[PATCH net-next v9 0/4] GVE Raw Addressing

[PATCH net-next v9 0/4] GVE Raw Addressing

[David Awogbemila]

Patchset description:
This  patchset introduces "raw addressing" mode to the GVE driver.
Previously (in "queue_page_list" or "qpl" mode), the driver would
pre-allocate and dma_map buffers to be used on egress and ingress.
On egress, it would copy data from the skb provided to the
pre-allocated buffers - this was expensive.
In raw addressing mode, the driver can avoid this copy and simply
dma_map the skb's data so that the NIC can use it.
On ingress, the driver passes buffers up to the networking stack and
then frees and reallocates buffers when necessary instead of using
skb_copy_to_linear_data.
Patch 3 separates the page refcount tracking mechanism
into a function gve_rx_can_recycle_buffer which uses get_page - this will
be changed in a future patch to eliminate the use of get_page in tracking
page refcounts.

Changes from v8:
Patch 4: Free skbs that are not sent in gve_tx().

Catherine Sullivan (3):
  gve: Add support for raw addressing device option
  gve: Add support for raw addressing to the rx path
  gve: Add support for raw addressing in the tx path

David Awogbemila (1):
  gve: Rx Buffer Recycling

 drivers/net/ethernet/google/gve/gve.h         |  39 +-
 drivers/net/ethernet/google/gve/gve_adminq.c  |  89 ++++-
 drivers/net/ethernet/google/gve/gve_adminq.h  |  15 +-
 drivers/net/ethernet/google/gve/gve_desc.h    |  19 +-
 drivers/net/ethernet/google/gve/gve_ethtool.c |   2 +
 drivers/net/ethernet/google/gve/gve_main.c    |  11 +-
 drivers/net/ethernet/google/gve/gve_rx.c      | 364 +++++++++++++-----
 drivers/net/ethernet/google/gve/gve_tx.c      | 202 ++++++++--
 8 files changed, 577 insertions(+), 164 deletions(-)

-- 
2.29.2.576.ga3fc446d84-goog

[PATCH net-next v9 1/4] gve: Add support for raw addressing device option

[David Awogbemila]

From: Catherine Sullivan <csully@google.com>

Add support to describe device for parsing device options. As
the first device option, add raw addressing.

"Raw Addressing" mode (as opposed to the current "qpl" mode) is an
operational mode which allows the driver avoid bounce buffer copies
which it currently performs using pre-allocated qpls (queue_page_lists)
when sending and receiving packets.
For egress packets, the provided skb data addresses will be dma_map'ed and
passed to the device, allowing the NIC can perform DMA directly - the
driver will not have to copy the buffer content into pre-allocated
buffers/qpls (as in qpl mode).
For ingress packets, copies are also eliminated as buffers are handed to
the networking stack and then recycled or re-allocated as
necessary, avoiding the use of skb_copy_to_linear_data().

This patch only introduces the option to the driver.
Subsequent patches will add the ingress and egress functionality.

Reviewed-by: Yangchun Fu <yangchun@google.com>
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David Awogbemila <awogbemila@google.com>
---
 drivers/net/ethernet/google/gve/gve.h        |  1 +
 drivers/net/ethernet/google/gve/gve_adminq.c | 71 ++++++++++++++++++++
 drivers/net/ethernet/google/gve/gve_adminq.h | 15 +++--
 drivers/net/ethernet/google/gve/gve_main.c   |  9 +++
 4 files changed, 92 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
index f5c80229ea96..782e2791ee11 100644
--- a/drivers/net/ethernet/google/gve/gve.h
+++ b/drivers/net/ethernet/google/gve/gve.h
@@ -199,6 +199,7 @@ struct gve_priv {
 	u64 num_registered_pages; /* num pages registered with NIC */
 	u32 rx_copybreak; /* copy packets smaller than this */
 	u16 default_num_queues; /* default num queues to set up */
+	u8 raw_addressing; /* 1 if this dev supports raw addressing, 0 otherwise */
 
 	struct gve_queue_config tx_cfg;
 	struct gve_queue_config rx_cfg;
diff --git a/drivers/net/ethernet/google/gve/gve_adminq.c b/drivers/net/ethernet/google/gve/gve_adminq.c
index 24ae6a28a806..608a8806f8c8 100644
--- a/drivers/net/ethernet/google/gve/gve_adminq.c
+++ b/drivers/net/ethernet/google/gve/gve_adminq.c
@@ -14,6 +14,57 @@
 #define GVE_ADMINQ_SLEEP_LEN		20
 #define GVE_MAX_ADMINQ_EVENT_COUNTER_CHECK	100
 
+#define GVE_DEVICE_OPTION_ERROR_FMT "%s option error:\n" \
+"Expected: length=%d, feature_mask=%x.\n" \
+"Actual: length=%d, feature_mask=%x.\n"
+
+static
+struct gve_device_option *gve_get_next_option(struct gve_device_descriptor *descriptor,
+					      struct gve_device_option *option)
+{
+	void *option_end, *descriptor_end;
+
+	option_end = (void *)(option + 1) + be16_to_cpu(option->option_length);
+	descriptor_end = (void *)descriptor + be16_to_cpu(descriptor->total_length);
+
+	return option_end > descriptor_end ? NULL : (struct gve_device_option *)option_end;
+}
+
+static
+void gve_parse_device_option(struct gve_priv *priv,
+			     struct gve_device_descriptor *device_descriptor,
+			     struct gve_device_option *option)
+{
+	u16 option_length = be16_to_cpu(option->option_length);
+	u16 option_id = be16_to_cpu(option->option_id);
+
+	switch (option_id) {
+	case GVE_DEV_OPT_ID_RAW_ADDRESSING:
+		/* If the length or feature mask doesn't match,
+		 * continue without enabling the feature.
+		 */
+		if (option_length != GVE_DEV_OPT_LEN_RAW_ADDRESSING ||
+		    option->feat_mask != cpu_to_be32(GVE_DEV_OPT_FEAT_MASK_RAW_ADDRESSING)) {
+			dev_warn(&priv->pdev->dev, GVE_DEVICE_OPTION_ERROR_FMT, "Raw Addressing",
+				 GVE_DEV_OPT_LEN_RAW_ADDRESSING,
+				 cpu_to_be32(GVE_DEV_OPT_FEAT_MASK_RAW_ADDRESSING),
+				 option_length, option->feat_mask);
+			priv->raw_addressing = 0;
+		} else {
+			dev_info(&priv->pdev->dev,
+				 "Raw addressing device option enabled.\n");
+			priv->raw_addressing = 1;
+		}
+		break;
+	default:
+		/* If we don't recognize the option just continue
+		 * without doing anything.
+		 */
+		dev_dbg(&priv->pdev->dev, "Unrecognized device option 0x%hx not enabled.\n",
+			option_id);
+	}
+}
+
 int gve_adminq_alloc(struct device *dev, struct gve_priv *priv)
 {
 	priv->adminq = dma_alloc_coherent(dev, PAGE_SIZE,
@@ -460,11 +511,14 @@ int gve_adminq_destroy_rx_queues(struct gve_priv *priv, u32 num_queues)
 int gve_adminq_describe_device(struct gve_priv *priv)
 {
 	struct gve_device_descriptor *descriptor;
+	struct gve_device_option *dev_opt;
 	union gve_adminq_command cmd;
 	dma_addr_t descriptor_bus;
+	u16 num_options;
 	int err = 0;
 	u8 *mac;
 	u16 mtu;
+	int i;
 
 	memset(&cmd, 0, sizeof(cmd));
 	descriptor = dma_alloc_coherent(&priv->pdev->dev, PAGE_SIZE,
@@ -518,6 +572,23 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 		priv->rx_desc_cnt = priv->rx_pages_per_qpl;
 	}
 	priv->default_num_queues = be16_to_cpu(descriptor->default_num_queues);
+	dev_opt = (void *)(descriptor + 1);
+
+	num_options = be16_to_cpu(descriptor->num_device_options);
+	for (i = 0; i < num_options; i++) {
+		struct gve_device_option *next_opt;
+
+		next_opt = gve_get_next_option(descriptor, dev_opt);
+		if (!next_opt) {
+			dev_err(&priv->dev->dev,
+				"options exceed device_descriptor's total length.\n");
+			err = -EINVAL;
+			goto free_device_descriptor;
+		}
+
+		gve_parse_device_option(priv, descriptor, dev_opt);
+		dev_opt = next_opt;
+	}
 
 free_device_descriptor:
 	dma_free_coherent(&priv->pdev->dev, sizeof(*descriptor), descriptor,
diff --git a/drivers/net/ethernet/google/gve/gve_adminq.h b/drivers/net/ethernet/google/gve/gve_adminq.h
index 281de8326bc5..af5f586167bd 100644
--- a/drivers/net/ethernet/google/gve/gve_adminq.h
+++ b/drivers/net/ethernet/google/gve/gve_adminq.h
@@ -79,12 +79,17 @@ struct gve_device_descriptor {
 
 static_assert(sizeof(struct gve_device_descriptor) == 40);
 
-struct device_option {
-	__be32 option_id;
-	__be32 option_length;
+struct gve_device_option {
+	__be16 option_id;
+	__be16 option_length;
+	__be32 feat_mask;
 };
 
-static_assert(sizeof(struct device_option) == 8);
+static_assert(sizeof(struct gve_device_option) == 8);
+
+#define GVE_DEV_OPT_ID_RAW_ADDRESSING 0x1
+#define GVE_DEV_OPT_LEN_RAW_ADDRESSING 0x0
+#define GVE_DEV_OPT_FEAT_MASK_RAW_ADDRESSING 0x0
 
 struct gve_adminq_configure_device_resources {
 	__be64 counter_array;
@@ -111,6 +116,8 @@ struct gve_adminq_unregister_page_list {
 
 static_assert(sizeof(struct gve_adminq_unregister_page_list) == 4);
 
