neigh: increase queue_len_bytes to match wmem_default
authorEric Dumazet <edumazet@google.com>
Tue, 29 Aug 2017 22:16:01 +0000 (15:16 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 29 Aug 2017 23:10:50 +0000 (16:10 -0700)
commiteaa72dc47488d599439cd0fd0f8c4f1bcb3906bb
treed35829277414b465b199dcada5af2e70f6183533
parent0dd5759dbb1c9a862e7d90c09d6cf398c45f1100
neigh: increase queue_len_bytes to match wmem_default

Florian reported UDP xmit drops that could be root caused to the
too small neigh limit.

Current limit is 64 KB, meaning that even a single UDP socket would hit
it, since its default sk_sndbuf comes from net.core.wmem_default
(~212992 bytes on 64bit arches).

Once ARP/ND resolution is in progress, we should allow a little more
packets to be queued, at least for one producer.

Once neigh arp_queue is filled, a rogue socket should hit its sk_sndbuf
limit and either block in sendmsg() or return -EAGAIN.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Documentation/networking/ip-sysctl.txt
include/net/sock.h
net/core/sock.c
net/decnet/dn_neigh.c
net/ipv4/arp.c
net/ipv4/tcp_input.c
net/ipv6/ndisc.c