[PATCH bpf-next] xsk: build skb by page

This patch is used to construct skb based on page to save memory copy
overhead.

This has one problem:

We construct the skb by fill the data page as a frag into the skb. In
this way, the linear space is empty, and the header information is also
in the frag, not in the linear space, which is not allowed for some
network cards. For example, Mellanox Technologies MT27710 Family
[ConnectX-4 Lx] will get the following error message:

    mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
    00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
    WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
    00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
    00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb

I also tried to use build_skb to construct skb, but because of the
existence of skb_shinfo, it must be behind the linear space, so this
method is not working. We can't put skb_shinfo on desc->addr, it will be
exposed to users, this is not safe.

Finally, I added a feature NETIF_F_SKB_NO_LINEAR to identify whether the
network card supports the header information of the packet in the frag
and not in the linear space.

---------------- Performance Testing ------------

The test environment is Aliyun ECS server.
Test cmd:

xdpsock -i eth0 -t  -S -s <msg size>

Test result data:

size    64      512     1024    1500
copy    1916747 1775988 1600203 1440054
page    1974058 1953655 1945463 1904478
percent 3.0%    10.0%   21.58%  32.3%

Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
---
 drivers/net/virtio_net.c        |   2 +-
 include/linux/netdev_features.h |   5 +-
 net/ethtool/common.c            |   1 +
 net/xdp/xsk.c                   | 108 +++++++++++++++++++++++++++++++++-------
 4 files changed, 97 insertions(+), 19 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 4ecccb8..841a331 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -2985,7 +2985,7 @@ static int virtnet_probe(struct virtio_device *vdev)
 	/* Set up network device as normal. */
 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
 	dev->netdev_ops = &virtnet_netdev;
-	dev->features = NETIF_F_HIGHDMA;
+	dev->features = NETIF_F_HIGHDMA | NETIF_F_SKB_NO_LINEAR;
 
 	dev->ethtool_ops = &virtnet_ethtool_ops;
 	SET_NETDEV_DEV(dev, &vdev->dev);

diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
index 934de56..8dd28e2 100644
--- a/include/linux/netdev_features.h
+++ b/include/linux/netdev_features.h

@@ -85,9 +85,11 @@ enum {
 
 	NETIF_F_HW_MACSEC_BIT,		/* Offload MACsec operations */
 
+	NETIF_F_SKB_NO_LINEAR_BIT,	/* Allow skb linear is empty */
+
 	/*
 	 * Add your fresh new feature above and remember to update
-	 * netdev_features_strings[] in net/core/ethtool.c and maybe
+	 * netdev_features_strings[] in net/ethtool/common.c and maybe
 	 * some feature mask #defines below. Please also describe it
 	 * in Documentation/networking/netdev-features.rst.
 	 */

@@ -157,6 +159,7 @@ enum {
 #define NETIF_F_GRO_FRAGLIST	__NETIF_F(GRO_FRAGLIST)
 #define NETIF_F_GSO_FRAGLIST	__NETIF_F(GSO_FRAGLIST)
 #define NETIF_F_HW_MACSEC	__NETIF_F(HW_MACSEC)
+#define NETIF_F_SKB_NO_LINEAR	__NETIF_F(SKB_NO_LINEAR)
 
 /* Finds the next feature with the highest number of the range of start till 0.
  */

diff --git a/net/ethtool/common.c b/net/ethtool/common.c
index 24036e3..2f3d309 100644
--- a/net/ethtool/common.c
+++ b/net/ethtool/common.c

@@ -68,6 +68,7 @@
 	[NETIF_F_HW_TLS_RX_BIT] =	 "tls-hw-rx-offload",
 	[NETIF_F_GRO_FRAGLIST_BIT] =	 "rx-gro-list",
 	[NETIF_F_HW_MACSEC_BIT] =	 "macsec-hw-offload",
+	[NETIF_F_SKB_NO_LINEAR_BIT] =	 "skb-no-linear",
 };
 
 const char

diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index 8037b04..94d17dc 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c

@@ -430,6 +430,95 @@ static void xsk_destruct_skb(struct sk_buff *skb)
 	sock_wfree(skb);
 }
 