+#define GVE_RAW_ADDRESSING_QPL_ID 0xFFFFFFFF
+
 struct gve_adminq_create_tx_queue {
 	__be32 queue_id;
 	__be32 reserved;
diff --git a/drivers/net/ethernet/google/gve/gve_main.c b/drivers/net/ethernet/google/gve/gve_main.c
index 48a433154ce0..70685c10db0e 100644
--- a/drivers/net/ethernet/google/gve/gve_main.c
+++ b/drivers/net/ethernet/google/gve/gve_main.c
@@ -678,6 +678,10 @@ static int gve_alloc_qpls(struct gve_priv *priv)
 	int i, j;
 	int err;
 
+	/* Raw addressing means no QPLs */
+	if (priv->raw_addressing)
+		return 0;
+
 	priv->qpls = kvzalloc(num_qpls * sizeof(*priv->qpls), GFP_KERNEL);
 	if (!priv->qpls)
 		return -ENOMEM;
@@ -718,6 +722,10 @@ static void gve_free_qpls(struct gve_priv *priv)
 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
 	int i;
 
+	/* Raw addressing means no QPLs */
+	if (priv->raw_addressing)
+		return;
+
 	kvfree(priv->qpl_cfg.qpl_id_map);
 
 	for (i = 0; i < num_qpls; i++)
@@ -1078,6 +1086,7 @@ static int gve_init_priv(struct gve_priv *priv, bool skip_describe_device)
 	if (skip_describe_device)
 		goto setup_device;
 
+	priv->raw_addressing = false;
 	/* Get the initial information we need from the device */
 	err = gve_adminq_describe_device(priv);
 	if (err) {
-- 
2.29.2.576.ga3fc446d84-goog

[PATCH net-next v9 2/4] gve: Add support for raw addressing to the rx path

[David Awogbemila]

From: Catherine Sullivan <csully@google.com>

Add support to use raw dma addresses in the rx path. Due to this new
support we can alloc a new buffer instead of making a copy.

RX buffers are handed to the networking stack and are
re-allocated as needed, avoiding the need to use
skb_copy_to_linear_data() as in "qpl" mode.

Reviewed-by: Yangchun Fu <yangchun@google.com>
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David Awogbemila <awogbemila@google.com>
---
 drivers/net/ethernet/google/gve/gve.h        |  12 +-
 drivers/net/ethernet/google/gve/gve_adminq.c |  14 +-
 drivers/net/ethernet/google/gve/gve_desc.h   |  11 +-
 drivers/net/ethernet/google/gve/gve_main.c   |   2 +-
 drivers/net/ethernet/google/gve/gve_rx.c     | 205 ++++++++++++++-----
 5 files changed, 182 insertions(+), 62 deletions(-)

diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
index 782e2791ee11..d8bba0ba34e3 100644
--- a/drivers/net/ethernet/google/gve/gve.h
+++ b/drivers/net/ethernet/google/gve/gve.h
@@ -38,6 +38,8 @@
 #define NIC_TX_STATS_REPORT_NUM	0
 #define NIC_RX_STATS_REPORT_NUM	4
 
+#define GVE_DATA_SLOT_ADDR_PAGE_MASK (~(PAGE_SIZE - 1))
+
 /* 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 */
@@ -49,7 +51,7 @@ struct gve_rx_desc_queue {
 struct gve_rx_slot_page_info {
 	struct page *page;
 	void *page_address;
-	u32 page_offset; /* offset to write to in page */
+	u8 page_offset; /* flipped to second half? */
 };
 
 /* A list of pages registered with the device during setup and used by a queue
@@ -64,10 +66,11 @@ struct gve_queue_page_list {
 
 /* Each slot in the data ring has a 1:1 mapping to a slot in the desc ring */
 struct gve_rx_data_queue {
-	struct gve_rx_data_slot *data_ring; /* read by NIC */
+	union gve_rx_data_slot *data_ring; /* read by NIC */
 	dma_addr_t data_bus; /* dma mapping of the slots */
 	struct gve_rx_slot_page_info *page_info; /* page info of the buffers */
 	struct gve_queue_page_list *qpl; /* qpl assigned to this queue */
+	u8 raw_addressing; /* use raw_addressing? */
 };
 
 struct gve_priv;
@@ -82,6 +85,7 @@ struct gve_rx_ring {
 	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 */
+	u32 db_threshold; /* threshold for posting new buffs and descs */
 	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 */
@@ -194,7 +198,7 @@ struct gve_priv {
 	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 rx_pages_per_qpl; /* rx buffer length */
+	u16 rx_data_slot_cnt; /* rx buffer length */
 	u64 max_registered_pages;
 	u64 num_registered_pages; /* num pages registered with NIC */
 	u32 rx_copybreak; /* copy packets smaller than this */
@@ -444,7 +448,7 @@ static inline u32 gve_num_tx_qpls(struct gve_priv *priv)
  */
 static inline u32 gve_num_rx_qpls(struct gve_priv *priv)
 {
-	return priv->rx_cfg.num_queues;
+	return priv->raw_addressing ? 0 : priv->rx_cfg.num_queues;
 }
 
 /* Returns a pointer to the next available tx qpl in the list of qpls
diff --git a/drivers/net/ethernet/google/gve/gve_adminq.c b/drivers/net/ethernet/google/gve/gve_adminq.c
index 608a8806f8c8..a1b9370db1e4 100644
--- a/drivers/net/ethernet/google/gve/gve_adminq.c
+++ b/drivers/net/ethernet/google/gve/gve_adminq.c
@@ -408,8 +408,10 @@ static int gve_adminq_create_rx_queue(struct gve_priv *priv, u32 queue_index)
 {
 	struct gve_rx_ring *rx = &priv->rx[queue_index];
 	union gve_adminq_command cmd;
+	u32 qpl_id;
 	int err;
 
+	qpl_id = priv->raw_addressing ? GVE_RAW_ADDRESSING_QPL_ID : rx->data.qpl->id;
 	memset(&cmd, 0, sizeof(cmd));
 	cmd.opcode = cpu_to_be32(GVE_ADMINQ_CREATE_RX_QUEUE);
 	cmd.create_rx_queue = (struct gve_adminq_create_rx_queue) {
@@ -420,7 +422,7 @@ static int gve_adminq_create_rx_queue(struct gve_priv *priv, u32 queue_index)
 		.queue_resources_addr = cpu_to_be64(rx->q_resources_bus),
 		.rx_desc_ring_addr = cpu_to_be64(rx->desc.bus),
 		.rx_data_ring_addr = cpu_to_be64(rx->data.data_bus),
-		.queue_page_list_id = cpu_to_be32(rx->data.qpl->id),
+		.queue_page_list_id = cpu_to_be32(qpl_id),
 	};
 
 	err = gve_adminq_issue_cmd(priv, &cmd);
@@ -565,11 +567,11 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 	mac = descriptor->mac;
 	dev_info(&priv->pdev->dev, "MAC addr: %pM\n", mac);
 	priv->tx_pages_per_qpl = be16_to_cpu(descriptor->tx_pages_per_qpl);
-	priv->rx_pages_per_qpl = be16_to_cpu(descriptor->rx_pages_per_qpl);
-	if (priv->rx_pages_per_qpl < priv->rx_desc_cnt) {
-		dev_err(&priv->pdev->dev, "rx_pages_per_qpl cannot be smaller than rx_desc_cnt, setting rx_desc_cnt down to %d.\n",
-			priv->rx_pages_per_qpl);
-		priv->rx_desc_cnt = priv->rx_pages_per_qpl;
+	priv->rx_data_slot_cnt = be16_to_cpu(descriptor->rx_pages_per_qpl);
+	if (priv->rx_data_slot_cnt < priv->rx_desc_cnt) {
+		dev_err(&priv->pdev->dev, "rx_data_slot_cnt cannot be smaller than rx_desc_cnt, setting rx_desc_cnt down to %d.\n",
+			priv->rx_data_slot_cnt);
+		priv->rx_desc_cnt = priv->rx_data_slot_cnt;
 	}
 	priv->default_num_queues = be16_to_cpu(descriptor->default_num_queues);
 	dev_opt = (void *)(descriptor + 1);
diff --git a/drivers/net/ethernet/google/gve/gve_desc.h b/drivers/net/ethernet/google/gve/gve_desc.h
index 54779871d52e..a1c0aaa60139 100644
--- a/drivers/net/ethernet/google/gve/gve_desc.h
+++ b/drivers/net/ethernet/google/gve/gve_desc.h
@@ -72,12 +72,15 @@ struct gve_rx_desc {
 } __packed;
 static_assert(sizeof(struct gve_rx_desc) == 64);
 
-/* As with the Tx ring format, the qpl_offset entries below are offsets into an
- * ordered list of registered pages.
+/* If the device supports raw dma addressing then the addr in data slot is
+ * the dma address of the buffer.
+ * If the device only supports registered segments then the addr is a byte
+ * offset into the registered segment (an ordered list of pages) where the
+ * buffer is.
  */
-struct gve_rx_data_slot {
-	/* byte offset into the rx registered segment of this slot */
+union gve_rx_data_slot {
 	__be64 qpl_offset;
+	__be64 addr;
 };
 
