-------
This patch enhances networking support for the simulation under Linux.
Includes updated support for Linux TUN/TAP, and the addition of support for
Linux bridge devices.
CHANGES
-------
o Check to see if the d_txavail callback is present before calling it in
the arp send code. This prevents a segfault when simulating the telnetd
daemon with arp send enabled.
o Adjust the simulation's netdriver_loop() so it will detect and respond to
ARP requests.
o Do not attempt to take the tap device's hardware address for use by the
simulation. That hardware address belongs to the host end of the link,
not the simulation end. Generate a randomized MAC address instead.
o Do not assign an IP address to the interface on the host side of the TAP
link.
+ Provide two modes: "host route" and "bridge".
+ In host route mode, maintain a host route that points any traffic for the
simulation's IP address to the tap device. In this mode, so long as the
simulation's IP is a free address in the same subnet as the host, no
additional configuration will be required to talk to it from the host.
Note that address changes are handled automatically if they follow the
rule of if-down/set-address/if-up, which everything seems to.
+ In bridge mode, add the tap device to the specified bridge instance. See
configs/sim/NETWORK-LINUX.txt for information and usage examples. This
enables much more flexible configurations (with fewer headaches), such as
running multiple simulations on a single host, all of which can access
the network the host is connected to.
o Refresh configurations in configs/sim where CONFIG_NET=y. They default
to "host route" mode.
o Add configs/sim/NETWORK-LINUX.txt
CAVEATS
-------
- The MAC address generation code is extremely simplistic, and does not
check for potential conflicts on the network. Probably not an issue, but
something to be aware of.
- I was careful to leave it in a state where Cygwin/pcap should still work,
but I don't have a Windows environment to test in. This should be
checked.
- I don't know if this was ever intended to work with OS X. I didn't even
try to test it there.
NOTES
-----
- Was able to get telnetd working and simulate nsh over telnet, but only so
long as listen backlogs were disabled.
There appears to be a bug in the backlog code where sockets are being
returned in SYN_RCVD state instead of waiting until they're ESTABLISHED;
if you perform an immediate send after accepting the connection, it will
confuse the stack and the send will hang; additionally, the connection
will never reach ESTABLISHED state.
Can be worked around by adding a sleep(1) after the accept in telnetd. I
don't have the necessary knowledge of the IP stack to know what the
correct fix is.
If there is no active connection (e.g. it is waiting in accept), then
the connection object, which doesn't yet exist, should not be cleaned
up when the socket is closed.
During a write, if there is no more buffer space for the user data,
return the amount that was written instead of waiting until there
is free space. If nothing has been written yet, then block as before.
This solves a deadlock that occurs if the user data is too large to
fit in the available buffer: the write thread will block before any
data is added to the write queue, leaving no possibility that more
buffers will free up when they are ACKed (since they have not yet been
sent). The write thread will then block forever and hold all of the
buffers.
TCP uses two semaphores for buffers, one for the entire buffer pool,
and one for the number of allowed buffers for readahead. To avoid
double taking a semaphore under heavy load, release the semaphore
after waiting before trying to acquire both again.
It is possible for a socket to have no connection, for example if a TCP
socket is waiting in accept. This checks for this condition in case the
socket is closed from a different thread.
When a socket is closed, it should make sure that any pending write
data is sent before the FIN is sent. It already would wait for all
sent data to be acked, however it would discard any pending write
data that had not been sent at least once.
This change adds a check for pending write data in addition to unacked
data. However, to be able to actually send any new data, the send
callback must be left. The callback should be freed later when the socket
is actually destroyed.
When a poll requesting POLLOUT happens, the poll should return
immediately if a write will not block. This change adds that, as
opposed to the old behaviour of blocking until a timer from the
Ethernet driver eventually triggers the poll to complete.
This is only implemented for buffered TCP. Unbuffered TCP should
behave as before.
Gregory Nutt