+static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
+					      struct xdp_desc *desc)
+{
+	u32 len, offset, copy, copied;
+	struct sk_buff *skb;
+	struct page *page;
+	char *buffer;
+	int err, i;
+	u64 addr;
+
+	skb = sock_alloc_send_skb(&xs->sk, 0, 1, &err);
+	if (unlikely(!skb))
+		return NULL;
+
+	addr = desc->addr;
+	len = desc->len;
+
+	buffer = xsk_buff_raw_get_data(xs->pool, addr);
+	offset = offset_in_page(buffer);
+	addr = buffer - (char *)xs->pool->addrs;
+
+	for (copied = 0, i = 0; copied < len; ++i) {
+		page = xs->pool->umem->pgs[addr >> PAGE_SHIFT];
+
+		get_page(page);
+
+		copy = min((u32)(PAGE_SIZE - offset), len - copied);
+
+		skb_fill_page_desc(skb, i, page, offset, copy);
+
+		copied += copy;
+		addr += copy;
+		offset = 0;
+	}
+
+	skb->len += len;
+	skb->data_len += len;
+	skb->truesize += len;
+
+	refcount_add(len, &xs->sk.sk_wmem_alloc);
+
+	return skb;
+}
+
+static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
+				     struct xdp_desc *desc, int *err)
+{
+	struct sk_buff *skb;
+
+	if (xs->dev->features & NETIF_F_SKB_NO_LINEAR) {
+		skb = xsk_build_skb_zerocopy(xs, desc);
+		if (unlikely(!skb)) {
+			*err = -ENOMEM;
+			return NULL;
+		}
+	} else {
+		char *buffer;
+		u64 addr;
+		u32 len;
+		int err;
+
+		len = desc->len;
+		skb = sock_alloc_send_skb(&xs->sk, len, 1, &err);
+		if (unlikely(!skb)) {
+			*err = -ENOMEM;
+			return NULL;
+		}
+
+		skb_put(skb, len);
+		addr = desc->addr;
+		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
+		err = skb_store_bits(skb, 0, buffer, len);
+
+		if (unlikely(err)) {
+			kfree_skb(skb);
+			*err = -EINVAL;
+			return NULL;
+		}
+	}
+
+	skb->dev = xs->dev;
+	skb->priority = xs->sk.sk_priority;
+	skb->mark = xs->sk.sk_mark;
+	skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
+	skb->destructor = xsk_destruct_skb;
+
+	return skb;
+}
+
 static int xsk_generic_xmit(struct sock *sk)
 {
 	struct xdp_sock *xs = xdp_sk(sk);

@@ -446,43 +535,28 @@ static int xsk_generic_xmit(struct sock *sk)
 		goto out;
 
 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
-		char *buffer;
-		u64 addr;
-		u32 len;
-
 		if (max_batch-- == 0) {
 			err = -EAGAIN;
 			goto out;
 		}
 
-		len = desc.len;
-		skb = sock_alloc_send_skb(sk, len, 1, &err);
+		skb = xsk_build_skb(xs, &desc, &err);
 		if (unlikely(!skb))
 			goto out;
 
-		skb_put(skb, len);
-		addr = desc.addr;
-		buffer = xsk_buff_raw_get_data(xs->pool, addr);
-		err = skb_store_bits(skb, 0, buffer, len);
 		/* This is the backpressure mechanism for the Tx path.
 		 * Reserve space in the completion queue and only proceed
 		 * if there is space in it. This avoids having to implement
 		 * any buffering in the Tx path.
 		 */
 		spin_lock_irqsave(&xs->pool->cq_lock, flags);
-		if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) {
+		if (xskq_prod_reserve(xs->pool->cq)) {
 			spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 			kfree_skb(skb);
 			goto out;
 		}
 		spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 
-		skb->dev = xs->dev;
-		skb->priority = sk->sk_priority;
-		skb->mark = sk->sk_mark;
-		skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
-		skb->destructor = xsk_destruct_skb;
-
 		err = __dev_direct_xmit(skb, xs->queue_id);
 		if  (err == NETDEV_TX_BUSY) {
 			/* Tell user-space to retry the send */

