On Sat, Aug 27, 2016 at 07:37:54AM -0700, Eric Dumazet wrote:
> From: Eric Dumazet <eduma...@google.com>
>
> When TCP operates in lossy environments (between 1 and 10 % packet
> losses), many SACK blocks can be exchanged, and I noticed we could
> drop them on busy senders, if these SACK blocks have to be queued
> into the socket backlog.
>
> While the main cause is the poor performance of RACK/SACK processing,
> we can try to avoid these drops of valuable information that can lead to
> spurious timeouts and retransmits.
>
> Cause of the drops is the skb->truesize overestimation caused by :
>
> - drivers allocating ~2048 (or more) bytes as a fragment to hold an
> Ethernet frame.
>
> - various pskb_may_pull() calls bringing the headers into skb->head
> might have pulled all the frame content, but skb->truesize could
> not be lowered, as the stack has no idea of each fragment truesize.
>
> The backlog drops are also more visible on bidirectional flows, since
> their sk_rmem_alloc can be quite big.
>
> Let's add some room for the backlog, as only the socket owner
> can selectively take action to lower memory needs, like collapsing
> receive queues or partial ofo pruning.
>
> Signed-off-by: Eric Dumazet <eduma...@google.com>
> Cc: Yuchung Cheng <ych...@google.com>
> Cc: Neal Cardwell <ncardw...@google.com>
> ---
> include/net/tcp.h | 1 +
> net/ipv4/tcp_ipv4.c | 33 +++++++++++++++++++++++++++++----
> net/ipv6/tcp_ipv6.c | 5 +----
> 3 files changed, 31 insertions(+), 8 deletions(-)
>
> diff --git a/include/net/tcp.h b/include/net/tcp.h
> index
> 25d64f6de69e1f639ed1531bf2d2df3f00fd76a2..5f5f09f6e019682ef29c864d2f43a8f247fcdd9a
> 100644
> --- a/include/net/tcp.h
> +++ b/include/net/tcp.h
> @@ -1163,6 +1163,7 @@ static inline void tcp_prequeue_init(struct tcp_sock
> *tp)
> }
>
> bool tcp_prequeue(struct sock *sk, struct sk_buff *skb);
> +bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb);
>
> #undef STATE_TRACE
>
> diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> index
> ad41e8ecf796bba1bd6d9ed155ca4a57ced96844..53e80cd004b6ce401c3acbb4b243b243c5c3c4a3
> 100644
> --- a/net/ipv4/tcp_ipv4.c
> +++ b/net/ipv4/tcp_ipv4.c
> @@ -1532,6 +1532,34 @@ bool tcp_prequeue(struct sock *sk, struct sk_buff *skb)
> }
> EXPORT_SYMBOL(tcp_prequeue);
>
> +bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb)
> +{
> + u32 limit = sk->sk_rcvbuf + sk->sk_sndbuf;
> +
> + /* Only socket owner can try to collapse/prune rx queues
> + * to reduce memory overhead, so add a little headroom here.
> + * Few sockets backlog are possibly concurrently non empty.
> + */
> + limit += 64*1024;
> +
> + /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
> + * we can fix skb->truesize to its real value to avoid future drops.
> + * This is valid because skb is not yet charged to the socket.
> + * It has been noticed pure SACK packets were sometimes dropped
> + * (if cooked by drivers without copybreak feature).
> + */
> + if (!skb->data_len)
> + skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
Shouldn't __pskb_pull_tail() already fix this? As it seems the expected
behavior and it would have a more global effect then. For drivers not
using copybreak, that's needed here anyway, but maybe this help other
protocols/situations too.
Thanks,
Marcelo
