Ben Pfaff [Tue, 29 Apr 2008 01:16:05 +0000 (18:16 -0700)]
Make dumping large numbers of flows possible.
This changes the kernel switch implementation to use the kernel Netlink
"dump" interface to allow flow stats that don't fit in the socket buffer
to be dumped gradually as the caller drains the socket buffer.
One of the changes here is a bug fix for nla_unreserve. Because Netlink
attributes' lengths are rounded up to a multiple of 4 bytes, reducing
the length of the payload by N bytes doesn't necessarily reduce the
length of the skb by N bytes. Instead, we need to know the original
length and final length of the attribute. This means that using 'len'
as a difference in bytes doesn't really make sense, so this changes
'len' to be the new length of the attribute payload and renames the
function to nla_shrink to (IMO) better reflect what it is now doing.
Since we have to release the RCU read lock between calls to the dump
function, we need table iterators that persist across RCU epochs. One
way to do this would be to add new "iterator_save" and "iterator_restore"
functions, but this seemed like overkill since there would then be a
total of 5 iterator functions that have only one user (flow stats dumping).
Instead, this patch refactors table iteration into a single "iterate"
function that takes a callback. This simplifies the table iteration code
significantly.
This change also modifies dpctl to understand the new format of flow
stats.
Tolerate ENOBUFS from kernel netlink code in second call to recvmsg.
The Linux kernel returns ENOBUFS to userspace code to indicate that
a netlink message that the kernel wanted to send could not be due to
a lack of buffer space. We checked for this and dealt with it in
one recvmsg call but not in another. This change tolerates it in
both places.
Should fix a problem encountered on the OF6k under "hping3 --flood".
Ben Pfaff [Wed, 23 Apr 2008 21:25:27 +0000 (14:25 -0700)]
Support NETLINK_ADD_MEMBERSHIP even when compiled with old kernel headers.
Commit 0a9024a2d7, "Changes based on feedback from GigaFin," added support
for compiling on a system that lacked up-to-date Linux kernel headers, by
dropping support for a feature only in newer kernels on such a system.
But the code already supported such a system properly; only the lack of
a macro definition kept it from compiling. So define the macro ourselves
instead of disabling the feature.
To elaborate, there are two issues here.
The first is kernel support for NETLINK_ADD_MEMBERSHIP. Older
kernels don't support this. But NETLINK_ADD_MEMBERSHIP was added
at the same time as support for more than 32 netlink multicast
groups. Netlink multicast groups numbered less than 32 can be
subscribed using a different mechanism, and in fact we do that
for such groups. Thus, attempting to use NETLINK_ADD_MEMBERSHIP
on an older kernel basically can't happen, and if it did it would
be a bug in the OpenFlow kernel module for telling userspace to
subscribe to a multicast group that cannot exist on that kernel.
The second is the kernel headers that the userspace programs are
compiled with. Without the change quoted below, the userspace
programs will silently fail if they are compiled with 2.4.x
headers and then run on a 2.6.x kernel and given a multicast
group numbered above 32, because they will not even try to
subscribe to that group. With the change, the userspace programs
will work properly when run on 2.4.x or 2.6.x kernels regardless
of the multicast group number and regardless of whether the
headers they were compiled against were from 2.4.x or 2.6.x.
Ben Pfaff [Fri, 4 Apr 2008 23:17:46 +0000 (16:17 -0700)]
Implement OpenFlow statistics in switches and in dpctl.
This has two notable omissions. First, only at most 4k of flow statistics
are reported. Second, aggregate statistics are not yet supported. Both
of these are fairly easily fixable, just not fixed yet.
Ben Pfaff [Thu, 3 Apr 2008 18:40:16 +0000 (11:40 -0700)]
Break data_hello and control_hello messages into multiple messages.
We want to allow multiple OpenFlow connections to a switch, so that
dpctl and other management tools can also connect to a switch that
is connecte to the controller. These tools will want to fetch the
information in the data_hello message, but currently this can only
be done by sending a control_hello message, which changes the
switch configuration state. So breaking the hello messages up into
multiple different messages, only some of which affect configuration
state, cures the problem.
Also, clean up the kernel datapath a bit, by adding helper functions
for allocating and resizing Generic Netlink messages that encapsulate
OpenFlow messages.
Ben Pfaff [Fri, 4 Apr 2008 23:20:28 +0000 (16:20 -0700)]
Allow secchan to listen for management connections also.
This is needed in the userspace switch code also. That
change is implemented in later change "Send replies to
OpenFlow requests only to the sender". Really that part
of the change should be moved into this one, but intervening
changes are intertwined enough with the same code to
make that difficult.
Ben Pfaff [Fri, 28 Mar 2008 21:12:15 +0000 (14:12 -0700)]
Get rid of mac.h, migrating its users to packets.h.
As a consequence, this fixes tests for multicast and local Ethernet
addresses, which formerly tested for the high bits but should
have tested for the low bits of the first octet.
Justin Pettit [Sun, 30 Mar 2008 08:41:42 +0000 (01:41 -0700)]
- Fixed endian issue with action type. Thanks, Ben!
- Removed BUG() call that could be maliciously/accidentally triggered by a faulty controller.
- Switched from __constant_htons to htons, which seems preferred.
Ben Pfaff [Fri, 28 Mar 2008 00:34:05 +0000 (17:34 -0700)]
Set l4 pointer properly in buffers that contain IP options.
The important part of this change is handling of IP options. The
rest is intended to ensure that packet->l3 and packet->l4 never
point outside the valid range of a packet in the presence of a
minimum-length Ethernet frame. Other code (particularly in forward.c)
assumes that if the IP protocol is TCP or UDP, then packet->l4 points
to a tcp_header or udp_header. Whether this is a good idea is
debatable--it probably warrants clean-up, at least--but that what's
there now.
Ben Pfaff [Thu, 27 Mar 2008 22:16:19 +0000 (15:16 -0700)]
Add the ability to connect to a vconn asynchronously.
Until now, vconn_connect() has always completed the connection
synchronously, blocking as necessary. In the userspace
switch, we want to be able to continue forwarding packets even
if the connection to the controller drops. Thus, this change set
that makes that possible.
The approach taken is perhaps more ambitious than needed, as it
actually adds a new high-level mechanism for polling on arbitrary
file descriptors. This necessitates quite a bit of change to
each of the userspace programs that use vconns, but it also has
the effect of simplifying them. The new structure of these programs
is a lot less fragile than the old one (which tended to end up
livelocking or hanging when something wasn't quite right), so it
seems like the changes are worth it.
projects / openvswitch / log