Merge branch 'gve-jumbo-frame'

Jeroen de Borst says:

====================
gve: Add jumbo-frame support for GQ

This patchset introduces jumbo-frame support for the GQ queue format.
The device already supports jumbo-frames on TX. This introduces
multi-descriptor RX packets using a packet continuation bit.

A widely deployed driver has a bug with causes it to fail to load
when a MTU greater than 2048 bytes is configured. A jumbo-frame device
option is introduced to pass a jumbo-frame MTU only to drivers that
support it.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

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

@@ -142,6 +142,19 @@ struct gve_index_list {
 	s16 tail;
 };
 
+/* A single received packet split across multiple buffers may be
+ * reconstructed using the information in this structure.
+ */
+struct gve_rx_ctx {
+	/* head and tail of skb chain for the current packet or NULL if none */
+	struct sk_buff *skb_head;
+	struct sk_buff *skb_tail;
+	u16 total_expected_size;
+	u8 expected_frag_cnt;
+	u8 curr_frag_cnt;
+	u8 reuse_frags;
+};
+
 /* Contains datapath state used to represent an RX queue. */
 struct gve_rx_ring {
 	struct gve_priv *gve;
@@ -153,6 +166,7 @@ struct gve_rx_ring {
 
 			/* threshold for posting new buffs and descs */
 			u32 db_threshold;
+			u16 packet_buffer_size;
 		};
 
 		/* DQO fields. */
@@ -200,15 +214,16 @@ struct gve_rx_ring {
 	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 */
+	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 into skb linear portion */
 	u32 q_num; /* queue index */
 	u32 ntfy_id; /* notification block index */
 	struct gve_queue_resources *q_resources; /* head and tail pointer idx */
 	dma_addr_t q_resources_bus; /* dma address for the queue resources */
 	struct u64_stats_sync statss; /* sync stats for 32bit archs */
 
-	/* head and tail of skb chain for the current packet or NULL if none */
-	struct sk_buff *skb_head;
-	struct sk_buff *skb_tail;
+	struct gve_rx_ctx ctx; /* Info for packet currently being processed in this ring. */
 };
 
 /* A TX desc ring entry */

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

@@ -38,7 +38,8 @@ void gve_parse_device_option(struct gve_priv *priv,
 			     struct gve_device_option *option,
 			     struct gve_device_option_gqi_rda **dev_op_gqi_rda,
 			     struct gve_device_option_gqi_qpl **dev_op_gqi_qpl,
-			     struct gve_device_option_dqo_rda **dev_op_dqo_rda)
+			     struct gve_device_option_dqo_rda **dev_op_dqo_rda,
+			     struct gve_device_option_jumbo_frames **dev_op_jumbo_frames)
 {
 	u32 req_feat_mask = be32_to_cpu(option->required_features_mask);
 	u16 option_length = be16_to_cpu(option->option_length);
@@ -111,6 +112,24 @@ void gve_parse_device_option(struct gve_priv *priv,
 		}
 		*dev_op_dqo_rda = (void *)(option + 1);
 		break;
+	case GVE_DEV_OPT_ID_JUMBO_FRAMES:
+		if (option_length < sizeof(**dev_op_jumbo_frames) ||
+		    req_feat_mask != GVE_DEV_OPT_REQ_FEAT_MASK_JUMBO_FRAMES) {
+			dev_warn(&priv->pdev->dev, GVE_DEVICE_OPTION_ERROR_FMT,
+				 "Jumbo Frames",
+				 (int)sizeof(**dev_op_jumbo_frames),
+				 GVE_DEV_OPT_REQ_FEAT_MASK_JUMBO_FRAMES,
+				 option_length, req_feat_mask);
+			break;
+		}
+
+		if (option_length > sizeof(**dev_op_jumbo_frames)) {
+			dev_warn(&priv->pdev->dev,
+				 GVE_DEVICE_OPTION_TOO_BIG_FMT,
+				 "Jumbo Frames");
+		}
+		*dev_op_jumbo_frames = (void *)(option + 1);
+		break;
 	default:
 		/* If we don't recognize the option just continue
 		 * without doing anything.
@@ -126,7 +145,8 @@ gve_process_device_options(struct gve_priv *priv,
 			   struct gve_device_descriptor *descriptor,
 			   struct gve_device_option_gqi_rda **dev_op_gqi_rda,
 			   struct gve_device_option_gqi_qpl **dev_op_gqi_qpl,
-			   struct gve_device_option_dqo_rda **dev_op_dqo_rda)
+			   struct gve_device_option_dqo_rda **dev_op_dqo_rda,
+			   struct gve_device_option_jumbo_frames **dev_op_jumbo_frames)
 {
 	const int num_options = be16_to_cpu(descriptor->num_device_options);
 	struct gve_device_option *dev_opt;
@@ -146,7 +166,7 @@ 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_dqo_rda, dev_op_jumbo_frames);
 		dev_opt = next_opt;
 	}
 
@@ -530,6 +550,7 @@ static int gve_adminq_create_rx_queue(struct gve_priv *priv, u32 queue_index)
 			cpu_to_be64(rx->data.data_bus),
 		cmd.create_rx_queue.index = cpu_to_be32(queue_index);
 		cmd.create_rx_queue.queue_page_list_id = cpu_to_be32(qpl_id);
+		cmd.create_rx_queue.packet_buffer_size = cpu_to_be16(rx->packet_buffer_size);
 	} else {
 		cmd.create_rx_queue.rx_ring_size =
 			cpu_to_be16(priv->rx_desc_cnt);
@@ -660,12 +681,31 @@ gve_set_desc_cnt_dqo(struct gve_priv *priv,
 	return 0;
 }
 
+static void gve_enable_supported_features(struct gve_priv *priv,
+					  u32 supported_features_mask,
+					  const struct gve_device_option_jumbo_frames
+						  *dev_op_jumbo_frames)
+{
+	/* Before control reaches this point, the page-size-capped max MTU from
+	 * the gve_device_descriptor field has already been stored in
+	 * priv->dev->max_mtu. We overwrite it with the true max MTU below.
+	 */
+	if (dev_op_jumbo_frames &&
+	    (supported_features_mask & GVE_SUP_JUMBO_FRAMES_MASK)) {
+		dev_info(&priv->pdev->dev,
+			 "JUMBO FRAMES device option enabled.\n");
+		priv->dev->max_mtu = be16_to_cpu(dev_op_jumbo_frames->max_mtu);
+	}
+}
+
 int gve_adminq_describe_device(struct gve_priv *priv)
 {
+	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;
 	struct gve_device_option_dqo_rda *dev_op_dqo_rda = NULL;
 	struct gve_device_descriptor *descriptor;
+	u32 supported_features_mask = 0;
 	union gve_adminq_command cmd;
 	dma_addr_t descriptor_bus;
 	int err = 0;
@@ -689,7 +729,8 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 		goto free_device_descriptor;
 
 	err = gve_process_device_options(priv, descriptor, &dev_op_gqi_rda,
-					 &dev_op_gqi_qpl, &dev_op_dqo_rda);
+					 &dev_op_gqi_qpl, &dev_op_dqo_rda,
+					 &dev_op_jumbo_frames);
 	if (err)
 		goto free_device_descriptor;
 
@@ -704,12 +745,19 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 		priv->queue_format = GVE_DQO_RDA_FORMAT;
 		dev_info(&priv->pdev->dev,
 			 "Driver is running with DQO RDA queue format.\n");
+		supported_features_mask =
+			be32_to_cpu(dev_op_dqo_rda->supported_features_mask);
 	} else if (dev_op_gqi_rda) {
 		priv->queue_format = GVE_GQI_RDA_FORMAT;
 		dev_info(&priv->pdev->dev,
 			 "Driver is running with GQI RDA queue format.\n");
+		supported_features_mask =
+			be32_to_cpu(dev_op_gqi_rda->supported_features_mask);
 	} else {
 		priv->queue_format = GVE_GQI_QPL_FORMAT;
+		if (dev_op_gqi_qpl)
+			supported_features_mask =
+				be32_to_cpu(dev_op_gqi_qpl->supported_features_mask);
 		dev_info(&priv->pdev->dev,
 			 "Driver is running with GQI QPL queue format.\n");
 	}
@@ -746,6 +794,9 @@ 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);
+
 free_device_descriptor:
 	dma_free_coherent(&priv->pdev->dev, PAGE_SIZE, descriptor,
 			  descriptor_bus);

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

@@ -108,6 +108,14 @@ struct gve_device_option_dqo_rda {
 
 static_assert(sizeof(struct gve_device_option_dqo_rda) == 8);
 
+struct gve_device_option_jumbo_frames {
+	__be32 supported_features_mask;
+	__be16 max_mtu;
+	u8 padding[2];
+};
+
+static_assert(sizeof(struct gve_device_option_jumbo_frames) == 8);
+
 /* Terminology:
  *
  * RDA - Raw DMA Addressing - Buffers associated with SKBs are directly DMA
@@ -121,6 +129,7 @@ enum gve_dev_opt_id {
 	GVE_DEV_OPT_ID_GQI_RDA = 0x2,
 	GVE_DEV_OPT_ID_GQI_QPL = 0x3,
 	GVE_DEV_OPT_ID_DQO_RDA = 0x4,
+	GVE_DEV_OPT_ID_JUMBO_FRAMES = 0x8,
 };
 
 enum gve_dev_opt_req_feat_mask {
@@ -128,6 +137,11 @@ enum gve_dev_opt_req_feat_mask {
 	GVE_DEV_OPT_REQ_FEAT_MASK_GQI_RDA = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_GQI_QPL = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_DQO_RDA = 0x0,
+	GVE_DEV_OPT_REQ_FEAT_MASK_JUMBO_FRAMES = 0x0,
+};
+
+enum gve_sup_feature_mask {
+	GVE_SUP_JUMBO_FRAMES_MASK = 1 << 2,
 };
 
 #define GVE_DEV_OPT_LEN_GQI_RAW_ADDRESSING 0x0

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

@@ -90,12 +90,13 @@ union gve_rx_data_slot {
 
 /* GVE Recive Packet Descriptor Flags */
 #define GVE_RXFLG(x)	cpu_to_be16(1 << (3 + (x)))
-#define	GVE_RXF_FRAG	GVE_RXFLG(3)	/* IP Fragment			*/
-#define	GVE_RXF_IPV4	GVE_RXFLG(4)	/* IPv4				*/
-#define	GVE_RXF_IPV6	GVE_RXFLG(5)	/* IPv6				*/
-#define	GVE_RXF_TCP	GVE_RXFLG(6)	/* TCP Packet			*/
-#define	GVE_RXF_UDP	GVE_RXFLG(7)	/* UDP Packet			*/
-#define	GVE_RXF_ERR	GVE_RXFLG(8)	/* Packet Error Detected	*/
+#define	GVE_RXF_FRAG		GVE_RXFLG(3)	/* IP Fragment			*/
+#define	GVE_RXF_IPV4		GVE_RXFLG(4)	/* IPv4				*/
+#define	GVE_RXF_IPV6		GVE_RXFLG(5)	/* IPv6				*/
+#define	GVE_RXF_TCP		GVE_RXFLG(6)	/* TCP Packet			*/
+#define	GVE_RXF_UDP		GVE_RXFLG(7)	/* UDP Packet			*/
+#define	GVE_RXF_ERR		GVE_RXFLG(8)	/* Packet Error Detected	*/
+#define	GVE_RXF_PKT_CONT	GVE_RXFLG(10)	/* Multi Fragment RX packet	*/
 
