Merge branch 'gve-desc'

Rushil Gupta says:

====================
gve: Add QPL mode for DQO descriptor format

GVE supports QPL ("queue-page-list") mode where
all data is communicated through a set of pre-registered
pages. Adding this mode to DQO.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Documentation/networking/device_drivers/ethernet/google/gve.rst

@@ -52,6 +52,15 @@ Descriptor Formats
 GVE supports two descriptor formats: GQI and DQO. These two formats have
 entirely different descriptors, which will be described below.
 
+Addressing Mode
+------------------
+GVE supports two addressing modes: QPL and RDA.
+QPL ("queue-page-list") mode communicates data through a set of
+pre-registered pages.
+
+For RDA ("raw DMA addressing") mode, the set of pages is dynamic.
+Therefore, the packet buffers can be anywhere in guest memory.
+
 Registers
 ---------
 All registers are MMIO.

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

@@ -52,6 +52,26 @@
 
 #define GVE_GQ_TX_MIN_PKT_DESC_BYTES 182
 
+#define DQO_QPL_DEFAULT_TX_PAGES 512
+#define DQO_QPL_DEFAULT_RX_PAGES 2048
+
+/* Maximum TSO size supported on DQO */
+#define GVE_DQO_TX_MAX	0x3FFFF
+
+#define GVE_TX_BUF_SHIFT_DQO 11
+
+/* 2K buffers for DQO-QPL */
+#define GVE_TX_BUF_SIZE_DQO BIT(GVE_TX_BUF_SHIFT_DQO)
+#define GVE_TX_BUFS_PER_PAGE_DQO (PAGE_SIZE >> GVE_TX_BUF_SHIFT_DQO)
+#define GVE_MAX_TX_BUFS_PER_PKT (DIV_ROUND_UP(GVE_DQO_TX_MAX, GVE_TX_BUF_SIZE_DQO))
+
+/* If number of free/recyclable buffers are less than this threshold; driver
+ * allocs and uses a non-qpl page on the receive path of DQO QPL to free
+ * up buffers.
+ * Value is set big enough to post at least 3 64K LRO packet via 2K buffer to NIC.
+ */
+#define GVE_DQO_QPL_ONDEMAND_ALLOC_THRESHOLD 96
+
 /* Each slot in the desc ring has a 1:1 mapping to a slot in the data ring */
 struct gve_rx_desc_queue {
 	struct gve_rx_desc *desc_ring; /* the descriptor ring */
@@ -218,6 +238,15 @@ struct gve_rx_ring {
 			 * which cannot be reused yet.
 			 */
 			struct gve_index_list used_buf_states;
+
+			/* qpl assigned to this queue */
+			struct gve_queue_page_list *qpl;
+
+			/* index into queue page list */
+			u32 next_qpl_page_idx;
+
+			/* track number of used buffers */
+			u16 used_buf_states_cnt;
 		} dqo;
 	};
 
@@ -329,8 +358,14 @@ struct gve_tx_pending_packet_dqo {
 	 * All others correspond to `skb`'s frags and should be unmapped with
 	 * `dma_unmap_page`.
 	 */
-	DEFINE_DMA_UNMAP_ADDR(dma[MAX_SKB_FRAGS + 1]);
-	DEFINE_DMA_UNMAP_LEN(len[MAX_SKB_FRAGS + 1]);
+	union {
+		struct {
+			DEFINE_DMA_UNMAP_ADDR(dma[MAX_SKB_FRAGS + 1]);
+			DEFINE_DMA_UNMAP_LEN(len[MAX_SKB_FRAGS + 1]);
+		};
+		s16 tx_qpl_buf_ids[GVE_MAX_TX_BUFS_PER_PKT];
+	};
+
 	u16 num_bufs;
 
