SSH config overwritten with VS Code Coder config

[aaronlehmann]

Both myself and and another user experienced ~/.ssh/config being completely overwritten to include just a START CODER VSCODE block with the corresponding wildcard Host entry. This happened to me using version 1.8.0 of the extension with multiple remote development sessions active.

[johnstcn]

Hey Aaron,

I found that, a missing #--- END CODER VSCODE .* --- comment in combination with more than one #--- START CODER VSCODE .* --- comment could cause any content between the first START and the first END to be removed.

Example:

Host beforeconfig
  HostName before.config.tld
  User before

# --- START CODER VSCODE dev.coder.com ---
Host coder-vscode.dev.coder.com--*
  ConnectTimeout 0
  LogLevel ERROR
  ProxyCommand some-command-here
  StrictHostKeyChecking no
  UserKnownHostsFile /dev/null

Host donotdelete
  HostName dont.delete.me
  User please

# --- START CODER VSCODE dev.coder.com ---
Host coder-vscode.dev.coder.com--*
  ConnectTimeout 0
  LogLevel ERROR
  ProxyCommand some-command-here
  StrictHostKeyChecking no
  UserKnownHostsFile /dev/null
# --- END CODER VSCODE dev.coder.com ---

Host afterconfig
  HostName after.config.tld
  User after 

With coder config-ssh, the above will currently (correctly) cause an error. However, the VSCode extension will silently remove Host donotdelete.

#510 should fix this.

Does this sound plausible as a root cause to your issue?

Sidenote: there's also a potential issue with concurrent file writes. Since you mentioned having multiple remote development sessions it's also possible there was some kind of race condition with writing the updated config. config-ssh does do an atomic write, but currently the VSCode extension doesn't. Filed #511 for this separately.

[aaronlehmann]

My guess is that the issue was non-atomicity. This happened seemingly randomly, so I wouldn't expect a deterministic failure mode like a missing #--- END CODER VSCODE .* --- to be the issue. Also, the other person this happened to tends to have a lot of VS Code windows open at once.

[aaronlehmann]

Somewhat related: we've seen a race condition in the past updating the coder binary: #347

[bjornrobertsson]

Similar experience when testing, although never lost everything. Everything between "# ------------START-CODER-----------" and it's end "# ------------END-CODER------------" got removed.
I initially pegged this on the ProxyCommand (pointed to different coder binary, the homebrew one) and possibly now being out-dated, due to coder-* entry or historical changes in behaviour? But perhaps more perplexing is that the command doesn't warn about this, although it becomes apparent by using -n/--dry-run exactly what will happen.

I.e.

$ /opt/homebrew/bin/coder config-ssh -n
Dry run, the following changes would be made to your SSH configuration:

  * Update the coder section in /Users/$USER/.ssh/config

--- /Users/$USER/.ssh/config
+++ /Users/$USER/.ssh/config
@@ -25,398 +25,12 @@
 # You should not hand-edit this section unless you are removing it, all
 # changes will be lost when running "coder config-ssh".
 #
-Host coder.CertTwo
-	HostName coder.CertTwo
+Host coder.*
 	ConnectTimeout=0
 	StrictHostKeyChecking=no
 	UserKnownHostsFile=/dev/null
 	LogLevel ERROR
-	ProxyCommand /opt/homebrew/bin/coder --global-config "/Users/$USER/Library/Application Support/coderv2" ssh --stdio CertTwo
-Host coder.CertTwo.main

But without -n:

$ /opt/homebrew/bin/coder config-ssh
> The following changes will be made to your SSH configuration:

    * Update the coder section in /Users/$USER/.ssh/config

  Continue? (yes/no) yes

Updated "/Users/$USER/.ssh/config"
You should now be able to ssh into your workspace.

In this case I feel the text 'Update the coder section' a little underwhelming. Updating the 'caveat' warning and potentially if this replacement is known, add to the list of Updates that will happen would work.

[aaronlehmann]

BTW, another one of our users got hit by this today. I have no idea why it only started happening recently - perhaps our upgrade to Coder 2.21 affected timings enough to expose this race. Really hoping the atomicity fix does the trick!

[johnstcn]

This should be fixed in v1.9.1 of the extension.