 /* GVE IRQ */
 #define GVE_IRQ_ACK	BIT(31)

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

@@ -43,6 +43,7 @@ static const char gve_gstrings_main_stats[][ETH_GSTRING_LEN] = {
 
 static const char gve_gstrings_rx_stats[][ETH_GSTRING_LEN] = {
 	"rx_posted_desc[%u]", "rx_completed_desc[%u]", "rx_bytes[%u]",
+	"rx_cont_packet_cnt[%u]", "rx_frag_flip_cnt[%u]", "rx_frag_copy_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]",
@@ -265,6 +266,9 @@ 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++] = rx->rx_cont_packet_cnt;
+			data[i++] = rx->rx_frag_flip_cnt;
+			data[i++] = rx->rx_frag_copy_cnt;
 			/* rx dropped packets */
 			data[i++] = tmp_rx_skb_alloc_fail +
 				tmp_rx_buf_alloc_fail +

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

@@ -1371,14 +1371,6 @@ static int gve_init_priv(struct gve_priv *priv, bool skip_describe_device)
 			"Could not get device information: err=%d\n", err);
 		goto err;
 	}
-	if (gve_is_gqi(priv) && priv->dev->max_mtu > PAGE_SIZE) {
-		priv->dev->max_mtu = PAGE_SIZE;
-		err = gve_adminq_set_mtu(priv, priv->dev->mtu);
-		if (err) {
-			dev_err(&priv->pdev->dev, "Could not set mtu");
-			goto err;
-		}
-	}
 	priv->dev->mtu = priv->dev->max_mtu;
 	num_ntfy = pci_msix_vec_count(priv->pdev);
 	if (num_ntfy <= 0) {

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

@@ -143,6 +143,16 @@ static int gve_prefill_rx_pages(struct gve_rx_ring *rx)
 	return err;
 }
 
+static void gve_rx_ctx_clear(struct gve_rx_ctx *ctx)
+{
+	ctx->curr_frag_cnt = 0;
+	ctx->total_expected_size = 0;
+	ctx->expected_frag_cnt = 0;
+	ctx->skb_head = NULL;
+	ctx->skb_tail = NULL;
+	ctx->reuse_frags = false;
+}
+
 static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 {
 	struct gve_rx_ring *rx = &priv->rx[idx];
@@ -209,6 +219,12 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 	rx->cnt = 0;
 	rx->db_threshold = priv->rx_desc_cnt / 2;
 	rx->desc.seqno = 1;
+
+	/* Allocating half-page buffers allows page-flipping which is faster
+	 * than copying or allocating new pages.
+	 */
+	rx->packet_buffer_size = PAGE_SIZE / 2;
+	gve_rx_ctx_clear(&rx->ctx);
 	gve_rx_add_to_block(priv, idx);
 
 	return 0;
@@ -275,18 +291,28 @@ static enum pkt_hash_types gve_rss_type(__be16 pkt_flags)
 	return PKT_HASH_TYPE_L2;
 }
 
+static u16 gve_rx_ctx_padding(struct gve_rx_ctx *ctx)
+{
+	return (ctx->curr_frag_cnt == 0) ? GVE_RX_PAD : 0;
+}
+
 static struct sk_buff *gve_rx_add_frags(struct napi_struct *napi,
 					struct gve_rx_slot_page_info *page_info,
-					u16 len)
+					u16 packet_buffer_size, u16 len,
+					struct gve_rx_ctx *ctx)
 {
-	struct sk_buff *skb = napi_get_frags(napi);
+	u32 offset = page_info->page_offset +  gve_rx_ctx_padding(ctx);
+	struct sk_buff *skb;
+
+	if (!ctx->skb_head)
+		ctx->skb_head = napi_get_frags(napi);
 
-	if (unlikely(!skb))
+	if (unlikely(!ctx->skb_head))
 		return NULL;
 
-	skb_add_rx_frag(skb, 0, page_info->page,
-			page_info->page_offset +
-			GVE_RX_PAD, len, PAGE_SIZE / 2);
+	skb = ctx->skb_head;
+	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page_info->page,
+			offset, len, packet_buffer_size);
 
 	return skb;
 }
@@ -300,12 +326,6 @@ static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info, __be64 *sl
 	*(slot_addr) ^= offset;
 }
 
-static bool gve_rx_can_flip_buffers(struct net_device *netdev)
-{
-	return PAGE_SIZE >= 4096
-		? netdev->mtu + GVE_RX_PAD + ETH_HLEN <= PAGE_SIZE / 2 : false;
-}
-
 static int gve_rx_can_recycle_buffer(struct gve_rx_slot_page_info *page_info)
 {
 	int pagecount = page_count(page_info->page);
@@ -325,11 +345,11 @@ static struct sk_buff *
 gve_rx_raw_addressing(struct device *dev, struct net_device *netdev,
 		      struct gve_rx_slot_page_info *page_info, u16 len,
 		      struct napi_struct *napi,
-		      union gve_rx_data_slot *data_slot)
+		      union gve_rx_data_slot *data_slot,
+		      u16 packet_buffer_size, struct gve_rx_ctx *ctx)
 {
-	struct sk_buff *skb;
+	struct sk_buff *skb = gve_rx_add_frags(napi, page_info, packet_buffer_size, len, ctx);
 
