bibtex2html and compile plugins

Answer to anarcat mentionning compile in the bibtex2html documentation.

I do not think that the compile plugin can replace the bibtex2html plugin right now: the compile plugin compiles a document, and renders an HTML link to the compiled document, whereas (if I am not wrong), the bibtex2html plugin compiles a document, and renders the content of the compiled document (which happens to be some HTML code).

Right now, it is not possible for the compile plugin to render the content of the compiled document. This could be done by providing a DESTCONTENT template variable, containing the content of the compiled document. This should not be hard to implement.

-- Louis (author of compile)

Interesting! I am thinking of writing a simpler plugin (maybe called "exec") that would be a merge of the two approaches. There would be an htmlize hook to render arbitrary page extensions (based on the configuration) into anything with predefined pipelines. Then a preprocess hook would allo directives to inject links to documents or simply inline them like bibtex2html does. The plugin could be secure insofar as the pipelines configured are secure as well. Should that be merged in compile or be a separate plugin? --anarcat

This "arbitrary executable" stuff scares me, and I'm not going to merge anything like that without a relatively paranoid review. As a result, it could take a while.

At some point when I have more time and energy I should try to write down what ikiwiki's security model is. The short version is that the plugins shipped with ikiwiki should never let wiki editors execute arbitrary code, even if they have direct VCS access or can alter "safe"-flagged setup options. The ability to alter non-"safe" setup options is equivalent to access to the uid running the wiki, and so is the ability to alter the plugins that the wiki uses.

Defining pipelines or compilation commands in the setup file does not directly contradict that, because the setup option would not be flagged as safe, but it does provide an easy way to cause a huge increase in attack surface, particularly when shell scripts are known to be a difficult thing to write securely. If people want arbitrary compilation, Perl or XML-RPC (e.g. Python) plugins are probably safer (even if they call external commands with IPC::Run or subprocess!), and they would mean that the authors of the arbitrary-compilation code can't have any illusions about "oh, this isn't all that security-sensitive, I'm just writing an ad-hoc command".

I hope that ImageTragick is still fresh in everyone's minds - many of the individual vulnerabilities there involved ImageMagick and GraphicsMagick running arbitrary shell pipelines from delegates.xml that turned out not to be hardened against invocation by a hostile user. --smcv

The exec plugin would definitely not be marked as "safe": it allows wiki administrators to execute arbitrary code as the wiki user.

That said, I believe it is possible to craft a configuration that would be safe to use for users in general. Of course, we can't exclude foot-shooting here: if an admin misconfigures the plugin, they can introduce new attack vectors. But default and sample configurations should be secure.

It is with that model in mind that I wrote the bibtex2html plugin: it doesn't use the shell to execute bibtex2html:

my @bibtex_cmd = (qw[bibtex2html -noheader -nofooter -nobibsource -nodoc -q -o -], $near); open(PIPE, "-|", @bibtex_cmd) || error "can't open pipe to @bibtex_cmd: $!";

I specifically tried to address that case, to make sure users can't execute arbitrary code even if the plugin is enabled.

Still: it is tricky! The above pipeline could still be vulnerable to certain attacks if bibtex2html does some dangerous stuff on its own. For example, it could call other executables with the shell without checking arguments, and then the filename would be expanded into hostile shell commands. Even worse and trickier, the filename could be something like -oclobberfile and one file would be destroyed!

bibtex2html is probably vulnerable to such an attack right now. We should check attachments for weird filenames and restrict what is allowed to upload and pass to the plugin.

In case you haven't reviewed the compile plugin in detail, what struck me as an interesting model is the way it allows admin to configure extensions to pipeline mappings. What I had in mind was something like this:

- bib: 'bibtex2html -o- %s'
- svg: 'inkscape -o- %s'
- tex:
  - 'pdflatex %s'
  - 'bibtex %s'
  - 'pdflatex %s'
  - 'pdflatex %s'

(forgive my YAML cluelessness, the idea above is to define a hash of extension -> (command) mapping.) The command would be broken up on spaces into an array and the %s element would be replaced by the source file, which would be forbidden to use shell metacharacters, including prefixed dashes. I believe such a plugin could be crafted to be secure with proper configuration

Of course, it's better if there's a native plugin for everything. For bibtex, we need to use Text::Bibtex, for example. But that basically means rewriting bibtex2html in Perl, which not something any user can do easily. And it's an even worse problems for documents like Word spreadsheets or Latex documents. Only the native commands can do the right thing.

A clever admin can certainly find out about such a command and having a way for that admin to easily hook that into ikiwiki would be a powerful tool, with all that implies. --anarcat

Concerning the ability to run arbitrary commands, a discussion was started by someone who wanted a secure version of this plugin. The idea I had (which has some similarities with what is being discussed here) was to provide a compile_secure boolean option to restrict what the user can do (if false, users can run arbitrary commands; if true, users can only run a set of predefined commands). However, since fr33domlover, who started the discussion, did not answer, nothing was implemented.

Concerning arbitrary commands, I do not know Perl, but I think it can run commands using something similar to exec, which prevents (?) shell injections. This adds the burden of manipulating arrays instead of strings, but security should be improved.

But none of those ideas solve the problems you mentionned, being that external commands can do nasty things (the -oclobberfile option of bibtex2html) or contain bugs (like ImageMagick).

If we want to merge this plugin and compile, I think a better idea than the one I proposed at the beginning of the discussion would be to provide two different directives: a [[!compile ""]] would compile the file and render it as a link to the compiled file (what the compile plugin does right now), while [[!render ""]] would compile the file, and render its content in the current page (whath the bibtex2html plugin does). In fact, providing this [[!render ...]] directive (without the security considerations) seems easy enough to implement, and I might implement it some day (soon, if it solves anarcat problem and closes the discussion).

While I am really happy to see that my plugin sparks some interest, I fear I won't be able to implement what is discussed here, apart from the quick feature I mentionned in the previous paragraph (I have a baby at home, I am moving to another city in a few weeks, and the only code I ever wrote in Perl was to contribute to IkiWiki). However, you have my blessing for making whatever you want with my code: contribute, write a version 2 of it, write a new plugin that makes it obsolete, copy the good ideas and dump the rest, etc.