 /* GVE Recive Packet Descriptor Seq No */
diff --git a/drivers/net/ethernet/google/gve/gve_main.c b/drivers/net/ethernet/google/gve/gve_main.c
index 70685c10db0e..b7dae4f7b2fd 100644
--- a/drivers/net/ethernet/google/gve/gve_main.c
+++ b/drivers/net/ethernet/google/gve/gve_main.c
@@ -694,7 +694,7 @@ static int gve_alloc_qpls(struct gve_priv *priv)
 	}
 	for (; i < num_qpls; i++) {
 		err = gve_alloc_queue_page_list(priv, i,
-						priv->rx_pages_per_qpl);
+						priv->rx_data_slot_cnt);
 		if (err)
 			goto free_qpls;
 	}
diff --git a/drivers/net/ethernet/google/gve/gve_rx.c b/drivers/net/ethernet/google/gve/gve_rx.c
index 008fa897a3e6..596772f5e29a 100644
--- a/drivers/net/ethernet/google/gve/gve_rx.c
+++ b/drivers/net/ethernet/google/gve/gve_rx.c
@@ -16,12 +16,39 @@ static void gve_rx_remove_from_block(struct gve_priv *priv, int queue_idx)
 	block->rx = NULL;
 }
 
+static void gve_rx_free_buffer(struct device *dev,
+			       struct gve_rx_slot_page_info *page_info,
+			       union gve_rx_data_slot *data_slot)
+{
+	dma_addr_t dma = (dma_addr_t)(be64_to_cpu(data_slot->addr) &
+				      GVE_DATA_SLOT_ADDR_PAGE_MASK);
+
+	gve_free_page(dev, page_info->page, dma, DMA_FROM_DEVICE);
+}
+
+static void gve_rx_unfill_pages(struct gve_priv *priv, struct gve_rx_ring *rx)
+{
+	if (rx->data.raw_addressing) {
+		u32 slots = rx->mask + 1;
+		int i;
+
+		for (i = 0; i < slots; i++)
+			gve_rx_free_buffer(&priv->pdev->dev, &rx->data.page_info[i],
+					   &rx->data.data_ring[i]);
+	} else {
+		gve_unassign_qpl(priv, rx->data.qpl->id);
+		rx->data.qpl = NULL;
+	}
+	kvfree(rx->data.page_info);
+	rx->data.page_info = NULL;
+}
+
 static void gve_rx_free_ring(struct gve_priv *priv, int idx)
 {
 	struct gve_rx_ring *rx = &priv->rx[idx];
 	struct device *dev = &priv->pdev->dev;
+	u32 slots = rx->mask + 1;
 	size_t bytes;
-	u32 slots;
 
 	gve_rx_remove_from_block(priv, idx);
 
@@ -33,11 +60,8 @@ static void gve_rx_free_ring(struct gve_priv *priv, int idx)
 			  rx->q_resources, rx->q_resources_bus);
 	rx->q_resources = NULL;
 
-	gve_unassign_qpl(priv, rx->data.qpl->id);
-	rx->data.qpl = NULL;
-	kvfree(rx->data.page_info);
+	gve_rx_unfill_pages(priv, rx);
 
-	slots = rx->mask + 1;
 	bytes = sizeof(*rx->data.data_ring) * slots;
 	dma_free_coherent(dev, bytes, rx->data.data_ring,
 			  rx->data.data_bus);
@@ -46,19 +70,35 @@ static void gve_rx_free_ring(struct gve_priv *priv, int idx)
 }
 
 static void gve_setup_rx_buffer(struct gve_rx_slot_page_info *page_info,
-				struct gve_rx_data_slot *slot,
-				dma_addr_t addr, struct page *page)
+			     dma_addr_t addr, struct page *page, __be64 *slot_addr)
 {
 	page_info->page = page;
 	page_info->page_offset = 0;
 	page_info->page_address = page_address(page);
-	slot->qpl_offset = cpu_to_be64(addr);
+	*slot_addr = cpu_to_be64(addr);
+}
+
+static int gve_rx_alloc_buffer(struct gve_priv *priv, struct device *dev,
+			       struct gve_rx_slot_page_info *page_info,
+			       union gve_rx_data_slot *data_slot)
+{
+	struct page *page;
+	dma_addr_t dma;
+	int err;
+
+	err = gve_alloc_page(priv, dev, &page, &dma, DMA_FROM_DEVICE);
+	if (err)
+		return err;
+
+	gve_setup_rx_buffer(page_info, dma, page, &data_slot->addr);
+	return 0;
 }
 
 static int gve_prefill_rx_pages(struct gve_rx_ring *rx)
 {
 	struct gve_priv *priv = rx->gve;
 	u32 slots;
+	int err;
 	int i;
 
 	/* Allocate one page per Rx queue slot. Each page is split into two
@@ -71,17 +111,30 @@ static int gve_prefill_rx_pages(struct gve_rx_ring *rx)
 	if (!rx->data.page_info)
 		return -ENOMEM;
 
-	rx->data.qpl = gve_assign_rx_qpl(priv);
-
+	if (!rx->data.raw_addressing)
+		rx->data.qpl = gve_assign_rx_qpl(priv);
 	for (i = 0; i < slots; i++) {
-		struct page *page = rx->data.qpl->pages[i];
-		dma_addr_t addr = i * PAGE_SIZE;
+		if (!rx->data.raw_addressing) {
+			struct page *page = rx->data.qpl->pages[i];
+			dma_addr_t addr = i * PAGE_SIZE;
 
-		gve_setup_rx_buffer(&rx->data.page_info[i],
-				    &rx->data.data_ring[i], addr, page);
+			gve_setup_rx_buffer(&rx->data.page_info[i], addr, page,
+					    &rx->data.data_ring[i].qpl_offset);
+			continue;
+		}
+		err = gve_rx_alloc_buffer(priv, &priv->pdev->dev, &rx->data.page_info[i],
+					  &rx->data.data_ring[i]);
+		if (err)
+			goto alloc_err;
 	}
 
 	return slots;
+alloc_err:
+	while (i--)
+		gve_rx_free_buffer(&priv->pdev->dev,
+				   &rx->data.page_info[i],
+				   &rx->data.data_ring[i]);
+	return err;
 }
 
 static void gve_rx_add_to_block(struct gve_priv *priv, int queue_idx)
@@ -110,8 +163,9 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 	rx->gve = priv;
 	rx->q_num = idx;
 
-	slots = priv->rx_pages_per_qpl;
+	slots = priv->rx_data_slot_cnt;
 	rx->mask = slots - 1;
+	rx->data.raw_addressing = priv->raw_addressing;
 
 	/* alloc rx data ring */
 	bytes = sizeof(*rx->data.data_ring) * slots;
@@ -156,8 +210,8 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 		err = -ENOMEM;
 		goto abort_with_q_resources;
 	}
-	rx->mask = slots - 1;
 	rx->cnt = 0;
+	rx->db_threshold = priv->rx_desc_cnt / 2;
 	rx->desc.seqno = 1;
 	gve_rx_add_to_block(priv, idx);
 
@@ -168,7 +222,7 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 			  rx->q_resources, rx->q_resources_bus);
 	rx->q_resources = NULL;
 abort_filled:
-	kvfree(rx->data.page_info);
+	gve_rx_unfill_pages(priv, rx);
 abort_with_slots:
 	bytes = sizeof(*rx->data.data_ring) * slots;
 	dma_free_coherent(hdev, bytes, rx->data.data_ring, rx->data.data_bus);
@@ -233,7 +287,7 @@ static struct sk_buff *gve_rx_copy(struct gve_rx_ring *rx,
 {
 	struct sk_buff *skb = napi_alloc_skb(napi, len);
 	void *va = page_info->page_address + GVE_RX_PAD +
-		   page_info->page_offset;
+		   (page_info->page_offset ? PAGE_SIZE / 2 : 0);
 
 	if (unlikely(!skb))
 		return NULL;
@@ -251,8 +305,7 @@ static struct sk_buff *gve_rx_copy(struct gve_rx_ring *rx,
 	return skb;
 }
 
-static struct sk_buff *gve_rx_add_frags(struct net_device *dev,
-					struct napi_struct *napi,
+static struct sk_buff *gve_rx_add_frags(struct napi_struct *napi,
 					struct gve_rx_slot_page_info *page_info,
 					u16 len)
 {
@@ -262,20 +315,19 @@ static struct sk_buff *gve_rx_add_frags(struct net_device *dev,
 		return NULL;
 
 	skb_add_rx_frag(skb, 0, page_info->page,
-			page_info->page_offset +
+			(page_info->page_offset ? PAGE_SIZE / 2 : 0) +
 			GVE_RX_PAD, len, PAGE_SIZE / 2);
 
 	return skb;
 }
 
-static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info,
-			     struct gve_rx_data_slot *data_ring)
+static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info, __be64 *slot_addr)
 {
-	u64 addr = be64_to_cpu(data_ring->qpl_offset);
+	const __be64 offset = cpu_to_be64(PAGE_SIZE / 2);
 
-	page_info->page_offset ^= PAGE_SIZE / 2;
-	addr ^= PAGE_SIZE / 2;
-	data_ring->qpl_offset = cpu_to_be64(addr);
+	/* "flip" to other packet buffer on this page */
+	page_info->page_offset ^= 0x1;
+	*(slot_addr) ^= offset;
 }
 
 static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
@@ -285,7 +337,9 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 	struct gve_priv *priv = rx->gve;
 	struct napi_struct *napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
 	struct net_device *dev = priv->dev;
-	struct sk_buff *skb;
+	union gve_rx_data_slot *data_slot;
+	struct sk_buff *skb = NULL;
+	dma_addr_t page_bus;
 	int pagecount;
 	u16 len;
 