-	skb = gve_rx_add_frags(napi, page_info, len);
 	if (!skb)
 		return NULL;
 
@@ -348,6 +368,7 @@ gve_rx_qpl(struct device *dev, struct net_device *netdev,
 	   u16 len, struct napi_struct *napi,
 	   union gve_rx_data_slot *data_slot)
 {
+	struct gve_rx_ctx *ctx = &rx->ctx;
 	struct sk_buff *skb;
 
 	/* if raw_addressing mode is not enabled gvnic can only receive into
@@ -355,8 +376,8 @@ gve_rx_qpl(struct device *dev, struct net_device *netdev,
 	 * choice is to copy the data out of it so that we can return it to the
 	 * device.
 	 */
-	if (page_info->can_flip) {
-		skb = gve_rx_add_frags(napi, page_info, len);
+	if (ctx->reuse_frags) {
+		skb = gve_rx_add_frags(napi, page_info, rx->packet_buffer_size, len, ctx);
 		/* No point in recycling if we didn't get the skb */
 		if (skb) {
 			/* Make sure that the page isn't freed. */
@@ -364,114 +385,247 @@ gve_rx_qpl(struct device *dev, struct net_device *netdev,
 			gve_rx_flip_buff(page_info, &data_slot->qpl_offset);
 		}
 	} else {
-		skb = gve_rx_copy(netdev, napi, page_info, len, GVE_RX_PAD);
+		const u16 padding = gve_rx_ctx_padding(ctx);
+
+		skb = gve_rx_copy(netdev, napi, page_info, len, padding, ctx);
 		if (skb) {
 			u64_stats_update_begin(&rx->statss);
-			rx->rx_copied_pkt++;
+			rx->rx_frag_copy_cnt++;
 			u64_stats_update_end(&rx->statss);
 		}
 	}
 	return skb;
 }
 
-static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
-		   netdev_features_t feat, u32 idx)
+#define GVE_PKTCONT_BIT_IS_SET(x) (GVE_RXF_PKT_CONT & (x))
+static u16 gve_rx_get_fragment_size(struct gve_rx_ctx *ctx, struct gve_rx_desc *desc)
 {
+	return be16_to_cpu(desc->len) - gve_rx_ctx_padding(ctx);
+}
+
+static bool gve_rx_ctx_init(struct gve_rx_ctx *ctx, struct gve_rx_ring *rx)
+{
+	bool qpl_mode = !rx->data.raw_addressing, packet_size_error = false;
+	bool buffer_error = false, desc_error = false, seqno_error = false;
 	struct gve_rx_slot_page_info *page_info;
 	struct gve_priv *priv = rx->gve;
-	struct napi_struct *napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
-	struct net_device *dev = priv->dev;
-	union gve_rx_data_slot *data_slot;
-	struct sk_buff *skb = NULL;
-	dma_addr_t page_bus;
-	void *va;
-	u16 len;
+	u32 idx = rx->cnt & rx->mask;
+	bool reuse_frags, can_flip;
+	struct gve_rx_desc *desc;
+	u16 packet_size = 0;
+	u16 n_frags = 0;
+	int recycle;
 
-	/* Prefetch two packet pages ahead, we will need it soon. */
-	page_info = &rx->data.page_info[(idx + 2) & rx->mask];
-	va = page_info->page_address + GVE_RX_PAD +
-		page_info->page_offset;
+	/** In QPL mode, we only flip buffers when all buffers containing the packet
+	 * can be flipped. RDA can_flip decisions will be made later, per frag.
+	 */
+	can_flip = qpl_mode;
+	reuse_frags = can_flip;
+	do {
+		u16 frag_size;
+
+		n_frags++;
+		desc = &rx->desc.desc_ring[idx];
+		desc_error = unlikely(desc->flags_seq & GVE_RXF_ERR) || desc_error;
+		if (GVE_SEQNO(desc->flags_seq) != rx->desc.seqno) {
+			seqno_error = true;
+			netdev_warn(priv->dev,
+				    "RX seqno error: want=%d, got=%d, dropping packet and scheduling reset.",
+				    rx->desc.seqno, GVE_SEQNO(desc->flags_seq));
+		}
+		frag_size = be16_to_cpu(desc->len);
+		packet_size += frag_size;
+		if (frag_size > rx->packet_buffer_size) {
+			packet_size_error = true;
+			netdev_warn(priv->dev,
+				    "RX fragment error: packet_buffer_size=%d, frag_size=%d, droping packet.",
+				    rx->packet_buffer_size, be16_to_cpu(desc->len));
+		}
+		page_info = &rx->data.page_info[idx];
+		if (can_flip) {
+			recycle = gve_rx_can_recycle_buffer(page_info);
+			reuse_frags = reuse_frags && recycle > 0;
+			buffer_error = buffer_error || unlikely(recycle < 0);
+		}
+		idx = (idx + 1) & rx->mask;
+		rx->desc.seqno = gve_next_seqno(rx->desc.seqno);
+	} while (GVE_PKTCONT_BIT_IS_SET(desc->flags_seq));
 
-	prefetch(page_info->page); /* Kernel page struct. */
-	prefetch(va);              /* Packet header. */
-	prefetch(va + 64);         /* Next cacheline too. */
+	prefetch(rx->desc.desc_ring + idx);
 
