Match error in ssh_sftp:write_file when connection gets closed concurrently

[pguillory]

Describe the bug
A match error is sometimes raised in ssh_sftp:write_file when the SSH connection or channel is closed by a separate process.

To Reproduce
Sorry about the Elixir repro code.

{:ok, channel} = :ssh_sftp.start_channel(ssh)

spawn(fn ->
  Process.sleep(50)
  :ssh_sftp.stop_channel(channel)
end)

spawn(fn ->
  :ssh_sftp.write_file(channel, '/tmp/foo', "bar")
end)

The timing matters.

• With a shorter sleep, write_file returns {error, closed}.
• With a longer sleep, it returns ok.
• With just the right sleep, it raises {badmatch, {error, closed}} on line 1079 of ssh_sftp.erl.

Expected behavior
I would expect it to return {error, closed} if the connection/channel becomes closed before the operation completes.

Affected versions
27

Additional context
I suspect that ssh_sftp:read_file has a similar problem. It looks like the actual match error is in the do_handle_call for send_window, and recv_window is almost identical.

[u3s]

yes. this problem exists for ssh_sftp module. I'm afraid it affects more functions than listed in description.

which exact OTP version you were using when receiving badmatch on line 1079.

[benjreinhart]

@pguillory and I work together and I spotted this error on my machine, which is running OTP 27.2.4.

[u3s]

what's the use case for closing channel in parallel to processing a file operation over same channel?

[benjreinhart]

@u3s @pguillory is more familiar with this code than I, but my 2c: we are using SSH to control a fleet of "sandbox" servers that run remote dev environments. Users (and AI) can make changes to the code in our Phoenix app / web frontend.

We have stateful erlang processes responsible for keeping the SSH connection open to the sandbox server. There are a few places in the app that can result in code updates, some of which eventually call write_file to push the updated file to the sandbox server. Separate parts of the app can terminate the connection to the sandbox for a variety of reasons. These can happen concurrently.

[u3s]

fixed in related PRs.
planned to be included in upcoming emergency patches.