[PATCH net-next v3 1/2] net: mhi: Add support for non-linear MBIM skb processing

[PATCH net-next v3 1/2] net: mhi: Add support for non-linear MBIM skb processing

[Loic Poulain]

Currently, if skb is non-linear, due to MHI skb chaining, it is
linearized in MBIM RX handler prior MBIM decoding, causing extra
allocation and copy that can be as large as the maximum MBIM frame
size (32K).

This change introduces MBIM decoding for non-linear skb, allowing to
process 'large' non-linear MBIM packets without skb linearization.
The IP packets are simply extracted from the MBIM frame using the
skb_copy_bits helper.

Signed-off-by: Loic Poulain <loic.poulain@linaro.org>
---
 v2: added to the series for reducing chained skb processing overhead 
 v3: Fix missing net_err_ratelimited print param (ndev->name).

 drivers/net/mhi/proto_mbim.c | 51 ++++++++++++++++++++++----------------------
 1 file changed, 25 insertions(+), 26 deletions(-)

diff --git a/drivers/net/mhi/proto_mbim.c b/drivers/net/mhi/proto_mbim.c
index 75b5484..5a176c3 100644
--- a/drivers/net/mhi/proto_mbim.c
+++ b/drivers/net/mhi/proto_mbim.c
@@ -91,20 +91,11 @@ static int mbim_rx_verify_nth16(struct sk_buff *skb)
 	return le16_to_cpu(nth16->wNdpIndex);
 }
 
-static int mbim_rx_verify_ndp16(struct sk_buff *skb, int ndpoffset)
+static int mbim_rx_verify_ndp16(struct sk_buff *skb, struct usb_cdc_ncm_ndp16 *ndp16)
 {
 	struct mhi_net_dev *dev = netdev_priv(skb->dev);
-	struct usb_cdc_ncm_ndp16 *ndp16;
 	int ret;
 
-	if (ndpoffset + sizeof(struct usb_cdc_ncm_ndp16) > skb->len) {
-		netif_dbg(dev, rx_err, dev->ndev, "invalid NDP offset  <%u>\n",
-			  ndpoffset);
-		return -EINVAL;
-	}
-
-	ndp16 = (struct usb_cdc_ncm_ndp16 *)(skb->data + ndpoffset);
-
 	if (le16_to_cpu(ndp16->wLength) < USB_CDC_NCM_NDP16_LENGTH_MIN) {
 		netif_dbg(dev, rx_err, dev->ndev, "invalid DPT16 length <%u>\n",
 			  le16_to_cpu(ndp16->wLength));
@@ -130,9 +121,6 @@ static void mbim_rx(struct mhi_net_dev *mhi_netdev, struct sk_buff *skb)
 	struct net_device *ndev = mhi_netdev->ndev;
 	int ndpoffset;
 
-	if (skb_linearize(skb))
-		goto error;
-
 	/* Check NTB header and retrieve first NDP offset */
 	ndpoffset = mbim_rx_verify_nth16(skb);
 	if (ndpoffset < 0) {
@@ -142,12 +130,19 @@ static void mbim_rx(struct mhi_net_dev *mhi_netdev, struct sk_buff *skb)
 
 	/* Process each NDP */
 	while (1) {
-		struct usb_cdc_ncm_ndp16 *ndp16;
-		struct usb_cdc_ncm_dpe16 *dpe16;
-		int nframes, n;
+		struct usb_cdc_ncm_ndp16 ndp16;
+		struct usb_cdc_ncm_dpe16 dpe16;
+		int nframes, n, dpeoffset;
+
+		if (skb_copy_bits(skb, ndpoffset, &ndp16, sizeof(ndp16))) {
+			net_err_ratelimited("%s: Incorrect NDP offset (%u)\n",
+					    ndev->name, ndpoffset);
+			__mbim_length_errors_inc(mhi_netdev);
+			goto error;
+		}
 
 		/* Check NDP header and retrieve number of datagrams */
-		nframes = mbim_rx_verify_ndp16(skb, ndpoffset);
+		nframes = mbim_rx_verify_ndp16(skb, &ndp16);
 		if (nframes < 0) {
 			net_err_ratelimited("%s: Incorrect NDP16\n", ndev->name);
 			__mbim_length_errors_inc(mhi_netdev);
@@ -155,8 +150,7 @@ static void mbim_rx(struct mhi_net_dev *mhi_netdev, struct sk_buff *skb)
 		}
 
