up_cnt;
 	data[i++] = priv->interface_down_cnt;
 	data[i++] = priv->reset_cnt;
@@ -277,6 +288,7 @@ gve_get_ethtool_stats(struct net_device *netdev,
 				start =
 				  u64_stats_fetch_begin(&priv->rx[ring].statss);
 				tmp_rx_bytes = rx->rbytes;
+				tmp_rx_hsplit_bytes = rx->rx_hsplit_bytes;
 				tmp_rx_skb_alloc_fail = rx->rx_skb_alloc_fail;
 				tmp_rx_buf_alloc_fail = rx->rx_buf_alloc_fail;
 				tmp_rx_desc_err_dropped_pkt =
@@ -284,6 +296,7 @@ gve_get_ethtool_stats(struct net_device *netdev,
 			} while (u64_stats_fetch_retry(&priv->rx[ring].statss,
 						       start));
 			data[i++] = tmp_rx_bytes;
+			data[i++] = tmp_rx_hsplit_bytes;
 			data[i++] = rx->rx_cont_packet_cnt;
 			data[i++] = rx->rx_frag_flip_cnt;
 			data[i++] = rx->rx_frag_copy_cnt;
@@ -480,6 +493,29 @@ static void gve_get_ringparam(struct net_device *netdev,
 	cmd->tx_max_pending = priv->tx_desc_cnt;
 	cmd->rx_pending = priv->rx_desc_cnt;
 	cmd->tx_pending = priv->tx_desc_cnt;
+
+	if (!gve_header_split_supported(priv))
+		kernel_cmd->tcp_data_split = ETHTOOL_TCP_DATA_SPLIT_UNKNOWN;
+	else if (priv->header_split_enabled)
+		kernel_cmd->tcp_data_split = ETHTOOL_TCP_DATA_SPLIT_ENABLED;
+	else
+		kernel_cmd->tcp_data_split = ETHTOOL_TCP_DATA_SPLIT_DISABLED;
+}
+
+static int gve_set_ringparam(struct net_device *netdev,
+			     struct ethtool_ringparam *cmd,
+			     struct kernel_ethtool_ringparam *kernel_cmd,
+			     struct netlink_ext_ack *extack)
+{
+	struct gve_priv *priv = netdev_priv(netdev);
+
+	if (priv->tx_desc_cnt != cmd->tx_pending ||
+	    priv->rx_desc_cnt != cmd->rx_pending) {
+		dev_info(&priv->pdev->dev, "Modify ring size is not supported.\n");
+		return -EOPNOTSUPP;
+	}
+
+	return gve_set_hsplit_config(priv, kernel_cmd->tcp_data_split);
 }
 
 static int gve_user_reset(struct net_device *netdev, u32 *flags)
@@ -655,6 +691,7 @@ static int gve_set_coalesce(struct net_device *netdev,
 
 const struct ethtool_ops gve_ethtool_ops = {
 	.supported_coalesce_params = ETHTOOL_COALESCE_USECS,
+	.supported_ring_params = ETHTOOL_RING_USE_TCP_DATA_SPLIT,
 	.get_drvinfo = gve_get_drvinfo,
 	.get_strings = gve_get_strings,
 	.get_sset_count = gve_get_sset_count,
@@ -667,6 +704,7 @@ const struct ethtool_ops gve_ethtool_ops = {
 	.get_coalesce = gve_get_coalesce,
 	.set_coalesce = gve_set_coalesce,
 	.get_ringparam = gve_get_ringparam,
+	.set_ringparam = gve_set_ringparam,
 	.reset = gve_user_reset,
 	.get_tunable = gve_get_tunable,
 	.set_tunable = gve_set_tunable,

drivers/net/ethernet/google/gve/gve_main.c

@@ -1307,9 +1307,13 @@ static void gve_rx_get_curr_alloc_cfg(struct gve_priv *priv,
 	cfg->qcfg = &priv->rx_cfg;
 	cfg->qcfg_tx = &priv->tx_cfg;
 	cfg->raw_addressing = !gve_is_qpl(priv);
+	cfg->enable_header_split = priv->header_split_enabled;
 	cfg->qpls = priv->qpls;
 	cfg->qpl_cfg = &priv->qpl_cfg;
 	cfg->ring_size = priv->rx_desc_cnt;
+	cfg->packet_buffer_size = gve_is_gqi(priv) ?
+				  GVE_DEFAULT_RX_BUFFER_SIZE :
+				  priv->data_buffer_size_dqo;
 	cfg->rx = priv->rx;
 }
 
