patch[RFC] setup.c: make bare repo discovery optional

From: Glen Choo <chooglen@google.com>

Add a config variable, `safe.barerepository`, that tells Git whether or not to recognize bare repositories when it is trying to discover the repository. This only affects repository discovery, thus it has no effect if discovery was not done (e.g. `--git-dir` was passed).

This is motivated by the fact that some workflows don't use bare repositories at all, and users may prefer to opt out of bare repository discovery altogether:

- An easy assumption for a user to make is that Git commands run
  anywhere inside a repository's working tree will use the same
  repository. However, if the working tree contains a bare repository
  below the root-level (".git" is preferred at the root-level), any
  operations inside that bare repository use the bare repository
  instead.

  In the worst case, attackers can use this confusion to trick users
  into running arbitrary code (see [1] for a deeper discussion). But
  even in benign situations (e.g. a user renames ".git/" to ".git.old/"
  and commits it for archival purposes), disabling bare repository
  discovery can be a simpler mode of operation (e.g. because the user
  doesn't actually want to use ".git.old/") [2].

- Git won't "accidentally" recognize a directory that wasn't meant to be
  a bare repository, but happens to resemble one. While such accidents
  are probably very rare in practice, this lets users reduce the chance
  to zero.

This config is designed to be used like an allow-list, but it is not yet clear what a good format for this allow-list would be. As such, this patch limits the config value to a tri-state of [true|false|unset]:

- [*|(unset)] recognize all bare repositories (like Git does today)
- (empty) recognize no bare repositories

and leaves the full format to be determined later.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: I don't personally know anyone who does this as part of their normal workflow, but a cursory search on GitHub suggests that there is a not insubstantial number of people who munge ".git" in order to store its contents.

https://github.com/search?l=&o=desc&p=1&q=ref+size%3A%3C1000+filename%3AHEAD&s=indexed&type=Code (aka search for the text "ref", size:<1000, filename:HEAD)

Signed-off-by: Glen Choo <chooglen@google.com>
---
    RFC setup.c: make bare repo discovery optional
    
    (Forgive the non-standard RFC tag, I haven't figured out how to send as
    RFC using GGG. I also didn't realize that /preview would also respect
    CC...)
    
    = Description
    
    A relatively easy win that came out of the discussions around embedded
    bare repos [1], is to just let users opt-out of discovering bare repos.
    This patch does exactly that, by adding a 'boolean' config variable,
    safe.barerepository.
    
    safe.barerepository is presented to users as an allow-list of
    directories that Git will recognize as a bare repository during the
    repository discovery process (much like safe.directory), but this patch
    only implements (and permits) boolean behavior (i.e. on, off and unset).
    Hopefully, this gives us some room to discuss and experiment with
    possible formats.
    
    Thanks to Taylor for suggesting the allow-list idea :)
    
    I think the core concept of letting users toggle bare repo discovery is
    solid, but I'm sending this as RFC for the following reasons:
    
     * I don't love the name safe.barerepository, because it feels like Git
       is saying that bare repos are unsafe and consequently, that bare repo
       users are behaving unsafely. On the other hand, this is quite similar
       to safe.directory in a few respects, so it might make sense for the
       naming to reflect that.
    
     * The *-gcc CI jobs don't pass. I haven't discerned any kind of pattern
       yet.
    
    = How this relates to embedded bare repos
    
    This does not change the default behavior (i.e. Git will still discover
    all bare repos by default) because that would be catastrophic for bare
    repo users [2]. As such, this patch isn't intended to solve the problem
    of embedded bare repos for all users once and for all, but I think it
    does improve the our stance on the matter:
    
     * In the short-term, users who know they won't need bare repos (or
       those who are willing to set GIT_DIR for all of their bare repos) can
       opt-in to a safer, easier to reason about mode of operation.
    
     * In the longer-term, we might identify a usable-enough default that we
       can give opt-out protection that works for the vast majority of
       users.
    
    = Other questions/Concerns
    
     * Maybe it's more informative for the user if we die() (or warn()) when
       we find a bare repo instead of silently ignoring it?
    
     * I wonder if it makes sense to separate the toggle for bare repo
       discovery and the allow-list of bare repositories. Something like
       core.barediscovery or discovery.barerepository has a lot less baggage
       than safe.*, and boolean enable/disable is a lot simpler, but this
       isn't good from an extensibility perspective.
    
     * Is there any reason why safe.barerepository shouldn't use the same
       format as (its obvious inspiration) safe.directory?
    
     * Are the docs clear enough? I found those hard to put into words, so
       I'd especially appreciate wording suggestions :)
    
    = Future work
    
     * Like safe.directory, safe.barerepository is only read from system and
       global config. I anticipate that this is too restrictive; there has
       already been some discussion of adding a GIT_SAFE_DIRECTORIES
       environment variable for safe.directory [3], and it would be useful
       to have the same thing for safe.barerepository.
    
    [1]
    https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com
    [2] In https://lore.kernel.org/git/xmqqh76ucdg6.fsf@gitster.g, Junio
    experimented with switching off bare repo discovery altogether and
    relying solely on GIT_DIR. The resulting fallout was deemed too big to
    be feasible. [3] https://lore.kernel.org/git/xmqqee1il09v.fsf@gitster.g/

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v1
Pull-Request: https://github.com/git/git/pull/1261

 Documentation/config/safe.txt | 24 +++++++++++++++
 setup.c                       | 36 ++++++++++++++++++++++-
 t/t1510-repo-setup.sh         | 55 +++++++++++++++++++++++++++++++++++
 3 files changed, 114 insertions(+), 1 deletion(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

"maliciously bare"? "malicious bare" probably.

"mean to be" -> "meant to be".

This is in a loop, which will go up and try the parent directory if the body of this block is not entered, so it is calling the new should_detect_bare() helper over and over if it returns false.

Not a very good idea.

Perhaps this would help?  I dunno.

static int should_detect_bare(void)
{
	static int should = -1; /* unknown yet */

	if (should < 0) {
		struct safe_bare_repository_data data = { 0 };
		read_very_early_config(safe_bare_repository_cb, &data);
		should = data.is_safe;
	}
	return should;
}

In any case, I very much appreciate the fact that this touches the setup_git_directory_gently_1() codepath only minimally, as we have other plans to update the code further soonish.

Thanks.

Hi Glen,

On Fri, May 06, 2022 at 06:30:10PM +0000, Glen Choo via GitGitGadget wrote:

Thanks for working on this! I'm excited to see some patches here, though I'm not totally convinced of this direction. More below.

To summarize, this proposal attempts to work around the problem of embedding bare repositories in non-bare checkouts by providing a way to opt-out of bare repository discovery (which is to only discover things that are listed in the safe.bareRepository configuration).

I agree that this would prevent the problem you're trying to solve, but I have significant concerns that this patch is going too far (at the risk of future damage to unrelated workflows) in order to accomplish that goal.

My concern is that if we ever flipped the default (i.e. that "safe.bareRepository" might someday be ""), that many legitimate cases of using bare repositories would be broken. I think there are many such legitimate use cases that _do_ rely on discovering bare repositories (i.e., Git invocations that do not have a `--git-dir` in their command-line). One such example would be forges, but I imagine that there are many other uses we don't even know about, and I would like to avoid breaking those if we ended up changing the default.

If it's possible to pursue a more targeted fix that leaves non-embedded bare repositories alone, I'd like to try and focus these efforts on a more narrow fix that would address just the case of embedded bare repositories. I think that the direction I outlined in:

    https://lore.kernel.org/git/Ylobp7sntKeWTLDX@nand.local/

could be a good place to start (see the paragraph beginning with "Here's an alternative approach" and below for the details).

One potential problem with that approach (that this patch doesn't suffer from) is that any discovery which finds a bare repository would have to continue up to the root of the volume in order to figure out whether or not that bare repository is embedded in another non-bare one. That is probably a non-starter due to performance, but I think you could easily work around with a top-level setting that controls whether or not you even _care_ about embedded bare repositories.

For example, if I set safe.bareRepository='*' in my top-level /etc/gitconfig, then we can avoid having to continue discovery for bare repositories altogether because we know we'll allow it anyway.

To pursue a change that targets just embedded bare repositories, I think you fundamentally have to do an exhaustive repository discovery in order to figure out whether the (bare) repository you're dealing with is embedded or not. So having an opt-out for users that either (a) don't care or (b) can't accept the performance degradation that Emily mentioned as a result of doing unbounded filesystem traversal would be sensible.

Playing devil's advocate for a moment, though, even if we had something like the proposal I outlined, flipping the top-level default from '*' to some value that implies we stop working in embedded bare repositories will break existing workflows. But that breakage would just be limited to embedded bare repositories, and not non-embedded ones. So I think on balance that breakage would affect fewer real-world users, while still being just as easy to recover from.

I did suggest an allow-list, but not this one ;-).

Yes, the concerns I outlined above are definitely echoing this sentiment. Another way to say it is that this feels like too big of a hammer (i.e., it is targeting _all_ bare repositories, not just embedded ones) for too small of a nail (embedded bare repositories). As you're probably sick of hearing me say by now, I would strongly prefer a more targeted solution (perhaps what I outlined, or perhaps something else, so long as it doesn't break non-embedded bare repositories if/ever we decided to change the default value of safe.bareRepository).

Interesting. I wouldn't expect this to be the case (since the default is to allow everything right now).

Perhaps, and I think if this were the case then I would feel differently about this patch. But I don't want us to paint ourselves into a corner, either. It would be unfortunate to, say, find ourselves in a position where the only protection against some novel embedded bare repository attack is to change a default that would break many existing workflows for _non_-embedded bare repositories.

We should definitely provide more feedback to the user. If I set `safe.bareRepository` to the empty string via a global config, and then execute a Git command in a non-embedded bare repository, I get:

    $ git.compile config --get --global --default='*' safe.bareRepository

    $ git.compile rev-parse --absolute-git-dir
    fatal: not a git repository (or any of the parent directories): .git

whereas on the last release of Git, I get instead:

    $ git rev-parse --absolute-git-dir
    /home/ttaylorr/repo.git

I'm still not convinced that just reading repository extensions while ignoring the rest of config and hooks is too confusing, so I'd be more in favor of something like:

    $ git.compile rev-parse --absolute-git-dir
    warning: ignoring repository config and hooks
    advice: to permit bare repository discovery (which
    advice: will read config and hooks), consider running:
    advice:
    advice:   $ git config --global --add safe.bareRepository /home/ttaylorr/repo.git
    /home/ttaylorr/repo.git

(though I still feel strongly that we should pursue a more targeted approach here).

Thanks, Taylor

Taylor Blau <me@ttaylorr.com> writes:

I think 99% of such use is to chdir into the directory with HEAD, refs/ and objects/ in it and let git recognise the cwd is a git directory. Am I mistaken, or are there tools that chdir into objects/08/ and rely on setup_git_directory_gently_1() to find the parent directory of that 'objects' directory to be a git directory?

I am wondering if another knob to help that particular use case easier may be sufficient. If you are a forge operator, you'd just set a boolean configuration variable to say "it is sufficient to chdir into a directory to use it a bare repository without exporting the environment variable GIT_DIR=."

It is likely that end-user human users would not want to enable such a variable, of course, but I wonder if a simple single knob would be sufficient to help other use cases you are worried about?

While I wish "extensions and nothing else", i.e. we use "degraded access", not "refuse to give access at all", were workable, I am pessimistic that it would work well in practice.

Saying "nothing else" is easy, but we do "if X exists, use it" for hook, and to implement "nothing else", you'd need to find such a code and say "even if X exists, because we are in this strange embedded bare thing, ignore this part of the logic" for every X. We've been casually saying "potentially risky config" and then started mixing "hooks" in the discussion, but who knows what other things are used from the repository by third-party tools that we need to yet add to the mix?

I do not think it would be confusing. I am worried about it being error prone.

Is the last line meant to be an output from "rev-parse --absolute-git-dir"? IOW, the warning says you are ignoring, but we are still recognising it as a repository?

By the way, do we need safe.bareRepository? Shouldn't safe.directory cover the same purpose?

If a directory is on the latter, you are saying that (1) the directory is OK to use as a repository, and (2) it is so even if the directory is owned by somebody else, not you.

Theoretically you can argue that there can be cases where you only want (1) and not (2), but as long as you control such a directory (like an embedded repository in your project's checkout) yourself, you do not have to worry about the "ok even if it is owned by somebody else" part.

On Mon, May 09, 2022 at 03:54:08PM -0700, Junio C Hamano wrote:

If you took this change, and then at some point in the future we changed the default value of safe.bareRepository to "", wouldn't that break that 99% of use cases you are talking about?

When I read your "I think 99% of such use is ...", it makes me think that this change won't disrupt bare repo discovery when we only traverse one layer above $CWD. But this change disrupts the case where we don't need to traverse at all to do discovery (i.e., when $CWD is the root of a bare repository).

Yes, GitHub would almost certainly set safe.bareRepository to "*" regardless of what Git's own default would be.

I'm not sure I agree that end-users wouldn't want to touch this knob. If they have embedded bare repositories that they rely on as test fixtures, for example, wouldn't safe.bareRepository need to be tweaked?

(On a separate but somewhat-related note, I still think that this setting should be read from the repository config, too, i.e., it seems odd that we'd force a user to set safe.bareRepository to some deeply nested repository (in the embedded case) via their global config.)

Yeah, on this and the above quoted hunk, I am fine if our behavior eventually became "call die()" for when we are in an embedded bare repository. But I do think this transition should be gradual, i.e., we should likely emit a warning in those cases that would be broken in the future to say "this will break, run this `git config` invocation if you want it to remain working".

In this example, yes. But again, I'm not so deeply attached to the idea that we *have* to run in those cases. So I would equally be OK with the above s/warning/fatal and minus the last line, too (i.e., that we call die(), obviously we'd have to emit the advice before calling die()).

I'm not sure yet, but will think more about it.

Thanks, Taylor

Taylor Blau <me@ttaylorr.com> writes:

Our spawning (e.g. "fetch" run_command()s "upload-pack" in a local repository, or "fetch" runs "upload-pack" over ssh connection, or http gateway runs "upload-pack" after learning which repository the request is fetching from) of subcommands can and should be fixed by exporting "GIT_DIR=." when we spawn them in the target directory, and such a fix should be more or less trivial. It must happen before such a switch of default happens (if it is what we plan to do, that is). Also, the trivial fix must be conveyed to third-party tool authors and give them time to adjust their ware.

That's part of the usual migration process, and I am not so worried about it.

If some third-party tool for whatever reason wants to start from a random subdirectory in a bare repository, that is a different story. Fixing such a third-party tool would be more involved than "more or less trivial".

By "this change" you mean what Glen proposes? I think it was designed to break the use case where you go there to signal that you want to use the directory as a repository.

And such a boolean, without safe.bareRepository setting, should be sufficient to cover that 99% of such use, because it disables that deliberate refusal of treating CWD as a repository without explicitly saying that is what you want with "GIT_DIR=.". One thing I wasn't sure about was if that 99% number is close to reality, hence my question.

And with such a boolean, I am hoping that GitHub do not have to make such a wildly open setting. Only $CWD that is the top of a repository, without allowing it to be any random subdirectory, would be allowed.

But not in the "My $CWD is always fine" knob, whose only reason is to simplify things without opening you up unnecessarily too widely for hosting sites.

Hi Taylor,

Taylor Blau <me@ttaylorr.com> writes:

Thanks again for the careful read. As I understand it, your concern is that making bare repository discovery configurable and then flipping the default to e.g. never detecting bare repositories is too disruptive to fix the embedded bare repository problem. And to avoid disrupting non-embedded bare repositories, you would prefer to pursue a more targeted fix.

If the problem statement were limited to embedded bare repositories, then I agree that this is way more than overkill, and that a targeted solution would be preferable.

More generally however, the problem of embedded bare repositories seems to suggest that bare repository discovery doesn't serve all users well, and in fact, may even be a net negative for a subset of users. I'd be interested in hearing your thoughts from that perspective, e.g.

- Should bare repository discovery should be configurable?
- What is a good default for bare repository discovery? (regardless of
  how feasible changing the default is)

This is a somewhat different direction from how the conversation started (I hope it doesn't look like I'm shifting the goal posts), but I think it's a good opportunity to step back and simplify something that we wished we got right in the beginning.

And even if we don't flip the default, shipping the config value still seems useful e.g. there's a good amount of interest in disabling bare repository discovery at $DAYJOB (and I think we'll get a lot of interesting results once we do).

Ah, yes. Oops. Sorry if it looked like I was putting words in your mouth.

What I really meant was that an allow-list (untethered from any specific purpose) seems like a useful 'UI primitive', so thanks for bringing up the option.

Ok, yeah I think safe.barerepository is a terrible way to achieve my purported goal of 'making bare repository discovery configurable/simpler/' - using the "safe." namespace makes it impossible to see this as anything other than protection against dangerous, unknown bare repositories. I'll drop the idea of safe-listing known bare repositories for now, that seems unproductive.

'Optionally disable bare repository discovery' still sounds like it's on the table though, but probably with a different kind of UX e.g. "discovery.barerepository" with the options:

- always: always discover bare repos
- never: never discover bare repos
- cwd-only: only discover bare repos if they are the cwd
- dotgit-only: only discover bare repos if they are a descendant of
  .git/

This might be a false alarm - I saw similar failures on an unrelated patch. I think my "master" is just out of date :(

Thanks all for the comments on v1, I've expanded this series somewhat to address them, notably:

 * The config value is now named discovery.bare and is an enum (instead of
   an allow-list). This hopefully moves us away from "bare repos are unsafe
   and need to be guarded against" and towards "bare repos can be made
   optional if it serves your needs better".
 * discovery.bare now causes git to die() instead of silently ignoring the
   bare repo.
 * Add an option that allows a bare repo if it is the CWD, since this is
   presumably a reasonable default for 99% of bare repo users [1].

= Questions/Concerns

 * die()-ing is necessary if we're trying to flip the default value of
   discovery.bare. We'd expect many bare repo users to be broken, and it's
   more helpful to fail loudly than to silently ignore the bare repo.
   
   But in the long term, long after we've flipped the default and users know
   that they need to opt into bare repo discovery, would it be a better UX
   to just silently ignore the bare repo?

= Patch organization

 * Patch 1 introduces discovery.bare with allowed values [always|never].

 * Patch 2 adds discover.bare=cwd, which is useful when users don't always
   set GIT_DIR e.g. their workflow really depends on it, they are in the
   midst of migration.

= Series history

Changes since v1:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

[1] I tried this 'cwd' setting on our test suite, with some pretty promising results.

https://github.com/chooglen/git/actions/runs/2321914777

Out of the 8 failing scripts:

 * 6 are of the form "make sure we 'do the right thing' inside a
   subdirectory of a bare repo" (which typically means .git) e.g.
   t9903-bash-prompt.sh. We should be setting discovery.bare=always for
   these tests, so this is a non-issue.
 * t5323-pack-redundant.sh can be rewritten to -C into the root of the bare
   repo instead of a subdirectory.
 * t3310-notes-merge-manual-resolve.sh: not sure what the test is checking
   in particular, but I think this can be rewritten.

IOW, I don't think we have any commands that require that CWD is a subdirectory of a bare repo, and we could use discovery.bare without much hassle.

Glen Choo (2):
  setup.c: make bare repo discovery optional
  setup.c: learn discovery.bareRepository=cwd

 Documentation/config/discovery.txt | 26 +++++++++
 setup.c                            | 89 ++++++++++++++++++++++++++++--
 t/t0034-discovery-bare.sh          | 69 +++++++++++++++++++++++
 3 files changed, 178 insertions(+), 6 deletions(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0034-discovery-bare.sh

base-commit: e8005e4871f130c4e402ddca2032c111252f070a
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v2
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v1:

 1:  3370258c4b3 ! 1:  22b10bf9da8 [RFC] setup.c: make bare repo discovery optional
     @@ Metadata
      Author: Glen Choo <chooglen@google.com>
      
       ## Commit message ##
     -    [RFC] setup.c: make bare repo discovery optional
     +    setup.c: make bare repo discovery optional
      
     -    Add a config variable, `safe.barerepository`, that tells Git whether or
     -    not to recognize bare repositories when it is trying to discover the
     -    repository. This only affects repository discovery, thus it has no
     +    Add a config variable, `discovery.bare`, that tells Git whether or not
     +    it should work with the bare repository it has discovered i.e. Git will
     +    die() if it discovers a bare repository, but it is not allowed by
     +    `discovery.bare`. This only affects repository discovery, thus it has no
          effect if discovery was not done (e.g. `--git-dir` was passed).
      
          This is motivated by the fact that some workflows don't use bare
     @@ Commit message
            are probably very rare in practice, this lets users reduce the chance
            to zero.
      
     -    This config is designed to be used like an allow-list, but it is not yet
     -    clear what a good format for this allow-list would be. As such, this
     -    patch limits the config value to a tri-state of [true|false|unset]:
     +    This config is an enum of:
      
     -    - [*|(unset)] recognize all bare repositories (like Git does today)
     -    - (empty) recognize no bare repositories
     +    - ["always"|(unset)]: always recognize bare repositories (like Git does
     +      today)
     +    - "never": never recognize bare repositories
      
     -    and leaves the full format to be determined later.
     +    More values are expected to be added later, and the default is expected
     +    to change (i.e. to something other than "always").
      