-	/* drop this packet */
-	if (unlikely(rx_desc->flags_seq & GVE_RXF_ERR)) {
+	ctx->curr_frag_cnt = 0;
+	ctx->total_expected_size = packet_size - GVE_RX_PAD;
+	ctx->expected_frag_cnt = n_frags;
+	ctx->skb_head = NULL;
+	ctx->reuse_frags = reuse_frags;
+
+	if (ctx->expected_frag_cnt > 1) {
 		u64_stats_update_begin(&rx->statss);
-		rx->rx_desc_err_dropped_pkt++;
+		rx->rx_cont_packet_cnt++;
+		u64_stats_update_end(&rx->statss);
+	}
+	if (ctx->total_expected_size > priv->rx_copybreak && !ctx->reuse_frags && qpl_mode) {
+		u64_stats_update_begin(&rx->statss);
+		rx->rx_copied_pkt++;
 		u64_stats_update_end(&rx->statss);
+	}
+
+	if (unlikely(buffer_error || seqno_error || packet_size_error)) {
+		gve_schedule_reset(priv);
 		return false;
 	}
 
-	len = be16_to_cpu(rx_desc->len) - GVE_RX_PAD;
-	page_info = &rx->data.page_info[idx];
+	if (unlikely(desc_error)) {
+		u64_stats_update_begin(&rx->statss);
+		rx->rx_desc_err_dropped_pkt++;
+		u64_stats_update_end(&rx->statss);
+		return false;
+	}
+	return true;
+}
 
-	data_slot = &rx->data.data_ring[idx];
-	page_bus = (rx->data.raw_addressing) ?
-			be64_to_cpu(data_slot->addr) & GVE_DATA_SLOT_ADDR_PAGE_MASK :
-			rx->data.qpl->page_buses[idx];
-	dma_sync_single_for_cpu(&priv->pdev->dev, page_bus,
-				PAGE_SIZE, DMA_FROM_DEVICE);
+static struct sk_buff *gve_rx_skb(struct gve_priv *priv, struct gve_rx_ring *rx,
+				  struct gve_rx_slot_page_info *page_info, struct napi_struct *napi,
+				  u16 len, union gve_rx_data_slot *data_slot)
+{
+	struct net_device *netdev = priv->dev;
+	struct gve_rx_ctx *ctx = &rx->ctx;
+	struct sk_buff *skb = NULL;
 
-	if (len <= priv->rx_copybreak) {
+	if (len <= priv->rx_copybreak && ctx->expected_frag_cnt == 1) {
 		/* Just copy small packets */
-		skb = gve_rx_copy(dev, napi, page_info, len, GVE_RX_PAD);
-		u64_stats_update_begin(&rx->statss);
-		rx->rx_copied_pkt++;
-		rx->rx_copybreak_pkt++;
-		u64_stats_update_end(&rx->statss);
+		skb = gve_rx_copy(netdev, napi, page_info, len, GVE_RX_PAD, ctx);
+		if (skb) {
+			u64_stats_update_begin(&rx->statss);
+			rx->rx_copied_pkt++;
+			rx->rx_frag_copy_cnt++;
+			rx->rx_copybreak_pkt++;
+		}	u64_stats_update_end(&rx->statss);
 	} else {
-		u8 can_flip = gve_rx_can_flip_buffers(dev);
-		int recycle = 0;
+		if (rx->data.raw_addressing) {
+			int recycle = gve_rx_can_recycle_buffer(page_info);
 
-		if (can_flip) {
-			recycle = gve_rx_can_recycle_buffer(page_info);
-			if (recycle < 0) {
-				if (!rx->data.raw_addressing)
-					gve_schedule_reset(priv);
-				return false;
+			if (unlikely(recycle < 0)) {
+				gve_schedule_reset(priv);
+				return NULL;
 			}
-		}
-
-		page_info->can_flip = can_flip && recycle;
-		if (rx->data.raw_addressing) {
-			skb = gve_rx_raw_addressing(&priv->pdev->dev, dev,
+			page_info->can_flip = recycle;
+			if (page_info->can_flip) {
+				u64_stats_update_begin(&rx->statss);
+				rx->rx_frag_flip_cnt++;
+				u64_stats_update_end(&rx->statss);
+			}
+			skb = gve_rx_raw_addressing(&priv->pdev->dev, netdev,
 						    page_info, len, napi,
-						    data_slot);
+						    data_slot,
+						    rx->packet_buffer_size, ctx);
 		} else {
-			skb = gve_rx_qpl(&priv->pdev->dev, dev, rx,
+			if (ctx->reuse_frags) {
+				u64_stats_update_begin(&rx->statss);
+				rx->rx_frag_flip_cnt++;
+				u64_stats_update_end(&rx->statss);
+			}
+			skb = gve_rx_qpl(&priv->pdev->dev, netdev, rx,
 					 page_info, len, napi, data_slot);
 		}
 	}
+	return skb;
+}
 
