[PATCH net-next] gve: support unreadable netmem

[PATCH net-next] gve: support unreadable netmem

[Harshitha Ramamurthy]

From: Mina Almasry <almasrymina@google.com>

Declare PP_FLAG_ALLOW_UNREADABLE_NETMEM to turn on unreadable netmem
support in GVE.

We also drop any net_iov packets where header split is not enabled.
We're unable to process packets where the header landed in unreadable
netmem.

Use page_pool_dma_sync_netmem_for_cpu in lieu of
dma_sync_single_range_for_cpu to correctly handle unreadable netmem
that should not be dma-sync'd.

Disable rx_copybreak optimization if payload is unreadable netmem as
that needs access to the payload.

Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Harshitha Ramamurthy <hramamurthy@google.com>
---
 .../ethernet/google/gve/gve_buffer_mgmt_dqo.c |  5 +++
 drivers/net/ethernet/google/gve/gve_rx_dqo.c  | 36 ++++++++++++++++---
 2 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/google/gve/gve_buffer_mgmt_dqo.c b/drivers/net/ethernet/google/gve/gve_buffer_mgmt_dqo.c
index 8f5021e59e0a..0e2b703c673a 100644
--- a/drivers/net/ethernet/google/gve/gve_buffer_mgmt_dqo.c
+++ b/drivers/net/ethernet/google/gve/gve_buffer_mgmt_dqo.c
@@ -260,6 +260,11 @@ struct page_pool *gve_rx_create_page_pool(struct gve_priv *priv,
 		.offset = xdp ? XDP_PACKET_HEADROOM : 0,
 	};
 
+	if (priv->header_split_enabled) {
+		pp.flags |= PP_FLAG_ALLOW_UNREADABLE_NETMEM;
+		pp.queue_idx = rx->q_num;
+	}
+
 	return page_pool_create(&pp);
 }
 
diff --git a/drivers/net/ethernet/google/gve/gve_rx_dqo.c b/drivers/net/ethernet/google/gve/gve_rx_dqo.c
index 7380c2b7a2d8..8c75a4d1e3e7 100644
--- a/drivers/net/ethernet/google/gve/gve_rx_dqo.c
+++ b/drivers/net/ethernet/google/gve/gve_rx_dqo.c
@@ -718,6 +718,24 @@ static int gve_rx_xsk_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 	return 0;
 }
 
+static void gve_dma_sync(struct gve_priv *priv, struct gve_rx_ring *rx,
+			 struct gve_rx_buf_state_dqo *buf_state, u16 buf_len)
+{
+	struct gve_rx_slot_page_info *page_info = &buf_state->page_info;
+
+	if (rx->dqo.page_pool) {
+		page_pool_dma_sync_netmem_for_cpu(rx->dqo.page_pool,
+						  page_info->netmem,
+						  page_info->page_offset,
+						  buf_len);
+	} else {
+		dma_sync_single_range_for_cpu(&priv->pdev->dev, buf_state->addr,
+					      page_info->page_offset +
+					      page_info->pad,
+					      buf_len, DMA_FROM_DEVICE);
+	}
+}
+
 /* Returns 0 if descriptor is completed successfully.
  * Returns -EINVAL if descriptor is invalid.
  * Returns -ENOMEM if data cannot be copied to skb.
@@ -793,13 +811,19 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 		rx->rx_hsplit_unsplit_pkt += unsplit;
 		rx->rx_hsplit_bytes += hdr_len;
 		u64_stats_update_end(&rx->statss);
+	} else if (!rx->ctx.skb_head && rx->dqo.page_pool &&
+		   netmem_is_net_iov(buf_state->page_info.netmem)) {
+		/* when header split is disabled, the header went to the packet
+		 * buffer. If the packet buffer is a net_iov, those can't be
+		 * easily mapped into the kernel space to access the header
+		 * required to process the packet.
+		 */
+		gve_free_buffer(rx, buf_state);
+		return -EFAULT;
 	}
 
 	/* Sync the portion of dma buffer for CPU to read. */
-	dma_sync_single_range_for_cpu(&priv->pdev->dev, buf_state->addr,
-				      buf_state->page_info.page_offset +
-				      buf_state->page_info.pad,
-				      buf_len, DMA_FROM_DEVICE);
+	gve_dma_sync(priv, rx, buf_state, buf_len);
 
 	/* Append to current skb if one exists. */
 	if (rx->ctx.skb_head) {
@@ -837,7 +861,9 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
 		u64_stats_update_end(&rx->statss);
 	}
 
-	if (eop && buf_len <= priv->rx_copybreak) {
+	if (eop && buf_len <= priv->rx_copybreak &&
+	    !(rx->dqo.page_pool &&
+	      netmem_is_net_iov(buf_state->page_info.netmem))) {
 		rx->ctx.skb_head = gve_rx_copy(priv->dev, napi,
 					       &buf_state->page_info, buf_len);
 		if (unlikely(!rx->ctx.skb_head))
-- 
2.50.0.727.gbf7dc18ff4-goog

Re: [PATCH net-next] gve: support unreadable netmem

[Simon Horman]

On Wed, Jul 23, 2025 at 10:28:29PM +0000, Harshitha Ramamurthy wrote:

...

> diff --git a/drivers/net/ethernet/google/gve/gve_rx_dqo.c b/drivers/net/ethernet/google/gve/gve_rx_dqo.c

...

> @@ -793,13 +811,19 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
>  		rx->rx_hsplit_unsplit_pkt += unsplit;
>  		rx->rx_hsplit_bytes += hdr_len;
>  		u64_stats_update_end(&rx->statss);
> +	} else if (!rx->ctx.skb_head && rx->dqo.page_pool &&
> +		   netmem_is_net_iov(buf_state->page_info.netmem)) {
> +		/* when header split is disabled, the header went to the packet
> +		 * buffer. If the packet buffer is a net_iov, those can't be
> +		 * easily mapped into the kernel space to access the header
> +		 * required to process the packet.
> +		 */
> +		gve_free_buffer(rx, buf_state);
> +		return -EFAULT;

nit: I think it would be nice to consistently handle error paths
     in this function using goto error.

>  	}
>  
>  	/* Sync the portion of dma buffer for CPU to read. */
> -	dma_sync_single_range_for_cpu(&priv->pdev->dev, buf_state->addr,
> -				      buf_state->page_info.page_offset +
> -				      buf_state->page_info.pad,
> -				      buf_len, DMA_FROM_DEVICE);
> +	gve_dma_sync(priv, rx, buf_state, buf_len);
>  
>  	/* Append to current skb if one exists. */
>  	if (rx->ctx.skb_head) {