@@ -294,18 +348,18 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 		u64_stats_update_begin(&rx->statss);
 		rx->rx_desc_err_dropped_pkt++;
 		u64_stats_update_end(&rx->statss);
-		return true;
+		return false;
 	}
 
 	len = be16_to_cpu(rx_desc->len) - GVE_RX_PAD;
 	page_info = &rx->data.page_info[idx];
-	dma_sync_single_for_cpu(&priv->pdev->dev, rx->data.qpl->page_buses[idx],
-				PAGE_SIZE, DMA_FROM_DEVICE);
 
-	/* gvnic can only receive into registered segments. If the buffer
-	 * can't be recycled, our only choice is to copy the data out of
-	 * it so that we can return it to the device.
-	 */
+	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);
 
 	if (PAGE_SIZE == 4096) {
 		if (len <= priv->rx_copybreak) {
@@ -316,6 +370,10 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 			u64_stats_update_end(&rx->statss);
 			goto have_skb;
 		}
+		if (rx->data.raw_addressing) {
+			skb = gve_rx_add_frags(napi, page_info, len);
+			goto have_skb;
+		}
 		if (unlikely(!gve_can_recycle_pages(dev))) {
 			skb = gve_rx_copy(rx, dev, napi, page_info, len);
 			goto have_skb;
@@ -326,17 +384,17 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 			 * the page fragment to a new SKB and pass it up the
 			 * stack.
 			 */
-			skb = gve_rx_add_frags(dev, napi, page_info, len);
+			skb = gve_rx_add_frags(napi, page_info, len);
 			if (!skb) {
 				u64_stats_update_begin(&rx->statss);
 				rx->rx_skb_alloc_fail++;
 				u64_stats_update_end(&rx->statss);
-				return true;
+				return false;
 			}
 			/* Make sure the kernel stack can't release the page */
 			get_page(page_info->page);
 			/* "flip" to other packet buffer on this page */
-			gve_rx_flip_buff(page_info, &rx->data.data_ring[idx]);
+			gve_rx_flip_buff(page_info, &rx->data.data_ring[idx].qpl_offset);
 		} else if (pagecount >= 2) {
 			/* We have previously passed the other half of this
 			 * page up the stack, but it has not yet been freed.
@@ -347,7 +405,10 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 			return false;
 		}
 	} else {
-		skb = gve_rx_copy(rx, dev, napi, page_info, len);
+		if (rx->data.raw_addressing)
+			skb = gve_rx_add_frags(napi, page_info, len);
+		else
+			skb = gve_rx_copy(rx, dev, napi, page_info, len);
 	}
 
 have_skb:
@@ -358,7 +419,7 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 		u64_stats_update_begin(&rx->statss);
 		rx->rx_skb_alloc_fail++;
 		u64_stats_update_end(&rx->statss);
-		return true;
+		return false;
 	}
 
 	if (likely(feat & NETIF_F_RXCSUM)) {
@@ -399,19 +460,48 @@ static bool gve_rx_work_pending(struct gve_rx_ring *rx)
 	return (GVE_SEQNO(flags_seq) == rx->desc.seqno);
 }
 
+static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
+{
+	int refill_target = rx->mask + 1;
+	u32 fill_cnt = rx->fill_cnt;
+
+	while (fill_cnt - rx->cnt < refill_target) {
+		struct gve_rx_slot_page_info *page_info;
+		struct device *dev = &priv->pdev->dev;
+		union gve_rx_data_slot *data_slot;
+		u32 idx = fill_cnt & rx->mask;
+
+		page_info = &rx->data.page_info[idx];
+		data_slot = &rx->data.data_ring[idx];
+		gve_rx_free_buffer(dev, page_info, data_slot);
+		page_info->page = NULL;
+		if (gve_rx_alloc_buffer(priv, dev, page_info, data_slot)) {
+			u64_stats_update_begin(&rx->statss);
+			rx->rx_buf_alloc_fail++;
+			u64_stats_update_end(&rx->statss);
+			break;
+		}
+		fill_cnt++;
+	}
+	rx->fill_cnt = fill_cnt;
+	return true;
+}
+
 bool gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
 		       netdev_features_t feat)
 {
 	struct gve_priv *priv = rx->gve;
+	u32 work_done = 0, packets = 0;
 	struct gve_rx_desc *desc;
 	u32 cnt = rx->cnt;
 	u32 idx = cnt & rx->mask;
-	u32 work_done = 0;
 	u64 bytes = 0;
 
 	desc = rx->desc.desc_ring + idx;
 	while ((GVE_SEQNO(desc->flags_seq) == rx->desc.seqno) &&
 	       work_done < budget) {
+		bool dropped;
+
 		netif_info(priv, rx_status, priv->dev,
 			   "[%d] idx=%d desc=%p desc->flags_seq=0x%x\n",
 			   rx->q_num, idx, desc, desc->flags_seq);
@@ -419,9 +509,11 @@ bool gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
 			   "[%d] seqno=%d rx->desc.seqno=%d\n",
 			   rx->q_num, GVE_SEQNO(desc->flags_seq),
 			   rx->desc.seqno);
-		bytes += be16_to_cpu(desc->len) - GVE_RX_PAD;
-		if (!gve_rx(rx, desc, feat, idx))
-			gve_schedule_reset(priv);
+		dropped = !gve_rx(rx, desc, feat, idx);
+		if (!dropped) {
+			bytes += be16_to_cpu(desc->len) - GVE_RX_PAD;
+			packets++;
+		}
 		cnt++;
 		idx = cnt & rx->mask;
 		desc = rx->desc.desc_ring + idx;
@@ -429,15 +521,34 @@ bool gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
 		work_done++;
 	}
 
-	if (!work_done)
+	if (!work_done && rx->fill_cnt - cnt > rx->db_threshold)
 		return false;
 
 	u64_stats_update_begin(&rx->statss);
-	rx->rpackets += work_done;
+	rx->rpackets += packets;
 	rx->rbytes += bytes;
 	u64_stats_update_end(&rx->statss);
 	rx->cnt = cnt;
-	rx->fill_cnt += work_done;
+
+	/* 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) {
+		/* In raw addressing mode buffs are only refilled if the avail
+		 * falls below a threshold.
+		 */
+		if (!gve_rx_refill_buffers(priv, rx))
+			return false;
+
+		/* 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) {
+			gve_rx_write_doorbell(priv, rx);
+			return true;
+		}
+	}
 
 	gve_rx_write_doorbell(priv, rx);
 	return gve_rx_work_pending(rx);
-- 
2.29.2.576.ga3fc446d84-goog

[PATCH net-next v9 3/4] gve: Rx Buffer Recycling

[David Awogbemila]

This patch lets the driver reuse buffers that have been freed by the
networking stack.

In the raw addressing case, this allows the driver avoid allocating new
buffers.
In the qpl case, the driver can avoid copies.

This patch separates the page refcount tracking mechanism
into a function gve_rx_can_recycle_buffer which uses get_page - this will
be changed in a future patch to entirely eliminate the use of get_page in
tracking page refcounts.

Signed-off-by: David Awogbemila <awogbemila@google.com>
---
 drivers/net/ethernet/google/gve/gve.h    |  10 +-
 drivers/net/ethernet/google/gve/gve_rx.c | 199 +++++++++++++++--------
 2 files changed, 135 insertions(+), 74 deletions(-)

diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
index d8bba0ba34e3..8aad4af2aa2b 100644
--- a/drivers/net/ethernet/google/gve/gve.h
+++ b/drivers/net/ethernet/google/gve/gve.h
@@ -52,6 +52,7 @@ struct gve_rx_slot_page_info {
 	struct page *page;
 	void *page_address;
 	u8 page_offset; /* flipped to second half? */
+	u8 can_flip;
 };
 
 /* A list of pages registered with the device during setup and used by a queue
@@ -502,15 +503,6 @@ static inline enum dma_data_direction gve_qpl_dma_dir(struct gve_priv *priv,
 		return DMA_FROM_DEVICE;
 }
 
-/* Returns true if the max mtu allows page recycling */
-static inline bool gve_can_recycle_pages(struct net_device *dev)
-{
-	/* We can't recycle the pages if we can't fit a packet into half a
-	 * page.
-	 */
-	return dev->max_mtu <= PAGE_SIZE / 2;
-}
-
 /* buffers */
 int gve_alloc_page(struct gve_priv *priv, struct device *dev,
 		   struct page **page, dma_addr_t *dma,
diff --git a/drivers/net/ethernet/google/gve/gve_rx.c b/drivers/net/ethernet/google/gve/gve_rx.c
index 596772f5e29a..bf123fe524c4 100644
--- a/drivers/net/ethernet/google/gve/gve_rx.c
+++ b/drivers/net/ethernet/google/gve/gve_rx.c
@@ -279,8 +279,7 @@ static enum pkt_hash_types gve_rss_type(__be16 pkt_flags)
 	return PKT_HASH_TYPE_L2;
 }
 
-static struct sk_buff *gve_rx_copy(struct gve_rx_ring *rx,
-				   struct net_device *dev,
+static struct sk_buff *gve_rx_copy(struct net_device *dev,
 				   struct napi_struct *napi,
 				   struct gve_rx_slot_page_info *page_info,
 				   u16 len)
@@ -298,10 +297,6 @@ static struct sk_buff *gve_rx_copy(struct gve_rx_ring *rx,
 
 	skb->protocol = eth_type_trans(skb, dev);
 
-	u64_stats_update_begin(&rx->statss);
-	rx->rx_copied_pkt++;
-	u64_stats_update_end(&rx->statss);
-
 	return skb;
 }
 
@@ -330,6 +325,79 @@ 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 page *page)
+{
+	int pagecount = page_count(page);
+
+	/* This page is not being used by any SKBs - reuse */
+	if (pagecount == 1)
+		return 1;
+	/* This page is still being used by an SKB - we can't reuse */
+	else if (pagecount >= 2)
+		return 0;
+	WARN(pagecount < 1, "Pagecount should never be < 1");
+	return -1;
+}
+
+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)
+{
+	struct sk_buff *skb;
+
+	skb = gve_rx_add_frags(napi, page_info, len);
+	if (!skb)
+		return NULL;
+
+	/* Optimistically stop the kernel from freeing the page by increasing
+	 * the page bias. We will check the refcount in refill to determine if
+	 * we need to alloc a new page.
+	 */
+	get_page(page_info->page);
+
+	return skb;
+}
+
+static struct sk_buff *
+gve_rx_qpl(struct device *dev, struct net_device *netdev,
+	   struct gve_rx_ring *rx, struct gve_rx_slot_page_info *page_info,
+	   u16 len, struct napi_struct *napi,
+	   union gve_rx_data_slot *data_slot)
+{
+	struct sk_buff *skb;
+
+	/* if raw_addressing mode is not enabled gvnic can only receive into
+	 * registered segments. If the buffer can't be recycled, our only
+	 * 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);
+		/* No point in recycling if we didn't get the skb */
+		if (skb) {
+			/* Make sure that the page isn't freed. */
+			get_page(page_info->page);
+			gve_rx_flip_buff(page_info, &data_slot->qpl_offset);
+		}
+	} else {
+		skb = gve_rx_copy(netdev, napi, page_info, len);
+		if (skb) {
+			u64_stats_update_begin(&rx->statss);
+			rx->rx_copied_pkt++;
+			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)
 {
@@ -340,7 +408,6 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 	union gve_rx_data_slot *data_slot;
 	struct sk_buff *skb = NULL;
 	dma_addr_t page_bus;
-	int pagecount;
 	u16 len;
 
 	/* drop this packet */
@@ -361,60 +428,37 @@ static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 	dma_sync_single_for_cpu(&priv->pdev->dev, page_bus,
 				PAGE_SIZE, DMA_FROM_DEVICE);
 
-	if (PAGE_SIZE == 4096) {
-		if (len <= priv->rx_copybreak) {
-			/* Just copy small packets */
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
-			u64_stats_update_begin(&rx->statss);
-			rx->rx_copybreak_pkt++;
-			u64_stats_update_end(&rx->statss);
-			goto have_skb;
-		}
-		if (rx->data.raw_addressing) {
-			skb = gve_rx_add_frags(napi, page_info, len);
-			goto have_skb;
-		}
-		if (unlikely(!gve_can_recycle_pages(dev))) {
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
-			goto have_skb;
-		}
-		pagecount = page_count(page_info->page);
-		if (pagecount == 1) {
-			/* No part of this page is used by any SKBs; we attach
-			 * the page fragment to a new SKB and pass it up the
-			 * stack.
-			 */
-			skb = gve_rx_add_frags(napi, page_info, len);
-			if (!skb) {
-				u64_stats_update_begin(&rx->statss);
-				rx->rx_skb_alloc_fail++;
-				u64_stats_update_end(&rx->statss);
+	if (len <= priv->rx_copybreak) {
+		/* Just copy small packets */
+		skb = gve_rx_copy(dev, napi, page_info, len);
+		u64_stats_update_begin(&rx->statss);
+		rx->rx_copied_pkt++;
+		rx->rx_copybreak_pkt++;
+		u64_stats_update_end(&rx->statss);
+	} else {
+		u8 can_flip = gve_rx_can_flip_buffers(dev);
+		int recycle = 0;
+
+		if (can_flip) {
+			recycle = gve_rx_can_recycle_buffer(page_info->page);
+			if (recycle < 0) {
+				if (!rx->data.raw_addressing)
+					gve_schedule_reset(priv);
 				return false;
 			}
-			/* Make sure the kernel stack can't release the page */
-			get_page(page_info->page);
-			/* "flip" to other packet buffer on this page */
-			gve_rx_flip_buff(page_info, &rx->data.data_ring[idx].qpl_offset);
-		} else if (pagecount >= 2) {
-			/* We have previously passed the other half of this
-			 * page up the stack, but it has not yet been freed.
-			 */
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
+		}
+
+		page_info->can_flip = can_flip && recycle;
+		if (rx->data.raw_addressing) {
+			skb = gve_rx_raw_addressing(&priv->pdev->dev, dev,
+						    page_info, len, napi,
+						    data_slot);
 		} else {
-			WARN(pagecount < 1, "Pagecount should never be < 1");
-			return false;
+			skb = gve_rx_qpl(&priv->pdev->dev, dev, rx,
+					 page_info, len, napi, data_slot);
 		}
-	} else {
-		if (rx->data.raw_addressing)
-			skb = gve_rx_add_frags(napi, page_info, len);
-		else
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
 	}
 
-have_skb:
-	/* We didn't manage to allocate an skb but we haven't had any
-	 * reset worthy failures.
-	 */
 	if (!skb) {
 		u64_stats_update_begin(&rx->statss);
 		rx->rx_skb_alloc_fail++;
@@ -467,19 +511,44 @@ static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
 
 	while (fill_cnt - rx->cnt < refill_target) {
 		struct gve_rx_slot_page_info *page_info;
-		struct device *dev = &priv->pdev->dev;
-		union gve_rx_data_slot *data_slot;
 		u32 idx = fill_cnt & rx->mask;
 
 		page_info = &rx->data.page_info[idx];
-		data_slot = &rx->data.data_ring[idx];
-		gve_rx_free_buffer(dev, page_info, data_slot);
-		page_info->page = NULL;
-		if (gve_rx_alloc_buffer(priv, dev, page_info, data_slot)) {
-			u64_stats_update_begin(&rx->statss);
-			rx->rx_buf_alloc_fail++;
-			u64_stats_update_end(&rx->statss);
-			break;
+		if (page_info->can_flip) {
+			/* The other half of the page is free because it was
+			 * free when we processed the descriptor. Flip to it.
+			 */
+			union gve_rx_data_slot *data_slot =
+						&rx->data.data_ring[idx];
+
+			gve_rx_flip_buff(page_info, &data_slot->addr);
+			page_info->can_flip = 0;
+		} else {
+			/* It is possible that the networking stack has already
+			 * finished processing all outstanding packets in the buffer
+			 * and it can be reused.
+			 * Flipping is unnecessary here - if the networking stack still
+			 * owns half the page it is impossible to tell which half. Either
+			 * the whole page is free or it needs to be replaced.
+			 */
+			int recycle = gve_rx_can_recycle_buffer(page_info->page);
+
+			if (recycle < 0) {
+				if (!rx->data.raw_addressing)
+					gve_schedule_reset(priv);
+				return false;
+			}
+			if (!recycle) {
+				/* We can't reuse the buffer - alloc a new one*/
+				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, data_slot))
+					break;
+			}
 		}
 		fill_cnt++;
 	}
-- 
2.29.2.576.ga3fc446d84-goog

[PATCH net-next v9 4/4] gve: Add support for raw addressing in the tx path

[David Awogbemila]

From: Catherine Sullivan <csully@google.com>

During TX, skbs' data addresses are dma_map'ed and passed to the NIC.
This means that the device can perform DMA directly from these addresses
and the driver does not have to copy the buffer content into
pre-allocated buffers/qpls (as in qpl mode).

Reviewed-by: Yangchun Fu <yangchun@google.com>
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David Awogbemila <awogbemila@google.com>
---
 drivers/net/ethernet/google/gve/gve.h         |  16 +-
 drivers/net/ethernet/google/gve/gve_adminq.c  |   4 +-
 drivers/net/ethernet/google/gve/gve_desc.h    |   8 +-
 drivers/net/ethernet/google/gve/gve_ethtool.c |   2 +
 drivers/net/ethernet/google/gve/gve_tx.c      | 197 ++++++++++++++----
 5 files changed, 185 insertions(+), 42 deletions(-)

diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
index 8aad4af2aa2b..9888fa92be86 100644
--- a/drivers/net/ethernet/google/gve/gve.h
+++ b/drivers/net/ethernet/google/gve/gve.h
@@ -112,12 +112,20 @@ struct gve_tx_iovec {
 	u32 iov_padding; /* padding associated with this segment */
 };
 
+struct gve_tx_dma_buf {
+	DEFINE_DMA_UNMAP_ADDR(dma);
+	DEFINE_DMA_UNMAP_LEN(len);
+};
+
 /* Tracks the memory in the fifo occupied by the skb. Mapped 1:1 to a desc
  * ring entry but only used for a pkt_desc not a seg_desc
  */
 struct gve_tx_buffer_state {
 	struct sk_buff *skb; /* skb for this pkt */
-	struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
+	union {
+		struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
+		struct gve_tx_dma_buf buf;
+	};
 };
 
 /* A TX buffer - each queue has one */
@@ -140,19 +148,23 @@ struct gve_tx_ring {
 	__be32 last_nic_done ____cacheline_aligned; /* NIC tail pointer */
 	u64 pkt_done; /* free-running - total packets completed */
 	u64 bytes_done; /* free-running - total bytes completed */
+	u32 dropped_pkt; /* free-running - total packets dropped */
 
 	/* Cacheline 2 -- Read-mostly fields */
 	union gve_tx_desc *desc ____cacheline_aligned;
 	struct gve_tx_buffer_state *info; /* Maps 1:1 to a desc */
 	struct netdev_queue *netdev_txq;
 	struct gve_queue_resources *q_resources; /* head and tail pointer idx */
+	struct device *dev;
 	u32 mask; /* masks req and done down to queue size */
+	u8 raw_addressing; /* use raw_addressing? */
 
