Skip to content
Commit 75c119af authored by Eric Dumazet's avatar Eric Dumazet Committed by David S. Miller
Browse files

tcp: implement rb-tree based retransmit queue



Using a linear list to store all skbs in write queue has been okay
for quite a while : O(N) is not too bad when N < 500.

Things get messy when N is the order of 100,000 : Modern TCP stacks
want 10Gbit+ of throughput even with 200 ms RTT flows.

40 ns per cache line miss means a full scan can use 4 ms,
blowing away CPU caches.

SACK processing often can use various hints to avoid parsing
whole retransmit queue. But with high packet losses and/or high
reordering, hints no longer work.

Sender has to process thousands of unfriendly SACK, accumulating
a huge socket backlog, burning a cpu and massively dropping packets.

Using an rb-tree for retransmit queue has been avoided for years
because it added complexity and overhead, but now is the time
to be more resistant and say no to quadratic behavior.

1) RTX queue is no longer part of the write queue : already sent skbs
are stored in one rb-tree.

2) Since reaching the head of write queue no longer needs
sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue

Tested:

 On receiver :
 netem on ingress : delay 150ms 200us loss 1
 GRO disabled to force stress and SACK storms.

for f in `seq 1 10`
do
 ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1
done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}'

Before patch :

323.87
351.48
339.59
338.62
306.72
204.07
304.93
291.88
202.47
176.88
   2840

After patch:

1700.83
2207.98
2070.17
1544.26
2114.76
2124.89
1693.14
1080.91
2216.82
1299.94
  18053

Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent f3319816
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment