geoffwilliams@home:~$

Anatomy of an Augeas Bug

Sometimes you get one of those bugs that requires a completely different mindset to the ones you normally get to fix:

  • No stack trace
  • No clear error message
  • Tons of Regular Expressions
  • An unfamiliar language
  • … No problem!

What is Augeas anyway?

Augeas is a neat system for parsing and updating files. This makes it a really handy tool when you need to selectivly edit files you don’t own.

Augeas is context-aware, so it understands things like the level of nesting in XML files and the propper grammar for files it’s editing. It does this through an extra bit of code called a lens.

Lenses are the bit of magic glue that makes the whole system work and 99% of the time you don’t need to know or care about anything to do with them… Unless you stumble on an unsupported file-type or a bug as in this case.

Discovery and fault isolation

The customer was getting an error from Augeas whenever he had a line like:

%sportshorseracingopssudoers ALL=(ALL) ALL, !DISALLOWED

in /etc/sudoers. Basically Augeas refused to process the file once this line was present and works again if the ! character is removed from the command alias. Problem isolated.

A bit more information was visible by getting Augeas to scan all files it knows about:

[root@master ~]# augtool print /augeas//error
/augeas/files/etc/sudoers/error = "parse_failed"
/augeas/files/etc/sudoers/error/pos = "3796"
/augeas/files/etc/sudoers/error/line = "117"
/augeas/files/etc/sudoers/error/char = "0"
/augeas/files/etc/sudoers/error/lens = "/opt/puppet/share/augeas/lenses/dist/sudoers.aug:529.10-.70:"
/augeas/files/etc/sudoers/error/message = "Iterated lens matched less than it should"

So this is all I had to go on.

Hazy memories from Back In The Day

As soon as I saw “matches less then it should”, I was immediately transported back to my carefree days at Uni which looked a bit like this:

Happy Dayss...

…And immediately thought how that sounded exactly like a regular expression that wasn’t consuming enough characters.

Finding the lens file

I’m a great believer in never bothering to remember paths to files since its too much like hard work. Instead I always use find, which is lightning fast on small VM filesystems:

[root@master vagrant]# find / -iname '*lens*'
/usr/share/augeas/lenses
/opt/puppet/share/augeas/lenses

Since this is a Puppet Enterprise system, I knew to look in /opt/puppet/share/augeas/lenses and from here it only took a couple of seconds to find sudoers lens at /opt/puppet/share/augeas/lenses/dist/sudoers.aug.

This is the same file referenced in the customer’s error message - I had been expecting to find a bunch of Ruby files that needed to be fixed but there weren’t any.

Inside the Lens file

Once I opened the lens file and started looking around even more memories came flooding back. This time about Context-Free Grammars.
For the uninitiated, context-free grammars define the rules by which a string or data structure is parsed and converted into a parse-tree.

They do so by providing a series of production rules that desribe how the entire parsing process in exacting detail. Basically, this is how you write a language compiler.

Inside the lens file, I saw this code::

let lns = ( empty | comment | includedir | alias | defaults | spec )*

Which is clearly a the top level entry point into the CFG.

What this line is saying is that the rule lns can be rewritten as the union of the following rules:

  • empty
  • comment
  • includedir
  • alias
  • defaults
  • spec

Lets have a look at the simplest one of these rules: empty

(* View: empty
Map empty lines *)
let empty   = [ del /[ \t]*#?[ \t]*\n/ "\n" ]

So far, so good. This looks like it matches and strips a bunch of whitespace characters. Note the newline \n at end of the line. I found out later that this is vital! Miss it off and nothing will work as the line isn’t treated as finished.

After having a look at some of the other rules, I eventually found the one I was interested in: spec which looked like this:

let spec = [ label "spec" . indent
               . alias_list "user" sto_to_com_user . sep_cont
               . Build.opt_list spec_list sep_col
               . comment_or_eol ]

Following the production rules, I eventually ended up at this fragment of code:

(* Variable: sto_to_com_cmnd
sto_to_com_cmnd does not begin or end with a space *)
let sto_to_com_cmnd =
      let alias = Rx.word - /(NO)?(PASSWD|EXEC|SETENV)/
   in let non_alias = /(!?[\/a-z]([^,:#()\n\\]|\\\\[=:,\\])*[^,=:#() \t\n\\])|[^,=:#() \t\n\\]/
   in store (alias | non_alias)

Which splits up the parse depending on whether we have encountered an alias or a non alias. It’s hard to see but words containing a capital can only be matched as aliases because the non_alias rule only matches lower-case letters.

Looking closely, we can see that the non aliases are allowed to start with an optional ! character but the aliases are not (we’re missing the /!?/ bit).

Now what?

It was time for a bit of research to see how to fix this properly in order to get the best chance of having a fix accepted upstream. I found this page which was really helpful and definitely worth a read. Basically Augeas uses test driven development to your expected to write a test and then fix a bug or introduce a feature.

This is the only sane way of producing a piece of software that relies on so many regular expressions.

