-----Original Message-----
From: Simon Horman <horms@kernel.org>
Sent: Friday, September 29, 2023 4:57 AM
To: Haiyang Zhang <haiyangz@microsoft.com>
Cc: linux-hyperv@vger.kernel.org; netdev@vger.kernel.org; Dexuan Cui
[off-list ref]; KY Srinivasan [off-list ref]; Paul Rosswurm
[off-list ref]; olaf@aepfle.de; vkuznets [off-list ref];
davem@davemloft.net; wei.liu@kernel.org; edumazet@google.com;
kuba@kernel.org; pabeni@redhat.com; leon@kernel.org; Long Li
[off-list ref]; ssengar@linux.microsoft.com; linux-
rdma@vger.kernel.org; daniel@iogearbox.net; john.fastabend@gmail.com;
bpf@vger.kernel.org; ast@kernel.org; Ajay Sharma
[off-list ref]; hawk@kernel.org; tglx@linutronix.de;
shradhagupta@linux.microsoft.com; linux-kernel@vger.kernel.org
Subject: Re: [PATCH net, 3/3] net: mana: Fix oversized sge0 for GSO packets

On Sat, Sep 23, 2023 at 06:31:47PM -0700, Haiyang Zhang wrote:

quoted

Handle the case when GSO SKB linear length is too large.

MANA NIC requires GSO packets to put only the header part to SGE0,
otherwise the TX queue may stop at the HW level.

So, use 2 SGEs for the skb linear part which contains more than the
packet header.

Fixes: ca9c54d2d6a5 ("net: mana: Add a driver for Microsoft Azure Network

Adapter (MANA)")

quoted

Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>

Hi Haiyang Zhang,

thanks for your patch.
Please find some feedback inline.

quoted

---
 drivers/net/ethernet/microsoft/mana/mana_en.c | 186 ++++++++++++------
 include/net/mana/mana.h                       |   5 +-
 2 files changed, 134 insertions(+), 57 deletions(-)

diff --git a/drivers/net/ethernet/microsoft/mana/mana_en.c

b/drivers/net/ethernet/microsoft/mana/mana_en.c

quoted

index 86e724c3eb89..0a3879163b56 100644

--- a/drivers/net/ethernet/microsoft/mana/mana_en.c
+++ b/drivers/net/ethernet/microsoft/mana/mana_en.c

@@ -91,63 +91,136 @@ static unsigned int mana_checksum_info(struct

sk_buff *skb)

quoted

 	return 0;
 }

+static inline void mana_add_sge(struct mana_tx_package *tp,
+				struct mana_skb_head *ash, int sg_i,
+				dma_addr_t da, int sge_len, u32 gpa_mkey)

Please don't use inline for code in .c files unless there
is a demonstrable reason to do so: in general, the compiler should be
left to inline code as it sees fit.

Sure, will remove the "inline".

quoted

+{
+	ash->dma_handle[sg_i] = da;
+	ash->size[sg_i] = sge_len;
+
+	tp->wqe_req.sgl[sg_i].address = da;
+	tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey;
+	tp->wqe_req.sgl[sg_i].size = sge_len;
+}
+
 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context

*apc,

quoted

-			struct mana_tx_package *tp)
+			struct mana_tx_package *tp, int gso_hs)
 {
 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
+	int hsg = 1; /* num of SGEs of linear part */
 	struct gdma_dev *gd = apc->ac->gdma_dev;
+	int skb_hlen = skb_headlen(skb);
+	int sge0_len, sge1_len = 0;
 	struct gdma_context *gc;
 	struct device *dev;
 	skb_frag_t *frag;
 	dma_addr_t da;
+	int sg_i;
 	int i;

 	gc = gd->gdma_context;
 	dev = gc->dev;
-	da = dma_map_single(dev, skb->data, skb_headlen(skb),

DMA_TO_DEVICE);

quoted

+	if (gso_hs && gso_hs < skb_hlen) {
+		sge0_len = gso_hs;
+		sge1_len = skb_hlen - gso_hs;
+	} else {
+		sge0_len = skb_hlen;
+	}
+
+	da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE);
 	if (dma_mapping_error(dev, da))
 		return -ENOMEM;

-	ash->dma_handle[0] = da;
-	ash->size[0] = skb_headlen(skb);
+	mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey);

-	tp->wqe_req.sgl[0].address = ash->dma_handle[0];
-	tp->wqe_req.sgl[0].mem_key = gd->gpa_mkey;
-	tp->wqe_req.sgl[0].size = ash->size[0];
+	if (sge1_len) {
+		sg_i = 1;
+		da = dma_map_single(dev, skb->data + sge0_len, sge1_len,
+				    DMA_TO_DEVICE);
+		if (dma_mapping_error(dev, da))
+			goto frag_err;
+
+		mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey);
+		hsg = 2;
+	}

