bugs

xorp源码hg

Known issues:

 - Routes appear in the RTM routing table but do not propagate to
   the FIB.
   This can be verified by comparing the output of 'route print' 
   with that of 'netsh routing ip show rtmroutes'.

 * Feedback from Microsoft is needed to fix this issue.

 - We use WriteFile() to broadcast data to multiple clients.
   If one of the clients does not read the data, then the server thread
   will hang as a result.
   The alternative is to use overlapped I/O, but this means
   recording lots of state for each write.

 * Therefore, clients MUST read from the pipe after issuing commands.
   This should not affect XORP, as we do not rely on the SO_USELOOPBACK
   socket option when dealing with routing sockets, and the socket is
   polled for input as part of the FEA's event loop.

 - The use of the BSD routing socket message format is not completely
   identical.

 * We only accept a DST, GATEWAY and NETMASK for RTM_ADD.
 * We only accept a DST and a NETMASK for RTM_DELETE.
 * Any other combinations are not acceptable.

 * RTM_IFANNOUNCE messages DO NOT contain the interface name for deletions.
   Microsoft's Router Manager does not provide this, only the index.

 * The InterfaceStatus() callback does not get called when link sense
   changes on the media, therefore, RTM_IFINFO messages are not sent
   when carrier is dropped or regained.

 * RTM_NEWADDR messages are sent whenever the interface address list
   changes. This is incomplete and the semantics are not the same as
   that of the BSD routing socket; it would be necessary to record all
   the state to just send the deltas, as InterfaceStatus() has the
   entire address list for each interface passed to it on a change.

 - The regression tests read the first 'reply' which is actually
   the echo of the command they sent. The next 'reply' is the kernel's.

 * PF_ROUTE is a broadcast domain in BSD, with specific semantics.
   The DLL does not always broadcast the client's inbound command message first.
   It should do so, and do it separately, following by broadcasting
   the and separately; followed by the kernel's reply. This maintains POLA.

   On BSD, the option SO_USELOOPBACK can be used to squelch the inbound
   command message from being echoed to the process which is sending it,
   thus maintaining the assumption of 'write once, read twice'.

   Given that the DLL can block if client's don't read, this is particularly
   important!

 - Routing and Remote Access will signal that all interfaces are going
   away when it shuts down. Client processes will see these notifications.

 - Currently we squelch notifications about all-1s broadcast and multicast
   destinations. This is because they show up as equal-cost multihop
   rules in the table, although this doesn't tell the whole story.

 * Separate DLLs for IPv4 and IPv6 shim support are needed because of
   how DLLs are loaded. The image is loaded once into the MS Router Manager.
   There is no copy-on-write for data sections because everything runs in
   a single NT process. A common global instance variable is needed. We
   can't use thread-local storage because we may be reentered by multiple
   pool threads, and we can't tell them apart usefully at initialization.

 * XXX: The following identifiers seem to be missing from the Windows SDK
   build 5284 headers:
	IPV6_ADDRESS_LEN_IN_BYTES
	ConvertAddressAndLengthToNetAddress()
	ConvertNetAddressToAddressAndLength()
   referenced by:
	RTM_IPV6_SET_ADDR_AND_LEN()
	RTM_IPV6_GET_ADDR_AND_LEN()
   There also seems to be no RTM_IPV6_MASK_FROM_LEN() function, which
   we'd need for converting our IPv6 prefix lengths to socket address
   structures for the modified BSD message format we currently use.