Problem interpreting signals emitted by terminal on BSD systems properly

[AndrewFarley]

NOTE: This issue's contents and title have been significantly changed as of Jan 15 because of some folks getting to the underlying issue.

The underlying issue is that when using a BSD-based system (like OS-X) this plugin sends some undesirable control signals. This appears to ONLY affect Macs, not Windows or Linux.

@pedros007 Believes this issue is related to the ControlSignals handled by a ShellSession

Reproduce steps which seems to only happen on a MacOS terminal:

• Start a shell session aws ssm start-session --target i-00112233445566778
• Once connected, press Ctrl-y
• Expected behavior: paste from the kill-ring (which is empty, so nothing should happen)
• Actual behavior: The session terminates with the log line

Cannot perform start session: read /dev/stdin: resource temporarily unavailable

The kludgy workaround currently is

• Run stty dsusp undef on your host.
• Start a shell session aws ssm start-session --target i-00112233445566778
• Press Ctrl-y to paste from the kill ring (which is empty, so nothing happens). This is what every other OS does, and what Linux normally does when you SSH.

For more details about this issue please see some of the comments towards the bottom of the thread/issue.

[justinmk3]

4. Only in JOE though, not in VI, or Nano.

Similer to #28 (comment) , I think the problem is likely that Joe is having trouble with the $TERM and/or terminfo/termcap found on the remote machine and it's just not handling it as gracefully as Vi / Nano.

You can check this by trying the following:

1. note the value of $TERM in the remote session
2. on your local machine or any other machine with an equivalent joe build, invoke joe with that TERM value, e.g.: TERM=xterm joe and see if it behaves similarly.

[AndrewFarley]

@justinmk3 Confirmed this formatting issue happens with TERM=xterm via SSM, but it does not happen locally, nor via SSH. I set my local TERM to xterm, and set it after SSM-ed into the session. The formatting works perfectly via SSH and locally, but when connected via SSM it fails to format things, often eliminating prefix whitespace, occasionally deleting return characters as well so some text appears on the same line somehow.

Confirm this happens on both on Terminal and iTerm on OS-X. And this has happened as far as back as I can remember using SSM. Only gotten around to reporting it now because of the increasing usage and adoption.

Another issue I've found @justinmk3 is that copy/paste doesn't work via joe. When I use other editors it works fine, but using joe it doesn't send return characters. I do suspect you're spot on on that its related to term or termcap, but I don't know what other adjustments I can make to fix this. And, like I said, this exact setup works perfectly via SSH. Any other ideas to try?

[AndrewFarley]

Example:

# VIA SSH...
$ TERM=xterm joe test # Then CTRL-K + V to pagedown to the bottom and I see...
                                                <img src="./externals/Landingpage/images/googleplus.png" alt="" />
                                        </a>
                                </li>
                                <li>
                                        <a target="_blank" href="http://website.com">
                                                <img src="./externals/Landingpage/images/twitter.png" alt="" />
                                        </a>
                                </li>
                                <li>
                                        <a target="_blank" href="http://website.com">
                                                <img src="./externals/Landingpage/images/facebook.png" alt="" />
                                        </a>
                                </li>
                        </ul>
                </div>
        </div>
</div>

# VIA SSM...
$ TERM=xterm joe test # Then CTRL-K + V to pagedown to the bottom and I see...
                                                <img src="./externals/Landingpage/images/googleplus.png" alt="" />
                                        </a>
                                </li>
                                <li>
                                        <a target="_blank" href="http://website.com">
                                                <img src="./externals/Landingpage/images/twitter.png" alt="" />
                                        </a>
                                </li>
<li>
    <a target="_blank" href="http://website.com">
<img src="./externals/Landingpage/images/facebook.png" alt="" />
</a>
</li>
</ul>
</div>
</div>
</div>

Note the bottom-half of the lines above are displayed without their space prefixes.

[justinmk3]

I'm not intimately familiar with SSM, but I think the next step would be to share more concrete details about your SSH config and how exactly you are invoking ssh, aws, and/or session-manager-plugin.

copy/paste doesn't work ... it doesn't send return characters
...
Note the bottom-half of the lines above are displayed without their space prefixes.

Something in the stack (SSM, SSH, your terminal, ...) is eating control codes (cursor-positioning in this case). I would expect SSM to strictly pass-through the data so the terminal can handle it.

Depends on how SSH+SSM is invoked/configured...

[AndrewFarley]

@justinmk3 I don't really know how/what to look into to give more information on this. My setup is simply...
OS-X Big Sur 11.6
iTerm or Terminal
Install AWS CLI 2.0

# Then run
AWS_DEFAULT_PROFILE=myprofile aws ssm start-session --target i-0123456789

Then open up joe and can replicate these bugs, but only via SSM, not via SSH. I will try this on a freshly installed computer with none of my bash_profile or any other modifications to check if this is something that is specific to my computer's setup. I'll start there, but if you have any other ideas of how I can provide further feedback would be great.

My belief is still that, for the most part, the only issues I have are with using joe via ssm but no other editor has any issues like described above. This leads me to believe there's some minor incompatibility between this editor and SSM, and I do not know which one to blame or to try to fix. I'd be happy to try to dig in and fix this issue, once I identify which component it is.

[justinmk3]

Thanks. I can reproduce the rendering issues (but not the exact CTRL-D behavior) by connecting that way. Forcing TERM=ansi at least fixes the rendering issues:

TERM=ansi joe

• For whatever reason, ssm sends TERM=xterm-255color to the remote session. But I don't think that's the whole problem.
• (Are you using screen or tmux, btw?)
• Maybe the termios properties are not setup correctly (by SSM, e.g. here)?

