• Guillaume Nault's avatar
    inet: frags: re-introduce skb coalescing for local delivery · 891584f4
    Guillaume Nault authored
    Before commit d4289fcc ("net: IP6 defrag: use rbtrees for IPv6
    defrag"), a netperf UDP_STREAM test[0] using big IPv6 datagrams (thus
    generating many fragments) and running over an IPsec tunnel, reported
    more than 6Gbps throughput. After that patch, the same test gets only
    9Mbps when receiving on a be2net nic (driver can make a big difference
    here, for example, ixgbe doesn't seem to be affected).
    
    By reusing the IPv4 defragmentation code, IPv6 lost fragment coalescing
    (IPv4 fragment coalescing was dropped by commit 14fe22e3 ("Revert
    "ipv4: use skb coalescing in defragmentation"")).
    
    Without fragment coalescing, be2net runs out of Rx ring entries and
    starts to drop frames (ethtool reports rx_drops_no_frags errors). Since
    the netperf traffic is only composed of UDP fragments, any lost packet
    prevents reassembly of the full datagram. Therefore, fragments which
    have no possibility to ever get reassembled pile up in the reassembly
    queue, until the memory accounting exeeds the threshold. At that point
    no fragment is accepted anymore, which effectively discards all
    netperf traffic.
    
    When reassembly timeout expires, some stale fragments are removed from
    the reassembly queue, so a few packets can be received, reassembled
    and delivered to the netperf receiver. But the nic still drops frames
    and soon the reassembly queue gets filled again with stale fragments.
    These long time frames where no datagram can be received explain why
    the performance drop is so significant.
    
    Re-introducing fragment coalescing is enough to get the initial
    performances again (6.6Gbps with be2net): driver doesn't drop frames
    anymore (no more rx_drops_no_frags errors) and the reassembly engine
    works at full speed.
    
    This patch is quite conservative and only coalesces skbs for local
    IPv4 and IPv6 delivery (in order to avoid changing skb geometry when
    forwarding). Coalescing could be extended in the future if need be, as
    more scenarios would probably benefit from it.
    
    [0]: Test configuration
    Sender:
    ip xfrm policy flush
    ip xfrm state flush
    ip xfrm state add src fc00:1::1 dst fc00:2::1 proto esp spi 0x1000 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:1::1 dst fc00:2::1
    ip xfrm policy add src fc00:1::1 dst fc00:2::1 dir in tmpl src fc00:1::1 dst fc00:2::1 proto esp mode transport action allow
    ip xfrm state add src fc00:2::1 dst fc00:1::1 proto esp spi 0x1001 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:2::1 dst fc00:1::1
    ip xfrm policy add src fc00:2::1 dst fc00:1::1 dir out tmpl src fc00:2::1 dst fc00:1::1 proto esp mode transport action allow
    netserver -D -L fc00:2::1
    
    Receiver:
    ip xfrm policy flush
    ip xfrm state flush
    ip xfrm state add src fc00:2::1 dst fc00:1::1 proto esp spi 0x1001 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:2::1 dst fc00:1::1
    ip xfrm policy add src fc00:2::1 dst fc00:1::1 dir in tmpl src fc00:2::1 dst fc00:1::1 proto esp mode transport action allow
    ip xfrm state add src fc00:1::1 dst fc00:2::1 proto esp spi 0x1000 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:1::1 dst fc00:2::1
    ip xfrm policy add src fc00:1::1 dst fc00:2::1 dir out tmpl src fc00:1::1 dst fc00:2::1 proto esp mode transport action allow
    netperf -H fc00:2::1 -f k -P 0 -L fc00:1::1 -l 60 -t UDP_STREAM -I 99,5 -i 5,5 -T5,5 -6
    Signed-off-by: default avatarGuillaume Nault <gnault@redhat.com>
    Acked-by: default avatarFlorian Westphal <fw@strlen.de>
    Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
    891584f4
inet_frag.h 4.73 KB