Linux kernel commit 576eb62598f10c8c7fd75703fe89010cdcfff596 ('bridge:
respect RFC2863 operational state') from 2012 introduced a regression
for using wpa_supplicant with EAPOL frames and a station interface in a
bridge. Since it does not look like this regression is going to get
fixed any time soon (it is already two years from that commit and over
1.5 from a discussion pointing out the regression), add a workaround in
wpa_supplicant to avoid this issue.
The wpa_supplicant workaround uses a secondary packet socket to capture
all frames (ETH_P_ALL) from the netdev that is in a bridge. This is
needed to avoid the kernel regression. However, this comes at the price
of more CPU load. Some of this is avoided with use of Linux socket
filter, but still, this is less efficient than a packet socket bound to
the specific EAPOL ethertype. The workaround gets disabled
automatically, if the main packet socket interface on the bridge
interface turns out to be working for RX (e.g., due to an old kernel
version being used or a new kernel version having a fix for the
regression). In addition, this workaround is only taken into use for the
special case of running wpa_supplicant with an interface in a bridge.
Signed-off-by: Jouni Malinen <j@w1.fi>
The sample code here ended up trying to register an eloop socket with fd
== -1. This was not really ever supposed to be used, but it is now also
hitting an assert in eloop. Skip the unnecessary
eloop_register_read_sock() to avoid this.
This was causing issues for hostapd since CONFIG_L2_PACKET is not set by
default. If CONFIG_RSN_PREAUTH=y was not used for CONFIG_L2_PACKET was
not set in .config explicitly, the defaul use of l2_packet_none.c ended
up hitting the newly added assert() in eloop.
Signed-off-by: Jouni Malinen <j@w1.fi>