-- 
1.8.3.1

Hi Xuan,

Thank you for the patch! Perhaps something to improve:

[auto build test WARNING on bpf-next/master]

url:    https://github.com/0day-ci/linux/commits/Xuan-Zhuo/xsk-build-skb-by-page/20210116-105116
base:   https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master
config: sh-allmodconfig (attached as .config)
compiler: sh4-linux-gcc (GCC) 9.3.0
reproduce (this is a W=1 build):
        wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
        chmod +x ~/bin/make.cross
        # https://github.com/0day-ci/linux/commit/ee139d2988e5c5945108889a7c95c751910c1877
        git remote add linux-review https://github.com/0day-ci/linux
        git fetch --no-tags linux-review Xuan-Zhuo/xsk-build-skb-by-page/20210116-105116
        git checkout ee139d2988e5c5945108889a7c95c751910c1877
        # save the attached .config to linux build tree
        COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=sh 

If you fix the issue, kindly add following tag as appropriate
Reported-by: kernel test robot <redacted>

All warnings (new ones prefixed by >>):

   net/xdp/xsk.c: In function 'xsk_build_skb':
   net/xdp/xsk.c:497:4: error: invalid type argument of unary '*' (have 'int')
     497 |    *err = -ENOMEM;
         |    ^~~~
   net/xdp/xsk.c:508:4: error: invalid type argument of unary '*' (have 'int')
     508 |    *err = -EINVAL;
         |    ^~~~

quoted

net/xdp/xsk.c:490:7: warning: variable 'addr' set but not used [-Wunused-but-set-variable]

     490 |   u64 addr;
         |       ^~~~

Kconfig warnings: (for reference only)
   WARNING: unmet direct dependencies detected for SND_ATMEL_SOC_PDC
   Depends on SOUND && !UML && SND && SND_SOC && SND_ATMEL_SOC && HAS_DMA
   Selected by
   - SND_ATMEL_SOC_SSC && SOUND && !UML && SND && SND_SOC && SND_ATMEL_SOC
   - SND_ATMEL_SOC_SSC_PDC && SOUND && !UML && SND && SND_SOC && SND_ATMEL_SOC && ATMEL_SSC


vim +/addr +490 net/xdp/xsk.c

   476	
   477	static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
   478					     struct xdp_desc *desc, int *err)
   479	{
   480		struct sk_buff *skb;
   481	
   482		if (xs->dev->features & NETIF_F_SKB_NO_LINEAR) {
   483			skb = xsk_build_skb_zerocopy(xs, desc);
   484			if (unlikely(!skb)) {
   485				*err = -ENOMEM;
   486				return NULL;
   487			}
   488		} else {
   489			char *buffer;
 > 490			u64 addr;
   491			u32 len;
   492			int err;
   493	
   494			len = desc->len;
   495			skb = sock_alloc_send_skb(&xs->sk, len, 1, &err);
   496			if (unlikely(!skb)) {
 > 497				*err = -ENOMEM;
   498				return NULL;
   499			}
   500	
   501			skb_put(skb, len);
   502			addr = desc->addr;
   503			buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
   504			err = skb_store_bits(skb, 0, buffer, len);
   505	
   506			if (unlikely(err)) {
   507				kfree_skb(skb);
   508				*err = -EINVAL;
   509				return NULL;
   510			}
   511		}
   512	
   513		skb->dev = xs->dev;
   514		skb->priority = xs->sk.sk_priority;
   515		skb->mark = xs->sk.sk_mark;
   516		skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
   517		skb->destructor = xsk_destruct_skb;
   518	
   519		return skb;
   520	}
   521	

---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org

Hi Xuan,

Thank you for the patch! Yet something to improve:

[auto build test ERROR on bpf-next/master]

url:    https://github.com/0day-ci/linux/commits/Xuan-Zhuo/xsk-build-skb-by-page/20210116-105116
base:   https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master
config: sh-allmodconfig (attached as .config)
compiler: sh4-linux-gcc (GCC) 9.3.0
reproduce (this is a W=1 build):
        wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
        chmod +x ~/bin/make.cross
        # https://github.com/0day-ci/linux/commit/ee139d2988e5c5945108889a7c95c751910c1877
        git remote add linux-review https://github.com/0day-ci/linux
        git fetch --no-tags linux-review Xuan-Zhuo/xsk-build-skb-by-page/20210116-105116
        git checkout ee139d2988e5c5945108889a7c95c751910c1877
        # save the attached .config to linux build tree
        COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=sh 

If you fix the issue, kindly add following tag as appropriate
Reported-by: kernel test robot <redacted>

All errors (new ones prefixed by >>):

   net/xdp/xsk.c: In function 'xsk_build_skb':

quoted

net/xdp/xsk.c:497:4: error: invalid type argument of unary '*' (have 'int')

     497 |    *err = -ENOMEM;
         |    ^~~~
   net/xdp/xsk.c:508:4: error: invalid type argument of unary '*' (have 'int')
     508 |    *err = -EINVAL;
         |    ^~~~
   net/xdp/xsk.c:490:7: warning: variable 'addr' set but not used [-Wunused-but-set-variable]
     490 |   u64 addr;
         |       ^~~~

Kconfig warnings: (for reference only)
   WARNING: unmet direct dependencies detected for SND_ATMEL_SOC_PDC
   Depends on SOUND && !UML && SND && SND_SOC && SND_ATMEL_SOC && HAS_DMA
   Selected by
   - SND_ATMEL_SOC_SSC && SOUND && !UML && SND && SND_SOC && SND_ATMEL_SOC
   - SND_ATMEL_SOC_SSC_PDC && SOUND && !UML && SND && SND_SOC && SND_ATMEL_SOC && ATMEL_SSC


vim +497 net/xdp/xsk.c

   476	
   477	static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
   478					     struct xdp_desc *desc, int *err)
   479	{
   480		struct sk_buff *skb;
   481	
   482		if (xs->dev->features & NETIF_F_SKB_NO_LINEAR) {
   483			skb = xsk_build_skb_zerocopy(xs, desc);
   484			if (unlikely(!skb)) {
   485				*err = -ENOMEM;
   486				return NULL;
   487			}
   488		} else {
   489			char *buffer;
   490			u64 addr;
   491			u32 len;
   492			int err;
   493	
   494			len = desc->len;
   495			skb = sock_alloc_send_skb(&xs->sk, len, 1, &err);
   496			if (unlikely(!skb)) {
 > 497				*err = -ENOMEM;
   498				return NULL;
   499			}
   500	
   501			skb_put(skb, len);
   502			addr = desc->addr;
   503			buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
   504			err = skb_store_bits(skb, 0, buffer, len);
   505	
   506			if (unlikely(err)) {
   507				kfree_skb(skb);
   508				*err = -EINVAL;
   509				return NULL;
   510			}
   511		}
   512	
   513		skb->dev = xs->dev;
   514		skb->priority = xs->sk.sk_priority;
   515		skb->mark = xs->sk.sk_mark;
   516		skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
   517		skb->destructor = xsk_destruct_skb;
   518	
   519		return skb;
   520	}
   521	

---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org

Xuan Zhuo wrote:

This patch is used to construct skb based on page to save memory copy
overhead.

This has one problem:

We construct the skb by fill the data page as a frag into the skb. In
this way, the linear space is empty, and the header information is also
in the frag, not in the linear space, which is not allowed for some
network cards. For example, Mellanox Technologies MT27710 Family
[ConnectX-4 Lx] will get the following error message:

    mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
    00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
    WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
    00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
    00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb

I also tried to use build_skb to construct skb, but because of the
existence of skb_shinfo, it must be behind the linear space, so this
method is not working. We can't put skb_shinfo on desc->addr, it will be
exposed to users, this is not safe.

Finally, I added a feature NETIF_F_SKB_NO_LINEAR to identify whether the
network card supports the header information of the packet in the frag
and not in the linear space.