 	/* Slow-path fields */
 	u32 q_num ____cacheline_aligned; /* queue idx */
 	u32 stop_queue; /* count of queue stops */
 	u32 wake_queue; /* count of queue wakes */
 	u32 ntfy_id; /* notification block index */
+	u32 dma_mapping_error; /* count of dma mapping errors */
 	dma_addr_t bus; /* dma address of the descr ring */
 	dma_addr_t q_resources_bus; /* dma address of the queue resources */
 	struct u64_stats_sync statss; /* sync stats for 32bit archs */
@@ -442,7 +454,7 @@ static inline u32 gve_rx_idx_to_ntfy(struct gve_priv *priv, u32 queue_idx)
  */
 static inline u32 gve_num_tx_qpls(struct gve_priv *priv)
 {
-	return priv->tx_cfg.num_queues;
+	return priv->raw_addressing ? 0 : priv->tx_cfg.num_queues;
 }
 
 /* Returns the number of rx queue page lists
diff --git a/drivers/net/ethernet/google/gve/gve_adminq.c b/drivers/net/ethernet/google/gve/gve_adminq.c
index a1b9370db1e4..53864f200599 100644
--- a/drivers/net/ethernet/google/gve/gve_adminq.c
+++ b/drivers/net/ethernet/google/gve/gve_adminq.c
@@ -369,8 +369,10 @@ static int gve_adminq_create_tx_queue(struct gve_priv *priv, u32 queue_index)
 {
 	struct gve_tx_ring *tx = &priv->tx[queue_index];
 	union gve_adminq_command cmd;
+	u32 qpl_id;
 	int err;
 
+	qpl_id = priv->raw_addressing ? GVE_RAW_ADDRESSING_QPL_ID : tx->tx_fifo.qpl->id;
 	memset(&cmd, 0, sizeof(cmd));
 	cmd.opcode = cpu_to_be32(GVE_ADMINQ_CREATE_TX_QUEUE);
 	cmd.create_tx_queue = (struct gve_adminq_create_tx_queue) {
@@ -379,7 +381,7 @@ static int gve_adminq_create_tx_queue(struct gve_priv *priv, u32 queue_index)
 		.queue_resources_addr =
 			cpu_to_be64(tx->q_resources_bus),
 		.tx_ring_addr = cpu_to_be64(tx->bus),
-		.queue_page_list_id = cpu_to_be32(tx->tx_fifo.qpl->id),
+		.queue_page_list_id = cpu_to_be32(qpl_id),
 		.ntfy_id = cpu_to_be32(tx->ntfy_id),
 	};
 
diff --git a/drivers/net/ethernet/google/gve/gve_desc.h b/drivers/net/ethernet/google/gve/gve_desc.h
index a1c0aaa60139..05ae6300e984 100644
--- a/drivers/net/ethernet/google/gve/gve_desc.h
+++ b/drivers/net/ethernet/google/gve/gve_desc.h
@@ -16,9 +16,11 @@
  * Base addresses encoded in seg_addr are not assumed to be physical
  * addresses. The ring format assumes these come from some linear address
  * space. This could be physical memory, kernel virtual memory, user virtual
- * memory. gVNIC uses lists of registered pages. Each queue is assumed
- * to be associated with a single such linear address space to ensure a
- * consistent meaning for seg_addrs posted to its rings.
+ * memory.
+ * If raw dma addressing is not supported then gVNIC uses lists of registered
+ * pages. Each queue is assumed to be associated with a single such linear
+ * address space to ensure a consistent meaning for seg_addrs posted to its
+ * rings.
  */
 
 struct gve_tx_pkt_desc {
diff --git a/drivers/net/ethernet/google/gve/gve_ethtool.c b/drivers/net/ethernet/google/gve/gve_ethtool.c
index 7b44769bd87c..07ec957ae091 100644
--- a/drivers/net/ethernet/google/gve/gve_ethtool.c
+++ b/drivers/net/ethernet/google/gve/gve_ethtool.c
@@ -50,6 +50,7 @@ static const char gve_gstrings_rx_stats[][ETH_GSTRING_LEN] = {
 static const char gve_gstrings_tx_stats[][ETH_GSTRING_LEN] = {
 	"tx_posted_desc[%u]", "tx_completed_desc[%u]", "tx_bytes[%u]",
 	"tx_wake[%u]", "tx_stop[%u]", "tx_event_counter[%u]",
+	"tx_dma_mapping_error[%u]",
 };
 
 static const char gve_gstrings_adminq_stats[][ETH_GSTRING_LEN] = {
@@ -322,6 +323,7 @@ gve_get_ethtool_stats(struct net_device *netdev,
 			data[i++] = tx->stop_queue;
 			data[i++] = be32_to_cpu(gve_tx_load_event_counter(priv,
 									  tx));
+			data[i++] = tx->dma_mapping_error;
 			/* stats from NIC */
 			if (skip_nic_stats) {
 				/* skip NIC tx stats */
diff --git a/drivers/net/ethernet/google/gve/gve_tx.c b/drivers/net/ethernet/google/gve/gve_tx.c
index d0244feb0301..6938f3a939d6 100644
--- a/drivers/net/ethernet/google/gve/gve_tx.c
+++ b/drivers/net/ethernet/google/gve/gve_tx.c
@@ -158,9 +158,11 @@ static void gve_tx_free_ring(struct gve_priv *priv, int idx)
 			  tx->q_resources, tx->q_resources_bus);
 	tx->q_resources = NULL;
 
-	gve_tx_fifo_release(priv, &tx->tx_fifo);
-	gve_unassign_qpl(priv, tx->tx_fifo.qpl->id);
-	tx->tx_fifo.qpl = NULL;
+	if (!tx->raw_addressing) {
+		gve_tx_fifo_release(priv, &tx->tx_fifo);
+		gve_unassign_qpl(priv, tx->tx_fifo.qpl->id);
+		tx->tx_fifo.qpl = NULL;
+	}
 
 	bytes = sizeof(*tx->desc) * slots;
 	dma_free_coherent(hdev, bytes, tx->desc, tx->bus);
@@ -206,11 +208,15 @@ static int gve_tx_alloc_ring(struct gve_priv *priv, int idx)
 	if (!tx->desc)
 		goto abort_with_info;
 
-	tx->tx_fifo.qpl = gve_assign_tx_qpl(priv);
+	tx->raw_addressing = priv->raw_addressing;
+	tx->dev = &priv->pdev->dev;
+	if (!tx->raw_addressing) {
+		tx->tx_fifo.qpl = gve_assign_tx_qpl(priv);
 
-	/* map Tx FIFO */
-	if (gve_tx_fifo_init(priv, &tx->tx_fifo))
-		goto abort_with_desc;
+		/* map Tx FIFO */
+		if (gve_tx_fifo_init(priv, &tx->tx_fifo))
+			goto abort_with_desc;
+	}
 
 	tx->q_resources =
 		dma_alloc_coherent(hdev,
@@ -228,7 +234,8 @@ static int gve_tx_alloc_ring(struct gve_priv *priv, int idx)
 	return 0;
 
 abort_with_fifo:
-	gve_tx_fifo_release(priv, &tx->tx_fifo);
+	if (!tx->raw_addressing)
+		gve_tx_fifo_release(priv, &tx->tx_fifo);
 abort_with_desc:
 	dma_free_coherent(hdev, bytes, tx->desc, tx->bus);
 	tx->desc = NULL;
@@ -301,27 +308,47 @@ static inline int gve_skb_fifo_bytes_required(struct gve_tx_ring *tx,
 	return bytes;
 }
 
-/* The most descriptors we could need are 3 - 1 for the headers, 1 for
- * the beginning of the payload at the end of the FIFO, and 1 if the
- * payload wraps to the beginning of the FIFO.
+/* The most descriptors we could need is MAX_SKB_FRAGS + 3 : 1 for each skb frag,
+ * +1 for the skb linear portion, +1 for when tcp hdr needs to be in separate descriptor,
+ * and +1 if the payload wraps to the beginning of the FIFO.
  */
-#define MAX_TX_DESC_NEEDED	3
+#define MAX_TX_DESC_NEEDED	(MAX_SKB_FRAGS + 3)
+static void gve_tx_unmap_buf(struct device *dev, struct gve_tx_buffer_state *info)
+{
+	if (info->skb) {
+		dma_unmap_single(dev, dma_unmap_addr(&info->buf, dma),
+				 dma_unmap_len(&info->buf, len),
+				 DMA_TO_DEVICE);
+		dma_unmap_len_set(&info->buf, len, 0);
+	} else {
+		dma_unmap_page(dev, dma_unmap_addr(&info->buf, dma),
+			       dma_unmap_len(&info->buf, len),
+			       DMA_TO_DEVICE);
+		dma_unmap_len_set(&info->buf, len, 0);
+	}
+}
 
 /* Check if sufficient resources (descriptor ring space, FIFO space) are
  * available to transmit the given number of bytes.
  */
 static inline bool gve_can_tx(struct gve_tx_ring *tx, int bytes_required)
 {
-	return (gve_tx_avail(tx) >= MAX_TX_DESC_NEEDED &&
-		gve_tx_fifo_can_alloc(&tx->tx_fifo, bytes_required));
+	bool can_alloc = true;
+
+	if (!tx->raw_addressing)
+		can_alloc = gve_tx_fifo_can_alloc(&tx->tx_fifo, bytes_required);
+
+	return (gve_tx_avail(tx) >= MAX_TX_DESC_NEEDED && can_alloc);
 }
 
 /* Stops the queue if the skb cannot be transmitted. */
 static int gve_maybe_stop_tx(struct gve_tx_ring *tx, struct sk_buff *skb)
 {
-	int bytes_required;
+	int bytes_required = 0;
+
+	if (!tx->raw_addressing)
+		bytes_required = gve_skb_fifo_bytes_required(tx, skb);
 