 		 /* Only IP data type supported, no DSS in MHI context */
-		ndp16 = (struct usb_cdc_ncm_ndp16 *)(skb->data + ndpoffset);
-		if ((ndp16->dwSignature & cpu_to_le32(MBIM_NDP16_SIGN_MASK))
+		if ((ndp16.dwSignature & cpu_to_le32(MBIM_NDP16_SIGN_MASK))
 				!= cpu_to_le32(USB_CDC_MBIM_NDP16_IPS_SIGN)) {
 			net_err_ratelimited("%s: Unsupported NDP type\n", ndev->name);
 			__mbim_errors_inc(mhi_netdev);
@@ -164,19 +158,24 @@ static void mbim_rx(struct mhi_net_dev *mhi_netdev, struct sk_buff *skb)
 		}
 
 		/* Only primary IP session 0 (0x00) supported for now */
-		if (ndp16->dwSignature & ~cpu_to_le32(MBIM_NDP16_SIGN_MASK)) {
+		if (ndp16.dwSignature & ~cpu_to_le32(MBIM_NDP16_SIGN_MASK)) {
 			net_err_ratelimited("%s: bad packet session\n", ndev->name);
 			__mbim_errors_inc(mhi_netdev);
 			goto next_ndp;
 		}
 
 		/* de-aggregate and deliver IP packets */
-		dpe16 = ndp16->dpe16;
-		for (n = 0; n < nframes; n++, dpe16++) {
-			u16 dgram_offset = le16_to_cpu(dpe16->wDatagramIndex);
-			u16 dgram_len = le16_to_cpu(dpe16->wDatagramLength);
+		dpeoffset = ndpoffset + sizeof(struct usb_cdc_ncm_ndp16);
+		for (n = 0; n < nframes; n++, dpeoffset += sizeof(dpe16)) {
+			u16 dgram_offset, dgram_len;
 			struct sk_buff *skbn;
 
+			if (skb_copy_bits(skb, dpeoffset, &dpe16, sizeof(dpe16)))
+				break;
+
+			dgram_offset = le16_to_cpu(dpe16.wDatagramIndex);
+			dgram_len = le16_to_cpu(dpe16.wDatagramLength);
+
 			if (!dgram_offset || !dgram_len)
 				break; /* null terminator */
 
@@ -185,7 +184,7 @@ static void mbim_rx(struct mhi_net_dev *mhi_netdev, struct sk_buff *skb)
 				continue;
 
 			skb_put(skbn, dgram_len);
-			memcpy(skbn->data, skb->data + dgram_offset, dgram_len);
+			skb_copy_bits(skb, dgram_offset, skbn->data, dgram_len);
 
 			switch (skbn->data[0] & 0xf0) {
 			case 0x40:
@@ -206,7 +205,7 @@ static void mbim_rx(struct mhi_net_dev *mhi_netdev, struct sk_buff *skb)
 		}
 next_ndp:
 		/* Other NDP to process? */
-		ndpoffset = (int)le16_to_cpu(ndp16->wNextNdpIndex);
+		ndpoffset = (int)le16_to_cpu(ndp16.wNextNdpIndex);
 		if (!ndpoffset)
 			break;
 	}
-- 
2.7.4

[PATCH net-next v3 2/2] net: mhi: Allow decoupled MTU/MRU

[Loic Poulain]

MBIM protocol makes the mhi network interface asymmetric, ingress data
received from MHI is MBIM protocol, possibly containing multiple
aggregated IP packets, while egress data received from network stack is
IP protocol.

This changes allows a 'protocol' to specify its own MRU, that when
specified is used to allocate MHI RX buffers (skb).

For MBIM, Set the default MTU to 1500, which is the usual network MTU
for WWAN IP packets, and MRU to 3.5K (for allocation efficiency),
allowing skb to fit in an usual 4K page (including padding,
skb_shared_info, ...).

