Default ssh command

Default ssh command

[Aaron Hall]

I was trying to push some binary files (PNGs) over ssh and found some
files inexplicably wouldn't successfully push.

I set, in my .gitconfig file:

[core]
        sshCommand = ssh -e none

And it works now for all files (that I can tell).

From a close examination of the output where I noted escape codes
showing up, I believe the problem to be ssh escape codes.

`\n~.` for example closes the ssh client.

Therefore I think the default ssh command should be `ssh -e none`
instead of `ssh` for the benefit of users who won't recognize the
problem, unless `git` is expected to make use of ssh escape codes.

I'm new to the mailing list, sorry if I missed some protocol here.
Thank you so much for working on `git` and making your work
 available to the rest of the world. Cheers!

Re: Default ssh command

[Junio C Hamano]

Aaron Hall <aaronchallmba@gmail.com> writes:

> I was trying to push some binary files (PNGs) over ssh and found some
> files inexplicably wouldn't successfully push.
>
> I set, in my .gitconfig file:
>
> [core]
>         sshCommand = ssh -e none
>
> And it works now for all files (that I can tell).
>
> From a close examination of the output where I noted escape codes
> showing up, I believe the problem to be ssh escape codes.
>
> `\n~.` for example closes the ssh client.
>
> Therefore I think the default ssh command should be `ssh -e none`
> instead of `ssh` for the benefit of users who won't recognize the
> problem, unless `git` is expected to make use of ssh escape codes.

Curious.  

When we use ssh to run the receive-pack over the wire while pushing,
we do not ask for pty, and because the "-e" option to set the escape
character is meaningful only when ssh runs with pty allocated, you
shouldn't have to do this, unless you force pty allocation by
e.g. passing "-t" yourself.

Re: Default ssh command

[Aaron Hall]

> When we use ssh to run the receive-pack over the wire while pushing,
> we do not ask for pty, and because the "-e" option to set the escape
> character is meaningful only when ssh runs with pty allocated, you
> shouldn't have to do this, unless you force pty allocation by
> e.g. passing "-t" yourself.

Then I wonder why I'm getting a pseudoterminal.

Note that I'm using git from git bash on Windows in a walled off
corporate environment.

The ssh manpage says:

> If no pseudo-terminal has been allocated, the session is transparent
> and can be used to reliably transfer binary data.  On most
> systems, setting the escape character to ``none'' will also
> make the session transparent even if a tty is used.

If we want the ssh to be binary transparent regardless
(and I think we do) then that makes my suggestion,
`ssh -e none`, still perhaps a good one?

On Tue, Jul 6, 2021 at 10:03 PM Aaron Hall <aaronchallmba@gmail.com> wrote:
>
> I'll reply on list.
>
> On Tue, Jul 6, 2021 at 9:58 PM Aaron Hall <aaronchallmba@gmail.com> wrote:
> >
> > My context is using git from git bash on Windows in a walled off environment.
> >
> > On Tue, Jul 6, 2021 at 9:48 PM Junio C Hamano <gitster@pobox.com> wrote:
> > >
> > > Aaron Hall <aaronchallmba@gmail.com> writes:
> > >
> > > > I was trying to push some binary files (PNGs) over ssh and found some
> > > > files inexplicably wouldn't successfully push.
> > > >
> > > > I set, in my .gitconfig file:
> > > >
> > > > [core]
> > > >         sshCommand = ssh -e none
> > > >
> > > > And it works now for all files (that I can tell).
> > > >
> > > > From a close examination of the output where I noted escape codes
> > > > showing up, I believe the problem to be ssh escape codes.
> > > >
> > > > `\n~.` for example closes the ssh client.
> > > >
> > > > Therefore I think the default ssh command should be `ssh -e none`
> > > > instead of `ssh` for the benefit of users who won't recognize the
> > > > problem, unless `git` is expected to make use of ssh escape codes.
> > >
> > > Curious.
> > >
> > > When we use ssh to run the receive-pack over the wire while pushing,
> > > we do not ask for pty, and because the "-e" option to set the escape
> > > character is meaningful only when ssh runs with pty allocated, you
> > > shouldn't have to do this, unless you force pty allocation by
> > > e.g. passing "-t" yourself.