          [1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com
          [2]: I don't personally know anyone who does this as part of their
     @@ Commit message
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
     - ## Documentation/config/safe.txt ##
     +    WIP setup.c: make discovery.bare die on failure
     +
     +    Signed-off-by: Glen Choo <chooglen@google.com>
     +
     + ## Documentation/config/discovery.txt (new) ##
      @@
     -+safe.barerepository::
     -+	This config entry specifies directories that Git can recognize as
     -+	a bare repository when looking for the repository (aka repository
     ++discovery.bare::
     ++	Specifies what kinds of directories Git can recognize as a bare
     ++	repository when looking for the repository (aka repository
      +	discovery). This has no effect if repository discovery is not
      +	performed e.g. the path to the repository is set via `--git-dir`
      +	(see linkgit:git[1]).
      ++
     -+It is recommended that you set this value so that Git will only use the bare
     -+repositories you intend it to. This prevents certain types of security and
     -+non-security problems, such as:
     -+
     -+* `git clone`-ing a repository containing a maliciously bare repository
     -+  inside it.
     -+* Git recognizing a directory that isn't mean to be a bare repository,
     -+  but happens to look like one.
     -++
     -+The currently supported values are `*` (Git recognizes all bare
     -+repositories) and the empty value (Git never recognizes bare repositories).
     -+Defaults to `*`.
     -++
      +This config setting is only respected when specified in a system or global
      +config, not when it is specified in a repository config or via the command
     -+line option `-c safe.barerepository=<path>`.
     ++line option `-c discovery.bare=<value>`.
     +++
     ++The currently supported values are `always` (Git always recognizes bare
     ++repositories) and `never` (Git never recognizes bare repositories).
     ++This defaults to `always`, but this default is likely to change.
     +++
     ++If your workflow does not rely on bare repositories, it is recommended that
     ++you set this value to `never`. This makes repository discovery easier to
     ++reason about and prevents certain types of security and non-security
     ++problems, such as:
      +
     - safe.directory::
     - 	These config entries specify Git-tracked directories that are
     - 	considered safe even if they are owned by someone other than the
     ++* `git clone`-ing a repository containing a malicious bare repository
     ++  inside it.
     ++* Git recognizing a directory that isn't meant to be a bare repository,
     ++  but happens to look like one.
      
       ## setup.c ##
     +@@
     + static int inside_git_dir = -1;
     + static int inside_work_tree = -1;
     + static int work_tree_config_is_bogus;
     ++enum discovery_bare_config {
     ++	DISCOVERY_BARE_UNKNOWN = -1,
     ++	DISCOVERY_BARE_NEVER = 0,
     ++	DISCOVERY_BARE_ALWAYS,
     ++};
     ++static enum discovery_bare_config discovery_bare_config =
     ++	DISCOVERY_BARE_UNKNOWN;
     + 
     + static struct startup_info the_startup_info;
     + struct startup_info *startup_info = &the_startup_info;
      @@ setup.c: static int ensure_valid_ownership(const char *path)
       	return data.is_safe;
       }
       
     -+/*
     -+ * This is similar to safe_directory_data, but only supports true/false.
     -+ */
     -+struct safe_bare_repository_data {
     -+	int is_safe;
     -+};
     -+
     -+static int safe_bare_repository_cb(const char *key, const char *value, void *d)
     ++static int discovery_bare_cb(const char *key, const char *value, void *d)
      +{
     -+	struct safe_bare_repository_data *data = d;
     -+
     -+	if (strcmp(key, "safe.barerepository"))
     ++	if (strcmp(key, "discovery.bare"))
      +		return 0;
      +
     -+	if (!value || !strcmp(value, "*")) {
     -+		data->is_safe = 1;
     ++	if (!strcmp(value, "never")) {
     ++		discovery_bare_config = DISCOVERY_BARE_NEVER;
      +		return 0;
      +	}
     -+	if (!*value) {
     -+		data->is_safe = 0;
     ++	if (!strcmp(value, "always")) {
     ++		discovery_bare_config = DISCOVERY_BARE_ALWAYS;
      +		return 0;
      +	}
      +	return -1;
      +}
      +
     -+static int should_detect_bare(void)
     ++static int check_bare_repo_allowed(void)
      +{
     -+	struct safe_bare_repository_data data;
     -+
     -+	read_very_early_config(safe_bare_repository_cb, &data);
     ++	if (discovery_bare_config == DISCOVERY_BARE_UNKNOWN) {
     ++		read_very_early_config(discovery_bare_cb, NULL);
     ++		/* We didn't find a value; use the default. */
     ++		if (discovery_bare_config == DISCOVERY_BARE_UNKNOWN)
     ++			discovery_bare_config = DISCOVERY_BARE_ALWAYS;
     ++	}
     ++	switch (discovery_bare_config) {
     ++	case DISCOVERY_BARE_NEVER:
     ++		return 0;
     ++	case DISCOVERY_BARE_ALWAYS:
     ++		return 1;
     ++	default:
     ++		BUG("invalid discovery_bare_config %d", discovery_bare_config);
     ++	}
     ++}
      +
     -+	return data.is_safe;
     ++static const char *discovery_bare_config_to_string(void)
     ++{
     ++	switch (discovery_bare_config) {
     ++	case DISCOVERY_BARE_NEVER:
     ++		return "never";
     ++	case DISCOVERY_BARE_ALWAYS:
     ++		return "always";
     ++	default:
     ++		BUG("invalid discovery_bare_config %d", discovery_bare_config);
     ++	}
      +}
      +
       enum discovery_result {
       	GIT_DIR_NONE = 0,
       	GIT_DIR_EXPLICIT,
     +@@ setup.c: enum discovery_result {
     + 	GIT_DIR_HIT_CEILING = -1,
     + 	GIT_DIR_HIT_MOUNT_POINT = -2,
     + 	GIT_DIR_INVALID_GITFILE = -3,
     +-	GIT_DIR_INVALID_OWNERSHIP = -4
     ++	GIT_DIR_INVALID_OWNERSHIP = -4,
     ++	GIT_DIR_DISALLOWED_BARE = -5
     + };
     + 
     + /*
      @@ setup.c: static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
     - 			return GIT_DIR_DISCOVERED;
       		}
       
     --		if (is_git_directory(dir->buf)) {
     -+		if (should_detect_bare() && is_git_directory(dir->buf)) {
     + 		if (is_git_directory(dir->buf)) {
     ++			if (!check_bare_repo_allowed())
     ++				return GIT_DIR_DISALLOWED_BARE;
       			if (!ensure_valid_ownership(dir->buf))
       				return GIT_DIR_INVALID_OWNERSHIP;
       			strbuf_addstr(gitdir, ".");
     +@@ setup.c: const char *setup_git_directory_gently(int *nongit_ok)
     + 		}
     + 		*nongit_ok = 1;
     + 		break;
     ++	case GIT_DIR_DISALLOWED_BARE:
     ++		if (!nongit_ok) {
     ++			die(_("cannot use bare repository '%s' (discovery.bare is '%s')"),
     ++			    dir.buf,
     ++			    discovery_bare_config_to_string());
     ++		}
     ++		*nongit_ok = 1;
     ++		break;
     + 	case GIT_DIR_NONE:
     + 		/*
     + 		 * As a safeguard against setup_git_directory_gently_1 returning
      
     - ## t/t1510-repo-setup.sh ##
     -@@ t/t1510-repo-setup.sh: test_expect_success '#16e: bareness preserved by --bare' '
     - 	)
     - '
     - 
     -+# Test the tri-state of [(unset)|""|"*"].
     -+test_expect_success '#16f: bare repo in worktree' '
     -+	test_when_finished "git config --global --unset safe.barerepository" &&
     -+	setup_repo 16f unset "" unset &&
     + ## t/t0034-discovery-bare.sh (new) ##
     +@@
     ++#!/bin/sh
      +
     -+	git init --bare 16f/default/bare &&
     -+	git init --bare 16f/default/bare/bare &&
     -+	try_case 16f/default/bare unset unset \
     -+		. "(null)" "$here/16f/default/bare" "(null)" &&
     -+	try_case 16f/default/bare/bare unset unset \
     -+		. "(null)" "$here/16f/default/bare/bare" "(null)" &&
     ++test_description='verify discovery.bare checks'
      +
     -+	git config --global safe.barerepository "*" &&
     -+	git init --bare 16f/all/bare &&
     -+	git init --bare 16f/all/bare/bare &&
     -+	try_case 16f/all/bare unset unset \
     -+		. "(null)" "$here/16f/all/bare" "(null)" &&
     -+	try_case 16f/all/bare/bare unset unset \
     -+		. "(null)" "$here/16f/all/bare/bare" "(null)" &&
     ++. ./test-lib.sh
      +
     -+	git config --global safe.barerepository "" &&
     -+	git init --bare 16f/never/bare &&
     -+	git init --bare 16f/never/bare/bare &&
     -+	try_case 16f/never/bare unset unset \
     -+		".git" "$here/16f" "$here/16f" "never/bare/" &&
     -+	try_case 16f/never/bare/bare unset unset \
     -+		".git" "$here/16f" "$here/16f" "never/bare/bare/"
     -+'
     ++pwd="$(pwd)"
      +
     -+test_expect_success '#16g: inside .git with safe.barerepository' '
     -+	test_when_finished "git config --global --unset safe.barerepository" &&
     ++expect_allowed () {
     ++	git rev-parse --absolute-git-dir >actual &&
     ++	echo "$pwd/outer-repo/bare-repo" >expected &&
     ++	test_cmp expected actual
     ++}
      +
     -+	# Omit the "default" case; it is covered by 16a.
     ++expect_rejected () {
     ++	test_must_fail git rev-parse --absolute-git-dir 2>err &&
     ++	grep "discovery.bare" err
     ++}
      +
     -+	git config --global safe.barerepository "*" &&
     -+	setup_repo 16g/all unset "" unset &&
     -+	mkdir -p 16g/all/.git/wt/sub &&
     -+	try_case 16g/all/.git unset unset \
     -+		. "(null)" "$here/16g/all/.git" "(null)" &&
     -+	try_case 16g/all/.git/wt unset unset \
     -+		"$here/16g/all/.git" "(null)" "$here/16g/all/.git/wt" "(null)" &&
     -+	try_case 16g/all/.git/wt/sub unset unset \
     -+		"$here/16g/all/.git" "(null)" "$here/16g/all/.git/wt/sub" "(null)" &&
     ++test_expect_success 'setup bare repo in worktree' '
     ++	git init outer-repo &&
     ++	git init --bare outer-repo/bare-repo
     ++'
     ++
     ++test_expect_success 'discovery.bare unset' '
     ++	(
     ++		cd outer-repo/bare-repo &&
     ++		expect_allowed &&
     ++		cd refs/ &&
     ++		expect_allowed
     ++	)
     ++'
     ++
     ++test_expect_success 'discovery.bare=always' '
     ++	git config --global discovery.bare always &&
     ++	(
     ++		cd outer-repo/bare-repo &&
     ++		expect_allowed &&
     ++		cd refs/ &&
     ++		expect_allowed
     ++	)
     ++'
      +
     -+	git config --global safe.barerepository "" &&
     -+	setup_repo 16g/never unset "" unset &&
     -+	mkdir -p 16g/never/.git/wt/sub &&
     -+	try_case 16g/never/.git unset unset \
     -+		".git" "$here/16g/never" "$here/16g/never" ".git/" &&
     -+	try_case 16g/never/.git/wt unset unset \
     -+		".git" "$here/16g/never" "$here/16g/never" ".git/wt/" &&
     -+	try_case 16g/never/.git/wt/sub unset unset \
     -+		".git" "$here/16g/never" "$here/16g/never" ".git/wt/sub/"
     ++test_expect_success 'discovery.bare=never' '
     ++	git config --global discovery.bare never &&
     ++	(
     ++		cd outer-repo/bare-repo &&
     ++		expect_rejected &&
     ++		cd refs/ &&
     ++		expect_rejected
     ++	) &&
     ++	(
     ++		GIT_DIR=outer-repo/bare-repo &&
     ++		export GIT_DIR &&
     ++		expect_allowed
     ++	)
      +'
      +
     - test_expect_success '#17: GIT_WORK_TREE without explicit GIT_DIR is accepted (bare case)' '
     - 	# Just like #16.
     - 	setup_repo 17a unset "" true &&
     ++test_done
 -:  ----------- > 2:  62070aab7eb setup.c: learn discovery.bareRepository=cwd

From: Glen Choo <chooglen@google.com>

Add a config variable, `discovery.bare`, that tells Git whether or not it should work with the bare repository it has discovered i.e. Git will die() if it discovers a bare repository, but it is not allowed by `discovery.bare`. This only affects repository discovery, thus it has no effect if discovery was not done (e.g. `--git-dir` was passed).

This is motivated by the fact that some workflows don't use bare repositories at all, and users may prefer to opt out of bare repository discovery altogether:

- An easy assumption for a user to make is that Git commands run
  anywhere inside a repository's working tree will use the same
  repository. However, if the working tree contains a bare repository
  below the root-level (".git" is preferred at the root-level), any
  operations inside that bare repository use the bare repository
  instead.

  In the worst case, attackers can use this confusion to trick users
  into running arbitrary code (see [1] for a deeper discussion). But
  even in benign situations (e.g. a user renames ".git/" to ".git.old/"
  and commits it for archival purposes), disabling bare repository
  discovery can be a simpler mode of operation (e.g. because the user
  doesn't actually want to use ".git.old/") [2].

- Git won't "accidentally" recognize a directory that wasn't meant to be
  a bare repository, but happens to resemble one. While such accidents
  are probably very rare in practice, this lets users reduce the chance
  to zero.

This config is an enum of:

- ["always"|(unset)]: always recognize bare repositories (like Git does
  today)
- "never": never recognize bare repositories

More values are expected to be added later, and the default is expected to change (i.e. to something other than "always").

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: I don't personally know anyone who does this as part of their normal workflow, but a cursory search on GitHub suggests that there is a not insubstantial number of people who munge ".git" in order to store its contents.

https://github.com/search?l=&o=desc&p=1&q=ref+size%3A%3C1000+filename%3AHEAD&s=indexed&type=Code (aka search for the text "ref", size:<1000, filename:HEAD)

Signed-off-by: Glen Choo <chooglen@google.com>

WIP setup.c: make discovery.bare die on failure

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/discovery.txt | 24 +++++++++++
 setup.c                            | 66 +++++++++++++++++++++++++++++-
 t/t0034-discovery-bare.sh          | 59 ++++++++++++++++++++++++++
 3 files changed, 148 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0034-discovery-bare.sh

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

The intended commit message ends here...

Ugh, dumb mistake (bad squash). Fortunately this was one of my more professional-sounding WIP commit messages.

On 5/13/2022 7:37 PM, Glen Choo via GitGitGadget wrote:

I think it is fine to include the "never" option for users to opt-in to this super-protected state, but I want to make it very clear that we should never move to it as a new default. This phrasing of 'something other than "always"' is key, but it might be good to point out that "never" is very unlikely to be that default.

Accidental concatenation of squashed commit?

Avoid "e.g." here.

	This has no effect if the repository is specified directly via
	the --git-dir command-line option or the GIT_DIR environment
	variable.

We are sprinkling config options that have these same restrictions throughout the config documentation. It might be time to define a term like "protected config" at the top of git-config.txt and then refer to that from these other locations.

This sentence structure is likely to change in the future, and as it stands will become complicated. A bulleted list will have easier edits in the future.

For now, I would say "but this default may change in the future." instead.

(You might need a "+" here.)

I think these last bits recommending the 'never' option are a bit distracting. It doesn't make repository discovery "easier to reason about" because we still discover the bare repo and die() instead of skipping it and looking higher for a non-bare repository in the parent directories. The case of an "accidentally-recognized bare repo" is so unlikely it is probably not worth mention in these docs.

Instead, I think something like this might be better:

  If you do not use bare repositories in your workflow, then it may
  be beneficial to set `discovery.bare` to `never` in your global
  config. This will protect you from attacks that involve cloning a
  repository that contains a bare repository and running a Git
  command within that directory.

This will add the third place where we use read_very_early_config(), adding to the existing calls in tr2_sysenv_load() and ensure_valid_ownership(). If I understand it correctly, that means that every Git execution in a bare repository will now parse the system and global config three times.

This doesn't count the check for uploadpack.packobjectshook in upload-pack.c that uses current_config_scope() to restrict its value to the system and global config.

We are probably at the point where we need to instead create a configset that stores this "protected config" and allow us to lookup config keys directly from that configset instead of iterating through these config files repeatedly.

This could also be done in advance of the config parsing by setting discovery_bare_config = DISCOVERY_BARE_ALWAYS before calling read_very_early_config(). Avoids an if and a comment here, which might be nice.

You return -1 in discovery_bare_cb when the key matches, but the value is not understood. Should we check the return value of read_very_early_config(), too?

In general, I'm not sure these BUG() statements are helpful, but they aren't hurting anything. I wonder if it would be better to use DISCOVERY_BARE_UNKNOWN instead of default, because then the compiler should notice that the switch needs updating when a new enum mode is added.

I think that you can add a comma at the end of this enum to avoid the changed line the next time the enum needs to be expanded.

Won't this fail if someone runs a Git command inside of a .git/ directory for a non-bare repository? I just want to be sure that we hit this error instead:

	fatal: this operation must be run in a work tree

I see that this error is tested in t0008-ignores.sh, but that's with the default "always" value. It would be good to explicitly check that this is the right error when using the "never" config.

Thanks, -Stolee

On Mon, May 16, 2022 at 02:46:55PM -0400, Derrick Stolee wrote:

I am confused, then.

What does a user who has some legitimate (non-embedded) bare repositories do if they are skeptical of other bare repositories? I suspect the best answer we would be able to provide with these patches is "use `--git-dir`".

What happens to a user who has a combination of legitimate bare repositories, embedded bare repositories that they trust, and other embedded bare repositories that they don't?

As far as I can tell, our recommendation with these tools would be to:

  - run `git config --global discovery.bare never`, and
  - include `--git-dir=$(pwd)` in any git invocations in bare
    repositories that they do trust

This gets at my concerns from [1] and [2] (mostly [2], in this case) that we're trying to close the embedded bare repos problem with an overly broad solution, at the expense of usability.

I can't shake the feeling that something like I described towards the bottom of [2] would give you all of the security guarantees you're after without compromising on usability for non-embedded bare repositories.

I'm happy to explore this direction more myself if you don't want to. I would just much rather see us adopt an approach that doesn't break more use-cases than it has to if such a thing can be avoided.

I cannot endorse these patches as-is.

Thanks, Taylor

[1]: https://lore.kernel.org/git/Ylobp7sntKeWTLDX@nand.local/ [2]: https://lore.kernel.org/git/YnmKwLoQCorBnMe2@nand.local/

Thanks for being thorough, I find it really helpful.

For brevity, I won't reply to comments that I think are obviously good, so you can assume I'll incorproate anything that isn't commented on.

Derrick Stolee <derrickstolee@github.com> writes:

Agree, and I think defining the term will be useful in future on-list discussions.

Looking at all of the read_very_early_config() calls,

- check_bare_repo_allowed() can use git_configset_get_string()
- ensure_valid_ownership() can use git_configset_get_value_multi()
- tr2_sysenv_load() reads every value with the "trace2." prefix. AFAICT
  configsets only support exact key lookups and I don't see an easy way
  teach configsets to support prefix lookups.

(I didn't look too closely at uploadpack.packobjectshook because I don't know enough about config scopes to comment.)

So using a configset, we'll still need to read the config files at least twice. That's better than thrice, but it doesn't cover the tr2_sysenv_load() use case, and we'll run into this yet again if add function that reads all config values with a given prefix.

An hacky alternative that covers all of these use cases would be to read all protected config in a single pass, e.g.

  static struct protected_config {
         struct safe_directory_data safe_directory_data;
         const char *discovery_bare;
         struct string_list tr2_sysenv;
  };

  static int protected_config_cb()
  {
    /* Parse EVERYTHING that belongs in protected_config. */
  }

but protected_config_cb() would have to parse too many unrelated things for my liking.

So I'll use the configset for the cases where the key is known, and perhaps we'll punt on tr2_sysenv_load().

This comment doesn't apply because unlike most other config reading functions, read_very_early_config() and read_early_config() die when the callback returns -1.

I'm not sure why this is the case though, and maybe you think there is value in having a non-die()-ing variant, e.g. read_very_early_config_gently()?

Yes, it will fail if run inside of a .git/ directory. "never" prevents you from working from inside .git/ unless you set GIT_DIR.

IIRC, we don't show "fatal: this operation must be run in a work tree" for every Git command, e.g. "git log" works just fine. It makes sense to show this warning when the CWD supports 'some, but not all' Git commands, but I don't think this is valuable if we forbid *all* Git commands.

Instead of trying to make "never" accomodate this use case, perhaps what we want is a "dotgit-only" option that allows a bare repository if it is below a .git/ directory. Since we forbid .git in the index, this seems somewhat safe, but I hadn't proposed this sooner because I don't know if we need it yet, and I'm certain that there are less secure edge cases that need to be thought through.

Glen Choo <chooglen@google.com> writes:

Since I'm trying to replace read_very_early_config() anyway, is this a good time to teach git to respect "-c safe.directory"?

My understanding of [1] is that we only ignore "-c safe.directory" because read_very_early_config() doesn't support it, but we would prefer to support it if we could.

[1] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/

From: Glen Choo <chooglen@google.com>

Add a 'cwd' option to discovery.bareRepository, which allows a bare repository to be used if and only if the cwd is the root of a bare repository. This covers the common case where a user works with a bare repository by cd-ing into the repository's root.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/discovery.txt |  6 ++++--
 setup.c                            | 27 ++++++++++++++++++++-------
 t/t0034-discovery-bare.sh          | 10 ++++++++++
 3 files changed, 34 insertions(+), 9 deletions(-)

On 5/13/2022 7:37 PM, Glen Choo via GitGitGadget wrote:

I don't consider this case valuable. In addition to allowing the most-common use case, it also allows the most-common route that an attacker would use to try to get a user to run a Git command in a malicious embedded bare repo. I think we are better off without it.

Thanks, -Stolee

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Would a middle-ground of giving a warning() message help? Can it be loud and annoying enough to knudge the users to adjust without breaking the functionality?

The longer-term default should be "cwd is allowed, but we do not bother going up from object/04 subdirectory of a bare repository", not "bare repositories should not be usable at all without GIT_DIR".

Missing comma before "i.e."

;-)

Hopefully "git fetch" over ssh:// and file:/// would run the other side with GIT_DIR explicitly set? As long as this recommendation does not break these use cases, I think we are OK, but I do not yet find these "problems, such as..." so convincing.

Junio C Hamano <gitster@pobox.com> writes:

Personally, when my tool changes its behavior, I would strongly prefer it to die than to "change behavior + warn". I'd feel more comfortable knowing that the tool did nothing as opposed to doing the wrong thing and only being informed after the fact. Also, I sometimes ignore warnings ;)

When we _do_ transition away from die(), ignore + warning() sounds like a good first step.

But if any of this flies in the face of the project's conventions, let me know as such.

Thanks.

Ah, I'll check this and get back to you.

What would be a convincing rationale to you? I'll capture that here.

I'm assuming that you already have such an rationale in mind when you say that the longer-term default is that "we respect bare repositories only if they are the cwd.". I'm also assuming that this rationale is something other than embedded bare repos, because "cwd-only" does not protect against that.

Perhaps "never" sounds better to folks who don't ever expect bare repositories and want to lock down the environment. Randall (cc-ed) suggests one such use case in [1].

(To Randall: Oops, I actually meant to cc you earlier, since you were the first to suggest a practical use case for never allowing bare repos. It must've slipped my mind).

[1] https://lore.kernel.org/git/005d01d84ad0$782e8fc0$688baf40$@nexbridge.com.

Glen Choo <chooglen@google.com> writes:

Heh, personally I would try very hard not to change the behaviour without explicitly asked by the users with configuration or command line option. Flipping the default has traditionally been done in two or three phases.

 (1) We start by giving a loud and annoying warning to those who
     haven't configured and tell them the default *will* change, how
     to keep the current behaviour forever, and how to live in the
     future by adopting the future default early.

 (2) After a while, we flip the default.  Those who haven't
     configured are given a notice that the default has changed, how
     to keep the old behaviour forever, and how to explicitly choose
     the same value as the default to squelch the notice.

 (3) After yet another while, we stop giving the notice.  If we
     omitted (2), here is where we flip the default.

Strictly speaking, we can have (1) in one release and then could directly jump to (3), but some distros may skip the releases that has (1), and (2) is an attempt to help users of such distros.

That is a wrong question. You are the one pushing for castrating the bare repositories.

No, I do not have such a "different" rationale to justify the change proposed in this patch. I was saying that the claim "embedded bare repos are risky", backed by your two examples, did not sound all that serious a problem. Presented with a more serious brekage scenario, it may make the description more convincing.

Thanks.

Junio C Hamano <gitster@pobox.com> writes:

Ah, that is very helpful. Thanks. It's pretty clear that I misunderstood what you meant by "giving a warning() message" - the warning() is there to prepare users in advance of the change; we don't actually want the warning() in the long term.

For something as disruptive as discovering bare repos, having all of (1), (2) and (3) sounds appropriate.

Let me clarify in case this wasn't received the way I intended. Earlier in the thread, you mentioned:

  The longer-term default should be "cwd is allowed, but we do not
  bother going up from object/04 subdirectory of a bare repository",
  [...]

which I took to mean "Junio thinks that, by default, Git should stop walking up to find a bare repo, and thinks this is better because of rationale X.", and not, "Junio does not think that the default needs to change, but is just suggesting a better default than Glen's".

If it is the former, then there is obviously some thought process here that is worth sharing.

If it the latter, then I'm in favor of taking Stolee's suggestion to drop "cwd", since nobody else finds it useful enough. (I like the 'simplification' story, but not enough to push "cwd" through, especially since it does quite little security-wise.)

Fair. I'll mull over this.

Glen Choo <chooglen@google.com> writes:

The X is "it would not break existing use case too badly, just to address a 'security' story whose severity is not so clearly expressed".

As long as you'll be there to answer the angry mob that complain loudly (and irritatingly enough, the only do so after a release is made to flip the default), I do not care too much either way ;-).

Thanks.

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

This number is already in use by an in-flight topic, if I am not mistaken. Please make it a habit to always check your topic works well when merged to 'next' and to 'seen'.

Thanks.

On 5/13/2022 7:37 PM, Glen Choo via GitGitGadget wrote:

Please include a full cover letter with each version, so reviewers can respond to the full series goals.

Your series here intends to start protecting against malicious embedded bare repositories by allowing users to opt-in to a more protected state. When the 'discovery.bare' option is set, then Git may die() on a bare repository that is discovered based on the current working directory (these protections are ignored if the user specifies the directory directly through --git-dir or $GIT_DIR).

The 'discovery.bare' option has these values at the end of your series:

* 'always' (default) allows all bare repos, matching the current
  behavior of Git.

* 'never' avoids operating in bare repositories altogether.

* 'cwd' operates in a bare repository only if the current directory
  is exactly the root of the bare repository.

It is important that we keep 'always' as the default at first, because we do not want to introduce a breaking change without warning (at least for an issue like this that has been around for a long time).

The 'never' option is a good one for very security-conscious users who really want to avoid problems. I don't anticipate that users who know about this option and set it themselves are the type that would fall for the social engineering required to attack using this vector, but I can imagine an IT department installing the value in system config across a fleet of machines.

I find the 'cwd' option to not be valuable. It unblocks most existing users, but also almost completely removes the protection that the option was supposed to provide.

I find neither the 'never' or 'cwd' options an acceptable choice for a future default.

I also think that this protection is too rigid: it restricts _all_ bare repositories, not just embedded ones. There is no check to see if the parent directory of the bare repository is inside a non-bare repository.

This leads to what I think would be a valuable replacement for the 'cwd' option:

* 'no-embedded' allows non-embedded bare repositories. An
  _embedded bare repository_ is a bare repository whose parent
  directory is contained in the worktree of a non-bare Git
  repository. When in this mode, embedded bare repositories are
  not allowed unless the parent non-bare Git repository has a
  'safe.embedded' config value storing the path to the current
  embedded bare repository.

That was certainly difficult to write, but here it is as pseudo-code to hopefully remove some doubt as to how this might work:

  if repo is bare:
    if value == "always":
       return ALLOWED
    if value == "never":
       return FORBIDDEN;

    path = get_parent_repo()

    if !path:
       return ALLOWED
    
    if config_file_has_value("{path}/.git/config", "safe.embedded", repo):
       return ALLOWED

    return FORBIDDEN

With this kind of option, we can protect users from these social engineering attacks while providing an opt-in protection for scenarios where embedded bare repos are currently being used (while also not breaking anyone using non-embedded bare repos).

I think Taylor was mentioning something like this in his previous replies, perhaps even to the previous thread on this topic.

This 'no-embedded' option is something that I could see as a potential new default, after it has proven itself in a released version of Git.

There are performance drawbacks to checking the parent path for a Git repo, which is why it is only done when in "no-embedded" mode.

I mentioned some other concerns in your PATCH 1 about how we are now adding the third use of read_very_early_config() and that we should probably refactor that before adding the third option, in order to avoid additional performance costs as well as it being difficult to audit which config options are only checked from these "protected" config files.

Thanks, -Stolee

On Mon, May 16, 2022 at 03:07:35PM -0400, Derrick Stolee wrote:

Thanks for the summary, which I think will be especially helpful to others looking at this series for the very first time.

When I first read this, I disagreed, since presumably that same crowd has legitimate bare repositories that they want to continue being able to operate in without having to pass `--git-dir` or `$GIT_DIR` in.

In fact...

...this resonates quite a bit more with me. "never" isn't a good option unless you aren't a user of bare repositories _and_ don't have any embedded bare repositories (either at all, or any ones that you trust).

This is indeed very similar to a proposal I had made upthread (which you note lower down in this email). One thing that's nice is that we only have to traverse up to the parent repo when in the "no-embedded" mode. That may be slow (since it's unbounded all the way up to the filesystem root or a ceiling directory, whichever we encounter first), but I think it's unavoidable if you need to distinguish between embedded and non-embedded bare repositories.

Yep, see: https://lore.kernel.org/git/Ylobp7sntKeWTLDX@nand.local/.a

I would be totally happy to see "no-embedded" become the default. It might be nice to issue a warning when the top-level config is unset, to give users a heads up about cases that may be broken, perhaps like:

    if repo is bare:
      switch (value) {
      case "always":
        return ALLOWED;
      case "never":
        return FORBIDDEN;
      case "no-embedded": # fallthrough
      case "":
        path = get_parent_repo()
        if !path
          return ALLOWED;

        if config_file_has_value("{path}/.git/config", "safe.embedded", repo)
          return ALLOWED;

        if value == "no-embedded":
          return FORBIDDEN;

        # otherwise, we're in an embedded bare repository with an unset
        # discovery.bare config.
        #
        # warn that this will break in the future...
        warning(_("%s is embedded within %s"), the_repository.path, path);
        advise(_("to allow discovery for this embedded repo, either run"));
        advise(_(""));
        advise(_("  $ git config --global discovery.bare always, or"));
        advise(_("  $ git -C '%s' config --local safe.embedded '%s'"),
               path, relpath(path, the_repository.path));

        # ...but allow the invocation for now until the default is
        # changed.
        return ALLOWED;
      default:
        die(_("unrecognized value of discovery.bare: '%s'"), value);
      }

...where relpath is similar to Go's path/filepath.Rel function.

With an appropriate deprecation period, I think we could even get away from the "continue executing, but don't read config+hooks", which in retrospect is more error-prone and difficult to reason about than I initially had given it credit for.

Thanks, Taylor

Derrick Stolee <derrickstolee@github.com> writes:

Sounds sensible. I wonder how expensive this will be in practice, but the behaviour seems well thought out.

Thanks, I think this has advanced the conversation quite a bit.

Derrick Stolee <derrickstolee@github.com> writes:

My mistake, I should have prepared this summary myself. Thanks again.

Yes.

Yes. Setting the 'never' option in a system config is the use case that motivated this.

Ok, I agree that it provides next-to-no protection. I'll drop it in this series; it's easy enough to reimplement if users really want it anyway.

[...]

I agree, this sounds like a good default that should work for most users.

That said, I don't think I will implement it, and even if I do, it won't be in this series. I have serious doubts that I'd be able to deliver it in a reasonable amount of time (I tried preparing patches to this effect and failed [1]), and 'never' is sufficient for $DAYJOB's current needs.

I would be very happy to see this come to fruition though. I have no objections to anyone preparing patches for this, and I'll gladly review those if that's helpful.

[1] The specific trouble I had was figuring out whether or not the
 'parent' repo was tracking the bare repo, since an untracked bare repo
 in the working tree isn't (in some sense) really "embedded" and it
 can't have come from a remote.

 But maybe the tracking check is unnecessary. We would break a few more
 users without it, but 'safe.embedded' is an easy enough way for a user
 to unbreak themselves.

Makes sense. I'll ask about specifics on that subthread.

Thanks everyone for the feedback. This round has two major changes compared to the last:

 * This feature is now purely motivated by the embedded bare repo attack;
   I've dropped all of language pertaining to 'simplifying' bare repository
   discovery.

 * Protected config gets a formal definition and is implemented using a
   config_set, which should hopefully address most of the performance
   concerns of using read_very_early_config() (see 2/5 for more discussion).
   
   With this new implementation, it's easy to teach protected config to
   include "-c". This round includes some patches (4-5/5) that do this but
   I'm ok to drop them if we don't agree that they are a good idea.

This round is newly rebased onto master because sg/safe-directory-tests-and-docs has been recently integrated and the "-c" patches directly contradict that.

= Description

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

This series implements a simple way of preventing such attacks: create a config option, discovery.bare, that tells Git whether or not to die when it finds a bare repository. discovery.bare has two values:

 * "always": always allow bare repositories (default), identical to current
   behavior
 * "never": never allow bare repositories

and users/system administrators who never expect to work with bare repositories can secure their environments using "never". discovery.bare has no effect if --git-dir or GIT_DIR is passed because we are confident that the user is not confused about which repository is being used.

This series does not change the default behavior, but in the long-run, a "no-embedded" option might be a safe and usable default [2]. "never" is too restrictive and unlikely to be the default.

For security reasons, discovery.bare cannot be read from repository-level config (because we would end up trusting the embedded bare repository that we aren't supposed to trust to begin with). Since this would introduce a 3rd variable that is only read from 'protected/trusted config' (the others are safe.directory and uploadpack.packObjectsHook) this series also defines and creates a shared implementation for 'protected config'

= Patch organization

 * Patches 1-2 define "protected config" and create a shared implementation.
 * Patch 3 introduces discovery.bare.
 * Patches 4-5 expand the definition of "protected config" to include the
   CLI option "-c" [3]. Since this is identical to how
   uploadpack.packObjectsHook currently behaves, it is refactored to a
   "protected config only" variable.

= Series history

Changes in v3:

 * Rebase onto a more recent 'master'
 * Reframe this feature in only in terms of the 'embedded bare repo' attack.
 * Other docs improvements (thanks Stolee in particular!)
 * Protected config no longer uses read_very_early_config() and is only read
   once
 * Protected config now includes "-c"
 * uploadpack.packObjectsHook now uses protected config instead of ignoring
   repo config using config scopes

Changes in v2:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

= Future work

 * This series does not implement the "no-embedded" option [2] and I won't
   work on it any time soon, but I'd be more than happy to review if someone
   sends patches.
 * With discovery.bare, if a builtin is marked RUN_SETUP_GENTLY, setup.c
   doesn't die() and we don't tell users why their repository was rejected,
   e.g. "git config" gives an opaque "fatal: not in a git directory". This
   isn't a new problem though, since safe.directory has the same issue.

[1] https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2] This was first suggested in https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com [3] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/ suggests that safe.directory doesn't need to ignore "-c", it just happened to be implemented that way.

Glen Choo (5):
  Documentation: define protected configuration
  config: read protected config with `git_protected_config()`
  setup.c: create `discovery.bare`
  config: include "-c" in protected config
  upload-pack: make uploadpack.packObjectsHook protected

 Documentation/config.txt            |  8 ++++
 Documentation/config/discovery.txt  | 19 ++++++++
 Documentation/config/safe.txt       | 19 ++++----
 Documentation/config/uploadpack.txt | 22 ++++------
 Documentation/glossary-content.txt  | 18 ++++++++
 config.c                            | 41 +++++++++++++++++
 config.h                            | 17 ++++++++
 repository.c                        |  5 +++
 repository.h                        |  8 ++++
 setup.c                             | 68 ++++++++++++++++++++++++++++-
 t/t0033-safe-directory.sh           | 24 +++++-----
 t/t0035-discovery-bare.sh           | 63 ++++++++++++++++++++++++++
 upload-pack.c                       | 17 +++++---
 13 files changed, 283 insertions(+), 46 deletions(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

base-commit: f9b95943b68b6b8ca5a6072f50a08411c6449b55
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v3
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v3
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v2:

 -:  ----------- > 1:  575676c760d Documentation: define protected configuration
 -:  ----------- > 2:  7499a280961 config: read protected config with `git_protected_config()`
 1:  22b10bf9da8 ! 3:  d5a3e9f9845 setup.c: make bare repo discovery optional
     @@ Metadata
      Author: Glen Choo <chooglen@google.com>
      
       ## Commit message ##
     -    setup.c: make bare repo discovery optional
     +    setup.c: create `discovery.bare`
      
     -    Add a config variable, `discovery.bare`, that tells Git whether or not
     -    it should work with the bare repository it has discovered i.e. Git will
     -    die() if it discovers a bare repository, but it is not allowed by
     -    `discovery.bare`. This only affects repository discovery, thus it has no
     -    effect if discovery was not done (e.g. `--git-dir` was passed).
     +    There is a known social engineering attack that takes advantage of the
     +    fact that a working tree can include an entire bare repository,
     +    including a config file. A user could run a Git command inside the bare
     +    repository thinking that the config file of the 'outer' repository would
     +    be used, but in reality, the bare repository's config file (which is
     +    attacker-controlled) is used, which may result in arbitrary code
     +    execution. See [1] for a fuller description and deeper discussion.
      
     -    This is motivated by the fact that some workflows don't use bare
     -    repositories at all, and users may prefer to opt out of bare repository
     -    discovery altogether:
     +    A simple mitigation is to forbid bare repositories unless specified via
     +    `--git-dir` or `GIT_DIR`. In environments that don't use bare
     +    repositories, this would be minimally disruptive.
      
     -    - An easy assumption for a user to make is that Git commands run
     -      anywhere inside a repository's working tree will use the same
     -      repository. However, if the working tree contains a bare repository
     -      below the root-level (".git" is preferred at the root-level), any
     -      operations inside that bare repository use the bare repository
     -      instead.
     -
     -      In the worst case, attackers can use this confusion to trick users
     -      into running arbitrary code (see [1] for a deeper discussion). But
     -      even in benign situations (e.g. a user renames ".git/" to ".git.old/"
     -      and commits it for archival purposes), disabling bare repository
     -      discovery can be a simpler mode of operation (e.g. because the user
     -      doesn't actually want to use ".git.old/") [2].
     -
     -    - Git won't "accidentally" recognize a directory that wasn't meant to be
     -      a bare repository, but happens to resemble one. While such accidents
     -      are probably very rare in practice, this lets users reduce the chance
     -      to zero.
     +    Create a config variable, `discovery.bare`, that tells Git whether or
     +    not to die() when it discovers a bare repository. This only affects
     +    repository discovery, thus it has no effect if discovery was not
     +    done (e.g. `--git-dir` was passed).
      
          This config is an enum of:
      
     -    - ["always"|(unset)]: always recognize bare repositories (like Git does
     -      today)
     -    - "never": never recognize bare repositories
     +    - "always": always allow bare repositories (this is the default)
     +    - "never": never allow bare repositories
      
     -    More values are expected to be added later, and the default is expected
     -    to change (i.e. to something other than "always").
     +    If we want to protect users from such attacks by default, neither value
     +    will suffice - "always" provides no protection, but "never" is
     +    impractical for bare repository users. A more usable default would be to
     +    allow only non-embedded bare repositories ([2] contains one such
     +    proposal), but detecting if a repository is embedded is potentially
     +    non-trivial, so this work is not implemented in this series.
      
          [1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com
     -    [2]: I don't personally know anyone who does this as part of their
     -    normal workflow, but a cursory search on GitHub suggests that there is a
     -    not insubstantial number of people who munge ".git" in order to store
     -    its contents.
     -
     -    https://github.com/search?l=&o=desc&p=1&q=ref+size%3A%3C1000+filename%3AHEAD&s=indexed&type=Code
     -    (aka search for the text "ref", size:<1000, filename:HEAD)
     +    [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
     -    WIP setup.c: make discovery.bare die on failure
     -
     -    Signed-off-by: Glen Choo <chooglen@google.com>
     + ## Documentation/config.txt ##
     +@@ Documentation/config.txt: include::config/diff.txt[]
     + 
     + include::config/difftool.txt[]
     + 
     ++include::config/discovery.txt[]
     ++
     + include::config/extensions.txt[]
     + 
     + include::config/fastimport.txt[]
      
       ## Documentation/config/discovery.txt (new) ##
      @@
      +discovery.bare::
     -+	Specifies what kinds of directories Git can recognize as a bare
     -+	repository when looking for the repository (aka repository
     -+	discovery). This has no effect if repository discovery is not
     -+	performed e.g. the path to the repository is set via `--git-dir`
     -+	(see linkgit:git[1]).
     ++	'(Protected config only)' Specifies whether Git will work with a
     ++	bare repository that it found during repository discovery. This
     ++	has no effect if the repository is specified directly via the
     ++	--git-dir command-line option or the GIT_DIR environment
     ++	variable (see linkgit:git[1]).
      ++
     -+This config setting is only respected when specified in a system or global
     -+config, not when it is specified in a repository config or via the command
     -+line option `-c discovery.bare=<value>`.
     ++The currently supported values are:
      ++
     -+The currently supported values are `always` (Git always recognizes bare
     -+repositories) and `never` (Git never recognizes bare repositories).
     -+This defaults to `always`, but this default is likely to change.
     ++* `always`: Git always works with bare repositories
     ++* `never`: Git never works with bare repositories
      ++
     -+If your workflow does not rely on bare repositories, it is recommended that
     -+you set this value to `never`. This makes repository discovery easier to
     -+reason about and prevents certain types of security and non-security
     -+problems, such as:
     -+
     -+* `git clone`-ing a repository containing a malicious bare repository
     -+  inside it.
     -+* Git recognizing a directory that isn't meant to be a bare repository,
     -+  but happens to look like one.
     ++This defaults to `always`, but this default may change in the future.
     +++
     ++If you do not use bare repositories in your workflow, then it may be
     ++beneficial to set `discovery.bare` to `never` in your global config.
     ++This will protect you from attacks that involve cloning a repository
     ++that contains a bare repository and running a Git command within that
     ++directory.
      
