[Dnsmasq-discuss] Announce: dnsmasq-2.28
Simon Kelley
simon at thekelleys.org.uk
Mon Apr 17 14:44:09 BST 2006
Dnsmasq 2.28 is now available: this fixes the problems which people have
been seeing with 2.27 on older Linux kernels, and the crashes when
linking with uclibc.
Cheers,
Simon.
Full changelog:
Eliminated all raw network access when running on
Linux. All DHCP network activity now goes through the IP
stack. Packet sockets are no longer required. Apart from
being a neat hack, this should also allow DHCP over IPsec
to work better. On *BSD and OS X, the old method of raw net
access through BPF is retained.
Simplified build options. Networking is now slimmed down
to a choice of "linux" or "other". Netlink is always used
under Linux. Since netlink has been available since 2.2
and non-optional in an IPv4-configured kernel since 2.4,
and the dnsmasq netlink code is now well tested, this
should work out fine.
Removed decayed build support for libc5 and Solaris.
Removed pselect code: use a pipe for race-free signal
handling instead, as this works everywhere.
No longer enable the ISC leasefile reading code in the
distributed sources. I doubt there are many people left
using this 1.x compatibility code. Those that are will
have to explicitly enable it in src/config.h.
Don't send the "DHCP maximum message size" option, even if
requested. RFC2131 says this is a "MUST NOT".
Support larger-than-minimum DHCP message. Dnsmasq is now
happy to get larger than 576-byte DHCP messages, and will
return large messages, if permitted by the "maximum
message size" option of the message to which it is
replying. There's now an arbitrary sanity limit of 16384
bytes.
Added --no-ping option. This fixes an RFC2131 "SHOULD".
Building on the 2.27 MAC-address changes, allow clients to
provide no MAC address at all, relying on the client-id as
a unique identifier. This should make things like DHCP for
USB come easier.
Fixed regression in netlink code under 2.2.x kernels which
occurred in 2.27. Erik Jan Tromp is the vintage kernel fan
who found this. P.S. It looks like this "netlink bind:
permission denied" problem occured in kernels at least as
late a 2.4.18. Good information from Alain Richoux.
Added a warning when it's impossible to give a host its
configured address because the address is leased
elsewhere. A sensible suggestion from Mircea Bardac.
Added minimal support for RFC 3046 DHCP relay agent-id
options. The DHCP server now echoes these back to the
relay, as required by the RFC. Also, RFC 3527 link selection
sub-options are honoured.
Set the process "dumpable" flag when running in debug
mode: this makes getting core dumps from root processes
much easier.
Fixed one-byte buffer overflow which seems to only cause
problems when dnsmasq is linked with uclibc. Thanks to
Eric House and Eric Spakman for help in chasing this down.
Tolerate configuration screwups which lead to the DHCP
server attemping to allocate its own address to a
client; eg setting the whole subnet range as a DHCP
range. Addresses in use by the server are now excluded
from use by clients.
Did some thinking about HAVE_BROKEN_RTC mode, and made it
much simpler and better. The key is to just keep lease
lengths in the lease file. Since these normally never
change, even as the lease is renewed, the lease file never
needs to change except when machines arrive on the network
or leave. This eliminates the code for timed writes, and
reduces the amount of wear on a flash filesystem to the
absolute minimum. Also re-did the basic time function in
this mode to use the portable times(), rather than parsing
/proc/uptime.
Believe the source port number when replying to unicast
DHCP requests and DHCP requests via a relay, instead of
always using the standard ports. This will allow relays on
non-standard ports and DHCPINFORM from unprivileged ports
to work. The source port sent by unconfigured clients is
still ignored, since this may be unreliable. This means that
a DHCP client must use the standard port to do full
configuration.
More information about the Dnsmasq-discuss
mailing list