net: ipv4: use a dedicated counter for icmp_v4 redirect packets
authorLorenzo Bianconi <lorenzo.bianconi@redhat.com>
Wed, 6 Feb 2019 18:18:04 +0000 (19:18 +0100)
committerDavid S. Miller <davem@davemloft.net>
Sat, 9 Feb 2019 05:50:15 +0000 (21:50 -0800)
commitc09551c6ff7fe16a79a42133bcecba5fc2fc3291
treebc1eb9cd1f89d8bd62e3b4544e5a34fd4373d802
parentb5bfc21af5cb3d53f9cee0ef82eaa43762a90f81
net: ipv4: use a dedicated counter for icmp_v4 redirect packets

According to the algorithm described in the comment block at the
beginning of ip_rt_send_redirect, the host should try to send
'ip_rt_redirect_number' ICMP redirect packets with an exponential
backoff and then stop sending them at all assuming that the destination
ignores redirects.
If the device has previously sent some ICMP error packets that are
rate-limited (e.g TTL expired) and continues to receive traffic,
the redirect packets will never be transmitted. This happens since
peer->rate_tokens will be typically greater than 'ip_rt_redirect_number'
and so it will never be reset even if the redirect silence timeout
(ip_rt_redirect_silence) has elapsed without receiving any packet
requiring redirects.

Fix it by using a dedicated counter for the number of ICMP redirect
packets that has been sent by the host

I have not been able to identify a given commit that introduced the
issue since ip_rt_send_redirect implements the same rate-limiting
algorithm from commit 1da177e4c3f4 ("Linux-2.6.12-rc2")

Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/inetpeer.h
net/ipv4/inetpeer.c
net/ipv4/route.c