-	bytes_required = gve_skb_fifo_bytes_required(tx, skb);
 	if (likely(gve_can_tx(tx, bytes_required)))
 		return 0;
 
@@ -395,17 +422,13 @@ static void gve_dma_sync_for_device(struct device *dev, dma_addr_t *page_buses,
 {
 	u64 last_page = (iov_offset + iov_len - 1) / PAGE_SIZE;
 	u64 first_page = iov_offset / PAGE_SIZE;
-	dma_addr_t dma;
 	u64 page;
 
-	for (page = first_page; page <= last_page; page++) {
-		dma = page_buses[page];
-		dma_sync_single_for_device(dev, dma, PAGE_SIZE, DMA_TO_DEVICE);
-	}
+	for (page = first_page; page <= last_page; page++)
+		dma_sync_single_for_device(dev, page_buses[page], PAGE_SIZE, DMA_TO_DEVICE);
 }
 
-static int gve_tx_add_skb(struct gve_tx_ring *tx, struct sk_buff *skb,
-			  struct device *dev)
+static int gve_tx_add_skb_copy(struct gve_priv *priv, struct gve_tx_ring *tx, struct sk_buff *skb)
 {
 	int pad_bytes, hlen, hdr_nfrags, payload_nfrags, l4_hdr_offset;
 	union gve_tx_desc *pkt_desc, *seg_desc;
@@ -447,7 +470,7 @@ static int gve_tx_add_skb(struct gve_tx_ring *tx, struct sk_buff *skb,
 	skb_copy_bits(skb, 0,
 		      tx->tx_fifo.base + info->iov[hdr_nfrags - 1].iov_offset,
 		      hlen);
-	gve_dma_sync_for_device(dev, tx->tx_fifo.qpl->page_buses,
+	gve_dma_sync_for_device(&priv->pdev->dev, tx->tx_fifo.qpl->page_buses,
 				info->iov[hdr_nfrags - 1].iov_offset,
 				info->iov[hdr_nfrags - 1].iov_len);
 	copy_offset = hlen;
@@ -463,7 +486,7 @@ static int gve_tx_add_skb(struct gve_tx_ring *tx, struct sk_buff *skb,
 		skb_copy_bits(skb, copy_offset,
 			      tx->tx_fifo.base + info->iov[i].iov_offset,
 			      info->iov[i].iov_len);
-		gve_dma_sync_for_device(dev, tx->tx_fifo.qpl->page_buses,
+		gve_dma_sync_for_device(&priv->pdev->dev, tx->tx_fifo.qpl->page_buses,
 					info->iov[i].iov_offset,
 					info->iov[i].iov_len);
 		copy_offset += info->iov[i].iov_len;
@@ -472,6 +495,94 @@ static int gve_tx_add_skb(struct gve_tx_ring *tx, struct sk_buff *skb,
 	return 1 + payload_nfrags;
 }
 
+static int gve_tx_add_skb_no_copy(struct gve_priv *priv, struct gve_tx_ring *tx,
+				  struct sk_buff *skb)
+{
+	const struct skb_shared_info *shinfo = skb_shinfo(skb);
+	int hlen, payload_nfrags, l4_hdr_offset;
+	union gve_tx_desc *pkt_desc, *seg_desc;
+	struct gve_tx_buffer_state *info;
+	bool is_gso = skb_is_gso(skb);
+	u32 idx = tx->req & tx->mask;
+	struct gve_tx_dma_buf *buf;
+	u64 addr;
+	u32 len;
+	int i;
+
+	info = &tx->info[idx];
+	pkt_desc = &tx->desc[idx];
+
+	l4_hdr_offset = skb_checksum_start_offset(skb);
+	/* If the skb is gso, then we want only up to the tcp header in the first segment
+	 * to efficiently replicate on each segment otherwise we want the linear portion
+	 * of the skb (which will contain the checksum because skb->csum_start and
+	 * skb->csum_offset are given relative to skb->head) in the first segment.
+	 */
+	hlen = is_gso ? l4_hdr_offset + tcp_hdrlen(skb) : skb_headlen(skb);
+	len = skb_headlen(skb);
+
+	info->skb =  skb;
+
+	addr = dma_map_single(tx->dev, skb->data, len, DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(tx->dev, addr))) {
+		tx->dma_mapping_error++;
+		goto drop;
+	}
+	buf = &info->buf;
+	dma_unmap_len_set(buf, len, len);
+	dma_unmap_addr_set(buf, dma, addr);
+
+	payload_nfrags = shinfo->nr_frags;
+	if (hlen < len) {
+		/* For gso the rest of the linear portion of the skb needs to
+		 * be in its own descriptor.
+		 */
+		payload_nfrags++;
+		gve_tx_fill_pkt_desc(pkt_desc, skb, is_gso, l4_hdr_offset,
+				     1 + payload_nfrags, hlen, addr);
+
+		len -= hlen;
+		addr += hlen;
+		idx = (tx->req + 1) & tx->mask;
+		seg_desc = &tx->desc[idx];
+		gve_tx_fill_seg_desc(seg_desc, skb, is_gso, len, addr);
+	} else {
+		gve_tx_fill_pkt_desc(pkt_desc, skb, is_gso, l4_hdr_offset,
+				     1 + payload_nfrags, hlen, addr);
+	}
+
+	for (i = 0; i < shinfo->nr_frags; i++) {
+		const skb_frag_t *frag = &shinfo->frags[i];
+
+		idx = (idx + 1) & tx->mask;
+		seg_desc = &tx->desc[idx];
+		len = skb_frag_size(frag);
+		addr = skb_frag_dma_map(tx->dev, frag, 0, len, DMA_TO_DEVICE);
+		if (unlikely(dma_mapping_error(tx->dev, addr))) {
+			tx->dma_mapping_error++;
+			goto unmap_drop;
+		}
+		buf = &tx->info[idx].buf;
+		tx->info[idx].skb = NULL;
+		dma_unmap_len_set(buf, len, len);
+		dma_unmap_addr_set(buf, dma, addr);
+
+		gve_tx_fill_seg_desc(seg_desc, skb, is_gso, len, addr);
+	}
+
+	return 1 + payload_nfrags;
+
+unmap_drop:
+	i += (payload_nfrags == shinfo->nr_frags ? 1 : 2);
+	while (i--) {
+		idx--;
+		gve_tx_unmap_buf(tx->dev, &tx->info[idx & tx->mask]);
+	}
+drop:
+	tx->dropped_pkt++;
+	return 0;
+}
+
 netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev)
 {
 	struct gve_priv *priv = netdev_priv(dev);
@@ -490,17 +601,26 @@ netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev)
 		gve_tx_put_doorbell(priv, tx->q_resources, tx->req);
 		return NETDEV_TX_BUSY;
 	}
-	nsegs = gve_tx_add_skb(tx, skb, &priv->pdev->dev);
-
-	netdev_tx_sent_queue(tx->netdev_txq, skb->len);
-	skb_tx_timestamp(skb);
-
-	/* give packets to NIC */
-	tx->req += nsegs;
+	if (tx->raw_addressing)
+		nsegs = gve_tx_add_skb_no_copy(priv, tx, skb);
+	else
+		nsegs = gve_tx_add_skb_copy(priv, tx, skb);
+
+	/* If the packet is getting sent, we need to update the skb */
+	if (nsegs) {
+		netdev_tx_sent_queue(tx->netdev_txq, skb->len);
+		skb_tx_timestamp(skb);
+		tx->req += nsegs;
+	} else {
+		dev_kfree_skb_any(skb);
+	}
 
 	if (!netif_xmit_stopped(tx->netdev_txq) && netdev_xmit_more())
 		return NETDEV_TX_OK;
 
+	/* Give packets to NIC. Even if this packet failed to send the doorbell
+	 * might need to be rung because of xmit_more.
+	 */
 	gve_tx_put_doorbell(priv, tx->q_resources, tx->req);
 	return NETDEV_TX_OK;
 }
@@ -525,24 +645,29 @@ static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
 		info = &tx->info[idx];
 		skb = info->skb;
 
+		/* Unmap the buffer */
+		if (tx->raw_addressing)
+			gve_tx_unmap_buf(tx->dev, info);
+		tx->done++;
 		/* Mark as free */
 		if (skb) {
 			info->skb = NULL;
 			bytes += skb->len;
 			pkts++;
 			dev_consume_skb_any(skb);
+			if (tx->raw_addressing)
+				continue;
 			/* FIFO free */
 			for (i = 0; i < ARRAY_SIZE(info->iov); i++) {
-				space_freed += info->iov[i].iov_len +
-					       info->iov[i].iov_padding;
+				space_freed += info->iov[i].iov_len + info->iov[i].iov_padding;
 				info->iov[i].iov_len = 0;
 				info->iov[i].iov_padding = 0;
 			}
 		}
-		tx->done++;
 	}
 
-	gve_tx_free_fifo(&tx->tx_fifo, space_freed);
+	if (!tx->raw_addressing)
+		gve_tx_free_fifo(&tx->tx_fifo, space_freed);
 	u64_stats_update_begin(&tx->statss);
 	tx->bytes_done += bytes;
 	tx->pkt_done += pkts;
-- 
2.29.2.576.ga3fc446d84-goog

Re: [PATCH net-next v9 4/4] gve: Add support for raw addressing in the tx path

[Alexander Duyck]