       ## setup.c ##
      @@
     @@ setup.c: static int ensure_valid_ownership(const char *path)
      +static int check_bare_repo_allowed(void)
      +{
      +	if (discovery_bare_config == DISCOVERY_BARE_UNKNOWN) {
     -+		read_very_early_config(discovery_bare_cb, NULL);
     -+		/* We didn't find a value; use the default. */
     -+		if (discovery_bare_config == DISCOVERY_BARE_UNKNOWN)
     -+			discovery_bare_config = DISCOVERY_BARE_ALWAYS;
     ++		discovery_bare_config = DISCOVERY_BARE_ALWAYS;
     ++		git_protected_config(discovery_bare_cb, NULL);
      +	}
      +	switch (discovery_bare_config) {
      +	case DISCOVERY_BARE_NEVER:
      +		return 0;
      +	case DISCOVERY_BARE_ALWAYS:
      +		return 1;
     -+	default:
     ++	case DISCOVERY_BARE_UNKNOWN:
      +		BUG("invalid discovery_bare_config %d", discovery_bare_config);
      +	}
     ++	return 0;
      +}
      +
      +static const char *discovery_bare_config_to_string(void)
     @@ setup.c: static int ensure_valid_ownership(const char *path)
      +		return "never";
      +	case DISCOVERY_BARE_ALWAYS:
      +		return "always";
     -+	default:
     ++	case DISCOVERY_BARE_UNKNOWN:
      +		BUG("invalid discovery_bare_config %d", discovery_bare_config);
      +	}
     ++	return NULL;
      +}
      +
       enum discovery_result {
     @@ setup.c: enum discovery_result {
       	GIT_DIR_INVALID_GITFILE = -3,
      -	GIT_DIR_INVALID_OWNERSHIP = -4
      +	GIT_DIR_INVALID_OWNERSHIP = -4,
     -+	GIT_DIR_DISALLOWED_BARE = -5
     ++	GIT_DIR_DISALLOWED_BARE = -5,
       };
       
       /*
     @@ setup.c: const char *setup_git_directory_gently(int *nongit_ok)
       		/*
       		 * As a safeguard against setup_git_directory_gently_1 returning
      
     - ## t/t0034-discovery-bare.sh (new) ##
     + ## t/t0035-discovery-bare.sh (new) ##
      @@
      +#!/bin/sh
      +
     @@ t/t0034-discovery-bare.sh (new)
      +
      +pwd="$(pwd)"
      +
     -+expect_allowed () {
     -+	git rev-parse --absolute-git-dir >actual &&
     -+	echo "$pwd/outer-repo/bare-repo" >expected &&
     -+	test_cmp expected actual
     -+}
     -+
      +expect_rejected () {
     -+	test_must_fail git rev-parse --absolute-git-dir 2>err &&
     ++	test_must_fail git rev-parse --git-dir 2>err &&
      +	grep "discovery.bare" err
      +}
      +
     @@ t/t0034-discovery-bare.sh (new)
      +test_expect_success 'discovery.bare unset' '
      +	(
      +		cd outer-repo/bare-repo &&
     -+		expect_allowed &&
     -+		cd refs/ &&
     -+		expect_allowed
     ++		git rev-parse --git-dir
      +	)
      +'
      +
     @@ t/t0034-discovery-bare.sh (new)
      +	git config --global discovery.bare always &&
      +	(
      +		cd outer-repo/bare-repo &&
     -+		expect_allowed &&
     -+		cd refs/ &&
     -+		expect_allowed
     ++		git rev-parse --git-dir
      +	)
      +'
      +
     @@ t/t0034-discovery-bare.sh (new)
      +	git config --global discovery.bare never &&
      +	(
      +		cd outer-repo/bare-repo &&
     -+		expect_rejected &&
     -+		cd refs/ &&
      +		expect_rejected
     -+	) &&
     ++	)
     ++'
     ++
     ++test_expect_success 'discovery.bare in the repository' '
     ++	(
     ++		cd outer-repo/bare-repo &&
     ++		# Temporarily set discovery.bare=always, otherwise git
     ++		# config fails with "fatal: not in a git directory"
     ++		# (like safe.directory)
     ++		git config --global discovery.bare always &&
     ++		git config discovery.bare always &&
     ++		git config --global discovery.bare never &&
     ++		expect_rejected
     ++	)
     ++'
     ++
     ++test_expect_success 'discovery.bare on the command line' '
     ++	git config --global discovery.bare never &&
      +	(
     -+		GIT_DIR=outer-repo/bare-repo &&
     -+		export GIT_DIR &&
     -+		expect_allowed
     ++		cd outer-repo/bare-repo &&
     ++		test_must_fail git -c discovery.bare=always rev-parse --git-dir 2>err &&
     ++		grep "discovery.bare" err
      +	)
      +'
      +
 2:  62070aab7eb < -:  ----------- setup.c: learn discovery.bareRepository=cwd
 -:  ----------- > 4:  66a0a208176 config: include "-c" in protected config
 -:  ----------- > 5:  e25d5907cd1 upload-pack: make uploadpack.packObjectsHook protected

From: Glen Choo <chooglen@google.com>

For security reasons, some config variables are only trusted when they are specified in so-called 'protected configuration' [1]. A future commit will introduce another such config variable, so this is a good time to standardize the documentation and implementation of 'protected configuration'.

Define 'protected configuration' as global and system-level config, and mark `safe.directory` 'Protected config only'. In a future commit, protected configuration will also include "-c".

The following variables are intentionally not marked 'Protected config only':

- `uploadpack.packObjectsHook` has the same security concerns as
  `safe.directory`, but due to a different implementation, it also
  respects the "-c" option.

  When protected configuration includes "-c", `upload.packObjectsHook`
  will be marked 'Protected config only'.

- `trace2.*` happens to read the same config as `safe.directory` because
  they share an implementation. However, this is not for security
  reasons; it is because we want to start tracing so early that
  repository-level config and "-c" are not available [2].

  This requirement is unique to `trace2.*`, so it does not makes sense
  for protected configuration to be subject to the same constraints.

[1] For example, https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/ [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  6 ++++++
 Documentation/config/safe.txt      | 19 ++++++++-----------
 Documentation/glossary-content.txt | 18 ++++++++++++++++++
 3 files changed, 32 insertions(+), 11 deletions(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

What's the SP in "only) '" doing?

The description is a bit unclear what "protected configuration" refers.

If it is the scopes (as in "git config --show-scope") Git can trust more, in other words, a statement like this

    safe.directory is honored only when it comes from a protected
    configuration.

is what you want to make easier to write by introducing a new phrase, perhaps use the word "scope" for more consistency? E.g.

    Only safe.directory that is defined in a trusted scope is
    honored.

I dunno.

It would make sense to give a rationale behind the seemingly arbitrary choice of what is and what is not "protected". Not necessarily in the glossary, but in the proposed log message of the commit that makes the decision. The rationale must help readers to be able to answer the following questions.

 - The system level is "protected" because?  Is it because we do not
   even try to protect ourselves from those who can write anywhere
   in /etc/ or other system directories?

 - The per-user config is "protected" because?  Is it because our
   primary interest in "protection" is to protect individual users
   from landmines laid in the filesystem by other users, and those
   who can already write into $HOME are not we try to guard against?

 - The per-repo config is not "protected" (i.e. "trusted"), because?
   If we are not honoring a configuration in the repository, why are
   we working in that repository in the first place?

 - The per invocation config is not "protected" (i.e. "trusted"),
   because?  If we cannot trusting our own command line, what
   prevents an attacker from mucking with our command line to say
   "sudo whatever" using the same attack vector?

Thanks.

On 5/27/2022 7:29 PM, Junio C Hamano wrote:

I think the answers to these two questions is "yes", so they can be turned into an affirmative sentence:

	We do not event try to protect ourselves from those who can
	write anywhere...

This requires an example:

	Some workflows use repositories stored in shared directories,
	which are writable by multiple unprivileged users.
 

With this argument, I agree that -c config can be considered protected. At the very least, it is visible to the user when they are running a command. This would unify our expectations with uploadPack.packObjectsHook, too.

Thanks, -Stolee

Derrick Stolee <derrickstolee@github.com> writes:

s/event/even/.

Hmph, "... and we do not trust these colleagues"? It might be true, but sounds a bit weak rationale, at least to me. A natural reaction coming form a devil's advocate naïve me would be "well, then I would not be directly interacting with such a repository; I'd work in a clone of it of my own, and pull and push as needed".

Isn't the reason more like "users may go spelunking random places in the filesystem, with PS1 settings and the like that causes some "git" command invoked automatically in their current directory, and we want to protect these users from getting harmed by a random repository with hostile contents in their configuration and hooks without even realizing they have wandered into such a repository"?

Yup, that matches my understanding.

In any case, I'd prefer to see not just the definition but the reasoning behind the decision that made some "protected" while leaving others not-"protected" clearly documented to help users.

Thanks.

Junio C Hamano <gitster@pobox.com> writes:

Hm, this is my understanding as well, i.e. `safe.directory` is meant to protect you from shared repositories that you didn't expect, but it lets you trust the shared repositories that you need (and there is no protection once you decide to trust the repo).

Junio C Hamano <gitster@pobox.com> writes:

Silly typo. Thanks for the catch :)

Good point. I think using scope would be a lot clearer, and maybe I will consider s/protected configuration/protected scope. I'm hesitant to call the scope "trusted", because I don't want to insinuate that repository config is "untrusted" since we _do_ trust it in most cases.

I don't think Documentation/git-config.txt has adequately defined what a 'scope' is though, even though scopes have been with us since 9acc591111 (config: add a notion of "scope", 2016-05-18). The best I could find is "--show-scope", introduced in 145d59f482 (config: add '--show-scope' to print the scope of a config value, 2020-02-10), which mentions scopes but doesn't link the idea back to the specific files or CLI options ("--system", "--global", etc).

So I'll see if I can improve the docs around scopes since that will help the language in this patch.

From: Glen Choo <chooglen@google.com>

Protected config is read using `read_very_early_config()`, which has several downsides:

- Every call to `read_very_early_config()` parses global and
  system-level config files anew, but this can be optimized by just
  parsing them once [1].
- Protected variables should respect "-c" because we can reasonably
  assume that it comes from the user. But, `read_very_early_config()`
  can't use "-c" because it is called so early that it does not have
  access to command line arguments.

Introduce `git_protected_config()`, which reads protected config and caches the values in `the_repository.protected_config`. Then, refactor `safe.directory` to use `git_protected_config()`.

This implementation can still be improved, however:

- `git_protected_config()` iterates through every variable in
  `the_repository.protected_config`, which may still be too expensive to
  be called in every "git" invocation. There exist constant time lookup
  functions for non-protected config (repo_config_get_*()), but for
  simplicity, this commit does not implement similar functions for
  protected config.

- Protected config is stored in `the_repository` so that we don't need
  to statically allocate it. But this might be confusing since protected
  config ignores repository config by definition.

[1] While `git_protected_config()` should save on file I/O, I wasn't able to measure a meaningful difference between that and `read_very_early_config()` on my machine (which has an SSD).

Signed-off-by: Glen Choo <chooglen@google.com>
---
 config.c     | 35 +++++++++++++++++++++++++++++++++++
 config.h     |  8 ++++++++
 repository.c |  5 +++++
 repository.h |  8 ++++++++
 setup.c      |  2 +-
 5 files changed, 57 insertions(+), 1 deletion(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Now we are talking about protected "variable". Is that a synonym for "config", or are there some distinctions between them?

Yes, it indeed is. Is it because we were over-eager when we introduced the "struct repository *repo" parameter to many functions and the configuration system wants you to have some repository, even when you know you are not reading from any repository?

I am wondering if it is a cleaner solution *not* to hang the protected config as a configset in the_repository, but keep the configset as a separate global variable, perhaps static to config.c and is meant to be only accessed via git_protected_config() and the like.

This becomes necessary only because each repository instance has protected_config, even though we need only one instance, no matter how many repositories we are accessing in this single invocation of Git, no?

How should "git config -l" interact with "protected config" and "protected variables", by the way? Should a user be able to tell which ones are coming from protected scope? Should we gain, next to --global, --system, etc., --protected option to list only the protected config/variable?

This is another thing that I find iffy on terminology. Should a random variable, like user.name, be a "protected config", if it is found in $HOME/.gitconfig? If it comes from there, surely we can trust its value, but unlike things like safe.directory, there is no code that wants to enforce that we pay attention only to user.name that came from trusted scopes. Should such a variable be called "protected variable"?

Thanks.

Junio C Hamano <gitster@pobox.com> writes:

Sorry, that's an old term I was toying with (this somehow snuck through my proofreading). I just meant "variable that is only read from protected config", aka a "protected config only variable".

A goal in this version was to introduce as little jargon as possible, so
- "protected config" refers to the set of config sources, and
- "protected config only" refers to config variables/settings that are
  only read from protected config.

Ah no, I was just trying to avoid yet-another global variable (since IIRC we want to move towards a more lib-like Git), and the_repository was a convenient global variable to (ab)use.

I think your suggestion to use a global variable is better, as much as I want to avoid another global variable. Protected config would affect any repositories that we work with in-core, so using a global sounds ok.

environment.c might be a better place since we already make a concerted effort to put global config variables there instead of config.c.

As an aside, I wonder how we could get rid of all of the globals in environment.c in the long term. Maybe we would have yet-another all encompassing global, the_environment, and then figure out which variables belong to the repository and which belong to the environment.

Yes.

I'll have to think about this some more. My initial thoughts are that we should do this if we formalize 'protected' as a scope-like concept, but I don't see the lack of "--protected" as a significant hindrance to users because they can use "--global" and "--system" (albeit in two invocations instead of one).

Ah.. I think it would be best to pretend that the "Protected variable" typo never happened. That term was destined to be confusing and meaningless.

Instead, we can use "protected config" to refer to the config and "protected config only" to refer to variables. Since "protected config" is defined as (global + system + CLI) config, then yes, we would say that it is "protected config". But since we do not enforce that "user.name" _must_ come from only protected config, it is not "protected config only".

Glen Choo <chooglen@google.com> writes:

OK. Let's have such a clear pair of definitions somewhere in the doc or at least in a proposed log message.

If this does not have to be known only inside config.c, until we introduce a more global bag of things, which may have the current the_repository as one of its components, I do not think it hurts to have a file-scope static there. Then, perhaps git_configset_get*() helper functions can recognize cs==NULL as a sign that the caller wants to grab from the "protected config", or something? If we do not want to expose the underying global variable to the public, that is.

I think we are on the same page, we'd probably need something called the_world ;-)

Very clear.  Thanks.

On 5/27/2022 5:09 PM, Glen Choo via GitGitGadget wrote:

I originally thought that we should jump to that "right" solution, but
the existing logic in ensure_valid_ownership() uses the iterator method,
mostly because it uses a multi-valued string. There are helpers that
allow iterating over a specific multi-valued key, but there is no reason
to complicate the current patch with that amount of refactoring. That
can be handled as a completely separate topic.
 

I agree with Junio's suggestion of keeping this as a static global in config.c, accessible only by the public methods from config.h. A future where we have "the_world" might be nice for inventory on all these globals. Definitely not something to hold up this series.

This loads the config from three files, including the xdg_config, which I wasn't thinking about before.

This implementation does not use the -c config yet, which you listed as a downside of read_very_early_config(). I see that you include that in your patch 4, but the commit message for this patch could list that as a step that will be handled by a later change.

(You could also do that as patch 3 and add a test near the existing safe.directory tests instead of waiting for discovery.bare.)

These two methods are clearly correct.

..._check_init() is an OK name. I've seen us use "prepare_...()" in other areas as a way of making sure that we have the proper state (see prepare_packed_git() and the like), so maybe a rename here to match would be worthwhile. Feel free to ignore.

This will have no equivalent when protected_config is left as a static global, but that is fine. It only goes out of scope with the end of the process, anyway.

Nice to have a very simple conversion here.

Thanks, -Stolee

From: Glen Choo <chooglen@google.com>

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

A simple mitigation is to forbid bare repositories unless specified via `--git-dir` or `GIT_DIR`. In environments that don't use bare repositories, this would be minimally disruptive.

Create a config variable, `discovery.bare`, that tells Git whether or not to die() when it discovers a bare repository. This only affects repository discovery, thus it has no effect if discovery was not done (e.g. `--git-dir` was passed).

This config is an enum of:

- "always": always allow bare repositories (this is the default)
- "never": never allow bare repositories

If we want to protect users from such attacks by default, neither value will suffice - "always" provides no protection, but "never" is impractical for bare repository users. A more usable default would be to allow only non-embedded bare repositories ([2] contains one such proposal), but detecting if a repository is embedded is potentially non-trivial, so this work is not implemented in this series.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  2 +
 Documentation/config/discovery.txt | 19 +++++++++
 setup.c                            | 66 +++++++++++++++++++++++++++++-
 t/t0035-discovery-bare.sh          | 64 +++++++++++++++++++++++++++++
 4 files changed, 150 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Can discovery_bare come from anywhere other than config?

I am wondering if both the variable and the type should be called "discovery_bare_allowed" instead. That it comes from the config is not the more important part. That it determines if it is allowed is.

OK, so the thing is initialized to "unknown", and the first time we want to use the value of it, we read from the file (or default to "always"). Makes sense.

And then ...

... this is being defensive; we know discovery_bare_cb() won't give UNKNOWN, but we want to make sure.

But this one feels strangely asymmetrical, as there is no inherent reason why one must be called before the other. I would expect it to either

 * take a parameter of type "enum discovery_bare" and return
   "never", "always", or "unset", without calling any BUG().

or

 * have the same "we lazily figure out the discovery_bare_config
   variable on demand" logic.

As both of these functions are file-scope static, we can live with it, though.

On 5/27/2022 5:09 PM, Glen Choo via GitGitGadget wrote:

Using this static global is fine, I think.

However, I do think that this _cb method could benefit from interpreting the 'd' pointer as a 'enum discovery_bare_config *' and assigning the value at the pointer. We can then pass the global to the git_protected_config() call below.

This is probably over-defensive future-proofing, but this kind of change would be necessary if we ever wanted to return the enum instead of simply an integer, as below:

With the recommended change to the _cb method, we could rewrite this as

static enum discovery_bare_config get_discovery_bare(void)
{
	enum discovery_bare_config result = DISCOVERY_BARE_ALWAYS;
	git_protected_config(discovery_bare_cb, &result);
	return result;
}

With this, we can drop the UNKNOWN and let the caller treat the response as a simple boolean.

I think this is simpler overall, but also makes it easier to extend in the future to have "discovery.bare=non-embedded" by adding a new mode and adjusting the consumer in setup_git_directory_gently_1() to use a switch() on the resurned enum.

This case should be a "default:" in case somehow an arbitrary integer value was placed in the variable. This could also take an enum as a parameter, to avoid being coupled to the global.

Should we make a simple "expect_accepted" helper in case we ever want to replace the "git rev-parse --git-dir" with anything else?

Ok, at the current place in the series, this test_must_fail matches expectation. If you reorder to have this patch after your current patch 4, then we can write this test immediately as a successful case.

We could also reuse some information from the expect_rejected helper by adding this:

expect_rejected () {
	test_must_fail git $* rev-parse --git-dir 2>err &&
	grep "discovery.bare" err
}

Then you can test the -c options in the tests as

	expect_rejected -c discovery.bare=always

Thanks, -Stolee

From: Glen Choo <chooglen@google.com>

Now that protected config includes "-c", "uploadpack.packObjectsHook" behaves identically to a 'Protected config only' variable. Refactor it to use git_protected_config() and mark it 'Protected config only'.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/uploadpack.txt | 22 +++++++++-------------
 upload-pack.c                       | 17 +++++++++++------
 2 files changed, 20 insertions(+), 19 deletions(-)

On 5/27/2022 5:09 PM, Glen Choo via GitGitGadget wrote:

I'm really glad to see this simplification at the end of your series.

...

This is much cleaner.

It's unfortunate that there are two places that need this change. Is it worth adding a static helper that executes these?

static void get_upload_pack_config(void *data)
{
	git_config(upload_pack_config, data);
	git_protected_config(upload_pack_protected_config, data);
}

Thanks, -Stolee

From: Glen Choo <chooglen@google.com>

Protected config should include the command line (aka "-c") because we can be quite certain that this config is specified by the user.

Introduce a function, `git_configset_add_parameters()`, that adds "-c" config to a config_set, and use it to add "-c" to protected config.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  6 +++---
 Documentation/glossary-content.txt |  2 +-
 config.c                           |  6 ++++++
 config.h                           |  9 +++++++++
 t/t0033-safe-directory.sh          | 24 ++++++++++--------------
 t/t0035-discovery-bare.sh          |  3 +--
 6 files changed, 30 insertions(+), 20 deletions(-)

On 5/27/2022 5:09 PM, Glen Choo via GitGitGadget wrote:
  

This one-line method could be inlined into the read_protected_config() method:

	git_config_from_parameters(config_set_callback, the_repository->protected_config);

...would be the way to inline it.

You do make it public here. I wonder if we can think of other consumers of this method that justify the addition to the API.

But this is also a nitpick. I don't feel strongly one way or another. The code definitely works as-is.

Thanks, -Stolee

Thanks again for the kind feedback, everyone :)

The motivation has remained the same as the last round; you can find it in the "Description" section in the previous cover letter [1].

This round doesn't introduce any major code changes. The most notable changes are:

 * I've reorganized the patches so that protected config includes "-c" from
   the beginning (3/5) (instead of trying to avoid changing safe.directory).
   As a result, uploadpack.packObjectsHook becomes the first 'protected
   config only' variable instead of safe.directory.
   
   Since we start the conversion with uploadpack.packObjectHook, I was
   curious whether we might be able to reuse its approach of "reading the
   full set of config, but checking the scope of each value", which might be
   nice because we could reuse the cache in the_repository->config instead
   of creating an entirely new configset. I didn't pursue it further, but
   I've noted this alternative in 3/5's commit message.

 * 'Protected configuration' is now defined as a set of configuration scopes
   (2/5). This follows a suggestion from Junio [2], which I thought read
   very clearly. We haven't done a great job at describing 'scopes' in
   Documentation/git-config.txt though, so I tried to remedy that by
   cleaning up a little and adding a SCOPES section (1/5).
   
   Frankly, I'm not very happy with the end result - it's not nearly as
   clear as I had hoped, and I think I might be introducing some confusion
   between the words "config" and "scope". I'd appreciate any feedback that
   helps us get to a good final wording.