Re: Default ssh command

[brian m. carlson]

[-- Attachment #1: Type: text/plain, Size: 2967 bytes --]

On 2021-07-07 at 02:13:34, Aaron Hall wrote:
> > When we use ssh to run the receive-pack over the wire while pushing,
> > we do not ask for pty, and because the "-e" option to set the escape
> > character is meaningful only when ssh runs with pty allocated, you
> > shouldn't have to do this, unless you force pty allocation by
> > e.g. passing "-t" yourself.
> 
> Then I wonder why I'm getting a pseudoterminal.
> 
> Note that I'm using git from git bash on Windows in a walled off
> corporate environment.

Do you maybe have RequestTTY set in some SSH configuration file, like
~/.ssh/config?  You can also run "ssh -vvv HOSTNAME" to see if there are
any config files.

For example, if I do "ssh -vvv git@github.com"[0], I see these lines:

  debug1: Reading configuration data /home/bmc/.ssh/config
  debug1: /home/bmc/.ssh/config line 44: Applying options for github.com
  debug1: /home/bmc/.ssh/config line 67: Applying options for *
  debug3: kex names ok: [sntrup4591761x25519-sha512@tinyssh.org,curve25519-sha256,curve25519-sha256@libssh.org,ecdh-sha2-nistp521,ecdh-sha2-nistp384,ecdh-sha2-nistp256]
  debug1: Reading configuration data /etc/ssh/ssh_config
  debug1: /etc/ssh/ssh_config line 19: include /etc/ssh/ssh_config.d/*.conf matched no files
  debug1: /etc/ssh/ssh_config line 21: Applying options for *

That tells me I have some configuration for github.com (and for all
hosts) in ~/.ssh/config and some for all hosts in /etc/ssh/ssh_config.
It doesn't tell me what they are, but I can look in the files to see.

Also, you may wish to run "env | grep '^GIT'" to see if anybody has
configured the GIT_SSH or GIT_SSH_COMMAND variables, and check with "git
config -l --show-origin" to see if someone has overridden the SSH
configuration.

> The ssh manpage says:
> 
> > If no pseudo-terminal has been allocated, the session is transparent
> > and can be used to reliably transfer binary data.  On most
> > systems, setting the escape character to ``none'' will also
> > make the session transparent even if a tty is used.
> 
> If we want the ssh to be binary transparent regardless
> (and I think we do) then that makes my suggestion,
> `ssh -e none`, still perhaps a good one?

The problem is that on some systems "ssh" is something other than
OpenSSH and we will likely break those systems.  That doesn't mean we
can't do it if it's necessary, but it does mean that if this is a
configuration issue, it would be better to fix the configuration if we
can rather than potentially break systems needlessly.

This problem will also occur for tools like rsync and various other
programs which expect the standard behavior for the ssh binary, so
making a change here in Git is also not the right place if we can avoid
it.

[0] Chosen simply because I know it works, it has configuration on my
system, and it's publicly accessible.
-- 
brian m. carlson (he/him or they/them)
Toronto, Ontario, CA

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 262 bytes --]

Re: Default ssh command

[Aaron Hall]

> Do you maybe have RequestTTY set in some SSH configuration file, like
> ~/.ssh/config?  You can also run "ssh -vvv HOSTNAME" to see if there are
> any config files.

That's it! Not sure why I put a wildcard RequestTTY force... I was probably
just going through the man page and putting down things that looked like
were what I wanted...

My urgency on this issue is gone. But unless git expects escape codes,
perhaps it still should set them to none.

> The problem is that on some systems "ssh" is something other than
> OpenSSH and we will likely break those systems.  That doesn't mean we
> can't do it if it's necessary, but it does mean that if this is a
> configuration issue, it would be better to fix the configuration if we
> can rather than potentially break systems needlessly.

I would expect it would be a compatible change but I don't have time
to research it.

> This problem will also occur for tools like rsync and various other
> programs which expect the standard behavior for the ssh binary, so
> making a change here in Git is also not the right place if we can avoid
> it.

To me, it's weird that an improperly configured tool breaks another
tool, and it seems that we could be more robust to that sort of
problem without breaking anyone, but as you say we would need to
verify we're not breaking anyone.

I suggest keeping this open as an issue, but that's up to you all.