 	/* Linked list index to next element in the list, or -1 if none */
@@ -385,6 +420,32 @@ struct gve_tx_ring {
 			 * set.
 			 */
 			u32 last_re_idx;
+
+			/* free running number of packet buf descriptors posted */
+			u16 posted_packet_desc_cnt;
+			/* free running number of packet buf descriptors completed */
+			u16 completed_packet_desc_cnt;
+
+			/* QPL fields */
+			struct {
+			       /* Linked list of gve_tx_buf_dqo. Index into
+				* tx_qpl_buf_next, or -1 if empty.
+				*
+				* This is a consumer list owned by the TX path. When it
+				* runs out, the producer list is stolen from the
+				* completion handling path
+				* (dqo_compl.free_tx_qpl_buf_head).
+				*/
+				s16 free_tx_qpl_buf_head;
+
+			       /* Free running count of the number of QPL tx buffers
+				* allocated
+				*/
+				u32 alloc_tx_qpl_buf_cnt;
+
+				/* Cached value of `dqo_compl.free_tx_qpl_buf_cnt` */
+				u32 free_tx_qpl_buf_cnt;
+			};
 		} dqo_tx;
 	};
 
@@ -428,6 +489,24 @@ struct gve_tx_ring {
 			 * reached a specified timeout.
 			 */
 			struct gve_index_list timed_out_completions;
+
+			/* QPL fields */
+			struct {
+				/* Linked list of gve_tx_buf_dqo. Index into
+				 * tx_qpl_buf_next, or -1 if empty.
+				 *
+				 * This is the producer list, owned by the completion
+				 * handling path. When the consumer list
+				 * (dqo_tx.free_tx_qpl_buf_head) is runs out, this list
+				 * will be stolen.
+				 */
+				atomic_t free_tx_qpl_buf_head;
+
+				/* Free running count of the number of tx buffers
+				 * freed
+				 */
+				atomic_t free_tx_qpl_buf_cnt;
+			};
 		} dqo_compl;
 	} ____cacheline_aligned;
 	u64 pkt_done; /* free-running - total packets completed */
@@ -454,6 +533,21 @@ struct gve_tx_ring {
 			s16 num_pending_packets;
 
 			u32 complq_mask; /* complq size is complq_mask + 1 */
+
+			/* QPL fields */
+			struct {
+				/* qpl assigned to this queue */
+				struct gve_queue_page_list *qpl;
+
+				/* Each QPL page is divided into TX bounce buffers
+				 * of size GVE_TX_BUF_SIZE_DQO. tx_qpl_buf_next is
+				 * an array to manage linked lists of TX buffers.
+				 * An entry j at index i implies that j'th buffer
+				 * is next on the list after i
+				 */
+				s16 *tx_qpl_buf_next;
+				u32 num_tx_qpl_bufs;
+			};
 		} dqo;
 	} ____cacheline_aligned;
 	struct netdev_queue *netdev_txq;
@@ -532,6 +626,7 @@ enum gve_queue_format {
 	GVE_GQI_RDA_FORMAT		= 0x1,
 	GVE_GQI_QPL_FORMAT		= 0x2,
 	GVE_DQO_RDA_FORMAT		= 0x3,
+	GVE_DQO_QPL_FORMAT		= 0x4,
 };
 
 struct gve_priv {
@@ -551,7 +646,8 @@ struct gve_priv {
 	u16 num_event_counters;
 	u16 tx_desc_cnt; /* num desc per ring */
 	u16 rx_desc_cnt; /* num desc per ring */
-	u16 tx_pages_per_qpl; /* tx buffer length */
+	u16 tx_pages_per_qpl; /* Suggested number of pages per qpl for TX queues by NIC */
+	u16 rx_pages_per_qpl; /* Suggested number of pages per qpl for RX queues by NIC */
 	u16 rx_data_slot_cnt; /* rx buffer length */
 	u64 max_registered_pages;
 	u64 num_registered_pages; /* num pages registered with NIC */
@@ -809,11 +905,17 @@ static inline u32 gve_rx_idx_to_ntfy(struct gve_priv *priv, u32 queue_idx)
 	return (priv->num_ntfy_blks / 2) + queue_idx;
 }
 