 * I added a test for "git config --show-scope" and the 'worktree' scope,
   since 'worktree' wasn't listed in Documentation/git-config.txt (6/5).

= Patch organization

 * Patch 1 add a section on config scopes to our docs
 * Patches 2-3 define 'protected config' and create a shared implementation.
 * Patch 4 refactors safe.directory to use protected config
 * Patch 5 adds discovery.bare

= Series history

Changes in v4:

 * 2/5's commit message now justifies what scopes are included in protected
   config
 * The global configset is now a file-scope static inside config.c
   (previously it was a member of the_repository).
 * Rename discovery_bare_config to discovery_bare_allowed
 * Make discovery_bare_allowed function-scoped (instead of global).
 * Add an expect_accepted helper to the discovery.bare tests.
 * Add a helper to "upload-pack" that reads the protected and non-protected
   config

Changes in v3:

 * Rebase onto a more recent 'master'
 * Reframe this feature in only in terms of the 'embedded bare repo' attack.
 * Other docs improvements (thanks Stolee in particular!)
 * Protected config no longer uses read_very_early_config() and is only read
   once
 * Protected config now includes "-c"
 * uploadpack.packObjectsHook now uses protected config instead of ignoring
   repo config using config scopes

Changes in v2:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

= Future work

 * This series does not implement the "no-embedded" option [3] and I won't
   work on it any time soon, but I'd be more than happy to review if someone
   sends patches.
 * With discovery.bare, if a builtin is marked RUN_SETUP_GENTLY, setup.c
   doesn't die() and we don't tell users why their repository was rejected,
   e.g. "git config" gives an opaque "fatal: not in a git directory". This
   isn't a new problem though, since safe.directory has the same issue.

[1] https://lore.kernel.org/git/pull.1261.v3.git.git.1653685761.gitgitgadget@gmail.com [2] https://lore.kernel.org/git/xmqqh75a1rmd.fsf@gitster.g [3] This was first suggested in https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Glen Choo (5):
  Documentation/git-config.txt: add SCOPES section
  Documentation: define protected configuration
  config: read protected config with `git_protected_config()`
  safe.directory: use git_protected_config()
  setup.c: create `discovery.bare`

 Documentation/config.txt            |  2 +
 Documentation/config/discovery.txt  | 19 +++++++
 Documentation/config/safe.txt       |  6 +--
 Documentation/config/uploadpack.txt |  6 +--
 Documentation/git-config.txt        | 77 +++++++++++++++++++++++------
 config.c                            | 51 +++++++++++++++++++
 config.h                            | 17 +++++++
 setup.c                             | 59 +++++++++++++++++++++-
 t/t0033-safe-directory.sh           | 24 ++++-----
 t/t0035-discovery-bare.sh           | 68 +++++++++++++++++++++++++
 t/t5544-pack-objects-hook.sh        |  7 ++-
 upload-pack.c                       | 27 ++++++----
 12 files changed, 316 insertions(+), 47 deletions(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

base-commit: f9b95943b68b6b8ca5a6072f50a08411c6449b55
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v4
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v4
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v3:

 1:  575676c760d < -:  ----------- Documentation: define protected configuration
 -:  ----------- > 1:  c0e27ab3b3e Documentation/git-config.txt: add SCOPES section
 5:  e25d5907cd1 ! 2:  a5a1dcb03e1 upload-pack: make uploadpack.packObjectsHook protected
     @@ Metadata
      Author: Glen Choo <chooglen@google.com>
      
       ## Commit message ##
     -    upload-pack: make uploadpack.packObjectsHook protected
     +    Documentation: define protected configuration
      
     -    Now that protected config includes "-c", "uploadpack.packObjectsHook"
     -    behaves identically to a 'Protected config only' variable. Refactor it
     -    to use git_protected_config() and mark it 'Protected config only'.
     +    For security reasons, there are config variables that are only trusted
     +    when they are specified in extra-trustworthy configuration scopes, which
     +    are sometimes referred to on-list as 'protected configuration' [1]. A
     +    future commit will introduce another such variable, so let's define our
     +    terms so that we can have consistent documentation and implementation.
     +
     +    In our documentation, define 'protected config' as the system, global
     +    and command config scopes. As a shorthand, I will refer to variables
     +    that are only respected in protected config as 'protected config only',
     +    but this term is not used in the documentation.
     +
     +    This definition of protected configuration is based on whether or not
     +    Git can reasonably protect the user by ignoring the configuration scope:
     +
     +    - System, global and command line config are considered protected
     +      because an attacker who has control over any of those can do plenty of
     +      harm without Git, so we gain very little by ignoring those scopes.
     +    - On the other hand, local (and similarly, worktree) config are not
     +      considered protected because it is relatively easy for an attacker to
     +      control local config, e.g.:
     +      - On some shared user environments, a non-admin attacker can create a
     +        repository high up the directory hierarchy (e.g. C:\.git on Windows),
     +        and a user may accidentally use it when their PS1 automatically
     +        invokes "git" commands.
     +
     +        `safe.directory` prevents attacks of this form by making sure that
     +        the user intended to use the shared repository. It obviously
     +        shouldn't be read from the repository, because that would end up
     +        trusting the repository that Git was supposed to reject.
     +      - "git upload-pack" is expected to run in repositories that may not be
     +        controlled by the user. We cannot ignore all config in that
     +        repository (because "git upload-pack" would fail), but we can limit
     +        the risks by ignoring `uploadpack.packObjectsHook`.
     +
     +    Only `uploadpack.packObjectsHook` is 'protected config only'. The
     +    following variables are intentionally excluded:
     +
     +    - `safe.directory` should be 'protected config only', but it does not
     +      technically fit the definition because it is not respected in the
     +      "command" scope. A future commit will fix this.
     +
     +    - `trace2.*` happens to read the same scopes as `safe.directory` because
     +      they share an implementation. However, this is not for security
     +      reasons; it is because we want to start tracing so early that
     +      repository-level config and "-c" are not available [2].
     +
     +      This requirement is unique to `trace2.*`, so it does not makes sense
     +      for protected configuration to be subject to the same constraints.
     +
     +    [1] For example,
     +    https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/
     +    [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
       ## Documentation/config/uploadpack.txt ##
     -@@ Documentation/config/uploadpack.txt: uploadpack.keepAlive::
     - 	disables keepalive packets entirely. The default is 5 seconds.
     - 
     - uploadpack.packObjectsHook::
     --	If this option is set, when `upload-pack` would run
     --	`git pack-objects` to create a packfile for a client, it will
     --	run this shell command instead.  The `pack-objects` command and
     --	arguments it _would_ have run (including the `git pack-objects`
     --	at the beginning) are appended to the shell command. The stdin
     --	and stdout of the hook are treated as if `pack-objects` itself
     --	was run. I.e., `upload-pack` will feed input intended for
     --	`pack-objects` to the hook, and expects a completed packfile on
     --	stdout.
     --+
     +@@ Documentation/config/uploadpack.txt: uploadpack.packObjectsHook::
     + 	`pack-objects` to the hook, and expects a completed packfile on
     + 	stdout.
     + +
      -Note that this configuration variable is ignored if it is seen in the
      -repository-level config (this is a safety measure against fetching from
      -untrusted repositories).
     -+	'(Protected config only)' If this option is set, when
     -+	`upload-pack` would run `git pack-objects` to create a packfile
     -+	for a client, it will run this shell command instead. The
     -+	`pack-objects` command and arguments it _would_ have run
     -+	(including the `git pack-objects` at the beginning) are appended
     -+	to the shell command. The stdin and stdout of the hook are
     -+	treated as if `pack-objects` itself was run. I.e., `upload-pack`
     -+	will feed input intended for `pack-objects` to the hook, and
     -+	expects a completed packfile on stdout.
     ++Note that this configuration variable is only respected when it is specified
     ++in protected config (see <<SCOPES>>). This is a safety measure against
     ++fetching from untrusted repositories.
       
       uploadpack.allowFilter::
       	If this option is set, `upload-pack` will support partial
      
     - ## upload-pack.c ##
     -@@ upload-pack.c: static int upload_pack_config(const char *var, const char *value, void *cb_data)
     - 		data->advertise_sid = git_config_bool(var, value);
     - 	}
     + ## Documentation/git-config.txt ##
     +@@ Documentation/git-config.txt: You can change the way options are read/written by specifying the path to a
     + file (`--file`), or by specifying a configuration scope (`--system`,
     + `--global`, `--local`, `--worktree`); see <<OPTIONS>> above.
       
     --	if (current_config_scope() != CONFIG_SCOPE_LOCAL &&
     --	    current_config_scope() != CONFIG_SCOPE_WORKTREE) {
     --		if (!strcmp("uploadpack.packobjectshook", var))
     --			return git_config_string(&data->pack_objects_hook, var, value);
     --	}
     --
     - 	if (parse_object_filter_config(var, value, data) < 0)
     - 		return -1;
     ++[[SCOPES]]
     + SCOPES
     + ------
       
     - 	return parse_hide_refs_config(var, value, "uploadpack");
     - }
     +@@ Documentation/git-config.txt: Most configuration options are respected regardless of the scope it is
     + defined in, but some options are only respected in certain scopes. See the
     + option's documentation for the full details.
       
     -+static int upload_pack_protected_config(const char *var, const char *value, void *cb_data)
     -+{
     -+	struct upload_pack_data *data = cb_data;
     ++Protected config
     ++~~~~~~~~~~~~~~~~
      +
     -+	if (!strcmp("uploadpack.packobjectshook", var))
     -+		return git_config_string(&data->pack_objects_hook, var, value);
     -+	return 0;
     -+}
     ++Protected config refers to the 'system', 'global', and 'command' scopes. Git
     ++considers these scopes to be especially trustworthy because they are likely
     ++to be controlled by the user or a trusted administrator. An attacker who
     ++controls these scopes can do substantial harm without using Git, so it is
     ++assumed that the user's environment protects these scopes against attackers.
      +
     - void upload_pack(const int advertise_refs, const int stateless_rpc,
     - 		 const int timeout)
     - {
     -@@ upload-pack.c: void upload_pack(const int advertise_refs, const int stateless_rpc,
     - 	upload_pack_data_init(&data);
     - 
     - 	git_config(upload_pack_config, &data);
     -+	git_protected_config(upload_pack_protected_config, &data);
     - 
     - 	data.stateless_rpc = stateless_rpc;
     - 	data.timeout = timeout;
     -@@ upload-pack.c: int upload_pack_v2(struct repository *r, struct packet_reader *request)
     - 	data.use_sideband = LARGE_PACKET_MAX;
     - 
     - 	git_config(upload_pack_config, &data);
     -+	git_protected_config(upload_pack_protected_config, &data);
     ++For security reasons, certain options are only respected when they are
     ++specified in protected config, and ignored otherwise.
     ++
     + ENVIRONMENT
     + -----------
       
     - 	while (state != FETCH_DONE) {
     - 		switch (state) {
 2:  7499a280961 ! 3:  94b40907e66 config: read protected config with `git_protected_config()`
     @@ Metadata
       ## Commit message ##
          config: read protected config with `git_protected_config()`
      
     -    Protected config is read using `read_very_early_config()`, which has
     -    several downsides:
     -
     -    - Every call to `read_very_early_config()` parses global and
     -      system-level config files anew, but this can be optimized by just
     -      parsing them once [1].
     -    - Protected variables should respect "-c" because we can reasonably
     -      assume that it comes from the user. But, `read_very_early_config()`
     -      can't use "-c" because it is called so early that it does not have
     -      access to command line arguments.
     -
     -    Introduce `git_protected_config()`, which reads protected config and
     -    caches the values in `the_repository.protected_config`. Then, refactor
     -    `safe.directory` to use `git_protected_config()`.
     -
     -    This implementation can still be improved, however:
     -
     -    - `git_protected_config()` iterates through every variable in
     -      `the_repository.protected_config`, which may still be too expensive to
     -      be called in every "git" invocation. There exist constant time lookup
     -      functions for non-protected config (repo_config_get_*()), but for
     -      simplicity, this commit does not implement similar functions for
     -      protected config.
     -
     -    - Protected config is stored in `the_repository` so that we don't need
     -      to statically allocate it. But this might be confusing since protected
     -      config ignores repository config by definition.
     -
     -    [1] While `git_protected_config()` should save on file I/O, I wasn't
     -    able to measure a meaningful difference between that and
     -    `read_very_early_config()` on my machine (which has an SSD).
     +    `uploadpack.packObjectsHook` is the only 'protected config only'
     +    variable today, but we've noted that `safe.directory` and the upcoming
     +    `discovery.bare` should also be 'protected config only'. So, for
     +    consistency, we'd like to have a single implementation for protected
     +    config.
     +
     +    The primary constraints are:
     +
     +    1. Reading from protected config should be as fast as possible. Nearly
     +       all "git" commands inside a bare repository will read both
     +       `safe.directory` and `discovery.bare`, so we cannot afford to be
     +       slow.
     +
     +    2. Protected config must be readable when the gitdir is not known.
     +       `safe.directory` and `discovery.bare` both affect repository
     +       discovery and the gitdir is not known at that point [1].
     +
     +    The chosen implementation in this commit is to read protected config and
     +    cache the values in a global configset. This is similar to the caching
     +    behavior we get with the_repository->config.
     +
     +    Introduce git_protected_config(), which reads protected config and
     +    caches them in the global configset protected_config. Then, refactor
     +    `uploadpack.packObjectsHook` to use git_protected_config().
     +
     +    The protected config functions are named similarly to their
     +    non-protected counterparts, e.g. git_protected_config_check_init() vs
     +    git_config_check_init().
     +
     +    In light of constraint 1, this implementation can still be improved
     +    since git_protected_config() iterates through every variable in
     +    protected_config, which may still be too expensive. There exist constant
     +    time lookup functions for non-protected config (repo_config_get_*()),
     +    but for simplicity, this commit does not implement similar functions for
     +    protected config.
     +
     +    An alternative that avoids introducing another configset is to continue
     +    to read all config using git_config(), but only accept values that have
     +    the correct config scope [2]. This technically fulfills constraint 2,
     +    because git_config() simply ignores the local and worktree config when
     +    the gitdir is not known. However, this would read incomplete config into
     +    the_repository->config, which would need to be reset when the gitdir is
     +    known and git_config() needs to read the local and worktree config.
     +    Resetting the_repository->config might be reasonable while we only have
     +    these 'protected config only' variables, but it's not clear whether this
     +    extends well to future variables.
     +
     +    [1] In this case, we do have a candidate gitdir though, so with a little
     +    refactoring, it might be possible to provide a gitdir.
     +    [2] This is how `uploadpack.packObjectsHook` was implemented prior to
     +    this commit.
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
       ## config.c ##
     +@@ config.c: static enum config_scope current_parsing_scope;
     + static int pack_compression_seen;
     + static int zlib_compression_seen;
     + 
     ++/*
     ++ * Config that comes from trusted sources, namely:
     ++ * - system config files (e.g. /etc/gitconfig)
     ++ * - global config files (e.g. $HOME/.gitconfig,
     ++ *   $XDG_CONFIG_HOME/git)
     ++ * - the command line.
     ++ *
     ++ * This is declared here for code cleanliness, but unlike the other
     ++ * static variables, this does not hold config parser state.
     ++ */
     ++static struct config_set protected_config;
     ++
     + static int config_file_fgetc(struct config_source *conf)
     + {
     + 	return getc_unlocked(conf->u.file);
     +@@ config.c: int git_configset_add_file(struct config_set *cs, const char *filename)
     + 	return git_config_from_file(config_set_callback, filename, cs);
     + }
     + 
     ++int git_configset_add_parameters(struct config_set *cs)
     ++{
     ++	return git_config_from_parameters(config_set_callback, cs);
     ++}
     ++
     + int git_configset_get_value(struct config_set *cs, const char *key, const char **value)
     + {
     + 	const struct string_list *values = NULL;
      @@ config.c: int repo_config_get_pathname(struct repository *repo,
       	return ret;
       }
       
     -+/* Read protected config into the_repository->protected_config. */
     ++/* Read values into protected_config. */
      +static void read_protected_config(void)
      +{
      +	char *xdg_config = NULL, *user_config = NULL, *system_config = NULL;
      +
     -+	CALLOC_ARRAY(the_repository->protected_config, 1);
     -+	git_configset_init(the_repository->protected_config);
     ++	git_configset_init(&protected_config);
      +
      +	system_config = git_system_config();
      +	git_global_config(&user_config, &xdg_config);
      +
     -+	git_configset_add_file(the_repository->protected_config, system_config);
     -+	git_configset_add_file(the_repository->protected_config, xdg_config);
     -+	git_configset_add_file(the_repository->protected_config, user_config);
     ++	git_configset_add_file(&protected_config, system_config);
     ++	git_configset_add_file(&protected_config, xdg_config);
     ++	git_configset_add_file(&protected_config, user_config);
     ++	git_configset_add_parameters(&protected_config);
      +
      +	free(system_config);
      +	free(xdg_config);
      +	free(user_config);
      +}
      +
     -+/* Ensure that the_repository->protected_config has been initialized. */
     ++/* Ensure that protected_config has been initialized. */
      +static void git_protected_config_check_init(void)
      +{
     -+	if (the_repository->protected_config &&
     -+	    the_repository->protected_config->hash_initialized)
     ++	if (protected_config.hash_initialized)
      +		return;
      +	read_protected_config();
      +}
     @@ config.c: int repo_config_get_pathname(struct repository *repo,
      +void git_protected_config(config_fn_t fn, void *data)
      +{
      +	git_protected_config_check_init();
     -+	configset_iter(the_repository->protected_config, fn, data);
     ++	configset_iter(&protected_config, fn, data);
      +}
      +
       /* Functions used historically to read configuration from 'the_repository' */
     @@ config.c: int repo_config_get_pathname(struct repository *repo,
       {
      
       ## config.h ##
     +@@ config.h: void git_configset_init(struct config_set *cs);
     +  */
     + int git_configset_add_file(struct config_set *cs, const char *filename);
     + 
     ++/**
     ++ * Parses command line options and environment variables, and adds the
     ++ * variable-value pairs to the `config_set`. Returns 0 on success, or -1
     ++ * if there is an error in parsing. The caller decides whether to free
     ++ * the incomplete configset or continue using it when the function
     ++ * returns -1.
     ++ */
     ++int git_configset_add_parameters(struct config_set *cs);
     ++
     + /**
     +  * Finds and returns the value list, sorted in order of increasing priority
     +  * for the configuration variable `key` and config set `cs`. When the
      @@ config.h: int repo_config_get_maybe_bool(struct repository *repo,
       int repo_config_get_pathname(struct repository *repo,
       			     const char *key, const char **dest);
     @@ config.h: int repo_config_get_maybe_bool(struct repository *repo,
        * Querying For Specific Variables
        * -------------------------------
      
     - ## repository.c ##
     -@@ repository.c: void repo_clear(struct repository *repo)
     - 		FREE_AND_NULL(repo->remote_state);
     - 	}
     - 
     -+	if (repo->protected_config) {
     -+		git_configset_clear(repo->protected_config);
     -+		FREE_AND_NULL(repo->protected_config);
     -+	}
     + ## t/t5544-pack-objects-hook.sh ##
     +@@ t/t5544-pack-objects-hook.sh: test_expect_success 'hook does not run from repo config' '
     + 	! grep "hook running" stderr &&
     + 	test_path_is_missing .git/hook.args &&
     + 	test_path_is_missing .git/hook.stdin &&
     +-	test_path_is_missing .git/hook.stdout
     ++	test_path_is_missing .git/hook.stdout &&
      +
     - 	repo_clear_path_cache(&repo->cached_paths);
     - }
     ++	# check that global config is used instead
     ++	test_config_global uploadpack.packObjectsHook ./hook &&
     ++	git clone --no-local . dst2.git 2>stderr &&
     ++	grep "hook running" stderr
     + '
       
     + test_expect_success 'hook works with partial clone' '
      
     - ## repository.h ##
     -@@ repository.h: struct repository {
     + ## upload-pack.c ##
     +@@ upload-pack.c: static int upload_pack_config(const char *var, const char *value, void *cb_data)
     + 		data->advertise_sid = git_config_bool(var, value);
     + 	}
       
     - 	struct repo_settings settings;
     +-	if (current_config_scope() != CONFIG_SCOPE_LOCAL &&
     +-	    current_config_scope() != CONFIG_SCOPE_WORKTREE) {
     +-		if (!strcmp("uploadpack.packobjectshook", var))
     +-			return git_config_string(&data->pack_objects_hook, var, value);
     +-	}
     +-
     + 	if (parse_object_filter_config(var, value, data) < 0)
     + 		return -1;
     + 
     + 	return parse_hide_refs_config(var, value, "uploadpack");
     + }
       
     -+	/*
     -+	 * Config that comes from trusted sources, namely
     -+	 * - system config files (e.g. /etc/gitconfig)
     -+	 * - global config files (e.g. $HOME/.gitconfig,
     -+	 *   $XDG_CONFIG_HOME/git)
     -+	 */
     -+	struct config_set *protected_config;
     ++static int upload_pack_protected_config(const char *var, const char *value, void *cb_data)
     ++{
     ++	struct upload_pack_data *data = cb_data;
     ++
     ++	if (!strcmp("uploadpack.packobjectshook", var))
     ++		return git_config_string(&data->pack_objects_hook, var, value);
     ++	return 0;
     ++}
     ++
     ++static void get_upload_pack_config(struct upload_pack_data *data)
     ++{
     ++	git_config(upload_pack_config, data);
     ++	git_protected_config(upload_pack_protected_config, data);
     ++}
      +
     - 	/* Subsystems */
     - 	/*
     - 	 * Repository's config which contains key-value pairs from the usual
     -
     - ## setup.c ##
     -@@ setup.c: static int ensure_valid_ownership(const char *path)
     - 	    is_path_owned_by_current_user(path))
     - 		return 1;
     + void upload_pack(const int advertise_refs, const int stateless_rpc,
     + 		 const int timeout)
     + {
     +@@ upload-pack.c: void upload_pack(const int advertise_refs, const int stateless_rpc,
     + 	struct upload_pack_data data;
       