@@ -1448,12 +1452,9 @@ static int gve_queues_start(struct gve_priv *priv,
 	if (err)
 		goto reset;
 
-	if (!gve_is_gqi(priv)) {
-		/* Hard code this for now. This may be tuned in the future for
-		 * performance.
-		 */
-		priv->data_buffer_size_dqo = GVE_DEFAULT_RX_BUFFER_SIZE;
-	}
+	priv->header_split_enabled = rx_alloc_cfg->enable_header_split;
+	priv->data_buffer_size_dqo = rx_alloc_cfg->packet_buffer_size;
+
 	err = gve_create_rings(priv);
 	if (err)
 		goto reset;
@@ -2065,6 +2066,56 @@ static void gve_tx_timeout(struct net_device *dev, unsigned int txqueue)
 	priv->tx_timeo_cnt++;
 }
 
+u16 gve_get_pkt_buf_size(const struct gve_priv *priv, bool enable_hsplit)
+{
+	if (enable_hsplit && priv->max_rx_buffer_size >= GVE_MAX_RX_BUFFER_SIZE)
+		return GVE_MAX_RX_BUFFER_SIZE;
+	else
+		return GVE_DEFAULT_RX_BUFFER_SIZE;
+}
+
+/* header-split is not supported on non-DQO_RDA yet even if device advertises it */
+bool gve_header_split_supported(const struct gve_priv *priv)
+{
+	return priv->header_buf_size && priv->queue_format == GVE_DQO_RDA_FORMAT;
+}
+
+int gve_set_hsplit_config(struct gve_priv *priv, u8 tcp_data_split)
+{
+	struct gve_tx_alloc_rings_cfg tx_alloc_cfg = {0};
+	struct gve_rx_alloc_rings_cfg rx_alloc_cfg = {0};
+	struct gve_qpls_alloc_cfg qpls_alloc_cfg = {0};
+	bool enable_hdr_split;
+	int err = 0;
+
+	if (tcp_data_split == ETHTOOL_TCP_DATA_SPLIT_UNKNOWN)
+		return 0;
+
+	if (!gve_header_split_supported(priv)) {
+		dev_err(&priv->pdev->dev, "Header-split not supported\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (tcp_data_split == ETHTOOL_TCP_DATA_SPLIT_ENABLED)
+		enable_hdr_split = true;
+	else
+		enable_hdr_split = false;
+
+	if (enable_hdr_split == priv->header_split_enabled)
+		return 0;
+
+	gve_get_curr_alloc_cfgs(priv, &qpls_alloc_cfg,
+				&tx_alloc_cfg, &rx_alloc_cfg);
+
+	rx_alloc_cfg.enable_header_split = enable_hdr_split;
+	rx_alloc_cfg.packet_buffer_size = gve_get_pkt_buf_size(priv, enable_hdr_split);
+
+	if (netif_running(priv->dev))
+		err = gve_adjust_config(priv, &qpls_alloc_cfg,
+					&tx_alloc_cfg, &rx_alloc_cfg);
+	return err;
+}
+
 static int gve_set_features(struct net_device *netdev,
 			    netdev_features_t features)
 {
@@ -2511,6 +2562,8 @@ static int gve_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 	priv->service_task_flags = 0x0;
 	priv->state_flags = 0x0;
 	priv->ethtool_flags = 0x0;
+	priv->data_buffer_size_dqo = GVE_DEFAULT_RX_BUFFER_SIZE;
+	priv->max_rx_buffer_size = GVE_DEFAULT_RX_BUFFER_SIZE;
 
 	gve_set_probe_in_progress(priv);
 	priv->gve_wq = alloc_ordered_workqueue("gve", 0);

drivers/net/ethernet/google/gve/gve_rx_dqo.c

@@ -199,6 +199,18 @@ static int gve_alloc_page_dqo(struct gve_rx_ring *rx,
 	return 0;
 }
 
