Hi,
I've got what I think is a reasonably stable version of dropbear's SSH
server working
in a non-posix, thread-based embedded system. I'd like to review the
changes I
made for anyone that may be considering doing this, but also to see if
any flags
get raised by the folks that have been using this for a while.
Since I don't really know if there's much interest in this within the
group, I'll skip
a lot of detail for this pass and just note the main points. Then if it
turns out that
this is of interest we can dig in more (and I'll probably learn
something)...
GOAL:
- A single ssh login at any given time, to connect to an embedded
command line interface
(no shell, no processes, no posix, etc..).
ASSUMPTIONS:
- Running a small kernel that supports threads;
- Using GCC for powerpc.
- I assume very little about any OS support for anything other than libc
(in my case I do
have an embedded FS, but that's easily bypassed). In other words: no
shell process,
no fork(), no exec(), no pipes, etc.. This includes no use of
fprintf(...) as well.
- I do have a TCP/IP stack; using a non-standard select (doesn't support
multiple FDs)
plus recv/send instead of read/write.
- I loosely base this on the "no-inetd" option; and I heavily chopped
away at things in
options.h (hopefully without breaking anything).
- Since there is no shell, this simply hooks to an internal command line
processor.
- Currently the server is built to run as if the following command line
were invoked:
./dropbear -s -F -b "yada yada" -r dropbear_rsa_host_key
and since I do have an FS, I created the dropbear_rsa_host_key file
using dropbearkey
on my host machine, and simply copied it to my embedded system's FS
for now. The need
for the FS could easily be eliminated.
DETAILS:
My build puts the two math directories into a library, and then builds
the server
using portions of ~25 of the ~65 .c files that are in the main dropbear
directory.
As a session starts up, I call loadhostkeys() and a hacked version of
seedrandom()
prior to svr_session(). The session_loop() function just blocks waiting
for incoming
packets from the network. User authentication is handled with a local
function
that simply verifies that the incoming login and password match; and the
majority of
the code that forks off the shell is just bypassed.
I simulate interaction with a shell by intercepting incoming characters
in common_recv_msg_channel_data(). Each line of text is simply passed
to a command
line parser. While that command line is being processed, all output
from that
embedded command is sent through the function ssh_putchar():
static void
ssh_putcharc(struct Channel *channel,char c)
{
CHECKCLEARTOWRITE();
buf_putbyte(ses.writepayload, SSH_MSG_CHANNEL_DATA);
buf_putint(ses.writepayload, channel->remotechan);
buf_putint(ses.writepayload, 1);
buf_putbyte(ses.writepayload, c);
encrypt_packet();
}
and one other important thing...
At the bottom of encrypt_packet(), I call write_packet() so that the data
is immediately pushed out the socket.
SUMMARY:
Thats about it in the nutshell. The two big gotchas with this were
issues that
would not necessarily be important in a process-based environment:
1. The use of dropbear_exit() for errors requires the use of
setjmp/longjmp because
its in a thread that needs to cleanup properly.
2. The heap is clean when exits are clean; but things get messy in a lot of
exception cases; hence, the need for a dropbear-specific heap which
allows me to force a clean heap when the session ends (simulating
the cleanup that
is automatically done when the process exits).
I'm guessing that this would be a mess to integrate back into the
distribution;
however, I'm up for it if folks wanna go for it. Perhaps a "cub"
version for embedded
systems would be a peer to the main source directory and that would
allow both the
process and thread based versions to use as much of the same code as
possible; but keep
the "mess" isolated a bit more.
Lemme know what you think,
Ed