     --	read_very_early_config(safe_directory_cb, &data);
     -+	git_protected_config(safe_directory_cb, &data);
     + 	upload_pack_data_init(&data);
     +-
     +-	git_config(upload_pack_config, &data);
     ++	get_upload_pack_config(&data);
       
     - 	return data.is_safe;
     - }
     + 	data.stateless_rpc = stateless_rpc;
     + 	data.timeout = timeout;
     +@@ upload-pack.c: int upload_pack_v2(struct repository *r, struct packet_reader *request)
     + 
     + 	upload_pack_data_init(&data);
     + 	data.use_sideband = LARGE_PACKET_MAX;
     +-
     +-	git_config(upload_pack_config, &data);
     ++	get_upload_pack_config(&data);
     + 
     + 	while (state != FETCH_DONE) {
     + 		switch (state) {
 4:  66a0a208176 ! 4:  156817966fa config: include "-c" in protected config
     @@ Metadata
      Author: Glen Choo <chooglen@google.com>
      
       ## Commit message ##
     -    config: include "-c" in protected config
     +    safe.directory: use git_protected_config()
      
     -    Protected config should include the command line (aka "-c") because we
     -    can be quite certain that this config is specified by the user.
     -
     -    Introduce a function, `git_configset_add_parameters()`, that adds "-c"
     -    config to a config_set, and use it to add "-c" to protected config.
     +    Use git_protected_config() to read `safe.directory` instead of
     +    read_very_early_config(), making it 'protected config only'. As a
     +    result, `safe.directory` now respects "-c", so update the tests and docs
     +    accordingly.
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
     - ## Documentation/config.txt ##
     -@@ Documentation/config.txt: names do not conflict with those that are used by Git itself and
     - other popular tools, and describe them in your documentation.
     - 
     - Variables marked with '(Protected config only)' are only respected when
     --they are specified in protected configuration. This includes global and
     --system-level config, and excludes repository config, the command line
     --option `-c`, and environment variables. For more details, see the
     -+they are specified in protected configuration. This includes global,
     -+system-level config, the command line option `-c`, and environment
     -+variables, and excludes repository config. For more details, see the
     - 'protected configuration' entry in linkgit:gitglossary[7].
     - 
     - include::config/advice.txt[]
     -
     - ## Documentation/glossary-content.txt ##
     -@@ Documentation/glossary-content.txt: Protected configuration includes:
     - - system-level config, e.g. `/etc/git/config`
     - - global config, e.g. `$XDG_CONFIG_HOME/git/config` and
     -   `$HOME/.gitconfig`
     -+- the command line option `-c` and its equivalent environment variables
     + ## Documentation/config/safe.txt ##
     +@@ Documentation/config/safe.txt: via `git config --add`. To reset the list of safe directories (e.g. to
     + override any such directories specified in the system config), add a
     + `safe.directory` entry with an empty value.
       +
     - Protected configuration excludes:
     +-This config setting is only respected when specified in a system or global
     +-config, not when it is specified in a repository config, via the command
     +-line option `-c safe.directory=<path>`, or in environment variables.
     ++This config setting is only respected in protected configuration (see
     ++<<SCOPES>>). This prevents the untrusted repository from tampering with this
     ++value.
       +
     - - repository config, e.g. `$GIT_DIR/config` and
     -   `$GIT_DIR/config.worktree`
     --- the command line option `-c` and its equivalent environment variables
     - 
     - [[def_reachable]]reachable::
     - 	All of the ancestors of a given <<def_commit,commit>> are said to be
     + The value of this setting is interpolated, i.e. `~/<path>` expands to a
     + path relative to the home directory and `%(prefix)/<path>` expands to a
      
     - ## config.c ##
     -@@ config.c: int git_configset_add_file(struct config_set *cs, const char *filename)
     - 	return git_config_from_file(config_set_callback, filename, cs);
     - }
     + ## setup.c ##
     +@@ setup.c: static int ensure_valid_ownership(const char *path)
     + 	    is_path_owned_by_current_user(path))
     + 		return 1;
       
     -+int git_configset_add_parameters(struct config_set *cs)
     -+{
     -+	return git_config_from_parameters(config_set_callback, cs);
     -+}
     -+
     - int git_configset_get_value(struct config_set *cs, const char *key, const char **value)
     - {
     - 	const struct string_list *values = NULL;
     -@@ config.c: static void read_protected_config(void)
     - 	git_configset_add_file(the_repository->protected_config, system_config);
     - 	git_configset_add_file(the_repository->protected_config, xdg_config);
     - 	git_configset_add_file(the_repository->protected_config, user_config);
     -+	git_configset_add_parameters(the_repository->protected_config);
     +-	read_very_early_config(safe_directory_cb, &data);
     ++	git_protected_config(safe_directory_cb, &data);
       
     - 	free(system_config);
     - 	free(xdg_config);
     -
     - ## config.h ##
     -@@ config.h: void git_configset_init(struct config_set *cs);
     -  */
     - int git_configset_add_file(struct config_set *cs, const char *filename);
     - 
     -+/**
     -+ * Parses command line options and environment variables, and adds the
     -+ * variable-value pairs to the `config_set`. Returns 0 on success, or -1
     -+ * if there is an error in parsing. The caller decides whether to free
     -+ * the incomplete configset or continue using it when the function
     -+ * returns -1.
     -+ */
     -+int git_configset_add_parameters(struct config_set *cs);
     -+
     - /**
     -  * Finds and returns the value list, sorted in order of increasing priority
     -  * for the configuration variable `key` and config set `cs`. When the
     + 	return data.is_safe;
     + }
      
       ## t/t0033-safe-directory.sh ##
      @@ t/t0033-safe-directory.sh: test_expect_success 'safe.directory is not set' '
     @@ t/t0033-safe-directory.sh: test_expect_success 'safe.directory is not set' '
       '
       
       test_expect_success 'ignoring safe.directory in repo config' '
     -
     - ## t/t0035-discovery-bare.sh ##
     -@@ t/t0035-discovery-bare.sh: test_expect_success 'discovery.bare on the command line' '
     - 	git config --global discovery.bare never &&
     - 	(
     - 		cd outer-repo/bare-repo &&
     --		test_must_fail git -c discovery.bare=always rev-parse --git-dir 2>err &&
     --		grep "discovery.bare" err
     -+		git -c discovery.bare=always rev-parse --git-dir
     - 	)
     - '
     - 
 3:  d5a3e9f9845 ! 5:  29053d029f8 setup.c: create `discovery.bare`
     @@ setup.c
       static int inside_git_dir = -1;
       static int inside_work_tree = -1;
       static int work_tree_config_is_bogus;
     -+enum discovery_bare_config {
     -+	DISCOVERY_BARE_UNKNOWN = -1,
     ++enum discovery_bare_allowed {
      +	DISCOVERY_BARE_NEVER = 0,
      +	DISCOVERY_BARE_ALWAYS,
      +};
     -+static enum discovery_bare_config discovery_bare_config =
     -+	DISCOVERY_BARE_UNKNOWN;
       
       static struct startup_info the_startup_info;
       struct startup_info *startup_info = &the_startup_info;
     @@ setup.c: static int ensure_valid_ownership(const char *path)
       
      +static int discovery_bare_cb(const char *key, const char *value, void *d)
      +{
     ++	enum discovery_bare_allowed *discovery_bare_allowed = d;
     ++
      +	if (strcmp(key, "discovery.bare"))
      +		return 0;
      +
      +	if (!strcmp(value, "never")) {
     -+		discovery_bare_config = DISCOVERY_BARE_NEVER;
     ++		*discovery_bare_allowed = DISCOVERY_BARE_NEVER;
      +		return 0;
      +	}
      +	if (!strcmp(value, "always")) {
     -+		discovery_bare_config = DISCOVERY_BARE_ALWAYS;
     ++		*discovery_bare_allowed = DISCOVERY_BARE_ALWAYS;
      +		return 0;
      +	}
      +	return -1;
      +}
      +
     -+static int check_bare_repo_allowed(void)
     ++static enum discovery_bare_allowed get_discovery_bare(void)
      +{
     -+	if (discovery_bare_config == DISCOVERY_BARE_UNKNOWN) {
     -+		discovery_bare_config = DISCOVERY_BARE_ALWAYS;
     -+		git_protected_config(discovery_bare_cb, NULL);
     -+	}
     -+	switch (discovery_bare_config) {
     -+	case DISCOVERY_BARE_NEVER:
     -+		return 0;
     -+	case DISCOVERY_BARE_ALWAYS:
     -+		return 1;
     -+	case DISCOVERY_BARE_UNKNOWN:
     -+		BUG("invalid discovery_bare_config %d", discovery_bare_config);
     -+	}
     -+	return 0;
     ++	enum discovery_bare_allowed result = DISCOVERY_BARE_ALWAYS;
     ++	git_protected_config(discovery_bare_cb, &result);
     ++	return result;
      +}
      +
     -+static const char *discovery_bare_config_to_string(void)
     ++static const char *discovery_bare_allowed_to_string(
     ++	enum discovery_bare_allowed discovery_bare_allowed)
      +{
     -+	switch (discovery_bare_config) {
     ++	switch (discovery_bare_allowed) {
      +	case DISCOVERY_BARE_NEVER:
      +		return "never";
      +	case DISCOVERY_BARE_ALWAYS:
      +		return "always";
     -+	case DISCOVERY_BARE_UNKNOWN:
     -+		BUG("invalid discovery_bare_config %d", discovery_bare_config);
     ++	default:
     ++		BUG("invalid discovery_bare_allowed %d",
     ++		    discovery_bare_allowed);
      +	}
      +	return NULL;
      +}
     @@ setup.c: static enum discovery_result setup_git_directory_gently_1(struct strbuf
       		}
       
       		if (is_git_directory(dir->buf)) {
     -+			if (!check_bare_repo_allowed())
     ++			if (!get_discovery_bare())
      +				return GIT_DIR_DISALLOWED_BARE;
       			if (!ensure_valid_ownership(dir->buf))
       				return GIT_DIR_INVALID_OWNERSHIP;
     @@ setup.c: const char *setup_git_directory_gently(int *nongit_ok)
      +		if (!nongit_ok) {
      +			die(_("cannot use bare repository '%s' (discovery.bare is '%s')"),
      +			    dir.buf,
     -+			    discovery_bare_config_to_string());
     ++			    discovery_bare_allowed_to_string(get_discovery_bare()));
      +		}
      +		*nongit_ok = 1;
      +		break;
     @@ t/t0035-discovery-bare.sh (new)
      +
      +pwd="$(pwd)"
      +
     ++expect_accepted () {
     ++	git "$@" rev-parse --git-dir
     ++}
     ++
      +expect_rejected () {
     -+	test_must_fail git rev-parse --git-dir 2>err &&
     ++	test_must_fail git "$@" rev-parse --git-dir 2>err &&
      +	grep "discovery.bare" err
      +}
      +
     @@ t/t0035-discovery-bare.sh (new)
      +test_expect_success 'discovery.bare unset' '
      +	(
      +		cd outer-repo/bare-repo &&
     -+		git rev-parse --git-dir
     ++		expect_accepted
      +	)
      +'
      +
     @@ t/t0035-discovery-bare.sh (new)
      +	git config --global discovery.bare always &&
      +	(
      +		cd outer-repo/bare-repo &&
     -+		git rev-parse --git-dir
     ++		expect_accepted
      +	)
      +'
      +
     @@ t/t0035-discovery-bare.sh (new)
      +	git config --global discovery.bare never &&
      +	(
      +		cd outer-repo/bare-repo &&
     -+		test_must_fail git -c discovery.bare=always rev-parse --git-dir 2>err &&
     -+		grep "discovery.bare" err
     ++		expect_accepted -c discovery.bare=always &&
     ++		expect_rejected -c discovery.bare=
      +	)
      +'
      +

From: Glen Choo <chooglen@google.com>

Use git_protected_config() to read `safe.directory` instead of read_very_early_config(), making it 'protected config only'. As a result, `safe.directory` now respects "-c", so update the tests and docs accordingly.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/safe.txt |  6 +++---
 setup.c                       |  2 +-
 t/t0033-safe-directory.sh     | 24 ++++++++++--------------
 3 files changed, 14 insertions(+), 18 deletions(-)

From: Glen Choo <chooglen@google.com>

`uploadpack.packObjectsHook` is the only 'protected config only' variable today, but we've noted that `safe.directory` and the upcoming `discovery.bare` should also be 'protected config only'. So, for consistency, we'd like to have a single implementation for protected config.

The primary constraints are:

1. Reading from protected config should be as fast as possible. Nearly
   all "git" commands inside a bare repository will read both
   `safe.directory` and `discovery.bare`, so we cannot afford to be
   slow.

2. Protected config must be readable when the gitdir is not known.
   `safe.directory` and `discovery.bare` both affect repository
   discovery and the gitdir is not known at that point [1].

The chosen implementation in this commit is to read protected config and cache the values in a global configset. This is similar to the caching behavior we get with the_repository->config.

Introduce git_protected_config(), which reads protected config and caches them in the global configset protected_config. Then, refactor `uploadpack.packObjectsHook` to use git_protected_config().

The protected config functions are named similarly to their non-protected counterparts, e.g. git_protected_config_check_init() vs git_config_check_init().

In light of constraint 1, this implementation can still be improved since git_protected_config() iterates through every variable in protected_config, which may still be too expensive. There exist constant time lookup functions for non-protected config (repo_config_get_*()), but for simplicity, this commit does not implement similar functions for protected config.

An alternative that avoids introducing another configset is to continue to read all config using git_config(), but only accept values that have the correct config scope [2]. This technically fulfills constraint 2, because git_config() simply ignores the local and worktree config when the gitdir is not known. However, this would read incomplete config into the_repository->config, which would need to be reset when the gitdir is known and git_config() needs to read the local and worktree config. Resetting the_repository->config might be reasonable while we only have these 'protected config only' variables, but it's not clear whether this extends well to future variables.

[1] In this case, we do have a candidate gitdir though, so with a little refactoring, it might be possible to provide a gitdir. [2] This is how `uploadpack.packObjectsHook` was implemented prior to this commit.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 config.c                     | 51 ++++++++++++++++++++++++++++++++++++
 config.h                     | 17 ++++++++++++
 t/t5544-pack-objects-hook.sh |  7 ++++-
 upload-pack.c                | 27 ++++++++++++-------
 4 files changed, 91 insertions(+), 11 deletions(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

The lossage of this block is because this general git_config() callback routine that is used to read from any scope is no longer used to pick up the sensitive variable. Instead, we need to get it with a different API, namely, git_protected_config().

It is probably is good that in the new code we are not encouraging folks to write random comparisons on current_config_scope(), and instead uniformly use a git_protected_config(). That may promote consistency.

An obvious alternative to achieve the same consistency would be to introduce a helper, and rewrite (instead of removing) the above part like so:

	if (in_protected_scope()) {
		... parse sensitive variable ...
	}

We would not need any other change to this file in this patch if we go that route, I suspect.

Where we used to just do git_config(upload_pack_config), we now need to do a separate git_protected_config(). It feels a bit wasteful to iterate over the same configset twice, but it is not like we are doing the IO and text file parsing multiple times. This looks quite straight-forward.

Junio C Hamano <gitster@pobox.com> writes:

Yes, and as noted in the commit message, this approach seems to work for `safe.directory` and `discovery.bare` too.

Yeah it's not optimal, but at the very least, I think it's easy enough to understand that we could replace it with something more economical in the future.

From: Glen Choo <chooglen@google.com>

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

A simple mitigation is to forbid bare repositories unless specified via `--git-dir` or `GIT_DIR`. In environments that don't use bare repositories, this would be minimally disruptive.

Create a config variable, `discovery.bare`, that tells Git whether or not to die() when it discovers a bare repository. This only affects repository discovery, thus it has no effect if discovery was not done (e.g. `--git-dir` was passed).

This config is an enum of:

- "always": always allow bare repositories (this is the default)
- "never": never allow bare repositories

If we want to protect users from such attacks by default, neither value will suffice - "always" provides no protection, but "never" is impractical for bare repository users. A more usable default would be to allow only non-embedded bare repositories ([2] contains one such proposal), but detecting if a repository is embedded is potentially non-trivial, so this work is not implemented in this series.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  2 +
 Documentation/config/discovery.txt | 19 +++++++++
 setup.c                            | 57 ++++++++++++++++++++++++-
 t/t0035-discovery-bare.sh          | 68 ++++++++++++++++++++++++++++++
 4 files changed, 145 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Ugh, I forgot to update the docs for `discovery.bare`. This should be reworded to be consistent with `safe.directory` and `uploadpack.packObjectsHook`, e.g.

   discovery.bare::
   	Specifies whether Git will work with a bare repository that it found
   	during repository discovery. This has no effect if the repository is
   	specified directly via the --git-dir command-line option or the
   	GIT_DIR environment variable (see linkgit:git[1]).

    This config setting is only respected in protected configuration
    (see <<SCOPES>>). This prevents the untrusted repository from
    tampering with this value.

From: Glen Choo <chooglen@google.com>

For security reasons, there are config variables that are only trusted when they are specified in extra-trustworthy configuration scopes, which are sometimes referred to on-list as 'protected configuration' [1]. A future commit will introduce another such variable, so let's define our terms so that we can have consistent documentation and implementation.

In our documentation, define 'protected config' as the system, global and command config scopes. As a shorthand, I will refer to variables that are only respected in protected config as 'protected config only', but this term is not used in the documentation.

This definition of protected configuration is based on whether or not Git can reasonably protect the user by ignoring the configuration scope:

- System, global and command line config are considered protected
  because an attacker who has control over any of those can do plenty of
  harm without Git, so we gain very little by ignoring those scopes.
- On the other hand, local (and similarly, worktree) config are not
  considered protected because it is relatively easy for an attacker to
  control local config, e.g.:
  - On some shared user environments, a non-admin attacker can create a
    repository high up the directory hierarchy (e.g. C:\.git on Windows),
    and a user may accidentally use it when their PS1 automatically
    invokes "git" commands.

    `safe.directory` prevents attacks of this form by making sure that
    the user intended to use the shared repository. It obviously
    shouldn't be read from the repository, because that would end up
    trusting the repository that Git was supposed to reject.
  - "git upload-pack" is expected to run in repositories that may not be
    controlled by the user. We cannot ignore all config in that
    repository (because "git upload-pack" would fail), but we can limit
    the risks by ignoring `uploadpack.packObjectsHook`.

Only `uploadpack.packObjectsHook` is 'protected config only'. The following variables are intentionally excluded:

- `safe.directory` should be 'protected config only', but it does not
  technically fit the definition because it is not respected in the
  "command" scope. A future commit will fix this.

- `trace2.*` happens to read the same scopes as `safe.directory` because
  they share an implementation. However, this is not for security
  reasons; it is because we want to start tracing so early that
  repository-level config and "-c" are not available [2].

  This requirement is unique to `trace2.*`, so it does not makes sense
  for protected configuration to be subject to the same constraints.

[1] For example, https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/ [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/uploadpack.txt |  6 +++---
 Documentation/git-config.txt        | 13 +++++++++++++
 2 files changed, 16 insertions(+), 3 deletions(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Probably better to delete "extra-trustworthy", or at least "extra-" - it's better to explain why and how they're trustworthy, which you have already done in the commit message.

[snip]

Also "especially trustworthy" here.

Jonathan Tan <jonathantanmy@google.com> writes:

Hm, do you find it superfluous, misleading or something else entirely?

The use of "extra-" was quite intentional. I'm afraid that if we describe protected config as "trustworthy", we insinuate that local/worktree config is "untrustworthy" (but of course this isn't always true, Git usually uses repo config.)

From: Glen Choo <chooglen@google.com>

In a subsequent commit, we will introduce "protected config", which is easiest to describe in terms of configuration scopes (i.e. it's the union of the 'system', 'global', and 'command' scopes). This description is fine for ML discussions, but it's inadequate for end users because we don't provide a good description of "config scopes" in the public docs.

145d59f482 (config: add '--show-scope' to print the scope of a config value, 2020-02-10) introduced the word "scope" to our public docs, but that only enumerates the scopes and assumes the user can figure out those values mean.

Add a SCOPES section to Documentation/git-config.txt that describes the config scopes, their corresponding CLI options, and mentions that some configuration options are only respected in certain scopes. Then, use the word "scope" to simplify the FILES section and change some confusing wording.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/git-config.txt | 64 ++++++++++++++++++++++++++++--------
 1 file changed, 50 insertions(+), 14 deletions(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Forgot to mention when I was sending my comments on patch 2: we should standardize on "protected config" and not use "protected configuration" anywhere.

Thanks - I think this is a nice feature to have. Everything looks good except for some minor comments on text in patch 2, which I have sent.

One alternative design would have been to have separate configsets for protected config and non-protected config (or even better, separate configsets for trace2 config, protected config minus trace2 config, and non-protected config) but that doesn't have to block the submission of this patch set.

Jonathan Tan <jonathantanmy@google.com> writes:

Makes sense.

I suppose that the idea behind this is that we only parse and store each config file exactly once. It's a good goal, but the whole point of the configset is that we can query a single struct to figure out the value of a config variable. Having multiple configsets starts to shift more of the burden to the callers because they now have to query multiple configsets to find their desired config value, and we already start to see some of this unpleasantness in this series.