Signed-off-by: Loic Poulain <loic.poulain@linaro.org>
---
  v2: squashed 1/2 and 1/2 (decoupled mtu/mru + mbim mru)
     Reduced MRU from 32k to 3.5K
  v3: no change

 drivers/net/mhi/mhi.h        |  1 +
 drivers/net/mhi/net.c        |  4 +++-
 drivers/net/mhi/proto_mbim.c | 11 +++++++++++
 3 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/drivers/net/mhi/mhi.h b/drivers/net/mhi/mhi.h
index 12e7407..1d0c499 100644
--- a/drivers/net/mhi/mhi.h
+++ b/drivers/net/mhi/mhi.h
@@ -29,6 +29,7 @@ struct mhi_net_dev {
 	struct mhi_net_stats stats;
 	u32 rx_queue_sz;
 	int msg_enable;
+	unsigned int mru;
 };
 
 struct mhi_net_proto {
diff --git a/drivers/net/mhi/net.c b/drivers/net/mhi/net.c
index b1769fb..4bef7fe 100644
--- a/drivers/net/mhi/net.c
+++ b/drivers/net/mhi/net.c
@@ -270,10 +270,12 @@ static void mhi_net_rx_refill_work(struct work_struct *work)
 						      rx_refill.work);
 	struct net_device *ndev = mhi_netdev->ndev;
 	struct mhi_device *mdev = mhi_netdev->mdev;
-	int size = READ_ONCE(ndev->mtu);
 	struct sk_buff *skb;
+	unsigned int size;
 	int err;
 
+	size = mhi_netdev->mru ? mhi_netdev->mru : READ_ONCE(ndev->mtu);
+
 	while (!mhi_queue_is_full(mdev, DMA_FROM_DEVICE)) {
 		skb = netdev_alloc_skb(ndev, size);
 		if (unlikely(!skb))
diff --git a/drivers/net/mhi/proto_mbim.c b/drivers/net/mhi/proto_mbim.c
index 5a176c3..fc72b3f 100644
--- a/drivers/net/mhi/proto_mbim.c
+++ b/drivers/net/mhi/proto_mbim.c
@@ -26,6 +26,15 @@
 
 #define MBIM_NDP16_SIGN_MASK 0x00ffffff
 
+/* Usual WWAN MTU */
+#define MHI_MBIM_DEFAULT_MTU 1500
+
+/* 3500 allows to optimize skb allocation, the skbs will basically fit in
+ * one 4K page. Large MBIM packets will simply be split over several MHI
+ * transfers and chained by the MHI net layer (zerocopy).
+ */
+#define MHI_MBIM_DEFAULT_MRU 3500
+
 struct mbim_context {
 	u16 rx_seq;
 	u16 tx_seq;
@@ -281,6 +290,8 @@ static int mbim_init(struct mhi_net_dev *mhi_netdev)
 		return -ENOMEM;
 
 	ndev->needed_headroom = sizeof(struct mbim_tx_hdr);
+	ndev->mtu = MHI_MBIM_DEFAULT_MTU;
+	mhi_netdev->mru = MHI_MBIM_DEFAULT_MRU;
 
 	return 0;
 }
-- 
2.7.4

Re: [PATCH net-next v3 1/2] net: mhi: Add support for non-linear MBIM skb processing

[patchwork-bot+netdevbpf]

Hello:

This series was applied to netdev/net-next.git (refs/heads/master):

On Mon, 29 Mar 2021 17:39:31 +0200 you wrote:
> Currently, if skb is non-linear, due to MHI skb chaining, it is
> linearized in MBIM RX handler prior MBIM decoding, causing extra
> allocation and copy that can be as large as the maximum MBIM frame
> size (32K).
> 
> This change introduces MBIM decoding for non-linear skb, allowing to
> process 'large' non-linear MBIM packets without skb linearization.
> The IP packets are simply extracted from the MBIM frame using the
> skb_copy_bits helper.
> 
> [...]

Here is the summary with links:
  - [net-next,v3,1/2] net: mhi: Add support for non-linear MBIM skb processing
    https://git.kernel.org/netdev/net-next/c/d9f0713c9217
  - [net-next,v3,2/2] net: mhi: Allow decoupled MTU/MRU
    https://git.kernel.org/netdev/net-next/c/3af562a37b7f

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html