-	if (!skb) {
-		u64_stats_update_begin(&rx->statss);
-		rx->rx_skb_alloc_fail++;
-		u64_stats_update_end(&rx->statss);
-		return false;
+static bool gve_rx(struct gve_rx_ring *rx, netdev_features_t feat,
+		   u64 *packet_size_bytes, u32 *work_done)
+{
+	struct gve_rx_slot_page_info *page_info;
+	struct gve_rx_ctx *ctx = &rx->ctx;
+	union gve_rx_data_slot *data_slot;
+	struct gve_priv *priv = rx->gve;
+	struct gve_rx_desc *first_desc;
+	struct sk_buff *skb = NULL;
+	struct gve_rx_desc *desc;
+	struct napi_struct *napi;
+	dma_addr_t page_bus;
+	u32 work_cnt = 0;
+	void *va;
+	u32 idx;
+	u16 len;
+
+	idx = rx->cnt & rx->mask;
+	first_desc = &rx->desc.desc_ring[idx];
+	desc = first_desc;
+	napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
+
+	if (unlikely(!gve_rx_ctx_init(ctx, rx)))
+		goto skb_alloc_fail;
+
+	while (ctx->curr_frag_cnt < ctx->expected_frag_cnt) {
+		/* Prefetch two packet buffers ahead, we will need it soon. */
+		page_info = &rx->data.page_info[(idx + 2) & rx->mask];
+		va = page_info->page_address + page_info->page_offset;
+
+		prefetch(page_info->page); /* Kernel page struct. */
+		prefetch(va);              /* Packet header. */
+		prefetch(va + 64);         /* Next cacheline too. */
+
+		len = gve_rx_get_fragment_size(ctx, desc);
+
+		page_info = &rx->data.page_info[idx];
+		data_slot = &rx->data.data_ring[idx];
+		page_bus = rx->data.raw_addressing ?
+			   be64_to_cpu(data_slot->addr) - page_info->page_offset :
+			   rx->data.qpl->page_buses[idx];
+		dma_sync_single_for_cpu(&priv->pdev->dev, page_bus, PAGE_SIZE, DMA_FROM_DEVICE);
+
+		skb = gve_rx_skb(priv, rx, page_info, napi, len, data_slot);
+		if (!skb) {
+			u64_stats_update_begin(&rx->statss);
+			rx->rx_skb_alloc_fail++;
+			u64_stats_update_end(&rx->statss);
+			goto skb_alloc_fail;
+		}
+
+		ctx->curr_frag_cnt++;
+		rx->cnt++;
+		idx = rx->cnt & rx->mask;
+		work_cnt++;
+		desc = &rx->desc.desc_ring[idx];
 	}
 
 	if (likely(feat & NETIF_F_RXCSUM)) {
 		/* NIC passes up the partial sum */
-		if (rx_desc->csum)
+		if (first_desc->csum)
 			skb->ip_summed = CHECKSUM_COMPLETE;
 		else
 			skb->ip_summed = CHECKSUM_NONE;
-		skb->csum = csum_unfold(rx_desc->csum);
+		skb->csum = csum_unfold(first_desc->csum);
 	}
 
 	/* parse flags & pass relevant info up */
 	if (likely(feat & NETIF_F_RXHASH) &&
-	    gve_needs_rss(rx_desc->flags_seq))
-		skb_set_hash(skb, be32_to_cpu(rx_desc->rss_hash),
-			     gve_rss_type(rx_desc->flags_seq));
+	    gve_needs_rss(first_desc->flags_seq))
+		skb_set_hash(skb, be32_to_cpu(first_desc->rss_hash),
+			     gve_rss_type(first_desc->flags_seq));
 
+	*packet_size_bytes = skb->len + (skb->protocol ? ETH_HLEN : 0);
+	*work_done = work_cnt;
 	if (skb_is_nonlinear(skb))
 		napi_gro_frags(napi);
 	else
 		napi_gro_receive(napi, skb);
+
+	gve_rx_ctx_clear(ctx);
 	return true;
+
+skb_alloc_fail:
+	if (napi->skb)
+		napi_free_frags(napi);
+	*packet_size_bytes = 0;
+	*work_done = ctx->expected_frag_cnt;
+	while (ctx->curr_frag_cnt < ctx->expected_frag_cnt) {
+		rx->cnt++;
+		ctx->curr_frag_cnt++;
+	}
+	gve_rx_ctx_clear(ctx);
+	return false;
 }
 
 bool gve_rx_work_pending(struct gve_rx_ring *rx)
@@ -529,7 +683,6 @@ static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
 				union gve_rx_data_slot *data_slot =
 						&rx->data.data_ring[idx];
 				struct device *dev = &priv->pdev->dev;
-
 				gve_rx_free_buffer(dev, page_info, data_slot);
 				page_info->page = NULL;
 				if (gve_rx_alloc_buffer(priv, dev, page_info,
@@ -550,16 +703,17 @@ static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
 static int gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
 			     netdev_features_t feat)
 {
+	u32 work_done = 0, total_packet_cnt = 0, ok_packet_cnt = 0;
 	struct gve_priv *priv = rx->gve;
-	u32 work_done = 0, packets = 0;
+	u32 idx = rx->cnt & rx->mask;
 	struct gve_rx_desc *desc;
-	u32 cnt = rx->cnt;
-	u32 idx = cnt & rx->mask;
 	u64 bytes = 0;
 
-	desc = rx->desc.desc_ring + idx;
+	desc = &rx->desc.desc_ring[idx];
 	while ((GVE_SEQNO(desc->flags_seq) == rx->desc.seqno) &&
 	       work_done < budget) {
+		u64 packet_size_bytes = 0;
+		u32 work_cnt = 0;
 		bool dropped;
 
 		netif_info(priv, rx_status, priv->dev,
@@ -570,37 +724,32 @@ static int gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
 			   rx->q_num, GVE_SEQNO(desc->flags_seq),
 			   rx->desc.seqno);
 
-		/* prefetch two descriptors ahead */
-		prefetch(rx->desc.desc_ring + ((cnt + 2) & rx->mask));
-
-		dropped = !gve_rx(rx, desc, feat, idx);
+		dropped = !gve_rx(rx, feat, &packet_size_bytes, &work_cnt);
 		if (!dropped) {
-			bytes += be16_to_cpu(desc->len) - GVE_RX_PAD;
-			packets++;
+			bytes += packet_size_bytes;
+			ok_packet_cnt++;
 		}
-		cnt++;
-		idx = cnt & rx->mask;
-		desc = rx->desc.desc_ring + idx;
-		rx->desc.seqno = gve_next_seqno(rx->desc.seqno);
-		work_done++;
+		total_packet_cnt++;
+		idx = rx->cnt & rx->mask;
+		desc = &rx->desc.desc_ring[idx];
+		work_done += work_cnt;
 	}
 