An alternative that I'd been thinking about is to make a few changes to the git_config_* + configset API to allow us to use a single configset for all of our needs:

1. Keep track of what config we've read when reading into
   the_repository->config, i.e. instead of a boolean "all config has
   been [un]read", we can express "system and global config has been
   read, but not local or command config". Then, use this information to
   load config from sources as they become available. This will allow us
   to read incomplete config for trace2 and setup.c (discovery.bare and
   safe.directory), and only read what we need later on.

   This assumes that when Git reads config, that config is always valid
   later on. So this is broken if, e.g. we read global config file A
   during setup, but when we discover the repo, we discard A and read
   global config file B instead. I don't know if we do this or if we are
   planning to in the future.

2. Add an additional argument that specifies what scopes to respect when
   reading config (maybe as a set of flags). This gives us extra
   specificity when using the git_config*() functions, so we could get
   rid of git_protected_config() like so:

    /* Change enum config_scope into flags first... */

    #define WIP_SCOPES_PROTECTED = CONFIG_SCOPE_SYSTEM & \
      CONFIG_SCOPE_GLOBAL & CONFIG_SCOPE_COMMAND

    static enum discovery_bare_allowed get_discovery_bare(void)
    {
      enum discovery_bare_allowed result = DISCOVERY_BARE_ALWAYS;
      git_config(discovery_bare_cb, &result, WIP_SCOPES_PROTECTED);
      return result;
    }

   And as an added bonus, this gives us an easy way to implement the
   constant time git_config_*() functions for protected config. We could
   even do this without doing 1. first. I haven't looked into whether
   we could turn the enum into flags, but otherwise, I think this is
   pretty feasible.

Glen Choo <chooglen@google.com> writes:

Using a single word consistently does make sense, but why favor a non-word over a proper word ;-)?

Yes, I was worried about this, too. "parse and store exactly once" may merely be a performance thing, but it still matters, even though it is not worse than making duplicate callbacks to overwrite globals that have been already set earlier, which will affect correctness ;-)

That is not a bad direction to go, but are we sure that we always read in the right order (and there is one single right order) and stop at the right step?

config.c::do_git_config_sequence() reads the system and then the global before the local, the worktree, and the command line. We would allow the values of "protected" configuration variables to be inspected by stopping after the first two and inspecting the result before the local and the rest overrides them, but will we need *only* that kind of partial configuration reading that stops exactly there? Even with the proposed "protected" scheme, I thought we plan to honor the command line ones, so we may need to read system+global+command without reading anything else to grab the values only from the protected sources (ah, I like the application of the adjective "protected" to the source, not variables, because that is what we are really talking about---alternatively we could call it "safe"). But if we later read local and worktree ones lazily, unless we _insert_ them before what we read from the command line, we'll break the last-one-wins property, so we need to be careful. I guess each configuration value in the configset knows where it came from, so it probably is possible to insert the ones you read lazily later in the right spot.

Alternatively, we could make the callback aware of the scope for each var-value it is called and have it filter, but that would be a bigger surgery.

I think a new iterator git_config_in_scope(), instead of updating git_config(), would make sense. By definition, all existing git_config() callers do not need the scope specifiers, and "protected" may be the first one but will not be the last one that needs to read from particular scopes.

Junio C Hamano <gitster@pobox.com> writes:

Hm, I guess there's an argument that "config" is a term of art that specifically refers to things from "git config". From that lens, it's much less confusing to see the CONFIGURATION section in Documentation/git-config.txt. But the argument is a little flimsy because I don't think that's something we've stuck to anywhere.

I'll use "configuration" if it's not too unwieldy.

Exactly.

Yeah, last-one-wins makes this a lot trickier. I thought that it would be nice to have insert-with-priority because that also eliminates some of the correctness concerns in this series, i.e. that ensures protected config has the same priority as regular config, but that's a bigger undertaking and I'm not certain about the performance.

Makes sense. The signature of git_config() could stay the same, but we could refactor it to use git_config_in_scope().

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

[...]

Erratum: This patch was sent as gc/document-config-worktree-scope [1].
Patch 6/5 obviously doesn't exist.

[1] https://lore.kernel.org/git/pull.1274.git.git.1654637044966.gitgitgadget@gmail.com

The previous round of this series was picked up by the Google-hosted "Review Club" (the event is on http://tinyurl.com/gitcal). This round incorporates Jonathan Tan's feedback (thanks!) as well some feedback from Review Club itself.

This round only contains changes to commit messages and documentation. As requested in Review Club, I've included a full "Description" section in this cover letter for the convenience of new readers.

= Description

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

This series implements a simple way of preventing such attacks: create a config option, discovery.bare, that tells Git whether or not to die when it finds a bare repository. discovery.bare has two values:

 * "always": always allow bare repositories (default), identical to current
   behavior
 * "never": never allow bare repositories

and users/system administrators who never expect to work with bare repositories can secure their environments using "never". discovery.bare has no effect if --git-dir or GIT_DIR is passed because we are confident that the user is not confused about which repository is being used.

This series does not change the default behavior, but in the long-run, a "no-embedded" option might be a safe and usable default [2]. "never" is too restrictive and unlikely to be the default.

For security reasons, discovery.bare cannot be read from repository-level config (because we would end up trusting the embedded bare repository that we aren't supposed to trust to begin with). Since this would introduce a 3rd variable that is only read from 'protected/trusted configuration' (the others are safe.directory and uploadpack.packObjectsHook) this series also defines and creates a shared implementation for 'protected configuration'

= Patch organization

 * Patch 1 add a section on configuration scopes to our docs
 * Patches 2-3 define 'protected configuration' and create a shared
   implementation.
 * Patch 4 refactors safe.directory to use protected configuration
 * Patch 5 adds discovery.bare

= Series history

Changes in v5:

 * Standardize the usage of "protected configuration" instead of mixing
   "config" and "configuration". This required some unfortunate rewrapping.
 * Remove mentions of "trustworthiness" when discussing protected
   configuration and focus on what Git does instead.
   * The rationale of protected vs non-protected is still kept.
 * Fix the stale documentation entry for discovery.bare.
 * Include a fuller description of how discovery.bare and "--git-dir"
   interact instead of saying "has no effect".

Changes in v4:

 * 2/5's commit message now justifies what scopes are included in protected
   config
 * The global configset is now a file-scope static inside config.c
   (previously it was a member of the_repository).
 * Rename discovery_bare_config to discovery_bare_allowed
 * Make discovery_bare_allowed function-scoped (instead of global).
 * Add an expect_accepted helper to the discovery.bare tests.
 * Add a helper to "upload-pack" that reads the protected and non-protected
   config

Changes in v3:

 * Rebase onto a more recent 'master'
 * Reframe this feature in only in terms of the 'embedded bare repo' attack.
 * Other docs improvements (thanks Stolee in particular!)
 * Protected config no longer uses read_very_early_config() and is only read
   once
 * Protected config now includes "-c"
 * uploadpack.packObjectsHook now uses protected config instead of ignoring
   repo config using config scopes

Changes in v2:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

= Future work

 * This series does not implement the "no-embedded" option [2] and I won't
   work on it any time soon, but I'd be more than happy to review if someone
   sends patches.
 * With discovery.bare, if a builtin is marked RUN_SETUP_GENTLY, setup.c
   doesn't die() and we don't tell users why their repository was rejected,
   e.g. "git config" gives an opaque "fatal: not in a git directory". This
   isn't a new problem though, since safe.directory has the same issue.

[1] https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com

[2] This was first suggested in https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Glen Choo (5):
  Documentation/git-config.txt: add SCOPES section
  Documentation: define protected configuration
  config: learn `git_protected_config()`
  safe.directory: use git_protected_config()
  setup.c: create `discovery.bare`

 Documentation/config.txt            |  2 +
 Documentation/config/discovery.txt  | 23 +++++++++
 Documentation/config/safe.txt       |  6 +--
 Documentation/config/uploadpack.txt |  6 +--
 Documentation/git-config.txt        | 77 +++++++++++++++++++++++------
 config.c                            | 51 +++++++++++++++++++
 config.h                            | 17 +++++++
 setup.c                             | 59 +++++++++++++++++++++-
 t/t0033-safe-directory.sh           | 24 ++++-----
 t/t0035-discovery-bare.sh           | 68 +++++++++++++++++++++++++
 t/t5544-pack-objects-hook.sh        |  7 ++-
 upload-pack.c                       | 27 ++++++----
 12 files changed, 320 insertions(+), 47 deletions(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

base-commit: f770e9f396d48b567ef7b37d273e91ad570a3522
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v5
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v5
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v4:

 1:  c0e27ab3b3e ! 1:  ee9619f6ec0 Documentation/git-config.txt: add SCOPES section
     @@ Metadata
       ## Commit message ##
          Documentation/git-config.txt: add SCOPES section
      
     -    In a subsequent commit, we will introduce "protected config", which is
     -    easiest to describe in terms of configuration scopes (i.e. it's the
     -    union of the 'system', 'global', and 'command' scopes). This description
     -    is fine for ML discussions, but it's inadequate for end users because we
     -    don't provide a good description of "config scopes" in the public docs.
     +    In a subsequent commit, we will introduce "protected configuration",
     +    which is easiest to describe in terms of configuration scopes (i.e. it's
     +    the union of the 'system', 'global', and 'command' scopes). This
     +    description is fine for ML discussions, but it's inadequate for end
     +    users because we don't provide a good description of "configuration
     +    scopes" in the public docs.
      
          145d59f482 (config: add '--show-scope' to print the scope of a config
          value, 2020-02-10) introduced the word "scope" to our public docs, but
     @@ Commit message
          those values mean.
      
          Add a SCOPES section to Documentation/git-config.txt that describes the
     -    config scopes, their corresponding CLI options, and mentions that some
     -    configuration options are only respected in certain scopes. Then,
     +    configuration scopes, their corresponding CLI options, and mentions that
     +    some configuration options are only respected in certain scopes. Then,
          use the word "scope" to simplify the FILES section and change some
          confusing wording.
      
 2:  a5a1dcb03e1 ! 2:  43627c05c0b Documentation: define protected configuration
     @@ Commit message
          Documentation: define protected configuration
      
          For security reasons, there are config variables that are only trusted
     -    when they are specified in extra-trustworthy configuration scopes, which
     -    are sometimes referred to on-list as 'protected configuration' [1]. A
     -    future commit will introduce another such variable, so let's define our
     -    terms so that we can have consistent documentation and implementation.
     +    when they are specified in certain configuration scopes, which are
     +    sometimes referred to on-list as 'protected configuration' [1]. A future
     +    commit will introduce another such variable, so let's define our terms
     +    so that we can have consistent documentation and implementation.
      
     -    In our documentation, define 'protected config' as the system, global
     -    and command config scopes. As a shorthand, I will refer to variables
     -    that are only respected in protected config as 'protected config only',
     -    but this term is not used in the documentation.
     +    In our documentation, define 'protected configuration' as the system,
     +    global and command config scopes. As a shorthand, I will refer to
     +    variables that are only respected in protected config as 'protected
     +    configuration only', but this term is not used in the documentation.
      
     -    This definition of protected configuration is based on whether or not
     -    Git can reasonably protect the user by ignoring the configuration scope:
     +    This definition of protected config is based on whether or not Git can
     +    reasonably protect the user by ignoring the configuration scope:
      
          - System, global and command line config are considered protected
            because an attacker who has control over any of those can do plenty of
     @@ Commit message
            considered protected because it is relatively easy for an attacker to
            control local config, e.g.:
            - On some shared user environments, a non-admin attacker can create a
     -        repository high up the directory hierarchy (e.g. C:\.git on Windows),
     -        and a user may accidentally use it when their PS1 automatically
     -        invokes "git" commands.
     +        repository high up the directory hierarchy (e.g. C:\.git on
     +        Windows), and a user may accidentally use it when their PS1
     +        automatically invokes "git" commands.
      
              `safe.directory` prevents attacks of this form by making sure that
              the user intended to use the shared repository. It obviously
     @@ Commit message
              repository (because "git upload-pack" would fail), but we can limit
              the risks by ignoring `uploadpack.packObjectsHook`.
      
     -    Only `uploadpack.packObjectsHook` is 'protected config only'. The
     +    Only `uploadpack.packObjectsHook` is 'protected configuration only'. The
          following variables are intentionally excluded:
      
     -    - `safe.directory` should be 'protected config only', but it does not
     -      technically fit the definition because it is not respected in the
     +    - `safe.directory` should be 'protected configuration only', but it does
     +      not technically fit the definition because it is not respected in the
            "command" scope. A future commit will fix this.
      
          - `trace2.*` happens to read the same scopes as `safe.directory` because
     @@ Documentation/git-config.txt: Most configuration options are respected regardles
       defined in, but some options are only respected in certain scopes. See the
       option's documentation for the full details.
       
     -+Protected config
     -+~~~~~~~~~~~~~~~~
     -+
     -+Protected config refers to the 'system', 'global', and 'command' scopes. Git
     -+considers these scopes to be especially trustworthy because they are likely
     -+to be controlled by the user or a trusted administrator. An attacker who
     -+controls these scopes can do substantial harm without using Git, so it is
     -+assumed that the user's environment protects these scopes against attackers.
     ++Protected configuration
     ++~~~~~~~~~~~~~~~~~~~~~~~
      +
     ++Protected configuration refers to the 'system', 'global', and 'command' scopes.
      +For security reasons, certain options are only respected when they are
     -+specified in protected config, and ignored otherwise.
     ++specified in protected configuration, and ignored otherwise.
     ++
     ++Git treats these scopes as if they are controlled by the user or a trusted
     ++administrator. This is because an attacker who controls these scopes can do
     ++substantial harm without using Git, so it is assumed that the user's environment
     ++protects these scopes against attackers.
      +
       ENVIRONMENT
       -----------
 3:  94b40907e66 ! 3:  3efe282e6b9 config: read protected config with `git_protected_config()`
     @@ Metadata
      Author: Glen Choo <chooglen@google.com>
      
       ## Commit message ##
     -    config: read protected config with `git_protected_config()`
     +    config: learn `git_protected_config()`
      
     -    `uploadpack.packObjectsHook` is the only 'protected config only'
     +    `uploadpack.packObjectsHook` is the only 'protected configuration only'
          variable today, but we've noted that `safe.directory` and the upcoming
     -    `discovery.bare` should also be 'protected config only'. So, for
     +    `discovery.bare` should also be 'protected configuration only'. So, for
          consistency, we'd like to have a single implementation for protected
          config.
      
          The primary constraints are:
      
     -    1. Reading from protected config should be as fast as possible. Nearly
     -       all "git" commands inside a bare repository will read both
     +    1. Reading from protected configuration should be as fast as possible.
     +       Nearly all "git" commands inside a bare repository will read both
             `safe.directory` and `discovery.bare`, so we cannot afford to be
             slow.
      
     @@ Commit message
             `safe.directory` and `discovery.bare` both affect repository
             discovery and the gitdir is not known at that point [1].
      
     -    The chosen implementation in this commit is to read protected config and
     -    cache the values in a global configset. This is similar to the caching
     -    behavior we get with the_repository->config.
     +    The chosen implementation in this commit is to read protected
     +    configuration and cache the values in a global configset. This is
     +    similar to the caching behavior we get with the_repository->config.
      
     -    Introduce git_protected_config(), which reads protected config and
     -    caches them in the global configset protected_config. Then, refactor
     +    Introduce git_protected_config(), which reads protected configuration
     +    and caches them in the global configset protected_config. Then, refactor
          `uploadpack.packObjectsHook` to use git_protected_config().
      
     -    The protected config functions are named similarly to their
     +    The protected configuration functions are named similarly to their
          non-protected counterparts, e.g. git_protected_config_check_init() vs
          git_config_check_init().
      
          In light of constraint 1, this implementation can still be improved
          since git_protected_config() iterates through every variable in
          protected_config, which may still be too expensive. There exist constant
     -    time lookup functions for non-protected config (repo_config_get_*()),
     -    but for simplicity, this commit does not implement similar functions for
     -    protected config.
     +    time lookup functions for non-protected configuration
     +    (repo_config_get_*()), but for simplicity, this commit does not
     +    implement similar functions for protected configuration.
      
          An alternative that avoids introducing another configset is to continue
          to read all config using git_config(), but only accept values that have
     @@ Commit message
          the_repository->config, which would need to be reset when the gitdir is
          known and git_config() needs to read the local and worktree config.
          Resetting the_repository->config might be reasonable while we only have
     -    these 'protected config only' variables, but it's not clear whether this
     -    extends well to future variables.
     +    these 'protected configuration only' variables, but it's not clear
     +    whether this extends well to future variables.
      
          [1] In this case, we do have a candidate gitdir though, so with a little
          refactoring, it might be possible to provide a gitdir.
 4:  156817966fa ! 4:  ec925823414 safe.directory: use git_protected_config()
     @@ Commit message
          safe.directory: use git_protected_config()
      
          Use git_protected_config() to read `safe.directory` instead of
     -    read_very_early_config(), making it 'protected config only'. As a
     -    result, `safe.directory` now respects "-c", so update the tests and docs
     -    accordingly.
     +    read_very_early_config(), making it 'protected configuration only'.
     +
     +    As a result, `safe.directory` now respects "-c", so update the tests and
     +    docs accordingly. It used to ignore "-c" due to how it was implemented,
     +    not because of security or correctness concerns [1].
     +
     +    [1] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
 5:  29053d029f8 ! 5:  14411512783 setup.c: create `discovery.bare`
     @@ Commit message
          Create a config variable, `discovery.bare`, that tells Git whether or
          not to die() when it discovers a bare repository. This only affects
          repository discovery, thus it has no effect if discovery was not
     -    done (e.g. `--git-dir` was passed).
     +    done, e.g. if the user passes `--git-dir=my-dir`, discovery will be
     +    skipped and my-dir will be used as the repo regardless of the
     +    `discovery.bare` value.
      
          This config is an enum of:
      
     @@ Documentation/config.txt: include::config/diff.txt[]
       ## Documentation/config/discovery.txt (new) ##
      @@
      +discovery.bare::
     -+	'(Protected config only)' Specifies whether Git will work with a
     -+	bare repository that it found during repository discovery. This
     -+	has no effect if the repository is specified directly via the
     -+	--git-dir command-line option or the GIT_DIR environment
     -+	variable (see linkgit:git[1]).
     ++	Specifies whether Git will work with a bare repository that it
     ++	found during repository discovery. If the repository is
     ++	specified directly via the --git-dir command-line option or the
     ++	GIT_DIR environment variable (see linkgit:git[1]), Git will
     ++	always use the specified repository, regardless of this value.
     +++
     ++This config setting is only respected in protected configuration (see
     ++<<SCOPES>>). This prevents the untrusted repository from tampering with
     ++this value.
      ++
      +The currently supported values are:
      ++

From: Glen Choo <chooglen@google.com>

For security reasons, there are config variables that are only trusted when they are specified in certain configuration scopes, which are sometimes referred to on-list as 'protected configuration' [1]. A future commit will introduce another such variable, so let's define our terms so that we can have consistent documentation and implementation.

In our documentation, define 'protected configuration' as the system, global and command config scopes. As a shorthand, I will refer to variables that are only respected in protected config as 'protected configuration only', but this term is not used in the documentation.

This definition of protected config is based on whether or not Git can reasonably protect the user by ignoring the configuration scope:

- System, global and command line config are considered protected
  because an attacker who has control over any of those can do plenty of
  harm without Git, so we gain very little by ignoring those scopes.
- On the other hand, local (and similarly, worktree) config are not
  considered protected because it is relatively easy for an attacker to
  control local config, e.g.:
  - On some shared user environments, a non-admin attacker can create a
    repository high up the directory hierarchy (e.g. C:\.git on
    Windows), and a user may accidentally use it when their PS1
    automatically invokes "git" commands.

    `safe.directory` prevents attacks of this form by making sure that
    the user intended to use the shared repository. It obviously
    shouldn't be read from the repository, because that would end up
    trusting the repository that Git was supposed to reject.
  - "git upload-pack" is expected to run in repositories that may not be
    controlled by the user. We cannot ignore all config in that
    repository (because "git upload-pack" would fail), but we can limit
    the risks by ignoring `uploadpack.packObjectsHook`.

Only `uploadpack.packObjectsHook` is 'protected configuration only'. The following variables are intentionally excluded:

- `safe.directory` should be 'protected configuration only', but it does
  not technically fit the definition because it is not respected in the
  "command" scope. A future commit will fix this.

- `trace2.*` happens to read the same scopes as `safe.directory` because
  they share an implementation. However, this is not for security
  reasons; it is because we want to start tracing so early that
  repository-level config and "-c" are not available [2].

  This requirement is unique to `trace2.*`, so it does not makes sense
  for protected configuration to be subject to the same constraints.

[1] For example, https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/ [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/uploadpack.txt |  6 +++---
 Documentation/git-config.txt        | 13 +++++++++++++
 2 files changed, 16 insertions(+), 3 deletions(-)

From: Glen Choo <chooglen@google.com>

Use git_protected_config() to read `safe.directory` instead of read_very_early_config(), making it 'protected configuration only'.

As a result, `safe.directory` now respects "-c", so update the tests and docs accordingly. It used to ignore "-c" due to how it was implemented, not because of security or correctness concerns [1].

[1] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/safe.txt |  6 +++---
 setup.c                       |  2 +-
 t/t0033-safe-directory.sh     | 24 ++++++++++--------------
 3 files changed, 14 insertions(+), 18 deletions(-)

From: Glen Choo <chooglen@google.com>

`uploadpack.packObjectsHook` is the only 'protected configuration only' variable today, but we've noted that `safe.directory` and the upcoming `discovery.bare` should also be 'protected configuration only'. So, for consistency, we'd like to have a single implementation for protected config.

