Honor openssh out max packet size

[jquast]

OpenSSH client requests a value of 16384 for value of Channel.out_max_packet_size,
when opening a session channel, but method _set_remote_channel “clamps” this value by Transport._sanitize_packet_size, which coerces the value to 32768.

Because paramiko server is not honoring the client's requested value, and there is no protocol negotiation to notify the client as such, when a packet larger than 16384 is received by the OpenSSH client, it will write to stderr: channel 0: rcvd big packet 32704, maxpack 16384.

This is especially difficult for interactive programs that use cursor addressing, as the position of the client cursor then becomes indeterminate. A large-windowed terminal (size ~170x100) using a curses application is very likely to receive this message, which will cause corruption in the output display.

This value is specified in common.py:

# lower bound on the max packet size we'll accept from the remote host
# Minimum packet size is 32768 bytes according to
# http://www.ietf.org/rfc/rfc4254.txt
MIN_PACKET_SIZE = 2 ** 15

If you follow rfc4254 for this declaration,

Implementations are expected to have some limit on the SSH transport
layer packet size (any limit for received packets MUST be 32768 bytes
or larger, as described in [SSH-TRANS]).

Then, following SSH-TRANS, which is http://www.ietf.org/rfc/rfc4253.txt for section 6.1, it reads:

6.1. Maximum Packet Length

All implementations MUST be able to process packets with an
uncompressed payload length of 32768 bytes or less and ...

We read that it is perfectly legal to request a maximum packet size of 16384 as OpenSSH client does (“or less”).

If we modify demos/demo_server.py,

--- a/demos/demo_server.py
+++ b/demos/demo_server.py
@@ -164,6 +164,7 @@ try:
         sys.exit(1)

     chan.send('\r\n\r\nWelcome to my dorky little BBS!\r\n\r\n')
+    chan.send('X'*32705)

We can see the client emit this message: channel 0: rcvd big packet 32704, maxpack 16384. Furthermore, the demo_server.log writes a misleading message, [chan 1] Max packet out: 16384 bytes — that was the value received, but not the value set and later honored.

This patch suggests separating MIN_PACKET_SIZE to MIN_WINDOW_SIZE, and lowering the value of MIN_PACKET_SIZE to 4096.

[bitprophet]

Thanks for the detailed write-up, super appreciated!

I wonder if this was worsened by semi-recent work on cleaning up & exposing some of the window/packet size stuff in #372 - though that doesn't really impact how we go ahead with this, is just idle speculation.

Looks good to me offhand (& the test suite does pass, the 'fail' is the travis worker for the 3.4 interpreter timing out, which is their end not ours), I'd like to test it out with my real world workflows but otherwise I'm slotting this into a release.

[lndbrg]

val should be 32767, it is purposedly set not to match the correct value, to make sure the clamping works. :)

[jquast]

Thank you @lndbrg, I'll push such change shortly.

[jquast]

These were addressed. Travis passed again. Thanks again.

[lndbrg]

Cool! 👍