-	if (!work_done && rx->fill_cnt - cnt > rx->db_threshold)
+	if (!work_done && rx->fill_cnt - rx->cnt > rx->db_threshold)
 		return 0;
 
 	if (work_done) {
 		u64_stats_update_begin(&rx->statss);
-		rx->rpackets += packets;
+		rx->rpackets += ok_packet_cnt;
 		rx->rbytes += bytes;
 		u64_stats_update_end(&rx->statss);
-		rx->cnt = cnt;
 	}
 
 	/* restock ring slots */
 	if (!rx->data.raw_addressing) {
 		/* In QPL mode buffs are refilled as the desc are processed */
 		rx->fill_cnt += work_done;
-	} else if (rx->fill_cnt - cnt <= rx->db_threshold) {
+	} else if (rx->fill_cnt - rx->cnt <= rx->db_threshold) {
 		/* In raw addressing mode buffs are only refilled if the avail
 		 * falls below a threshold.
 		 */
@@ -610,14 +759,14 @@ static int gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
 		/* If we were not able to completely refill buffers, we'll want
 		 * to schedule this queue for work again to refill buffers.
 		 */
-		if (rx->fill_cnt - cnt <= rx->db_threshold) {
+		if (rx->fill_cnt - rx->cnt <= rx->db_threshold) {
 			gve_rx_write_doorbell(priv, rx);
 			return budget;
 		}
 	}
 
 	gve_rx_write_doorbell(priv, rx);
-	return work_done;
+	return total_packet_cnt;
 }
 
 int gve_rx_poll(struct gve_notify_block *block, int budget)

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

@@ -240,8 +240,8 @@ static int gve_rx_alloc_ring_dqo(struct gve_priv *priv, int idx)
 	rx->dqo.bufq.mask = buffer_queue_slots - 1;
 	rx->dqo.complq.num_free_slots = completion_queue_slots;
 	rx->dqo.complq.mask = completion_queue_slots - 1;
-	rx->skb_head = NULL;
-	rx->skb_tail = NULL;
+	rx->ctx.skb_head = NULL;
+	rx->ctx.skb_tail = NULL;
 
 	rx->dqo.num_buf_states = min_t(s16, S16_MAX, buffer_queue_slots * 4);
 	rx->dqo.buf_states = kvcalloc(rx->dqo.num_buf_states,
@@ -467,12 +467,12 @@ static void gve_rx_skb_hash(struct sk_buff *skb,
 
 static void gve_rx_free_skb(struct gve_rx_ring *rx)
 {
-	if (!rx->skb_head)
+	if (!rx->ctx.skb_head)
 		return;
 
-	dev_kfree_skb_any(rx->skb_head);
-	rx->skb_head = NULL;
-	rx->skb_tail = NULL;
+	dev_kfree_skb_any(rx->ctx.skb_head);
+	rx->ctx.skb_head = NULL;
+	rx->ctx.skb_tail = NULL;
 }
 
 /* Chains multi skbs for single rx packet.
@@ -483,7 +483,7 @@ static int gve_rx_append_frags(struct napi_struct *napi,
 			       u16 buf_len, struct gve_rx_ring *rx,
 			       struct gve_priv *priv)
 {
-	int num_frags = skb_shinfo(rx->skb_tail)->nr_frags;
+	int num_frags = skb_shinfo(rx->ctx.skb_tail)->nr_frags;
 
 	if (unlikely(num_frags == MAX_SKB_FRAGS)) {
 		struct sk_buff *skb;
@@ -492,17 +492,17 @@ static int gve_rx_append_frags(struct napi_struct *napi,
 		if (!skb)
 			return -1;
 
-		skb_shinfo(rx->skb_tail)->frag_list = skb;
-		rx->skb_tail = skb;
+		skb_shinfo(rx->ctx.skb_tail)->frag_list = skb;
+		rx->ctx.skb_tail = skb;
 		num_frags = 0;
 	}
-	if (rx->skb_tail != rx->skb_head) {
-		rx->skb_head->len += buf_len;
-		rx->skb_head->data_len += buf_len;
-		rx->skb_head->truesize += priv->data_buffer_size_dqo;
+	if (rx->ctx.skb_tail != rx->ctx.skb_head) {
+		rx->ctx.skb_head->len += buf_len;
+		rx->ctx.skb_head->data_len += buf_len;
+		rx->ctx.skb_head->truesize += priv->data_buffer_size_dqo;
 	}
 
-	skb_add_rx_frag(rx->skb_tail, num_frags,
+	skb_add_rx_frag(rx->ctx.skb_tail, num_frags,
 			buf_state->page_info.page,
 			buf_state->page_info.page_offset,
 			buf_len, priv->data_buffer_size_dqo);
@@ -556,7 +556,7 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 				      buf_len, DMA_FROM_DEVICE);
 
 	/* Append to current skb if one exists. */
-	if (rx->skb_head) {
+	if (rx->ctx.skb_head) {
 		if (unlikely(gve_rx_append_frags(napi, buf_state, buf_len, rx,
 						 priv)) != 0) {
 			goto error;
@@ -567,11 +567,11 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 	}
 
 	if (eop && buf_len <= priv->rx_copybreak) {
-		rx->skb_head = gve_rx_copy(priv->dev, napi,
-					   &buf_state->page_info, buf_len, 0);
-		if (unlikely(!rx->skb_head))
+		rx->ctx.skb_head = gve_rx_copy(priv->dev, napi,
+					       &buf_state->page_info, buf_len, 0, NULL);
+		if (unlikely(!rx->ctx.skb_head))
 			goto error;
-		rx->skb_tail = rx->skb_head;
+		rx->ctx.skb_tail = rx->ctx.skb_head;
 
 		u64_stats_update_begin(&rx->statss);
 		rx->rx_copied_pkt++;
@@ -583,12 +583,12 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 		return 0;
 	}
 