---------------- Performance Testing ------------

The test environment is Aliyun ECS server.
Test cmd:

xdpsock -i eth0 -t  -S -s <msg size>

Test result data:

size    64      512     1024    1500
copy    1916747 1775988 1600203 1440054
page    1974058 1953655 1945463 1904478
percent 3.0%    10.0%   21.58%  32.3%

Looks like a good perf bump. Some easy suggestions below

+static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
+				     struct xdp_desc *desc, int *err)
+{

Passing a 'int *err' here is ugly IMO use the ERR_PTR/PTR_ERR macros
and roll it into the return value.

or maybe use the out: pattern used in the kernel, but just doing direct
returns like now but with ERR_PTR() would also be fine.

+	struct sk_buff *skb ;

        struct sk_buff *skb = NULL;
        err = -ENOMEM;

+
+	if (xs->dev->features & NETIF_F_SKB_NO_LINEAR) {
+		skb = xsk_build_skb_zerocopy(xs, desc);
+		if (unlikely(!skb)) {

			goto out

+			*err = -ENOMEM;
+			return NULL;
+		}
+	} else {
+		char *buffer;
+		u64 addr;
+		u32 len;
+		int err;
+
+		len = desc->len;
+		skb = sock_alloc_send_skb(&xs->sk, len, 1, &err);
+		if (unlikely(!skb)) {

			goto out;

+			*err = -ENOMEM;
+			return NULL;
+		}
+
+		skb_put(skb, len);
+		addr = desc->addr;
+		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
+		err = skb_store_bits(skb, 0, buffer, len);
+
+		if (unlikely(err)) {
+			kfree_skb(skb);

			err = -EINVAL;
			goto out

+			*err = -EINVAL;
+			return NULL;
+		}
+	}
+
+	skb->dev = xs->dev;
+	skb->priority = xs->sk.sk_priority;
+	skb->mark = xs->sk.sk_mark;
+	skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
+	skb->destructor = xsk_destruct_skb;
+
+	return skb;

out:
	kfree_skb(skb)
	return ERR_PTR(err);

+}
+

Otherwise looks good thanks.

On Sat, Jan 16, 2021 at 3:47 AM Xuan Zhuo [off-list ref] wrote:

This patch is used to construct skb based on page to save memory copy
overhead.

This has one problem:

We construct the skb by fill the data page as a frag into the skb. In
this way, the linear space is empty, and the header information is also
in the frag, not in the linear space, which is not allowed for some
network cards. For example, Mellanox Technologies MT27710 Family
[ConnectX-4 Lx] will get the following error message:

    mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
    00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
    WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
    00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
    00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb

I also tried to use build_skb to construct skb, but because of the
existence of skb_shinfo, it must be behind the linear space, so this
method is not working. We can't put skb_shinfo on desc->addr, it will be
exposed to users, this is not safe.

Finally, I added a feature NETIF_F_SKB_NO_LINEAR to identify whether the
network card supports the header information of the packet in the frag
and not in the linear space.

Could you please break this new feature bit into its own patch and
produce a patch set. Patch 1 being the NETIF_F_SKB_NO_LINEAR bit
addition and patch 2 being the xsk part. I can then ack patch 2 since
I am one of the maintainers of that, and then the owner of the netdev
bits can ack that part (whoever that is?). The bit addition looks fine
to me, but I am not the owner of those feature bits, so I would feel
somewhat awkward to ack that.

Otherwise, good performance boost. I like it. Thank you!

On Sat, Jan 16, 2021 at 10:44:53AM +0800, Xuan Zhuo wrote:

This patch is used to construct skb based on page to save memory copy
overhead.

This has one problem:

We construct the skb by fill the data page as a frag into the skb. In
this way, the linear space is empty, and the header information is also
in the frag, not in the linear space, which is not allowed for some
network cards. For example, Mellanox Technologies MT27710 Family
[ConnectX-4 Lx] will get the following error message:

    mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
    00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
    WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
    00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
    00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb

I also tried to use build_skb to construct skb, but because of the
existence of skb_shinfo, it must be behind the linear space, so this
method is not working. We can't put skb_shinfo on desc->addr, it will be
exposed to users, this is not safe.