+static void gve_rx_free_hdr_bufs(struct gve_priv *priv, struct gve_rx_ring *rx)
+{
+	struct device *hdev = &priv->pdev->dev;
+	int buf_count = rx->dqo.bufq.mask + 1;
+
+	if (rx->dqo.hdr_bufs.data) {
+		dma_free_coherent(hdev, priv->header_buf_size * buf_count,
+				  rx->dqo.hdr_bufs.data, rx->dqo.hdr_bufs.addr);
+		rx->dqo.hdr_bufs.data = NULL;
+	}
+}
+
 void gve_rx_stop_ring_dqo(struct gve_priv *priv, int idx)
 {
 	int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
@@ -258,9 +270,24 @@ static void gve_rx_free_ring_dqo(struct gve_priv *priv, struct gve_rx_ring *rx,
 	kvfree(rx->dqo.buf_states);
 	rx->dqo.buf_states = NULL;
 
+	gve_rx_free_hdr_bufs(priv, rx);
+
 	netif_dbg(priv, drv, priv->dev, "freed rx ring %d\n", idx);
 }
 
+static int gve_rx_alloc_hdr_bufs(struct gve_priv *priv, struct gve_rx_ring *rx)
+{
+	struct device *hdev = &priv->pdev->dev;
+	int buf_count = rx->dqo.bufq.mask + 1;
+
+	rx->dqo.hdr_bufs.data = dma_alloc_coherent(hdev, priv->header_buf_size * buf_count,
+						   &rx->dqo.hdr_bufs.addr, GFP_KERNEL);
+	if (!rx->dqo.hdr_bufs.data)
+		return -ENOMEM;
+
+	return 0;
+}
+
 void gve_rx_start_ring_dqo(struct gve_priv *priv, int idx)
 {
 	int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
@@ -302,6 +329,11 @@ static int gve_rx_alloc_ring_dqo(struct gve_priv *priv,
 	if (!rx->dqo.buf_states)
 		return -ENOMEM;
 
+	/* Allocate header buffers for header-split */
+	if (cfg->enable_header_split)
+		if (gve_rx_alloc_hdr_bufs(priv, rx))
+			goto err;
+
 	/* Set up linked list of buffer IDs */
 	for (i = 0; i < rx->dqo.num_buf_states - 1; i++)
 		rx->dqo.buf_states[i].next = i + 1;
@@ -443,6 +475,10 @@ void gve_rx_post_buffers_dqo(struct gve_rx_ring *rx)
 		desc->buf_id = cpu_to_le16(buf_state - rx->dqo.buf_states);
 		desc->buf_addr = cpu_to_le64(buf_state->addr +
 					     buf_state->page_info.page_offset);
+		if (rx->dqo.hdr_bufs.data)
+			desc->header_buf_addr =
+				cpu_to_le64(rx->dqo.hdr_bufs.addr +
+					    priv->header_buf_size * bufq->tail);
 
 		bufq->tail = (bufq->tail + 1) & bufq->mask;
 		complq->num_free_slots--;
@@ -458,7 +494,7 @@ void gve_rx_post_buffers_dqo(struct gve_rx_ring *rx)
 static void gve_try_recycle_buf(struct gve_priv *priv, struct gve_rx_ring *rx,
 				struct gve_rx_buf_state_dqo *buf_state)
 {
-	const int data_buffer_size = priv->data_buffer_size_dqo;
+	const u16 data_buffer_size = priv->data_buffer_size_dqo;
 	int pagecount;
 
 	/* Can't reuse if we only fit one buffer per page */
@@ -645,13 +681,16 @@ static int gve_rx_append_frags(struct napi_struct *napi,
  */
 static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 		      const struct gve_rx_compl_desc_dqo *compl_desc,
-		      int queue_idx)
+		      u32 desc_idx, int queue_idx)
 {
 	const u16 buffer_id = le16_to_cpu(compl_desc->buf_id);
+	const bool hbo = compl_desc->header_buffer_overflow;
 	const bool eop = compl_desc->end_of_packet != 0;
+	const bool hsplit = compl_desc->split_header;
 	struct gve_rx_buf_state_dqo *buf_state;
 	struct gve_priv *priv = rx->gve;
 	u16 buf_len;
+	u16 hdr_len;
 
 	if (unlikely(buffer_id >= rx->dqo.num_buf_states)) {
 		net_err_ratelimited("%s: Invalid RX buffer_id=%u\n",
@@ -672,12 +711,35 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 	}
 
 	buf_len = compl_desc->packet_len;
+	hdr_len = compl_desc->header_len;
 