-	rx->skb_head = napi_get_frags(napi);
-	if (unlikely(!rx->skb_head))
+	rx->ctx.skb_head = napi_get_frags(napi);
+	if (unlikely(!rx->ctx.skb_head))
 		goto error;
-	rx->skb_tail = rx->skb_head;
+	rx->ctx.skb_tail = rx->ctx.skb_head;
 
-	skb_add_rx_frag(rx->skb_head, 0, buf_state->page_info.page,
+	skb_add_rx_frag(rx->ctx.skb_head, 0, buf_state->page_info.page,
 			buf_state->page_info.page_offset, buf_len,
 			priv->data_buffer_size_dqo);
 	gve_dec_pagecnt_bias(&buf_state->page_info);
@@ -635,27 +635,27 @@ static int gve_rx_complete_skb(struct gve_rx_ring *rx, struct napi_struct *napi,
 		rx->gve->ptype_lut_dqo->ptypes[desc->packet_type];
 	int err;
 
-	skb_record_rx_queue(rx->skb_head, rx->q_num);
+	skb_record_rx_queue(rx->ctx.skb_head, rx->q_num);
 
 	if (feat & NETIF_F_RXHASH)
-		gve_rx_skb_hash(rx->skb_head, desc, ptype);
+		gve_rx_skb_hash(rx->ctx.skb_head, desc, ptype);
 
 	if (feat & NETIF_F_RXCSUM)
-		gve_rx_skb_csum(rx->skb_head, desc, ptype);
+		gve_rx_skb_csum(rx->ctx.skb_head, desc, ptype);
 
 	/* RSC packets must set gso_size otherwise the TCP stack will complain
 	 * that packets are larger than MTU.
 	 */
 	if (desc->rsc) {
-		err = gve_rx_complete_rsc(rx->skb_head, desc, ptype);
+		err = gve_rx_complete_rsc(rx->ctx.skb_head, desc, ptype);
 		if (err < 0)
 			return err;
 	}
 
-	if (skb_headlen(rx->skb_head) == 0)
+	if (skb_headlen(rx->ctx.skb_head) == 0)
 		napi_gro_frags(napi);
 	else
-		napi_gro_receive(napi, rx->skb_head);
+		napi_gro_receive(napi, rx->ctx.skb_head);
 
 	return 0;
 }
@@ -717,18 +717,18 @@ int gve_rx_poll_dqo(struct gve_notify_block *block, int budget)
 		/* Free running counter of completed descriptors */
 		rx->cnt++;
 
-		if (!rx->skb_head)
+		if (!rx->ctx.skb_head)
 			continue;
 
 		if (!compl_desc->end_of_packet)
 			continue;
 
 		work_done++;
-		pkt_bytes = rx->skb_head->len;
+		pkt_bytes = rx->ctx.skb_head->len;
 		/* The ethernet header (first ETH_HLEN bytes) is snipped off
 		 * by eth_type_trans.
 		 */
-		if (skb_headlen(rx->skb_head))
+		if (skb_headlen(rx->ctx.skb_head))
 			pkt_bytes += ETH_HLEN;
 
 		/* gve_rx_complete_skb() will consume skb if successful */
@@ -741,8 +741,8 @@ int gve_rx_poll_dqo(struct gve_notify_block *block, int budget)
 		}
 
 		bytes += pkt_bytes;
-		rx->skb_head = NULL;
-		rx->skb_tail = NULL;
+		rx->ctx.skb_head = NULL;
+		rx->ctx.skb_tail = NULL;
 	}
 
 	gve_rx_post_buffers_dqo(rx);

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

@@ -50,20 +50,31 @@ void gve_rx_add_to_block(struct gve_priv *priv, int queue_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,
-			    u16 pad)
+			    u16 padding, struct gve_rx_ctx *ctx)
 {
-	struct sk_buff *skb = napi_alloc_skb(napi, len);
-	void *va = page_info->page_address + pad +
-		   page_info->page_offset;
-
-	if (unlikely(!skb))
-		return NULL;
-
+	void *va = page_info->page_address + padding + page_info->page_offset;
+	int skb_linear_offset = 0;
+	bool set_protocol = false;
+	struct sk_buff *skb;
+
+	if (ctx) {
+		if (!ctx->skb_head)
+			ctx->skb_head = napi_alloc_skb(napi, ctx->total_expected_size);
+
+		if (unlikely(!ctx->skb_head))
+			return NULL;
+		skb = ctx->skb_head;
+		skb_linear_offset = skb->len;
+		set_protocol = ctx->curr_frag_cnt == ctx->expected_frag_cnt - 1;
+	} else {
+		skb = napi_alloc_skb(napi, len);
+		set_protocol = true;
+	}
 	__skb_put(skb, len);
+	skb_copy_to_linear_data_offset(skb, skb_linear_offset, va, len);
 
-	skb_copy_to_linear_data(skb, va, len);
-
-	skb->protocol = eth_type_trans(skb, dev);
+	if (set_protocol)
+		skb->protocol = eth_type_trans(skb, dev);
 
 	return skb;
 }

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

@@ -19,7 +19,7 @@ void gve_rx_add_to_block(struct gve_priv *priv, int queue_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,
-			    u16 pad);
+			    u16 pad, struct gve_rx_ctx *ctx);
 
 /* Decrement pagecnt_bias. Set it back to INT_MAX if it reached zero. */
 void gve_dec_pagecnt_bias(struct gve_rx_slot_page_info *page_info);