The previous model was flawed. Consider the following case:
* The main loop looks as follows (pseudocode):
loop {
let _ = (tcp:1234).read_all()
wait(iface.poll())
}
* The remote end is continuously transmitting data and at some
point fills the window of (tcp:1234), stopping the transmission
afterwards.
* The local end processes the packets and, as a part of egress
routine, emits an ACK. That also updates the window, and
the socket's poll_at() routine returns None, since there is
nothing to transmit or retransmit.
* The local end now waits indefinitely even though it can start
processing the data in the socket buffers right now.
The rate of emission of neighbor discovery packets is already
limited at the level of the entire neighbor cache, but poll()
would uselessly spin until the answer arrives (if ever).
- Do not send ICMPv4 responses for packets with a broadcast destination
address.
- Do not send DstUnreachable with ProtoUnreachable on receipt of a
packet with an unknown protocol with a non-unicast destination
address.
- Do not send DstUnreachable with PortUnreachable on receipt of a
UDP packet when no sockets are listening on the destination port
and the destination address is a non-unicast address.
- Send the correct amount of the original datagram when sending Destination
Unreachable error responses.
- Do not assume that a ip datagram has a payload when sending a proto
unreachable ICMPv4 error response.
- Add tests to iface tests.
- Ensure ICMP error responses are correctly formed when the
datagram has no payload.
- Ensure ICMP error responses are correctly handled for UDP packets
when no socket is listening on the destination port.
- Ensure the correct amount of the original payload is returned in
Destination Unreachable responses.
- Add tests for the following
- ICMP error responses are not sent in response to broadcast requests
- ARP requests are responded to and inserted into the cache
- ARP requests for someone else are not responded to, but the sender
is still inserted in the cache
- Add the ttl member to the IpRepr
- Add the ttl member along with setters and getters to the tcp and udp
socket types
- Add unit tests for the new set_ttl parameter
- Update usage of IpRepr to include the ttl value
To be precise, I'm talking about IPX, AppleTalk and DECnet here,
not things like PPPoE, ATAoE, FCoE, or PTP, which make sense
to implement on top of EthernetInterface but do not work on
the same level on top of it as IP.
This reverts commit 51b2f18d1165bf7257de8894df101299cc93b094.
There's no throughput difference so far as I could measure, but
this greatly increases worst-case latency. At some later point
we could perhaps pass a deadline to the poll function, but for now
reverting this is simple enough.
Typically, the poll function is used as a part of a larger RTOS.
If we only dispatch one packet per socket per poll() call,
then we have to wait for a complete scheduler roundtrip,
which is potentially a lot of time, and as a result we are
not filling the peer's window efficiently.
Dan Robertson
whitequark
Jeremy Soller
Egor Karavaev
Philipp Oppermann
Dan Robertson
Steffen Butzer