+static inline bool gve_is_qpl(struct gve_priv *priv)
+{
+	return priv->queue_format == GVE_GQI_QPL_FORMAT ||
+		priv->queue_format == GVE_DQO_QPL_FORMAT;
+}
+
 /* Returns the number of tx queue page lists
  */
 static inline u32 gve_num_tx_qpls(struct gve_priv *priv)
 {
-	if (priv->queue_format != GVE_GQI_QPL_FORMAT)
+	if (!gve_is_qpl(priv))
 		return 0;
 
 	return priv->tx_cfg.num_queues + priv->num_xdp_queues;
@@ -833,7 +935,7 @@ static inline u32 gve_num_xdp_qpls(struct gve_priv *priv)
  */
 static inline u32 gve_num_rx_qpls(struct gve_priv *priv)
 {
-	if (priv->queue_format != GVE_GQI_QPL_FORMAT)
+	if (!gve_is_qpl(priv))
 		return 0;
 
 	return priv->rx_cfg.num_queues;

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

@@ -39,7 +39,8 @@ void gve_parse_device_option(struct gve_priv *priv,
 			     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_jumbo_frames **dev_op_jumbo_frames)
+			     struct gve_device_option_jumbo_frames **dev_op_jumbo_frames,
+			     struct gve_device_option_dqo_qpl **dev_op_dqo_qpl)
 {
 	u32 req_feat_mask = be32_to_cpu(option->required_features_mask);
 	u16 option_length = be16_to_cpu(option->option_length);
@@ -112,6 +113,22 @@ void gve_parse_device_option(struct gve_priv *priv,
 		}
 		*dev_op_dqo_rda = (void *)(option + 1);
 		break;
+	case GVE_DEV_OPT_ID_DQO_QPL:
+		if (option_length < sizeof(**dev_op_dqo_qpl) ||
+		    req_feat_mask != GVE_DEV_OPT_REQ_FEAT_MASK_DQO_QPL) {
+			dev_warn(&priv->pdev->dev, GVE_DEVICE_OPTION_ERROR_FMT,
+				 "DQO QPL", (int)sizeof(**dev_op_dqo_qpl),
+				 GVE_DEV_OPT_REQ_FEAT_MASK_DQO_QPL,
+				 option_length, req_feat_mask);
+			break;
+		}
+
+		if (option_length > sizeof(**dev_op_dqo_qpl)) {
+			dev_warn(&priv->pdev->dev,
+				 GVE_DEVICE_OPTION_TOO_BIG_FMT, "DQO QPL");
+		}
+		*dev_op_dqo_qpl = (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) {
@@ -146,7 +163,8 @@ gve_process_device_options(struct gve_priv *priv,
 			   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_jumbo_frames **dev_op_jumbo_frames)
+			   struct gve_device_option_jumbo_frames **dev_op_jumbo_frames,
+			   struct gve_device_option_dqo_qpl **dev_op_dqo_qpl)
 {
 	const int num_options = be16_to_cpu(descriptor->num_device_options);
 	struct gve_device_option *dev_opt;
@@ -166,7 +184,8 @@ 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_rda, dev_op_jumbo_frames,
+					dev_op_dqo_qpl);
 		dev_opt = next_opt;
 	}
 
@@ -505,12 +524,24 @@ static int gve_adminq_create_tx_queue(struct gve_priv *priv, u32 queue_index)
 
 		cmd.create_tx_queue.queue_page_list_id = cpu_to_be32(qpl_id);
 	} else {
+		u16 comp_ring_size;
+		u32 qpl_id = 0;
+
+		if (priv->queue_format == GVE_DQO_RDA_FORMAT) {
+			qpl_id = GVE_RAW_ADDRESSING_QPL_ID;
+			comp_ring_size =
+				priv->options_dqo_rda.tx_comp_ring_entries;
+		} else {
+			qpl_id = tx->dqo.qpl->id;
+			comp_ring_size = priv->tx_desc_cnt;
+		}
+		cmd.create_tx_queue.queue_page_list_id = cpu_to_be32(qpl_id);
 		cmd.create_tx_queue.tx_ring_size =
 			cpu_to_be16(priv->tx_desc_cnt);
 		cmd.create_tx_queue.tx_comp_ring_addr =
 			cpu_to_be64(tx->complq_bus_dqo);
 		cmd.create_tx_queue.tx_comp_ring_size =
-			cpu_to_be16(priv->options_dqo_rda.tx_comp_ring_entries);
+			cpu_to_be16(comp_ring_size);
 	}
 