 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
+		sg_i = hsg + i;
+
 		frag = &skb_shinfo(skb)->frags[i];
 		da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
 				      DMA_TO_DEVICE);
-
 		if (dma_mapping_error(dev, da))
 			goto frag_err;

-		ash->dma_handle[i + 1] = da;
-		ash->size[i + 1] = skb_frag_size(frag);
-
-		tp->wqe_req.sgl[i + 1].address = ash->dma_handle[i + 1];
-		tp->wqe_req.sgl[i + 1].mem_key = gd->gpa_mkey;
-		tp->wqe_req.sgl[i + 1].size = ash->size[i + 1];
+		mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag),
+			     gd->gpa_mkey);
 	}

 	return 0;

 frag_err:
-	for (i = i - 1; i >= 0; i--)
-		dma_unmap_page(dev, ash->dma_handle[i + 1], ash->size[i +

1],

quoted

+	for (i = sg_i - 1; i >= hsg; i--)
+		dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
 			       DMA_TO_DEVICE);

-	dma_unmap_single(dev, ash->dma_handle[0], ash->size[0],

DMA_TO_DEVICE);

quoted

+	for (i = hsg - 1; i >= 0; i--)
+		dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
+				 DMA_TO_DEVICE);

 	return -ENOMEM;
 }

+/* Handle the case when GSO SKB linear length is too large.
+ * MANA NIC requires GSO packets to put only the packet header to SGE0.
+ * So, we need 2 SGEs for the skb linear part which contains more than the
+ * header.
+ */
+static inline int mana_fix_skb_head(struct net_device *ndev,
+				    struct sk_buff *skb, int gso_hs,
+				    u32 *num_sge)
+{
+	int skb_hlen = skb_headlen(skb);
+
+	if (gso_hs < skb_hlen) {
+		*num_sge = 2 + skb_shinfo(skb)->nr_frags;
+	} else if (gso_hs > skb_hlen) {
+		if (net_ratelimit())
+			netdev_err(ndev,
+				   "TX nonlinear head: hs:%d, skb_hlen:%d\n",
+				   gso_hs, skb_hlen);
+
+		return -EINVAL;
+	}
+
+	return 0;

nit: I think it would be slightly nicer if the num_sge parameter of this
function was removed and it returned negative values on error (already
the case) and positive values, representing the number f segments, on success.

Will do.

quoted

+}
+
+/* Get the GSO packet's header size */
+static inline int mana_get_gso_hs(struct sk_buff *skb)
+{
+	int gso_hs;
+
+	if (skb->encapsulation) {
+		gso_hs = skb_inner_tcp_all_headers(skb);
+	} else {
+		if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
+			gso_hs = skb_transport_offset(skb) +
+				 sizeof(struct udphdr);
+		} else {
+			gso_hs = skb_tcp_all_headers(skb);
+		}
+	}
+
+	return gso_hs;
+}
+
 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 {
 	enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
 	struct mana_port_context *apc = netdev_priv(ndev);
+	int gso_hs = 0; /* zero for non-GSO pkts */
 	u16 txq_idx = skb_get_queue_mapping(skb);
 	struct gdma_dev *gd = apc->ac->gdma_dev;
 	bool ipv4 = false, ipv6 = false;

@@ -159,7 +232,6 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb,

struct net_device *ndev)

quoted

 	struct mana_txq *txq;
 	struct mana_cq *cq;
 	int err, len;
-	u16 ihs;

 	if (unlikely(!apc->port_is_up))
 		goto tx_drop;

@@ -209,19 +281,6 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb,

struct net_device *ndev)

quoted

 	pkg.wqe_req.client_data_unit = 0;

 	pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
-	WARN_ON_ONCE(pkg.wqe_req.num_sge >

MAX_TX_WQE_SGL_ENTRIES);

quoted

-
-	if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
-		pkg.wqe_req.sgl = pkg.sgl_array;
-	} else {
-		pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
-					    sizeof(struct gdma_sge),
-					    GFP_ATOMIC);
-		if (!pkg.sgl_ptr)
-			goto tx_drop_count;
-
-		pkg.wqe_req.sgl = pkg.sgl_ptr;
-	}

It is unclear to me why this logic has moved from here to further
down in this function. Is it to avoid some cases where
alloation has to be unwond on error (when mana_fix_skb_head() fails) ?
If so, this feels more like an optimisation than a fix.

mana_fix_skb_head() may add one more sge (success case) so the sgl 
allocation should be done later. Otherwise, we need to free / re-allocate 
the array later.

quoted

 	if (skb->protocol == htons(ETH_P_IP))
 		ipv4 = true;

