Runaway "git remote" if group definition contains a remote by the same name

Hi,

it is also a way to create a fork bomb out of the innocent tool on platforms where pressing Ctrl-C does not terminate subprocesses of the foreground process (like, of course, Windows).

To reproduce, run

   git -c remotes.origin='origin other' remote update origin

I just cannot look at it right now, and have to resolve to only reporting the problem to warn people. Something seems to resolve the remotes group definition over and over again.

FWIW, the issue is still present.

---------- Forwarded message ----------
From: Alex Riesen <raa.lkml@gmail.com>
Date: Wed, Nov 17, 2010 at 6:10 PM
Subject: Runaway "git remote" if group definition contains a remote by
the same name
To: Git Mailing List <git@vger.kernel.org>

Hi,

it is also a way to create a fork bomb out of the innocent tool on platforms where pressing Ctrl-C does not terminate subprocesses of the foreground process (like, of course, Windows).

To reproduce, run

   git -c remotes.origin='origin other' remote update origin

I just cannot look at it right now, and have to resolve to only reporting the problem to warn people. Something seems to resolve the remotes group definition over and over again.

On Sat, Dec 28, 2013 at 03:56:55PM +0100, Alex Riesen wrote:

Hmm. This is a pretty straightforward reference cycle. We expand origin to contain itself, so it recurses forever on expansion. As with most such problems, the cycle path may be longer than one:

  git -c remotes.foo='bar baz' -c remotes.bar='foo baz' fetch foo

Detecting the cycle can be done by keeping track of which names we've seen, or just by putting in a depth limit that no sane setup would hit. In either case, it's complicated slightly by the fact that we pass the expanded list to a sub-process (which then recurses on the expansion). So we'd have to communicate the depth (or the list of seen remotes) via the command line or the environment.

One alternative would be to have the parent "git fetch" recursively expand the list itself down to scalar entries, and then invoke sub-processes on the result (and tell them not to expand at all). That would also let us cull duplicates if a remote is found via multiple groups.

Interestingly, the problem does not happen with this:

  git -c remotes.foo=foo fetch foo

Fetch sees that foo expands only to a single item and says "oh, that must not be a group". And then treats it like a regular remote, rather than recursing. So it's not clear to me whether groups are meant to be recursive or not. They are in some cases:

  # fetch remotes 1-4
  git -c remotes.parent='child1 child2' \
      -c remotes.child1='remote1 remote2' \
      -c remotes.child2='remote3 remote4' \
      fetch parent

but not in others:

  # "foo" should be an alias for "bar", but it's not
  git -c remotes.foo=bar \
      -c remotes.bar='remote1 remote2' \
      fetch foo

If they are not allowed to recurse, the problem is much easier; the parent fetch simply tells all of the sub-invocations not to expand the arguments further. However, whether it was planned or not, it has been this way for a long time. I would not be surprised if somebody is relying on the recursion to help organize their groups.

So I think the sanest thing is probably:

  1. Teach "fetch" to expand recursively in a single process, and then
     tell sub-processes (via a new command-line option) not to expand
     any further.

  2. Teach "fetch" to detect cycles (probably just by a simple depth
     counter).

  3. Teach the group-reading code to detect groups more robustly, so
     that a single-item group like "remotes.foo=bar" correctly recurses
     to "bar".

  4. (Optional) Teach the expansion code from step 1 to cull duplicates,
     so that we do not try to fetch from the same remote twice (e.g., if
     it is mentioned as part of two groups, and both are specified on
     the command line).

I do not plan to work on this myself in the immediate future, but perhaps it is an interesting low-hanging fruit for somebody else.

-Peff

Jeff King <peff@peff.net> writes:

I suspect that the expansion code will just accumulate remotes found into a string-list (as part of 4. below), so deduping would be fairly easily done without a depth counter.

A single-item remote group is somewhat annoying, but expanding it only at some places while ignoring it at other places is even more annoying, so this sounds like a right thing to do.

On Mon, Dec 30, 2013 at 11:10:31AM -0800, Junio C Hamano wrote:

I don't think that will work (at least not naively). The end-product of step 1, and the string_list that is de-duped in step 4, is a list of the concrete remotes. The cycles occur between groups, which are not mentioned in the final list.

You can keep a separate list of the groups we visit, of course, but we do not otherwise need it.

One thing that does make such a list easier is that we do not need to care about order. E.g., in config like this:

  [remotes]
  a = c
  b = c
  c = d e

you can mark "c" as seen after visiting it via "a". It is not technically a cycle, but since we would want to suppress duplicates anyway, we can be overly broad.

The only configuration that I think would be negatively affected is something like:

  [remote]
  foo = foo
  [remote "foo"]
  url = ...

that silently works now, but would become broken (because we would complain about the cycle). I think that's OK; that config is clearly stupid and broken. If it were "remote.foo = foo bar", trying to expand the concrete "foo" and "bar", that might make some sense, but then it is already broken in the current code (that is the example that started the discussion).

-Peff