 	return gve_adminq_issue_cmd(priv, &cmd);
@@ -555,6 +586,18 @@ static int gve_adminq_create_rx_queue(struct gve_priv *priv, u32 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 {
+		u16 rx_buff_ring_entries;
+		u32 qpl_id = 0;
+
+		if (priv->queue_format == GVE_DQO_RDA_FORMAT) {
+			qpl_id = GVE_RAW_ADDRESSING_QPL_ID;
+			rx_buff_ring_entries =
+				priv->options_dqo_rda.rx_buff_ring_entries;
+		} else {
+			qpl_id = rx->dqo.qpl->id;
+			rx_buff_ring_entries = priv->rx_desc_cnt;
+		}
+		cmd.create_rx_queue.queue_page_list_id = cpu_to_be32(qpl_id);
 		cmd.create_rx_queue.rx_ring_size =
 			cpu_to_be16(priv->rx_desc_cnt);
 		cmd.create_rx_queue.rx_desc_ring_addr =
@@ -564,7 +607,7 @@ static int gve_adminq_create_rx_queue(struct gve_priv *priv, u32 queue_index)
 		cmd.create_rx_queue.packet_buffer_size =
 			cpu_to_be16(priv->data_buffer_size_dqo);
 		cmd.create_rx_queue.rx_buff_ring_size =
-			cpu_to_be16(priv->options_dqo_rda.rx_buff_ring_entries);
+			cpu_to_be16(rx_buff_ring_entries);
 		cmd.create_rx_queue.enable_rsc =
 			!!(priv->dev->features & NETIF_F_LRO);
 	}
@@ -675,9 +718,13 @@ gve_set_desc_cnt_dqo(struct gve_priv *priv,
 		     const struct gve_device_option_dqo_rda *dev_op_dqo_rda)
 {
 	priv->tx_desc_cnt = be16_to_cpu(descriptor->tx_queue_entries);
+	priv->rx_desc_cnt = be16_to_cpu(descriptor->rx_queue_entries);
+
+	if (priv->queue_format == GVE_DQO_QPL_FORMAT)
+		return 0;
+
 	priv->options_dqo_rda.tx_comp_ring_entries =
 		be16_to_cpu(dev_op_dqo_rda->tx_comp_ring_entries);
-	priv->rx_desc_cnt = be16_to_cpu(descriptor->rx_queue_entries);
 	priv->options_dqo_rda.rx_buff_ring_entries =
 		be16_to_cpu(dev_op_dqo_rda->rx_buff_ring_entries);
 
@@ -687,7 +734,9 @@ gve_set_desc_cnt_dqo(struct gve_priv *priv,
 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)
+					  *dev_op_jumbo_frames,
+					  const struct gve_device_option_dqo_qpl
+					  *dev_op_dqo_qpl)
 {
 	/* Before control reaches this point, the page-size-capped max MTU from
 	 * the gve_device_descriptor field has already been stored in
@@ -699,6 +748,18 @@ static void gve_enable_supported_features(struct gve_priv *priv,
 			 "JUMBO FRAMES device option enabled.\n");
 		priv->dev->max_mtu = be16_to_cpu(dev_op_jumbo_frames->max_mtu);
 	}
+
+	/* Override pages for qpl for DQO-QPL */
+	if (dev_op_dqo_qpl) {
+		priv->tx_pages_per_qpl =
+			be16_to_cpu(dev_op_dqo_qpl->tx_pages_per_qpl);
+		priv->rx_pages_per_qpl =
+			be16_to_cpu(dev_op_dqo_qpl->rx_pages_per_qpl);
+		if (priv->tx_pages_per_qpl == 0)
+			priv->tx_pages_per_qpl = DQO_QPL_DEFAULT_TX_PAGES;
+		if (priv->rx_pages_per_qpl == 0)
+			priv->rx_pages_per_qpl = DQO_QPL_DEFAULT_RX_PAGES;
+	}
 }
 
