On Linux 2.6 only, we would send frames received on OpenFlow devices to
the host TCP/IP stack if the dest Ethernet address matched the device's
Ethernet address (or under some other conditions). This caused confusion
when the host stack was configured for forwarding, because received IP
packets would be forwarded to other interfaces with IP addresses even if
the device on which they were received was configured with no IP address.
(This was discovered by Murphy. Thanks Murphy!)
This change drops this "feature". A more or less equivalent feature that
may be less prone to confusion and which works on any kernel version will
be added in a future commit via the of%d device.
This also allowed factoring out common code from the bridge hooks.
When the controller wishes to add a flow entry to the switch, it sends a request with a description of matching flows in a "match". The switch was adding these exactly as described into the flow tables. However, when packets arrived, all invalid fields (e.g., the TCP/UDP ports in a non-TCP/UDP packet) are set to zero. If the rule being added by the controller had not set these invalid fields to zero, the packets would never match. We now zero out these fields when extracing a "match".
-- Reported by Brandon H. --
Otherwise, we give the table implementations the opportunity to screw up
by distinguishing between two flows without wildcards but with apparently
different priorities that are required to be treated as the same priority.
NLMSG_DEFAULT_SIZE was introduced in 2.6.20-rc1 in commit 339bf98ff,
"[NETLINK]: Do precise netlink message allocations where possible", but
we only defined it for kernels 2.6.18 and earlier. This change fixes
up the discrepancy.
There is no need for a "safe" version of the RCU list iteration macros,
because the ordinary version of the macros is "safe". Linux 2.6.26-rc1
removed these "safe" versions, so this change modifies OpenFlow's uses
of these macros to use the ordinary versions. This also simplifies the
code because we don't need to use the list_entry macros any longer.
We need object files to depend on the build directories, so that the
build directories get created before we try to output files into them.
But making them depend directly on the build directories means that
whenever a build directory changes (including any change to the set of
files inside it) all the source files get recompiled. So, instead, make
them depend on dummy files inside the directories, which will only get
modified if someone does it intentionally.
Now the number of actions is limited, as a practical matter, to the size
of the buffer that the kernel provides for Netlink dumps, which is usually
4096 bytes. A flow statistics entry must fit in a single one of these
buffers. Actions are 8 bytes each, so this is about 500 actions
considering overhead.
The new build system supports using a build tree separate from the
source tree, which makes it easier to build for multiple targets
using a single source tree. It also factors out the means for specifying
the modules to be built and the source files that comprise those
modules into simple files named Modules.mk. This is intended to
make it easier to add support for separate hardware table modules in
an upcoming commit.
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.
Ben Pfaff
root
Justin Pettit
Ben Pfaff
Natasha Gude