From: Eric Dumazet <[email protected]>
commit 0f6925b3e8da0dbbb52447ca8a8b42b371aac7db upstream.
Xuan Zhuo reported that commit 3226b158e67c ("net: avoid 32 x truesize
under-estimation for tiny skbs") brought a ~10% performance drop.
The reason for the performance drop was that GRO was forced
to chain sk_buff (using skb_shinfo(skb)->frag_list), which
uses more memory but also cause packet consumers to go over
a lot of overhead handling all the tiny skbs.
It turns out that virtio_net page_to_skb() has a wrong strategy :
It allocates skbs with GOOD_COPY_LEN (128) bytes in skb->head, then
copies 128 bytes from the page, before feeding the packet to GRO stack.
This was suboptimal before commit 3226b158e67c ("net: avoid 32 x truesize
under-estimation for tiny skbs") because GRO was using 2 frags per MSS,
meaning we were not packing MSS with 100% efficiency.
Fix is to pull only the ethernet header in page_to_skb()
Then, we change virtio_net_hdr_to_skb() to pull the missing
headers, instead of assuming they were already pulled by callers.
This fixes the performance regression, but could also allow virtio_net
to accept packets with more than 128bytes of headers.
Many thanks to Xuan Zhuo for his report, and his tests/help.
Fixes: 3226b158e67c ("net: avoid 32 x truesize under-estimation for tiny skbs")
Reported-by: Xuan Zhuo <[email protected]>
Link: https://www.spinics.net/lists/netdev/msg731397.html
Co-Developed-by: Xuan Zhuo <[email protected]>
Signed-off-by: Xuan Zhuo <[email protected]>
Signed-off-by: Eric Dumazet <[email protected]>
Cc: "Michael S. Tsirkin" <[email protected]>
Cc: Jason Wang <[email protected]>
Cc: [email protected]
Acked-by: Jason Wang <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
---
drivers/net/virtio_net.c | 10 +++++++---
include/linux/virtio_net.h | 14 +++++++++-----
2 files changed, 16 insertions(+), 8 deletions(-)
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -406,9 +406,13 @@ static struct sk_buff *page_to_skb(struc
offset += hdr_padded_len;
p += hdr_padded_len;
- copy = len;
- if (copy > skb_tailroom(skb))
- copy = skb_tailroom(skb);
+ /* Copy all frame if it fits skb->head, otherwise
+ * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
+ */
+ if (len <= skb_tailroom(skb))
+ copy = len;
+ else
+ copy = ETH_HLEN + metasize;
skb_put_data(skb, p, copy);
if (metasize) {
--- a/include/linux/virtio_net.h
+++ b/include/linux/virtio_net.h
@@ -65,14 +65,18 @@ static inline int virtio_net_hdr_to_skb(
skb_reset_mac_header(skb);
if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
- u16 start = __virtio16_to_cpu(little_endian, hdr->csum_start);
- u16 off = __virtio16_to_cpu(little_endian, hdr->csum_offset);
+ u32 start = __virtio16_to_cpu(little_endian, hdr->csum_start);
+ u32 off = __virtio16_to_cpu(little_endian, hdr->csum_offset);
+ u32 needed = start + max_t(u32, thlen, off + sizeof(__sum16));
+
+ if (!pskb_may_pull(skb, needed))
+ return -EINVAL;
if (!skb_partial_csum_set(skb, start, off))
return -EINVAL;
p_off = skb_transport_offset(skb) + thlen;
- if (p_off > skb_headlen(skb))
+ if (!pskb_may_pull(skb, p_off))
return -EINVAL;
} else {
/* gso packets without NEEDS_CSUM do not set transport_offset.