 int gve_adminq_describe_device(struct gve_priv *priv)
@@ -707,6 +768,7 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 	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_option_dqo_qpl *dev_op_dqo_qpl = NULL;
 	struct gve_device_descriptor *descriptor;
 	u32 supported_features_mask = 0;
 	union gve_adminq_command cmd;
@@ -733,13 +795,14 @@ 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_jumbo_frames,
+					 &dev_op_dqo_qpl);
 	if (err)
 		goto free_device_descriptor;
 
 	/* If the GQI_RAW_ADDRESSING option is not enabled and the queue format
 	 * is not set to GqiRda, choose the queue format in a priority order:
-	 * DqoRda, GqiRda, GqiQpl. Use GqiQpl as default.
+	 * DqoRda, DqoQpl, GqiRda, GqiQpl. Use GqiQpl as default.
 	 */
 	if (dev_op_dqo_rda) {
 		priv->queue_format = GVE_DQO_RDA_FORMAT;
@@ -747,7 +810,11 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 			 "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) {
+	} else if (dev_op_dqo_qpl) {
+		priv->queue_format = GVE_DQO_QPL_FORMAT;
+		supported_features_mask =
+			be32_to_cpu(dev_op_dqo_qpl->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");
@@ -798,7 +865,7 @@ 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_jumbo_frames, dev_op_dqo_qpl);
 
 free_device_descriptor:
 	dma_free_coherent(&priv->pdev->dev, PAGE_SIZE, descriptor,

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

@@ -109,6 +109,14 @@ struct gve_device_option_dqo_rda {
 
 static_assert(sizeof(struct gve_device_option_dqo_rda) == 8);
 
+struct gve_device_option_dqo_qpl {
+	__be32 supported_features_mask;
+	__be16 tx_pages_per_qpl;
+	__be16 rx_pages_per_qpl;
+};
+
+static_assert(sizeof(struct gve_device_option_dqo_qpl) == 8);
+
 struct gve_device_option_jumbo_frames {
 	__be32 supported_features_mask;
 	__be16 max_mtu;
@@ -130,6 +138,7 @@ 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_DQO_QPL = 0x7,
 	GVE_DEV_OPT_ID_JUMBO_FRAMES = 0x8,
 };
 
@@ -139,6 +148,7 @@ enum gve_dev_opt_req_feat_mask {
 	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,
+	GVE_DEV_OPT_REQ_FEAT_MASK_DQO_QPL = 0x0,
 };
 
 enum gve_sup_feature_mask {

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

@@ -31,7 +31,6 @@
 
 // Minimum amount of time between queue kicks in msec (10 seconds)
 #define MIN_TX_TIMEOUT_GAP (1000 * 10)
-#define DQO_TX_MAX	0x3FFFF
 
 char gve_driver_name[] = "gve";
 const char gve_version_str[] = GVE_VERSION;
@@ -494,7 +493,7 @@ static int gve_setup_device_resources(struct gve_priv *priv)
 		goto abort_with_stats_report;
 	}
 
-	if (priv->queue_format == GVE_DQO_RDA_FORMAT) {
+	if (!gve_is_gqi(priv)) {
 		priv->ptype_lut_dqo = kvzalloc(sizeof(*priv->ptype_lut_dqo),
 					       GFP_KERNEL);
 		if (!priv->ptype_lut_dqo) {
@@ -1083,11 +1082,12 @@ static int gve_alloc_xdp_qpls(struct gve_priv *priv)
 static int gve_alloc_qpls(struct gve_priv *priv)
 {
 	int max_queues = priv->tx_cfg.max_queues + priv->rx_cfg.max_queues;
+	int page_count;
 	int start_id;
 	int i, j;
 	int err;
 
-	if (priv->queue_format != GVE_GQI_QPL_FORMAT)
+	if (!gve_is_qpl(priv))
 		return 0;
 
 	priv->qpls = kvcalloc(max_queues, sizeof(*priv->qpls), GFP_KERNEL);
@@ -1095,17 +1095,25 @@ static int gve_alloc_qpls(struct gve_priv *priv)
 		return -ENOMEM;
 
 	start_id = gve_tx_start_qpl_id(priv);
+	page_count = priv->tx_pages_per_qpl;
 	for (i = start_id; i < start_id + gve_num_tx_qpls(priv); i++) {
 		err = gve_alloc_queue_page_list(priv, i,
-						priv->tx_pages_per_qpl);
+						page_count);
 		if (err)
 			goto free_qpls;
 	}
 
 	start_id = gve_rx_start_qpl_id(priv);
+
+	/* For GQI_QPL number of pages allocated have 1:1 relationship with
+	 * number of descriptors. For DQO, number of pages required are
+	 * more than descriptors (because of out of order completions).
+	 */
+	page_count = priv->queue_format == GVE_GQI_QPL_FORMAT ?
+		priv->rx_data_slot_cnt : priv->rx_pages_per_qpl;
 	for (i = start_id; i < start_id + gve_num_rx_qpls(priv); i++) {
 		err = gve_alloc_queue_page_list(priv, i,
-						priv->rx_data_slot_cnt);
+						page_count);
 		if (err)
 			goto free_qpls;
 	}
@@ -2051,7 +2059,7 @@ static int gve_init_priv(struct gve_priv *priv, bool skip_describe_device)
 
 	/* Big TCP is only supported on DQ*/
 	if (!gve_is_gqi(priv))
-		netif_set_tso_max_size(priv->dev, DQO_TX_MAX);
+		netif_set_tso_max_size(priv->dev, GVE_DQO_TX_MAX);
 
 	priv->num_registered_pages = 0;
 	priv->rx_copybreak = GVE_DEFAULT_RX_COPYBREAK;

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

@@ -22,11 +22,13 @@ static int gve_buf_ref_cnt(struct gve_rx_buf_state_dqo *bs)
 }
 
 static void gve_free_page_dqo(struct gve_priv *priv,
-			      struct gve_rx_buf_state_dqo *bs)
+			      struct gve_rx_buf_state_dqo *bs,
+			      bool free_page)
 {
 	page_ref_sub(bs->page_info.page, bs->page_info.pagecnt_bias - 1);
-	gve_free_page(&priv->pdev->dev, bs->page_info.page, bs->addr,
-		      DMA_FROM_DEVICE);
+	if (free_page)
+		gve_free_page(&priv->pdev->dev, bs->page_info.page, bs->addr,
+			      DMA_FROM_DEVICE);
 	bs->page_info.page = NULL;
 }
 
@@ -130,12 +132,20 @@ gve_get_recycled_buf_state(struct gve_rx_ring *rx)
 	 */
 	for (i = 0; i < 5; i++) {
 		buf_state = gve_dequeue_buf_state(rx, &rx->dqo.used_buf_states);
-		if (gve_buf_ref_cnt(buf_state) == 0)
+		if (gve_buf_ref_cnt(buf_state) == 0) {
+			rx->dqo.used_buf_states_cnt--;
 			return buf_state;
+		}
 
 		gve_enqueue_buf_state(rx, &rx->dqo.used_buf_states, buf_state);
 	}
 
+	/* For QPL, we cannot allocate any new buffers and must
+	 * wait for the existing ones to be available.
+	 */
+	if (rx->dqo.qpl)
+		return NULL;
+
 	/* If there are no free buf states discard an entry from
 	 * `used_buf_states` so it can be used.
 	 */
@@ -144,23 +154,39 @@ gve_get_recycled_buf_state(struct gve_rx_ring *rx)
 		if (gve_buf_ref_cnt(buf_state) == 0)
 			return buf_state;
 