Finally, I added a feature NETIF_F_SKB_NO_LINEAR to identify whether the
network card supports the header information of the packet in the frag
and not in the linear space.

---------------- Performance Testing ------------

The test environment is Aliyun ECS server.
Test cmd:

xdpsock -i eth0 -t  -S -s <msg size>

Test result data:

size    64      512     1024    1500
copy    1916747 1775988 1600203 1440054
page    1974058 1953655 1945463 1904478
percent 3.0%    10.0%   21.58%  32.3%

Nice, but it looks like the patch presented wouldn't compile.
It's worth retesting after you actually make it compile.

Rather than passing int *err, you can return PTR_ERR.
Seems cleaner ...

+{
+	struct sk_buff *skb;
+
+	if (xs->dev->features & NETIF_F_SKB_NO_LINEAR) {
+		skb = xsk_build_skb_zerocopy(xs, desc);
+		if (unlikely(!skb)) {
+			*err = -ENOMEM;
+			return NULL;
+		}
+	} else {
+		char *buffer;
+		u64 addr;
+		u32 len;
+		int err;

So err is int here

+
+		len = desc->len;
+		skb = sock_alloc_send_skb(&xs->sk, len, 1, &err);
+		if (unlikely(!skb)) {
+			*err = -ENOMEM;

.. and you dereference it here

+			return NULL;
+		}
+
+		skb_put(skb, len);
+		addr = desc->addr;
+		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
+		err = skb_store_bits(skb, 0, buffer, len);
+
+		if (unlikely(err)) {
+			kfree_skb(skb);
+			*err = -EINVAL;

Same thing here ... how does it compile?

On 2021/1/16 10:44, Xuan Zhuo wrote:

This patch is used to construct skb based on page to save memory copy
overhead.

This has one problem:

We construct the skb by fill the data page as a frag into the skb. In
this way, the linear space is empty, and the header information is also
in the frag, not in the linear space, which is not allowed for some
network cards. For example, Mellanox Technologies MT27710 Family
[ConnectX-4 Lx] will get the following error message:

    mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
    00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
    WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
    00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
    00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb

I also tried to use build_skb to construct skb, but because of the
existence of skb_shinfo, it must be behind the linear space, so this
method is not working. We can't put skb_shinfo on desc->addr, it will be
exposed to users, this is not safe.

Finally, I added a feature NETIF_F_SKB_NO_LINEAR to identify whether the

Does it make sense to use ETHTOOL_TX_COPYBREAK tunable in ethtool to
configure if the data is copied or not?

From: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Date: Sat, 16 Jan 2021 10:44:53 +0800

I think the best would be if you will split this patch into three:
 - the first one will introduce NETI_F_SKB_NO_LINEAR;
 - the second will add this feature to virtio_net;
 - the third will do the rest.

Also, it would be nice if you'll mention (in the cover letter or
in the third patch) that in order to get a nice boost on non-ZC
XSK xmit developers can add a support for completely non-linear
skbs and advertise this new feature in their drivers. I think
there'll be enough folks wanting to do this.

As the others said, just use ERR_PTR() and PTR_ERR().
You also should have received the letters from kernel test robot
that the current version is non-compilable at all.

From: Yunsheng Lin <redacted>
Date: Mon, 18 Jan 2021 20:40:52 +0800

On 2021/1/16 10:44, Xuan Zhuo wrote:

quoted

This patch is used to construct skb based on page to save memory copy
overhead.

This has one problem:

We construct the skb by fill the data page as a frag into the skb. In
this way, the linear space is empty, and the header information is also
in the frag, not in the linear space, which is not allowed for some
network cards. For example, Mellanox Technologies MT27710 Family
[ConnectX-4 Lx] will get the following error message:

    mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
    00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
    WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
    00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
    00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
    00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb

I also tried to use build_skb to construct skb, but because of the
existence of skb_shinfo, it must be behind the linear space, so this
method is not working. We can't put skb_shinfo on desc->addr, it will be
exposed to users, this is not safe.

Finally, I added a feature NETIF_F_SKB_NO_LINEAR to identify whether the

Does it make sense to use ETHTOOL_TX_COPYBREAK tunable in ethtool to
configure if the data is copied or not?