On Wed, Dec 2, 2020 at 10:24 AM David Awogbemila <awogbemila@google.com> wrote:
>
> From: Catherine Sullivan <csully@google.com>
>
> During TX, skbs' data addresses are dma_map'ed and passed to the NIC.
> This means that the device can perform DMA directly from these addresses
> and the driver does not have to copy the buffer content into
> pre-allocated buffers/qpls (as in qpl mode).
>
> Reviewed-by: Yangchun Fu <yangchun@google.com>
> Signed-off-by: Catherine Sullivan <csully@google.com>
> Signed-off-by: David Awogbemila <awogbemila@google.com>
> ---
>  drivers/net/ethernet/google/gve/gve.h         |  16 +-
>  drivers/net/ethernet/google/gve/gve_adminq.c  |   4 +-
>  drivers/net/ethernet/google/gve/gve_desc.h    |   8 +-
>  drivers/net/ethernet/google/gve/gve_ethtool.c |   2 +
>  drivers/net/ethernet/google/gve/gve_tx.c      | 197 ++++++++++++++----
>  5 files changed, 185 insertions(+), 42 deletions(-)
>
> diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
> index 8aad4af2aa2b..9888fa92be86 100644
> --- a/drivers/net/ethernet/google/gve/gve.h
> +++ b/drivers/net/ethernet/google/gve/gve.h
> @@ -112,12 +112,20 @@ struct gve_tx_iovec {
>         u32 iov_padding; /* padding associated with this segment */
>  };
>
> +struct gve_tx_dma_buf {
> +       DEFINE_DMA_UNMAP_ADDR(dma);
> +       DEFINE_DMA_UNMAP_LEN(len);
> +};
> +
>  /* Tracks the memory in the fifo occupied by the skb. Mapped 1:1 to a desc
>   * ring entry but only used for a pkt_desc not a seg_desc
>   */
>  struct gve_tx_buffer_state {
>         struct sk_buff *skb; /* skb for this pkt */
> -       struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
> +       union {
> +               struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
> +               struct gve_tx_dma_buf buf;
> +       };
>  };
>
>  /* A TX buffer - each queue has one */
> @@ -140,19 +148,23 @@ struct gve_tx_ring {
>         __be32 last_nic_done ____cacheline_aligned; /* NIC tail pointer */
>         u64 pkt_done; /* free-running - total packets completed */
>         u64 bytes_done; /* free-running - total bytes completed */
> +       u32 dropped_pkt; /* free-running - total packets dropped */

Generally I would probably use a u64 for any count values. I'm not
sure what rate you will be moving packets at but if something goes
wrong you are better off with the counter not rolling over every few
minutes.

>         /* Cacheline 2 -- Read-mostly fields */
>         union gve_tx_desc *desc ____cacheline_aligned;
>         struct gve_tx_buffer_state *info; /* Maps 1:1 to a desc */
>         struct netdev_queue *netdev_txq;
>         struct gve_queue_resources *q_resources; /* head and tail pointer idx */
> +       struct device *dev;
>         u32 mask; /* masks req and done down to queue size */
> +       u8 raw_addressing; /* use raw_addressing? */
>
>         /* Slow-path fields */
>         u32 q_num ____cacheline_aligned; /* queue idx */
>         u32 stop_queue; /* count of queue stops */
>         u32 wake_queue; /* count of queue wakes */
>         u32 ntfy_id; /* notification block index */
> +       u32 dma_mapping_error; /* count of dma mapping errors */

Since this is a counter wouldn't it make more sense to place it up
with the other counters?

Looking over the rest of the patch it seems fine to me. The counters
were the only thing that had me a bit concerned.

Re: [PATCH net-next v9 0/4] GVE Raw Addressing

[Alexander Duyck]

On Wed, Dec 2, 2020 at 10:24 AM David Awogbemila <awogbemila@google.com> wrote:
>
> Patchset description:
> This  patchset introduces "raw addressing" mode to the GVE driver.
> Previously (in "queue_page_list" or "qpl" mode), the driver would
> pre-allocate and dma_map buffers to be used on egress and ingress.
> On egress, it would copy data from the skb provided to the
> pre-allocated buffers - this was expensive.
> In raw addressing mode, the driver can avoid this copy and simply
> dma_map the skb's data so that the NIC can use it.
> On ingress, the driver passes buffers up to the networking stack and
> then frees and reallocates buffers when necessary instead of using
> skb_copy_to_linear_data.
> Patch 3 separates the page refcount tracking mechanism
> into a function gve_rx_can_recycle_buffer which uses get_page - this will
> be changed in a future patch to eliminate the use of get_page in tracking
> page refcounts.
>
> Changes from v8:
> Patch 4: Free skbs that are not sent in gve_tx().
>
> Catherine Sullivan (3):
>   gve: Add support for raw addressing device option
>   gve: Add support for raw addressing to the rx path
>   gve: Add support for raw addressing in the tx path
>
> David Awogbemila (1):
>   gve: Rx Buffer Recycling
>
>  drivers/net/ethernet/google/gve/gve.h         |  39 +-
>  drivers/net/ethernet/google/gve/gve_adminq.c  |  89 ++++-
>  drivers/net/ethernet/google/gve/gve_adminq.h  |  15 +-
>  drivers/net/ethernet/google/gve/gve_desc.h    |  19 +-
>  drivers/net/ethernet/google/gve/gve_ethtool.c |   2 +
>  drivers/net/ethernet/google/gve/gve_main.c    |  11 +-
>  drivers/net/ethernet/google/gve/gve_rx.c      | 364 +++++++++++++-----
>  drivers/net/ethernet/google/gve/gve_tx.c      | 202 ++++++++--
>  8 files changed, 577 insertions(+), 164 deletions(-)
>
> --
> 2.29.2.576.ga3fc446d84-goog
>

Other than the few nits with counters being u32 values I didn't really
see much else.

Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>

Re: [PATCH net-next v9 4/4] gve: Add support for raw addressing in the tx path

[David Awogbemila]

On Thu, Dec 3, 2020 at 10:16 AM Alexander Duyck
<alexander.duyck@gmail.com> wrote:
>
> On Wed, Dec 2, 2020 at 10:24 AM David Awogbemila <awogbemila@google.com> wrote:
> >
> > From: Catherine Sullivan <csully@google.com>
> >
> > During TX, skbs' data addresses are dma_map'ed and passed to the NIC.
> > This means that the device can perform DMA directly from these addresses
> > and the driver does not have to copy the buffer content into
> > pre-allocated buffers/qpls (as in qpl mode).
> >
> > Reviewed-by: Yangchun Fu <yangchun@google.com>
> > Signed-off-by: Catherine Sullivan <csully@google.com>
> > Signed-off-by: David Awogbemila <awogbemila@google.com>
> > ---
> >  drivers/net/ethernet/google/gve/gve.h         |  16 +-
> >  drivers/net/ethernet/google/gve/gve_adminq.c  |   4 +-
> >  drivers/net/ethernet/google/gve/gve_desc.h    |   8 +-
> >  drivers/net/ethernet/google/gve/gve_ethtool.c |   2 +
> >  drivers/net/ethernet/google/gve/gve_tx.c      | 197 ++++++++++++++----
> >  5 files changed, 185 insertions(+), 42 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
> > index 8aad4af2aa2b..9888fa92be86 100644
> > --- a/drivers/net/ethernet/google/gve/gve.h
> > +++ b/drivers/net/ethernet/google/gve/gve.h
> > @@ -112,12 +112,20 @@ struct gve_tx_iovec {
> >         u32 iov_padding; /* padding associated with this segment */
> >  };
> >
> > +struct gve_tx_dma_buf {
> > +       DEFINE_DMA_UNMAP_ADDR(dma);
> > +       DEFINE_DMA_UNMAP_LEN(len);
> > +};
> > +
> >  /* Tracks the memory in the fifo occupied by the skb. Mapped 1:1 to a desc
> >   * ring entry but only used for a pkt_desc not a seg_desc
> >   */
> >  struct gve_tx_buffer_state {
> >         struct sk_buff *skb; /* skb for this pkt */
> > -       struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
> > +       union {
> > +               struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
> > +               struct gve_tx_dma_buf buf;
> > +       };
> >  };
> >
> >  /* A TX buffer - each queue has one */
> > @@ -140,19 +148,23 @@ struct gve_tx_ring {
> >         __be32 last_nic_done ____cacheline_aligned; /* NIC tail pointer */
> >         u64 pkt_done; /* free-running - total packets completed */
> >         u64 bytes_done; /* free-running - total bytes completed */
> > +       u32 dropped_pkt; /* free-running - total packets dropped */
>
> Generally I would probably use a u64 for any count values. I'm not
> sure what rate you will be moving packets at but if something goes
> wrong you are better off with the counter not rolling over every few
> minutes.

Ok, using a u64 here would be better - I'll do that, thanks.

>
> >         /* Cacheline 2 -- Read-mostly fields */
> >         union gve_tx_desc *desc ____cacheline_aligned;
> >         struct gve_tx_buffer_state *info; /* Maps 1:1 to a desc */
> >         struct netdev_queue *netdev_txq;
> >         struct gve_queue_resources *q_resources; /* head and tail pointer idx */
> > +       struct device *dev;
> >         u32 mask; /* masks req and done down to queue size */
> > +       u8 raw_addressing; /* use raw_addressing? */
> >
> >         /* Slow-path fields */
> >         u32 q_num ____cacheline_aligned; /* queue idx */
> >         u32 stop_queue; /* count of queue stops */
> >         u32 wake_queue; /* count of queue wakes */
> >         u32 ntfy_id; /* notification block index */
> > +       u32 dma_mapping_error; /* count of dma mapping errors */
>
> Since this is a counter wouldn't it make more sense to place it up
> with the other counters?
>
> Looking over the rest of the patch it seems fine to me. The counters
> were the only thing that had me a bit concerned.

Ok, I'll use a u64 here and place it with the counter above.