@@ -229,6 +288,23 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb,

struct net_device *ndev)

quoted

 		ipv6 = true;

 	if (skb_is_gso(skb)) {
+		gso_hs = mana_get_gso_hs(skb);
+
+		if (mana_fix_skb_head(ndev, skb, gso_hs,

&pkg.wqe_req.num_sge))

quoted

+			goto tx_drop_count;
+
+		if (skb->encapsulation) {
+			u64_stats_update_begin(&tx_stats->syncp);
+			tx_stats->tso_inner_packets++;
+			tx_stats->tso_inner_bytes += skb->len - gso_hs;
+			u64_stats_update_end(&tx_stats->syncp);
+		} else {
+			u64_stats_update_begin(&tx_stats->syncp);
+			tx_stats->tso_packets++;
+			tx_stats->tso_bytes += skb->len - gso_hs;
+			u64_stats_update_end(&tx_stats->syncp);
+		}

nit: I wonder if this could be slightly more succinctly written as:

		u64_stats_update_begin(&tx_stats->syncp);
		if (skb->encapsulation) {
			tx_stats->tso_inner_packets++;
			tx_stats->tso_inner_bytes += skb->len - gso_hs;
		} else {
			tx_stats->tso_packets++;
			tx_stats->tso_bytes += skb->len - gso_hs;
		}
		u64_stats_update_end(&tx_stats->syncp);

Yes it can be written this way:)

Also, it is unclear to me why the stats logic is moved here from
futher down in the same block. It feels more like a clean-up than a fix
(as, btw, is my suggestion immediately above).

Since we need to calculate the gso_hs and fix head earlier than the stats and 
some other work, I move it immediately after skb_is_gso(skb).
The gso_hs calculation was part of the tx_stats block, so the tx_stats is moved 
together to remain close to the gso_hs calculation to keep readability.

quoted

+
 		pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
 		pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;

@@ -252,26 +328,6 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb,

struct net_device *ndev)

quoted

 						 &ipv6_hdr(skb)->daddr, 0,
 						 IPPROTO_TCP, 0);
 		}
-
-		if (skb->encapsulation) {
-			ihs = skb_inner_tcp_all_headers(skb);
-			u64_stats_update_begin(&tx_stats->syncp);
-			tx_stats->tso_inner_packets++;
-			tx_stats->tso_inner_bytes += skb->len - ihs;
-			u64_stats_update_end(&tx_stats->syncp);
-		} else {
-			if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
-				ihs = skb_transport_offset(skb) + sizeof(struct

udphdr);

quoted

-			} else {
-				ihs = skb_tcp_all_headers(skb);
-			}
-
-			u64_stats_update_begin(&tx_stats->syncp);
-			tx_stats->tso_packets++;
-			tx_stats->tso_bytes += skb->len - ihs;
-			u64_stats_update_end(&tx_stats->syncp);
-		}
-
 	} else if (skb->ip_summed == CHECKSUM_PARTIAL) {
 		csum_type = mana_checksum_info(skb);

@@ -294,11 +350,25 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb,

struct net_device *ndev)

quoted

 		} else {
 			/* Can't do offload of this type of checksum */
 			if (skb_checksum_help(skb))
-				goto free_sgl_ptr;
+				goto tx_drop_count;
 		}
 	}

-	if (mana_map_skb(skb, apc, &pkg)) {
+	WARN_ON_ONCE(pkg.wqe_req.num_sge >

MAX_TX_WQE_SGL_ENTRIES);

quoted

+
+	if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
+		pkg.wqe_req.sgl = pkg.sgl_array;
+	} else {
+		pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
+					    sizeof(struct gdma_sge),
+					    GFP_ATOMIC);
+		if (!pkg.sgl_ptr)
+			goto tx_drop_count;
+
+		pkg.wqe_req.sgl = pkg.sgl_ptr;
+	}
+
+	if (mana_map_skb(skb, apc, &pkg, gso_hs)) {
 		u64_stats_update_begin(&tx_stats->syncp);
 		tx_stats->mana_map_err++;
 		u64_stats_update_end(&tx_stats->syncp);

@@ -1255,12 +1325,18 @@ static void mana_unmap_skb(struct sk_buff *skb,

struct mana_port_context *apc)

quoted

 {
 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
+	int hsg = 1; /* num of SGEs of linear part */
 	struct device *dev = gc->dev;
 	int i;

-	dma_unmap_single(dev, ash->dma_handle[0], ash->size[0],

DMA_TO_DEVICE);

quoted

+	if (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0])
+		hsg = 2;

nit: Maybe this is nicer?

	/* num of SGEs of linear part */
	hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;

Will do.

Thanks,
- Haiyang