As far as I can grep, only mlx4 supports this, and it has a different
meaning in that driver.
So I guess a new netdev_feature would be a better solution.

quoted

network card supports the header information of the packet in the frag
and not in the linear space.

---------------- Performance Testing ------------

The test environment is Aliyun ECS server.
Test cmd:

xdpsock -i eth0 -t  -S -s <msg size>

Test result data:

size    64      512     1024    1500
copy    1916747 1775988 1600203 1440054
page    1974058 1953655 1945463 1904478
percent 3.0%    10.0%   21.58%  32.3%

Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
---
 drivers/net/virtio_net.c        |   2 +-
 include/linux/netdev_features.h |   5 +-
 net/ethtool/common.c            |   1 +
 net/xdp/xsk.c                   | 108 +++++++++++++++++++++++++++++++++-------
 4 files changed, 97 insertions(+), 19 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 4ecccb8..841a331 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -2985,7 +2985,7 @@ static int virtnet_probe(struct virtio_device *vdev)
 	/* Set up network device as normal. */
 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
 	dev->netdev_ops = &virtnet_netdev;
-	dev->features = NETIF_F_HIGHDMA;
+	dev->features = NETIF_F_HIGHDMA | NETIF_F_SKB_NO_LINEAR;
 
 	dev->ethtool_ops = &virtnet_ethtool_ops;
 	SET_NETDEV_DEV(dev, &vdev->dev);

diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
index 934de56..8dd28e2 100644
--- a/include/linux/netdev_features.h
+++ b/include/linux/netdev_features.h

@@ -85,9 +85,11 @@ enum {
 
 	NETIF_F_HW_MACSEC_BIT,		/* Offload MACsec operations */
 
+	NETIF_F_SKB_NO_LINEAR_BIT,	/* Allow skb linear is empty */
+
 	/*
 	 * Add your fresh new feature above and remember to update
-	 * netdev_features_strings[] in net/core/ethtool.c and maybe
+	 * netdev_features_strings[] in net/ethtool/common.c and maybe
 	 * some feature mask #defines below. Please also describe it
 	 * in Documentation/networking/netdev-features.rst.
 	 */

@@ -157,6 +159,7 @@ enum {
 #define NETIF_F_GRO_FRAGLIST	__NETIF_F(GRO_FRAGLIST)
 #define NETIF_F_GSO_FRAGLIST	__NETIF_F(GSO_FRAGLIST)
 #define NETIF_F_HW_MACSEC	__NETIF_F(HW_MACSEC)
+#define NETIF_F_SKB_NO_LINEAR	__NETIF_F(SKB_NO_LINEAR)
 
 /* Finds the next feature with the highest number of the range of start till 0.
  */

diff --git a/net/ethtool/common.c b/net/ethtool/common.c
index 24036e3..2f3d309 100644
--- a/net/ethtool/common.c
+++ b/net/ethtool/common.c

@@ -68,6 +68,7 @@
 	[NETIF_F_HW_TLS_RX_BIT] =	 "tls-hw-rx-offload",
 	[NETIF_F_GRO_FRAGLIST_BIT] =	 "rx-gro-list",
 	[NETIF_F_HW_MACSEC_BIT] =	 "macsec-hw-offload",
+	[NETIF_F_SKB_NO_LINEAR_BIT] =	 "skb-no-linear",

I completely forgot to add that you'd better to mention in both
enumeration/feature and its Ethtool string that the feature applies
to Tx path.
Smth like:

NETIF_F_SKB_TX_NO_LINEAR{,_BIT}, "skb-tx-no-linear"
or
NETIF_F_TX_SKB_NO_LINEAR{,_BIT}, "tx-skb-no-linear"

Otherwise, it may be confusing for users and developers.

quoted

 };
 
 const char

diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index 8037b04..94d17dc 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c

@@ -430,6 +430,95 @@ static void xsk_destruct_skb(struct sk_buff *skb)
 	sock_wfree(skb);
 }
 