Writing the testcase

The site above goes into lots of detail about testcases in Augeas. For this bug, all I needed to do was edit the file at /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug and add a new testcase to the bottom of the file. Running the tests was as simple as typing:

augparse /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug

Once I’d written a testcase which seemed to be doing something by copying the other examples, I went back in the lens file and started making changes. The obvious place was the sto_to_com_cmnd rule and I was able to prepend an optional ! character to the start of the let alias rule.

The only problem was this didn’t work!

I got a ton of errors like this:

[root@master ~]# augparse /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug
Syntax error in lens definition
/opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug:329:8: Unexpected character \
/opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug:329.8-.33:syntax error, unexpected UIDENT, expecting '}'
/opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug:syntax error
[root@master ~]# vim /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug
[root@master ~]# augparse /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug
Syntax error in lens definition

and then this:

[root@master ~]# augparse /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug
/opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug:321.0-330.3:exception thrown in test
/opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug:321.5-.62:exception: Input string does not match at all
    Lens: /opt/puppet/share/augeas/lenses/dist/sudoers.aug:510.11-513.33:
    Error encountered at 1:0 (0 characters into string)
                               <|=|%opssudoers ALL=(ALL) ALL, !>

    Tree generated so far:


Syntax error in lens definition
Failed to load /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug

It turns out that these messages actually indicated that my tests were bad, not the changes I’d made to the lens file.

Testing

Once I’d recognised this I set about fixing up my tests.

Unit testing

Unit testing in Augeas is not like any other unit testing I’ve seen. Instead of testing the result of executing a function, your verifying that a correctly labelled parse-tree has been generated.

The parse-tree generated from parsing the input string must exactly match 1:1 with the one specified in your test case! This includes:

  • whitespace tokens
  • comments
  • newlines (\n)

The test I developed for this bug looks like this:

(* Test: Sudoers.spec
     https://github.com/hercules-team/augeas/issues/262:  Sudoers lens doesn't suppot `!` for command aliases *)
test Sudoers.spec get "%opssudoers ALL=(ALL) ALL, !BANNED\n" =
  { "spec"
    { "user" = "%opssudoers" }
    { "host_group"
      { "host" = "ALL" }
      { "command" = "ALL"
        { "runas_user" = "ALL" } }
      { "command" = "!BANNED" }
    }
  }

Briefly, this code code performs a partial parse on the input string using the Sudoer.spec rule that we looked at earlier. After the equals sign, we describe the exact partial parse tree - partial because we are only parsing against the spec rule, not the whole file rule.

Once I was able to match the generated tree with the testcase tree, the tests passed. You can tell your tests are passing when augparse gives no output, like this:

[root@master vagrant]# /opt/puppet/bin/augparse /opt/puppet/share/augeas/lenses/dist/tests/test_sudoers.aug
[root@master vagrant]#

The things I had to do to correct my testfiles were:

  • Understand that there are typically two parts to a test: Running it and then doing an exact parse tree comparison to make sure everything is matching correctly
  • Comments and whitespace will normally show up in the parse-tree so you have to take care to match them exactly
  • Ensure the line is ended with a \n sequence.

With unit tests passing, it was time to perform whole-system testing to make sure that the lens was indeed correctly working.

Test the parse tree

To test that a valid parse-tree was produced outside of the testcase, I used augtool with it’s ls command. This printed out pages of data with no visible errors:

[root@master augeas_sudo_hotfix]# /opt/puppet/bin/augtool ls /files/etc/sudoers
#comment[1] = # Sudoers allows particular users to run various commands as
#comment[2] = # the root user, without needing the root password.
#comment[3] = #
#comment[4] = # Examples are provided at the bottom of the file for collections
#comment[5] = # of related commands, which can then be delegated out to particular
#comment[6] = # users or groups.
...

Double check for error messages

Just to make sure I hadn’t missed any error messages in the pages of text from the parse-tree, I ran the following command to check for errors:

[root@master vagrant]# /opt/puppet/bin/augtool print /augeas//error
[root@master vagrant]#

Test file updating

The final part of testing was to ensure that I was now able to write a negated command alias in the /etc/sudoers file.

This is really easy using augtool so all I had to do was run:

[root@master ~]# augtool
augtool> set /files/etc/sudoers/spec[2]/host_group/command[3] !NETWORKING
save
Saved 1 file(s)
augtool> quit
[root@master ~]#

The numbers in square brackets indicate the position in the parse-tree being added or edited and your able to use tab to complete the parse-tree paths as you type them which is hugely helpful.

After checking that the entry was created correctly in /etc/sudoers:

[root@master vagrant]# grep \!NET /etc/sudoers
%sportshorseracingopssudoers ALL=(ALL) ALL, !ADMIN_CMDS , !NETWORKING

I was happy that this code was ready for release by creating a GitHub Pull Request.

Resolution

Hopefully this fix can be merged upstream soon. If your interested in a closer look look at the code that was produced, please see the Pull Request

Post comment