-		gve_free_page_dqo(rx->gve, buf_state);
+		gve_free_page_dqo(rx->gve, buf_state, true);
 		gve_free_buf_state(rx, buf_state);
 	}
 
 	return NULL;
 }
 
-static int gve_alloc_page_dqo(struct gve_priv *priv,
+static int gve_alloc_page_dqo(struct gve_rx_ring *rx,
 			      struct gve_rx_buf_state_dqo *buf_state)
 {
-	int err;
+	struct gve_priv *priv = rx->gve;
+	u32 idx;
 
-	err = gve_alloc_page(priv, &priv->pdev->dev, &buf_state->page_info.page,
-			     &buf_state->addr, DMA_FROM_DEVICE, GFP_ATOMIC);
-	if (err)
-		return err;
+	if (!rx->dqo.qpl) {
+		int err;
 
+		err = gve_alloc_page(priv, &priv->pdev->dev,
+				     &buf_state->page_info.page,
+				     &buf_state->addr,
+				     DMA_FROM_DEVICE, GFP_ATOMIC);
+		if (err)
+			return err;
+	} else {
+		idx = rx->dqo.next_qpl_page_idx;
+		if (idx >= priv->rx_pages_per_qpl) {
+			net_err_ratelimited("%s: Out of QPL pages\n",
+					    priv->dev->name);
+			return -ENOMEM;
+		}
+		buf_state->page_info.page = rx->dqo.qpl->pages[idx];
+		buf_state->addr = rx->dqo.qpl->page_buses[idx];
+		rx->dqo.next_qpl_page_idx++;
+	}
 	buf_state->page_info.page_offset = 0;
 	buf_state->page_info.page_address =
 		page_address(buf_state->page_info.page);
@@ -195,9 +221,13 @@ static void gve_rx_free_ring_dqo(struct gve_priv *priv, int idx)
 
 	for (i = 0; i < rx->dqo.num_buf_states; i++) {
 		struct gve_rx_buf_state_dqo *bs = &rx->dqo.buf_states[i];
-
+		/* Only free page for RDA. QPL pages are freed in gve_main. */
 		if (bs->page_info.page)
-			gve_free_page_dqo(priv, bs);
+			gve_free_page_dqo(priv, bs, !rx->dqo.qpl);
+	}
+	if (rx->dqo.qpl) {
+		gve_unassign_qpl(priv, rx->dqo.qpl->id);
+		rx->dqo.qpl = NULL;
 	}
 
 	if (rx->dqo.bufq.desc_ring) {
@@ -229,7 +259,8 @@ static int gve_rx_alloc_ring_dqo(struct gve_priv *priv, int idx)
 	int i;
 
 	const u32 buffer_queue_slots =
-		priv->options_dqo_rda.rx_buff_ring_entries;
+		priv->queue_format == GVE_DQO_RDA_FORMAT ?
+		priv->options_dqo_rda.rx_buff_ring_entries : priv->rx_desc_cnt;
 	const u32 completion_queue_slots = priv->rx_desc_cnt;
 
 	netif_dbg(priv, drv, priv->dev, "allocating rx ring DQO\n");
@@ -243,7 +274,9 @@ static int gve_rx_alloc_ring_dqo(struct gve_priv *priv, int idx)
 	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.num_buf_states = priv->queue_format == GVE_DQO_RDA_FORMAT ?
+		min_t(s16, S16_MAX, buffer_queue_slots * 4) :
+		priv->rx_pages_per_qpl;
 	rx->dqo.buf_states = kvcalloc(rx->dqo.num_buf_states,
 				      sizeof(rx->dqo.buf_states[0]),
 				      GFP_KERNEL);
@@ -275,6 +308,13 @@ static int gve_rx_alloc_ring_dqo(struct gve_priv *priv, int idx)
 	if (!rx->dqo.bufq.desc_ring)
 		goto err;
 
+	if (priv->queue_format != GVE_DQO_RDA_FORMAT) {
+		rx->dqo.qpl = gve_assign_rx_qpl(priv, rx->q_num);
+		if (!rx->dqo.qpl)
+			goto err;
+		rx->dqo.next_qpl_page_idx = 0;
+	}
+
 	rx->q_resources = dma_alloc_coherent(hdev, sizeof(*rx->q_resources),
 					     &rx->q_resources_bus, GFP_KERNEL);
 	if (!rx->q_resources)
@@ -352,7 +392,7 @@ void gve_rx_post_buffers_dqo(struct gve_rx_ring *rx)
 			if (unlikely(!buf_state))
 				break;
 