In vim, I notice that CTRL-D, CTRL-W, and other control chars do not seem to reach vim in their "raw" form. So I don't think this is an editor-specific issue. Just happens that the symptoms are different amongst editors.

Conclusion

IIUC, ssm when invoked via aws ssm start-session --target ... does not use ssh. And it looks like InitDisplayMode is supposed to setup the terminal, but the linux impl does nothing.

For interactive sessions, that function may need to set some termios properties.

Related

• SCP "connection corrupted" on large file transfer amazon-ssm-agent#274
• SSM and xterm issues with zsh amazon-ssm-agent#344
• Session Manager Lags and Slows Down when Copying/Pasting amazon-ssm-agent#149
• Corrupted terminal commands and performance issues when pasting from clipboard amazon-ssm-agent#166

[AndrewFarley]

@justinmk3 For replicating this issue I was not using screen or tmux, no. I do use those tools especially screen fairly regularly however, but haven't noticed much issues with that. Only/mostly with joe.

I do confirm the TERM=ansi does fix the rendering issues, so at least this makes SSM a bit more usable for me. That's much appreciated.

If I understand your conclusion and links, it seems like some setup in the init display mode for Unix-based sessions might need a few commands to help ensure all the control and cursor-related messages get sent/received in both directions. I have no knowledge at this specific area, but I could poke around and fiddle. And worst-case, maybe I/you/others could help refine what we've learned here into a very clear issue report for someone else to work out.

[pedros007]

TL; DR: Run stty dsusp undef prior to starting an SSM session to allow Ctrl-y to properly yank from the kill ring when connecting to a Linux instance from my Mac laptop.

More details:

A coworker and I ran into an issue where we'd use SSM to ssh to an instance:

aws ssm start-session --document-name AWS-StartInteractiveCommand --parameters '{"command": ["sudo su ec2-user"]}' --target i-00112233445566778

but when we would yank from the kill ring (via Ctrl-Y), a few seconds later the session ends and the following is written to the terminal:

Cannot perform start session: read /dev/stdin: resource temporarily unavailable

I have only experienced this connecting from a Mac laptop to a Linux EC2 instance. Can't reproduce it when connecting from a Linux machine to a Linux EC2 instance. We noticed that on Mac stty -a shows that ^Y is mapped to dsusp. According to GNU docs on signal characters:

The DSUSP (suspend) character ... sends a SIGTSTP signal, like the SUSP character, but not right away—only when the program tries to read it as input

We found that SSM sessions will not terminate if you first unmap the dsusp signal. Here's how to do that:

stty dsusp undef

Is there some way that the session manager plugin can be updated to not fire the dsusp signal when running on a Mac and a user enters Ctrl-y?

[pedros007]

On a BSD system like my Mac, perhaps the bytes emitted by SIGTSTP are different from the bytes sent by Linux, so the signal is never captured by the session manager plugin?

session-manager-plugin/src/sessionmanagerplugin/session/sessionutil/control_signals_unix.go

Lines 25 to 33 in c523002

	// All the signals to handles interrupt
	// SIGINT captures Ctrl+C
	// SIGQUIT captures Ctrl+\
	// SIGTSTP captures Ctrl+Z
	var SignalsByteMap = map[os.Signal]byte{
	syscall.SIGINT: '\003',
	syscall.SIGQUIT: '\x1c',
	syscall.SIGTSTP: '\032',
	}

[AndrewFarley]

@pedros007 Man, this is exactly what I've been suffering from for ages. I'm glad someone dug harder than I did to figure this out. This alone is preventing me from completely eliminating my usage of SSH. My cli editor regularly uses CTRL-Y during editing and it disconnects me. I have to try really hard not to use a command sequence which is embedded deep in my unix-ey brain from 20+ years of using it. Every time it happens I curse this silly feature and its silly bugs.

I don't know enough about the internal unix control signals to recommend what the actual fix is, but I hope someone does and can make the fix to this plugin permanently so I can enjoy uninterrupted shell sessions. I'm pleased to here there's a command I can run stty dsusp undef to prevent this issue. Don't know how I'll remember to run it every time since SSM-ing doesn't really run as-user bash/profile stuff. But at least after I forget to do it once, this command will prevent it from me doing it twice in a row.

[trharris78]

I put this command in my .zshrc on my mac so that the terminal where I run the commands is set up correctly, and that seems to do the trick. If you have any other problematic Ctrl+chars, see if they are mapped to something in the stty -a output. I think @pedros007 also had trouble with Ctrl+S in emacs and was able to fix it by unmapping that particular control character. Something like stty stop undef.

[pedros007]

@AndrewFarley I wonder if this issue should be renamed. The current title implies it's a problem specific to your editor.
However, it seems more like a problem interpreting signals emitted by the terminal on BSD systems. Perhaps the problem is related to the ControlSignals handled by a ShellSession? So maybe rename to something like Shell Session signal handling on BSD systems like MacOS?

Reproduce steps which seems to only happen on a MacOS terminal:

1. Start a shell session aws ssm start-session --target i-00112233445566778
2. Once connected, press Ctrl-y

Expected behavior: paste from the kill-ring (which is empty, so nothing should happen)

Actual behavior: The session terminates with the log line

Cannot perform start session: read /dev/stdin: resource temporarily unavailable

The kludgy workaround fix is to:

1. Run stty dsusp undef on your host.
2. Start a shell session aws ssm start-session --target i-00112233445566778
3. Press Ctrl-y to paste from the kill ring (which is empty, so nothing happens).