@@ -102,14 +106,14 @@ retry:
}
p_off = keys.control.thoff + thlen;
- if (p_off > skb_headlen(skb) ||
+ if (!pskb_may_pull(skb, p_off) ||
keys.basic.ip_proto != ip_proto)
return -EINVAL;
skb_set_transport_header(skb, keys.control.thoff);
} else if (gso_type) {
p_off = thlen;
- if (p_off > skb_headlen(skb))
+ if (!pskb_may_pull(skb, p_off))
return -EINVAL;
}
}
On Mon, Apr 12, 2021 at 10:39:29AM +0200, Greg Kroah-Hartman wrote:
> From: Eric Dumazet <[email protected]>
>
> commit 0f6925b3e8da0dbbb52447ca8a8b42b371aac7db upstream.
>
> Xuan Zhuo reported that commit 3226b158e67c ("net: avoid 32 x truesize
> under-estimation for tiny skbs") brought a ~10% performance drop.
>
> The reason for the performance drop was that GRO was forced
> to chain sk_buff (using skb_shinfo(skb)->frag_list), which
> uses more memory but also cause packet consumers to go over
> a lot of overhead handling all the tiny skbs.
>
> It turns out that virtio_net page_to_skb() has a wrong strategy :
> It allocates skbs with GOOD_COPY_LEN (128) bytes in skb->head, then
> copies 128 bytes from the page, before feeding the packet to GRO stack.
>
> This was suboptimal before commit 3226b158e67c ("net: avoid 32 x truesize
> under-estimation for tiny skbs") because GRO was using 2 frags per MSS,
> meaning we were not packing MSS with 100% efficiency.
>
> Fix is to pull only the ethernet header in page_to_skb()
>
> Then, we change virtio_net_hdr_to_skb() to pull the missing
> headers, instead of assuming they were already pulled by callers.
>
> This fixes the performance regression, but could also allow virtio_net
> to accept packets with more than 128bytes of headers.
>
> Many thanks to Xuan Zhuo for his report, and his tests/help.
>
> Fixes: 3226b158e67c ("net: avoid 32 x truesize under-estimation for tiny skbs")
> Reported-by: Xuan Zhuo <[email protected]>
> Link: https://www.spinics.net/lists/netdev/msg731397.html
> Co-Developed-by: Xuan Zhuo <[email protected]>
> Signed-off-by: Xuan Zhuo <[email protected]>
> Signed-off-by: Eric Dumazet <[email protected]>
> Cc: "Michael S. Tsirkin" <[email protected]>
> Cc: Jason Wang <[email protected]>
> Cc: [email protected]
> Acked-by: Jason Wang <[email protected]>
> Signed-off-by: David S. Miller <[email protected]>
> Signed-off-by: Greg Kroah-Hartman <[email protected]>
Note that an issue related to this patch was recently reported.
It's quite possible that the root cause is a bug elsewhere
in the kernel, but it probably makes sense to defer the backport
until we know more ...
> ---
> drivers/net/virtio_net.c | 10 +++++++---
> include/linux/virtio_net.h | 14 +++++++++-----
> 2 files changed, 16 insertions(+), 8 deletions(-)
>
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -406,9 +406,13 @@ static struct sk_buff *page_to_skb(struc
> offset += hdr_padded_len;
> p += hdr_padded_len;
>
> - copy = len;
> - if (copy > skb_tailroom(skb))
> - copy = skb_tailroom(skb);
> + /* Copy all frame if it fits skb->head, otherwise
> + * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
> + */
> + if (len <= skb_tailroom(skb))
> + copy = len;
> + else
> + copy = ETH_HLEN + metasize;
> skb_put_data(skb, p, copy);
>
> if (metasize) {
> --- a/include/linux/virtio_net.h
> +++ b/include/linux/virtio_net.h
> @@ -65,14 +65,18 @@ static inline int virtio_net_hdr_to_skb(
> skb_reset_mac_header(skb);
>
> if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
> - u16 start = __virtio16_to_cpu(little_endian, hdr->csum_start);
> - u16 off = __virtio16_to_cpu(little_endian, hdr->csum_offset);
> + u32 start = __virtio16_to_cpu(little_endian, hdr->csum_start);
> + u32 off = __virtio16_to_cpu(little_endian, hdr->csum_offset);
> + u32 needed = start + max_t(u32, thlen, off + sizeof(__sum16));
> +
> + if (!pskb_may_pull(skb, needed))
> + return -EINVAL;
>
> if (!skb_partial_csum_set(skb, start, off))
> return -EINVAL;
>
> p_off = skb_transport_offset(skb) + thlen;
> - if (p_off > skb_headlen(skb))
> + if (!pskb_may_pull(skb, p_off))
> return -EINVAL;
> } else {
> /* gso packets without NEEDS_CSUM do not set transport_offset.