-			if (unlikely(gve_alloc_page_dqo(priv, buf_state))) {
+			if (unlikely(gve_alloc_page_dqo(rx, buf_state))) {
 				u64_stats_update_begin(&rx->statss);
 				rx->rx_buf_alloc_fail++;
 				u64_stats_update_end(&rx->statss);
@@ -415,6 +455,7 @@ static void gve_try_recycle_buf(struct gve_priv *priv, struct gve_rx_ring *rx,
 
 mark_used:
 	gve_enqueue_buf_state(rx, &rx->dqo.used_buf_states, buf_state);
+	rx->dqo.used_buf_states_cnt++;
 }
 
 static void gve_rx_skb_csum(struct sk_buff *skb,
@@ -475,6 +516,43 @@ static void gve_rx_free_skb(struct gve_rx_ring *rx)
 	rx->ctx.skb_tail = NULL;
 }
 
+static bool gve_rx_should_trigger_copy_ondemand(struct gve_rx_ring *rx)
+{
+	if (!rx->dqo.qpl)
+		return false;
+	if (rx->dqo.used_buf_states_cnt <
+		     (rx->dqo.num_buf_states -
+		     GVE_DQO_QPL_ONDEMAND_ALLOC_THRESHOLD))
+		return false;
+	return true;
+}
+
+static int gve_rx_copy_ondemand(struct gve_rx_ring *rx,
+				struct gve_rx_buf_state_dqo *buf_state,
+				u16 buf_len)
+{
+	struct page *page = alloc_page(GFP_ATOMIC);
+	int num_frags;
+
+	if (!page)
+		return -ENOMEM;
+
+	memcpy(page_address(page),
+	       buf_state->page_info.page_address +
+	       buf_state->page_info.page_offset,
+	       buf_len);
+	num_frags = skb_shinfo(rx->ctx.skb_tail)->nr_frags;
+	skb_add_rx_frag(rx->ctx.skb_tail, num_frags, page,
+			0, buf_len, PAGE_SIZE);
+
+	u64_stats_update_begin(&rx->statss);
+	rx->rx_frag_alloc_cnt++;
+	u64_stats_update_end(&rx->statss);
+	/* Return unused buffer. */
+	gve_enqueue_buf_state(rx, &rx->dqo.recycled_buf_states, buf_state);
+	return 0;
+}
+
 /* Chains multi skbs for single rx packet.
  * Returns 0 if buffer is appended, -1 otherwise.
  */
@@ -502,12 +580,20 @@ static int gve_rx_append_frags(struct napi_struct *napi,
 		rx->ctx.skb_head->truesize += priv->data_buffer_size_dqo;
 	}
 
+	/* Trigger ondemand page allocation if we are running low on buffers */
+	if (gve_rx_should_trigger_copy_ondemand(rx))
+		return gve_rx_copy_ondemand(rx, buf_state, buf_len);
+
 	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);
 	gve_dec_pagecnt_bias(&buf_state->page_info);
 
+	/* Advances buffer page-offset if page is partially used.
+	 * Marks buffer as used if page is full.
+	 */
+	gve_try_recycle_buf(priv, rx, buf_state);
 	return 0;
 }
 
@@ -561,8 +647,6 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 						 priv)) != 0) {
 			goto error;
 		}
-
-		gve_try_recycle_buf(priv, rx, buf_state);
 		return 0;
 	}
 
@@ -588,6 +672,12 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 		goto error;
 	rx->ctx.skb_tail = rx->ctx.skb_head;
 
+	if (gve_rx_should_trigger_copy_ondemand(rx)) {
+		if (gve_rx_copy_ondemand(rx, buf_state, buf_len) < 0)
+			goto error;
+		return 0;
+	}
+
 	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);