 	/* Page might have not been used for awhile and was likely last written
 	 * by a different thread.
 	 */
 	prefetch(buf_state->page_info.page);
 
+	/* Copy the header into the skb in the case of header split */
+	if (hsplit) {
+		int unsplit = 0;
+
+		if (hdr_len && !hbo) {
+			rx->ctx.skb_head = gve_rx_copy_data(priv->dev, napi,
+							    rx->dqo.hdr_bufs.data +
+							    desc_idx * priv->header_buf_size,
+							    hdr_len);
+			if (unlikely(!rx->ctx.skb_head))
+				goto error;
+			rx->ctx.skb_tail = rx->ctx.skb_head;
+		} else {
+			unsplit = 1;
+		}
+		u64_stats_update_begin(&rx->statss);
+		rx->rx_hsplit_pkt++;
+		rx->rx_hsplit_unsplit_pkt += unsplit;
+		rx->rx_hsplit_bytes += hdr_len;
+		u64_stats_update_end(&rx->statss);
+	}
+
 	/* Sync the portion of dma buffer for CPU to read. */
 	dma_sync_single_range_for_cpu(&priv->pdev->dev, buf_state->addr,
 				      buf_state->page_info.page_offset,
@@ -820,7 +882,7 @@ int gve_rx_poll_dqo(struct gve_notify_block *block, int budget)
 		/* Do not read data until we own the descriptor */
 		dma_rmb();
 
-		err = gve_rx_dqo(napi, rx, compl_desc, rx->q_num);
+		err = gve_rx_dqo(napi, rx, compl_desc, complq->head, rx->q_num);
 		if (err < 0) {
 			gve_rx_free_skb(rx);
 			u64_stats_update_begin(&rx->statss);

drivers/net/ethernet/google/gve/gve_utils.c

@@ -64,11 +64,9 @@ void gve_rx_add_to_block(struct gve_priv *priv, int queue_idx)
 	rx->ntfy_id = ntfy_idx;
 }
 
-struct sk_buff *gve_rx_copy(struct net_device *dev, struct napi_struct *napi,
-			    struct gve_rx_slot_page_info *page_info, u16 len)
+struct sk_buff *gve_rx_copy_data(struct net_device *dev, struct napi_struct *napi,
+				 u8 *data, u16 len)
 {
-	void *va = page_info->page_address + page_info->page_offset +
-		page_info->pad;
 	struct sk_buff *skb;
 
 	skb = napi_alloc_skb(napi, len);
@@ -76,12 +74,21 @@ struct sk_buff *gve_rx_copy(struct net_device *dev, struct napi_struct *napi,
 		return NULL;
 
 	__skb_put(skb, len);
-	skb_copy_to_linear_data_offset(skb, 0, va, len);
+	skb_copy_to_linear_data_offset(skb, 0, data, len);
 	skb->protocol = eth_type_trans(skb, dev);
 
 	return skb;
 }
 
+struct sk_buff *gve_rx_copy(struct net_device *dev, struct napi_struct *napi,
+			    struct gve_rx_slot_page_info *page_info, u16 len)
+{
+	void *va = page_info->page_address + page_info->page_offset +
+		page_info->pad;
+
+	return gve_rx_copy_data(dev, napi, va, len);
+}
+
 void gve_dec_pagecnt_bias(struct gve_rx_slot_page_info *page_info)
 {
 	page_info->pagecnt_bias--;

drivers/net/ethernet/google/gve/gve_utils.h

@@ -19,6 +19,9 @@ bool gve_rx_was_added_to_block(struct gve_priv *priv, int queue_idx);
 void gve_rx_remove_from_block(struct gve_priv *priv, int queue_idx);
 void gve_rx_add_to_block(struct gve_priv *priv, int queue_idx);
 
+struct sk_buff *gve_rx_copy_data(struct net_device *dev, struct napi_struct *napi,
+				 u8 *data, u16 len);
+
 struct sk_buff *gve_rx_copy(struct net_device *dev, struct napi_struct *napi,
 			    struct gve_rx_slot_page_info *page_info, u16 len);