+static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
+					      struct xdp_desc *desc)
+{
+	u32 len, offset, copy, copied;
+	struct sk_buff *skb;
+	struct page *page;
+	char *buffer;
+	int err, i;
+	u64 addr;
+
+	skb = sock_alloc_send_skb(&xs->sk, 0, 1, &err);
+	if (unlikely(!skb))
+		return NULL;
+
+	addr = desc->addr;
+	len = desc->len;
+
+	buffer = xsk_buff_raw_get_data(xs->pool, addr);
+	offset = offset_in_page(buffer);
+	addr = buffer - (char *)xs->pool->addrs;
+
+	for (copied = 0, i = 0; copied < len; ++i) {
+		page = xs->pool->umem->pgs[addr >> PAGE_SHIFT];
+
+		get_page(page);
+
+		copy = min((u32)(PAGE_SIZE - offset), len - copied);
+
+		skb_fill_page_desc(skb, i, page, offset, copy);
+
+		copied += copy;
+		addr += copy;
+		offset = 0;
+	}
+
+	skb->len += len;
+	skb->data_len += len;
+	skb->truesize += len;
+
+	refcount_add(len, &xs->sk.sk_wmem_alloc);
+
+	return skb;
+}
+
+static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
+				     struct xdp_desc *desc, int *err)
+{
+	struct sk_buff *skb;
+
+	if (xs->dev->features & NETIF_F_SKB_NO_LINEAR) {
+		skb = xsk_build_skb_zerocopy(xs, desc);
+		if (unlikely(!skb)) {
+			*err = -ENOMEM;
+			return NULL;
+		}
+	} else {
+		char *buffer;
+		u64 addr;
+		u32 len;
+		int err;
+
+		len = desc->len;
+		skb = sock_alloc_send_skb(&xs->sk, len, 1, &err);
+		if (unlikely(!skb)) {
+			*err = -ENOMEM;
+			return NULL;
+		}
+
+		skb_put(skb, len);
+		addr = desc->addr;
+		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
+		err = skb_store_bits(skb, 0, buffer, len);
+
+		if (unlikely(err)) {
+			kfree_skb(skb);
+			*err = -EINVAL;
+			return NULL;
+		}
+	}
+
+	skb->dev = xs->dev;
+	skb->priority = xs->sk.sk_priority;
+	skb->mark = xs->sk.sk_mark;
+	skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
+	skb->destructor = xsk_destruct_skb;
+
+	return skb;
+}
+
 static int xsk_generic_xmit(struct sock *sk)
 {
 	struct xdp_sock *xs = xdp_sk(sk);

@@ -446,43 +535,28 @@ static int xsk_generic_xmit(struct sock *sk)
 		goto out;
 
 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
-		char *buffer;
-		u64 addr;
-		u32 len;
-
 		if (max_batch-- == 0) {
 			err = -EAGAIN;
 			goto out;
 		}
 
-		len = desc.len;
-		skb = sock_alloc_send_skb(sk, len, 1, &err);
+		skb = xsk_build_skb(xs, &desc, &err);
 		if (unlikely(!skb))
 			goto out;
 
-		skb_put(skb, len);
-		addr = desc.addr;
-		buffer = xsk_buff_raw_get_data(xs->pool, addr);
-		err = skb_store_bits(skb, 0, buffer, len);
 		/* This is the backpressure mechanism for the Tx path.
 		 * Reserve space in the completion queue and only proceed
 		 * if there is space in it. This avoids having to implement
 		 * any buffering in the Tx path.
 		 */
 		spin_lock_irqsave(&xs->pool->cq_lock, flags);
-		if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) {
+		if (xskq_prod_reserve(xs->pool->cq)) {
 			spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 			kfree_skb(skb);
 			goto out;
 		}
 		spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 
-		skb->dev = xs->dev;
-		skb->priority = sk->sk_priority;
-		skb->mark = sk->sk_mark;
-		skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
-		skb->destructor = xsk_destruct_skb;
-
 		err = __dev_direct_xmit(skb, xs->queue_id);
 		if  (err == NETDEV_TX_BUSY) {
 			/* Tell user-space to retry the send */