Amazon-ssm-agent: Session manager dropping characters from output

Created on 18 Dec 2018 · 5Comments · Source: aws/amazon-ssm-agent

A simple test case is the best way to illustrate the bug. For the following shell snippet (on Linux, although I assume Windows has a similar problem):

printf '\xc8'; sleep 1; printf '\x80 123'; sleep 1; printf '456\n'

Running on my local computer produces the expected output:

Ȁ 123456

Running inside a shell in SSM session manager:

Ȁ 12456

Note \xc8\x80 is the UTF-8 representation of \u0200 or Ȁ

I've tested this against SSM agents with versions 2.3.274.0 and 2.3.344.0 (latest) and both exhibit the bug.

The culprit appears to be writePump in https://github.com/aws/amazon-ssm-agent/blob/c7f64dd6682154ee2a710d0dccf0034d868c5037/agent/session/plugins/shell/shell.go#L246
In particular:

if it encounters an unfinished UTF-8 sequence, it will delay processing it until the next iteration of the loop when there is more data (this is fine)
BUT it does all of its bookkeeping based on n, which is the number of bytes read from the pty, however this is incorrect if there is any data carried over from the previous iteration (i.e. due to an unfinished UTF-8 sequence). So, if there are x bytes carried over from the previous iteration and n bytes read in the current iteration, then there are x+n bytes in total, but writePump only processes the first n bytes and drops the remaining x.

The example given includes sleeps between the various printfs to force SSM to read \xc8 and \x80 123 in separate iterations of the loop. You can achieve the same thing without sleep by printing enough data to fill the read buffer instead (which appears to be 1024 bytes).

Source