> @@ -102,14 +106,14 @@ retry:
> }
>
> p_off = keys.control.thoff + thlen;
> - if (p_off > skb_headlen(skb) ||
> + if (!pskb_may_pull(skb, p_off) ||
> keys.basic.ip_proto != ip_proto)
> return -EINVAL;
>
> skb_set_transport_header(skb, keys.control.thoff);
> } else if (gso_type) {
> p_off = thlen;
> - if (p_off > skb_headlen(skb))
> + if (!pskb_may_pull(skb, p_off))
> return -EINVAL;
> }
> }
>
On Mon, Apr 12, 2021 at 05:11:40AM -0400, Michael S. Tsirkin wrote:
> On Mon, Apr 12, 2021 at 10:39:29AM +0200, Greg Kroah-Hartman wrote:
> > From: Eric Dumazet <[email protected]>
> >
> > commit 0f6925b3e8da0dbbb52447ca8a8b42b371aac7db upstream.
> >
> > Xuan Zhuo reported that commit 3226b158e67c ("net: avoid 32 x truesize
> > under-estimation for tiny skbs") brought a ~10% performance drop.
> >
> > The reason for the performance drop was that GRO was forced
> > to chain sk_buff (using skb_shinfo(skb)->frag_list), which
> > uses more memory but also cause packet consumers to go over
> > a lot of overhead handling all the tiny skbs.
> >
> > It turns out that virtio_net page_to_skb() has a wrong strategy :
> > It allocates skbs with GOOD_COPY_LEN (128) bytes in skb->head, then
> > copies 128 bytes from the page, before feeding the packet to GRO stack.
> >
> > This was suboptimal before commit 3226b158e67c ("net: avoid 32 x truesize
> > under-estimation for tiny skbs") because GRO was using 2 frags per MSS,
> > meaning we were not packing MSS with 100% efficiency.
> >
> > Fix is to pull only the ethernet header in page_to_skb()
> >
> > Then, we change virtio_net_hdr_to_skb() to pull the missing
> > headers, instead of assuming they were already pulled by callers.
> >
> > This fixes the performance regression, but could also allow virtio_net
> > to accept packets with more than 128bytes of headers.
> >
> > Many thanks to Xuan Zhuo for his report, and his tests/help.
> >
> > Fixes: 3226b158e67c ("net: avoid 32 x truesize under-estimation for tiny skbs")
> > Reported-by: Xuan Zhuo <[email protected]>
> > Link: https://www.spinics.net/lists/netdev/msg731397.html
> > Co-Developed-by: Xuan Zhuo <[email protected]>
> > Signed-off-by: Xuan Zhuo <[email protected]>
> > Signed-off-by: Eric Dumazet <[email protected]>
> > Cc: "Michael S. Tsirkin" <[email protected]>
> > Cc: Jason Wang <[email protected]>
> > Cc: [email protected]
> > Acked-by: Jason Wang <[email protected]>
> > Signed-off-by: David S. Miller <[email protected]>
> > Signed-off-by: Greg Kroah-Hartman <[email protected]>
>
>
>
> Note that an issue related to this patch was recently reported.
> It's quite possible that the root cause is a bug elsewhere
> in the kernel, but it probably makes sense to defer the backport
> until we know more ...
Thanks, I'll go drop it from all 4 queues. If you all find out that all
is good, and it should be added back, please let us at stable@vger know
about it.
thanks,
greg k-h