The primary constraints are:

1. Reading from protected configuration should be as fast as possible.
   Nearly all "git" commands inside a bare repository will read both
   `safe.directory` and `discovery.bare`, so we cannot afford to be
   slow.

2. Protected config must be readable when the gitdir is not known.
   `safe.directory` and `discovery.bare` both affect repository
   discovery and the gitdir is not known at that point [1].

The chosen implementation in this commit is to read protected configuration and cache the values in a global configset. This is similar to the caching behavior we get with the_repository->config.

Introduce git_protected_config(), which reads protected configuration and caches them in the global configset protected_config. Then, refactor `uploadpack.packObjectsHook` to use git_protected_config().

The protected configuration functions are named similarly to their non-protected counterparts, e.g. git_protected_config_check_init() vs git_config_check_init().

In light of constraint 1, this implementation can still be improved since git_protected_config() iterates through every variable in protected_config, which may still be too expensive. There exist constant time lookup functions for non-protected configuration (repo_config_get_*()), but for simplicity, this commit does not implement similar functions for protected configuration.

An alternative that avoids introducing another configset is to continue to read all config using git_config(), but only accept values that have the correct config scope [2]. This technically fulfills constraint 2, because git_config() simply ignores the local and worktree config when the gitdir is not known. However, this would read incomplete config into the_repository->config, which would need to be reset when the gitdir is known and git_config() needs to read the local and worktree config. Resetting the_repository->config might be reasonable while we only have these 'protected configuration only' variables, but it's not clear whether this extends well to future variables.

[1] In this case, we do have a candidate gitdir though, so with a little refactoring, it might be possible to provide a gitdir. [2] This is how `uploadpack.packObjectsHook` was implemented prior to this commit.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 config.c                     | 51 ++++++++++++++++++++++++++++++++++++
 config.h                     | 17 ++++++++++++
 t/t5544-pack-objects-hook.sh |  7 ++++-
 upload-pack.c                | 27 ++++++++++++-------
 4 files changed, 91 insertions(+), 11 deletions(-)

From: Glen Choo <chooglen@google.com>

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

A simple mitigation is to forbid bare repositories unless specified via `--git-dir` or `GIT_DIR`. In environments that don't use bare repositories, this would be minimally disruptive.

Create a config variable, `discovery.bare`, that tells Git whether or not to die() when it discovers a bare repository. This only affects repository discovery, thus it has no effect if discovery was not done, e.g. if the user passes `--git-dir=my-dir`, discovery will be skipped and my-dir will be used as the repo regardless of the `discovery.bare` value.

This config is an enum of:

- "always": always allow bare repositories (this is the default)
- "never": never allow bare repositories

If we want to protect users from such attacks by default, neither value will suffice - "always" provides no protection, but "never" is impractical for bare repository users. A more usable default would be to allow only non-embedded bare repositories ([2] contains one such proposal), but detecting if a repository is embedded is potentially non-trivial, so this work is not implemented in this series.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  2 +
 Documentation/config/discovery.txt | 23 ++++++++++
 setup.c                            | 57 ++++++++++++++++++++++++-
 t/t0035-discovery-bare.sh          | 68 ++++++++++++++++++++++++++++++
 4 files changed, 149 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

On Mon, Jun 27 2022, Glen Choo via GitGitGadget wrote:

You're missing a:

	TEST_PASSES_SANITIZE_LEAK=true

Above this line:

Which tells us that this new test doesn't leak (yay!)

I think we can do away with this helper, we use the argument support once, and for the rest we can inline the trivial command...

grep -F ?

This helper is less trivial, but more obvious would be a "run command and assirt xyz about the output" helper, see e.g. test_stdout_line_count.

Also: Odd to use a sub-shell when the helper takes -C...

...ditto...

Drop the sub-shell and use test_config?

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

Ah, thanks! Hooray.

That is true, having fewer test helpers can be a good idea. Though in this case, the helper wins out slightly (IMO at least) because of the readability/refactoring benefit.

This takes precedent from t0033, which does the same "run command and grep the result". And just as I typed this out, I remembered that t0033's corresponding test helper was made more specific in f62563988f (t0033-safe-directory: check the error message without matching the trash dir, 2022-04-27), because just grep-ing for the config variable masked some errors.

It turns out the same thing is happening in the last test - I forgot that "-c" doesn't unset the variable (it sets the value to ''), and the test_must_fail passes because we fail to parse "discovery.bare", _not_ because we forbade the repo.

So besides -F, I think the only change here would be to grep on the specific "cannot use bare repository" message (instead of grepping for "discovery.bare").

Ok, I'll drop the sub-shell.

Oh, I was so focused on t0033 that I hadn't realized that we had test_config_global. Thanks :)

From: Glen Choo <chooglen@google.com>

In a subsequent commit, we will introduce "protected configuration", which is easiest to describe in terms of configuration scopes (i.e. it's the union of the 'system', 'global', and 'command' scopes). This description is fine for ML discussions, but it's inadequate for end users because we don't provide a good description of "configuration scopes" in the public docs.

145d59f482 (config: add '--show-scope' to print the scope of a config value, 2020-02-10) introduced the word "scope" to our public docs, but that only enumerates the scopes and assumes the user can figure out those values mean.

Add a SCOPES section to Documentation/git-config.txt that describes the configuration scopes, their corresponding CLI options, and mentions that some configuration options are only respected in certain scopes. Then, use the word "scope" to simplify the FILES section and change some confusing wording.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/git-config.txt | 64 ++++++++++++++++++++++++++++--------
 1 file changed, 50 insertions(+), 14 deletions(-)

This is a quick re-roll to address Ævar's comments on the tests (thanks!).

This version is mostly simple refactoring, and I also removed a useless test assertion (see "Series history").

= Description

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

This series implements a simple way of preventing such attacks: create a config option, discovery.bare, that tells Git whether or not to die when it finds a bare repository. discovery.bare has two values:

 * "always": always allow bare repositories (default), identical to current
   behavior
 * "never": never allow bare repositories

and users/system administrators who never expect to work with bare repositories can secure their environments using "never". discovery.bare has no effect if --git-dir or GIT_DIR is passed because we are confident that the user is not confused about which repository is being used.

This series does not change the default behavior, but in the long-run, a "no-embedded" option might be a safe and usable default [2]. "never" is too restrictive and unlikely to be the default.

For security reasons, discovery.bare cannot be read from repository-level config (because we would end up trusting the embedded bare repository that we aren't supposed to trust to begin with). Since this would introduce a 3rd variable that is only read from 'protected/trusted configuration' (the others are safe.directory and uploadpack.packObjectsHook) this series also defines and creates a shared implementation for 'protected configuration'

= Patch organization

 * Patch 1 add a section on configuration scopes to our docs
 * Patches 2-3 define 'protected configuration' and create a shared
   implementation.
 * Patch 4 refactors safe.directory to use protected configuration
 * Patch 5 adds discovery.bare

= Series history

Changes in v6:

 * Add TEST_PASSES_SANITIZE_LEAK=true
 * Replace all sub-shells with -C and use test_config_global
 * Change the expect_rejected helper to use "grep -F" with a more specific
   message.
   * This reveals that the "-c discovery.bare=" assertion in the last test
     was passing for the wrong reason (because '' is an invalid value for
     "discovery.bare"). I removed it because it wasn't doing anything useful
     anyway - I was trying to make discovery.bare unset in the command line,
     but the whole point of that test is to assert that we respect the CLI
     arg.

Changes in v5:

 * Standardize the usage of "protected configuration" instead of mixing
   "config" and "configuration". This required some unfortunate rewrapping.
 * Remove mentions of "trustworthiness" when discussing protected
   configuration and focus on what Git does instead.
   * The rationale of protected vs non-protected is still kept.
 * Fix the stale documentation entry for discovery.bare.
 * Include a fuller description of how discovery.bare and "--git-dir"
   interact instead of saying "has no effect".

Changes in v4:

 * 2/5's commit message now justifies what scopes are included in protected
   config
 * The global configset is now a file-scope static inside config.c
   (previously it was a member of the_repository).
 * Rename discovery_bare_config to discovery_bare_allowed
 * Make discovery_bare_allowed function-scoped (instead of global).
 * Add an expect_accepted helper to the discovery.bare tests.
 * Add a helper to "upload-pack" that reads the protected and non-protected
   config

Changes in v3:

 * Rebase onto a more recent 'master'
 * Reframe this feature in only in terms of the 'embedded bare repo' attack.
 * Other docs improvements (thanks Stolee in particular!)
 * Protected config no longer uses read_very_early_config() and is only read
   once
 * Protected config now includes "-c"
 * uploadpack.packObjectsHook now uses protected config instead of ignoring
   repo config using config scopes

Changes in v2:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

= Future work

 * This series does not implement the "no-embedded" option [2] and I won't
   work on it any time soon, but I'd be more than happy to review if someone
   sends patches.
 * With discovery.bare, if a builtin is marked RUN_SETUP_GENTLY, setup.c
   doesn't die() and we don't tell users why their repository was rejected,
   e.g. "git config" gives an opaque "fatal: not in a git directory". This
   isn't a new problem though, since safe.directory has the same issue.

[1] https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com

[2] This was first suggested in https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Glen Choo (5):
  Documentation/git-config.txt: add SCOPES section
  Documentation: define protected configuration
  config: learn `git_protected_config()`
  safe.directory: use git_protected_config()
  setup.c: create `discovery.bare`

 Documentation/config.txt            |  2 +
 Documentation/config/discovery.txt  | 23 +++++++++
 Documentation/config/safe.txt       |  6 +--
 Documentation/config/uploadpack.txt |  6 +--
 Documentation/git-config.txt        | 77 +++++++++++++++++++++++------
 config.c                            | 51 +++++++++++++++++++
 config.h                            | 17 +++++++
 setup.c                             | 59 +++++++++++++++++++++-
 t/t0033-safe-directory.sh           | 24 ++++-----
 t/t0035-discovery-bare.sh           | 52 +++++++++++++++++++
 t/t5544-pack-objects-hook.sh        |  7 ++-
 upload-pack.c                       | 27 ++++++----
 12 files changed, 304 insertions(+), 47 deletions(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

base-commit: f770e9f396d48b567ef7b37d273e91ad570a3522
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v6
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v6
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v5:

 1:  ee9619f6ec0 = 1:  ee9619f6ec0 Documentation/git-config.txt: add SCOPES section
 2:  43627c05c0b = 2:  43627c05c0b Documentation: define protected configuration
 3:  3efe282e6b9 = 3:  3efe282e6b9 config: learn `git_protected_config()`
 4:  ec925823414 = 4:  ec925823414 safe.directory: use git_protected_config()
 5:  14411512783 ! 5:  a1323d963f9 setup.c: create `discovery.bare`
     @@ t/t0035-discovery-bare.sh (new)
      +
      +test_description='verify discovery.bare checks'
      +
     ++TEST_PASSES_SANITIZE_LEAK=true
      +. ./test-lib.sh
      +
      +pwd="$(pwd)"
     @@ t/t0035-discovery-bare.sh (new)
      +
      +expect_rejected () {
      +	test_must_fail git "$@" rev-parse --git-dir 2>err &&
     -+	grep "discovery.bare" err
     ++	grep -F "cannot use bare repository" err
      +}
      +
      +test_expect_success 'setup bare repo in worktree' '
     @@ t/t0035-discovery-bare.sh (new)
      +'
      +
      +test_expect_success 'discovery.bare unset' '
     -+	(
     -+		cd outer-repo/bare-repo &&
     -+		expect_accepted
     -+	)
     ++	expect_accepted -C outer-repo/bare-repo
      +'
      +
      +test_expect_success 'discovery.bare=always' '
     -+	git config --global discovery.bare always &&
     -+	(
     -+		cd outer-repo/bare-repo &&
     -+		expect_accepted
     -+	)
     ++	test_config_global discovery.bare always &&
     ++	expect_accepted -C outer-repo/bare-repo
      +'
      +
      +test_expect_success 'discovery.bare=never' '
     -+	git config --global discovery.bare never &&
     -+	(
     -+		cd outer-repo/bare-repo &&
     -+		expect_rejected
     -+	)
     ++	test_config_global discovery.bare never &&
     ++	expect_rejected -C outer-repo/bare-repo
      +'
      +
      +test_expect_success 'discovery.bare in the repository' '
     -+	(
     -+		cd outer-repo/bare-repo &&
     -+		# Temporarily set discovery.bare=always, otherwise git
     -+		# config fails with "fatal: not in a git directory"
     -+		# (like safe.directory)
     -+		git config --global discovery.bare always &&
     -+		git config discovery.bare always &&
     -+		git config --global discovery.bare never &&
     -+		expect_rejected
     -+	)
     ++	# discovery.bare must not be "never", otherwise git config fails
     ++	# with "fatal: not in a git directory" (like safe.directory)
     ++	test_config -C outer-repo/bare-repo discovery.bare always &&
     ++	test_config_global discovery.bare never &&
     ++	expect_rejected -C outer-repo/bare-repo
      +'
      +
      +test_expect_success 'discovery.bare on the command line' '
     -+	git config --global discovery.bare never &&
     -+	(
     -+		cd outer-repo/bare-repo &&
     -+		expect_accepted -c discovery.bare=always &&
     -+		expect_rejected -c discovery.bare=
     -+	)
     ++	test_config_global discovery.bare never &&
     ++	expect_accepted -C outer-repo/bare-repo \
     ++		-c discovery.bare=always
      +'
      +
      +test_done

From: Glen Choo <chooglen@google.com>

In a subsequent commit, we will introduce "protected configuration", which is easiest to describe in terms of configuration scopes (i.e. it's the union of the 'system', 'global', and 'command' scopes). This description is fine for ML discussions, but it's inadequate for end users because we don't provide a good description of "configuration scopes" in the public docs.

145d59f482 (config: add '--show-scope' to print the scope of a config value, 2020-02-10) introduced the word "scope" to our public docs, but that only enumerates the scopes and assumes the user can figure out those values mean.

Add a SCOPES section to Documentation/git-config.txt that describes the configuration scopes, their corresponding CLI options, and mentions that some configuration options are only respected in certain scopes. Then, use the word "scope" to simplify the FILES section and change some confusing wording.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/git-config.txt | 64 ++++++++++++++++++++++++++++--------
 1 file changed, 50 insertions(+), 14 deletions(-)

On Thu, Jun 30, 2022 at 06:13:55PM +0000, Glen Choo via GitGitGadget wrote:

Thanks, I think that "scope" is an appropriate term here. When I originally read this patch, I was thinking that "origin" would be more appropriate, since I was recalling the `--show-origin` option to `git config`. But that shows the file name, and `--show-scope` is a separate option entirely.

The latter is definitely more appropriate here, so I think this choice of naming is good and makes sense.

Nit: the last sentence is a little awkwardly worded. Perhaps just:
"Note that neither case produces an error message".

Should we mention that this is the same as the "local" scope below?

I think this paragraph could be slightly more descriptive about what `--file` does while still linking out to <<OPTIONS>> above for more detailed information. In the pre-image, we say:

    If not set explicitly with `--file`, there are four files will `git
    config will search`.

So I wonder if something more descriptive in this section might be:

    You can limit which configuration sources are read to or written
    from by specifying the path of a file with the `--file` option, or
    by specifying a scope with `--system`, `--global`, `--local`, or
    `--worktree`. For more, see <<OPTIONS>> above.

I don't think that's so different form what you wrote, but I think it's a little clearer particularly what `--file` does (instead of "change the way options are read/written" it "limit[s] which configuration sources are read to or written from").

I think a colon after "option" is more appropriate than a single "-" dash character, but this is definitely a trivial matter that I have no strong opinion on.

One thing that this reminds me of (which I don't think is worth taking up here, but perhaps in a future series, or as #leftoverbits) would be promoting these scopes behind a single option. Back in the day, you could ask for values out of `git config` by specifying their type with `--int`, `--bool`, or similar. In e3e042b185 (Merge branch 'tb/config-type', 2018-05-08), we changed to `--type=<int|bool|color|etc>`, which unified things and made it clearer which options were grouped together by a single concept.

I think a similar change would make sense here, that is to replace `--system`, `--global` (and so on) with `--scope=system`, `--scope=global`, etc.

But that's not material to this series, and just something to think about for later on if you end up thinking it's a good idea.

I assume "the option's" is referring to whichever configuration variable we're talking about. So it may be clearer to say "See the *respective* option's documentation for more information" or similar.

Thanks, Taylor

Hi! Thanks so much for lending your attention to this version again, I really appreciate this wording feedback in particular, because the Review Club reviewers and I agonized a lot over the wording and couldn't come up with great alternatives to what I wrote in the patch, and your suggestions are super helpful.

Taylor Blau <me@ttaylorr.com> writes:

Good suggestion. I didn't change this sentence, but I agree that it's worth improving.

Also a good idea.

I think this is _much_ clearer, actually. Thanks!

This sounds like a great idea, actually. I agree that `--scope` is probably a lot easier to reason about than having N scope flags, and that this probably belongs in a future series.

Good idea. Thanks again!

From: Glen Choo <chooglen@google.com>

For security reasons, there are config variables that are only trusted when they are specified in certain configuration scopes, which are sometimes referred to on-list as 'protected configuration' [1]. A future commit will introduce another such variable, so let's define our terms so that we can have consistent documentation and implementation.

In our documentation, define 'protected configuration' as the system, global and command config scopes. As a shorthand, I will refer to variables that are only respected in protected config as 'protected configuration only', but this term is not used in the documentation.

This definition of protected config is based on whether or not Git can reasonably protect the user by ignoring the configuration scope:

- System, global and command line config are considered protected
  because an attacker who has control over any of those can do plenty of
  harm without Git, so we gain very little by ignoring those scopes.
- On the other hand, local (and similarly, worktree) config are not
  considered protected because it is relatively easy for an attacker to
  control local config, e.g.:
  - On some shared user environments, a non-admin attacker can create a
    repository high up the directory hierarchy (e.g. C:\.git on
    Windows), and a user may accidentally use it when their PS1
    automatically invokes "git" commands.

    `safe.directory` prevents attacks of this form by making sure that
    the user intended to use the shared repository. It obviously
    shouldn't be read from the repository, because that would end up
    trusting the repository that Git was supposed to reject.
  - "git upload-pack" is expected to run in repositories that may not be
    controlled by the user. We cannot ignore all config in that
    repository (because "git upload-pack" would fail), but we can limit
    the risks by ignoring `uploadpack.packObjectsHook`.

Only `uploadpack.packObjectsHook` is 'protected configuration only'. The following variables are intentionally excluded:

- `safe.directory` should be 'protected configuration only', but it does
  not technically fit the definition because it is not respected in the
  "command" scope. A future commit will fix this.

- `trace2.*` happens to read the same scopes as `safe.directory` because
  they share an implementation. However, this is not for security
  reasons; it is because we want to start tracing so early that
  repository-level config and "-c" are not available [2].

  This requirement is unique to `trace2.*`, so it does not makes sense
  for protected configuration to be subject to the same constraints.

[1] For example, https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/ [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/uploadpack.txt |  6 +++---
 Documentation/git-config.txt        | 13 +++++++++++++
 2 files changed, 16 insertions(+), 3 deletions(-)

On Thu, Jun 30, 2022 at 06:13:56PM +0000, Glen Choo via GitGitGadget wrote:

I think this description is a good starting point, but I think I would have liked to see some more from the commit description make it into the documentation here.

One thing that I didn't see mentioned in either is that the list of protected configuration is far from exhaustive. There are dozens upon dozens of configuration values that Git will happily execute as a subprocess (core.editor, core.pager, core.alternateRefsCommand, to name just a few).

I don't think we should try and enumerate every possible path from configuration to command execution. But it is worth noting in the documentation that the list of configuration values which are only read in the protected context is non-exhaustive and best-effort only.

Thanks, Taylor

Taylor Blau <me@ttaylorr.com> writes:

Yeah, there's a bit of a tradeoff here. Glossing over some of the details helps keep the documentation briefer and easier to understand for the less experienced/invested, but is bound to frustrate others. I'd appreciate any wording suggestions if you have any.

By referencing command execution, I think you are alluding to Stolee's Security Boundary discussion thread [1], and in particular, the "Example Security Boundary Question: Unprotected Config and Executing Code" section?

That section discusses the problem of arbitrary command execution based on repository-local config, and how protected configuration might give us a way to prevent that. That's a reasonable extension to this series, though it seems a little premature to include allusions to command execution, especially since I don't think we're anywhere close to a long-term direction on what should/shouldn't be inside protected configuration. For example, Stolee noted that, most of the command execution options really do want per-repository customization, so if we want to continue to support that, we'll need to use protected configuration in a somewhat sophisticated manner (and not, e.g. only respect command execution options in protected configuration). Perhaps we could shelve this wording change until we've committed to such a direction.

[1] https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com

From: Glen Choo <chooglen@google.com>

`uploadpack.packObjectsHook` is the only 'protected configuration only' variable today, but we've noted that `safe.directory` and the upcoming `discovery.bare` should also be 'protected configuration only'. So, for consistency, we'd like to have a single implementation for protected config.

The primary constraints are:

1. Reading from protected configuration should be as fast as possible.
   Nearly all "git" commands inside a bare repository will read both
   `safe.directory` and `discovery.bare`, so we cannot afford to be
   slow.

2. Protected config must be readable when the gitdir is not known.
   `safe.directory` and `discovery.bare` both affect repository
   discovery and the gitdir is not known at that point [1].

The chosen implementation in this commit is to read protected configuration and cache the values in a global configset. This is similar to the caching behavior we get with the_repository->config.

Introduce git_protected_config(), which reads protected configuration and caches them in the global configset protected_config. Then, refactor `uploadpack.packObjectsHook` to use git_protected_config().

The protected configuration functions are named similarly to their non-protected counterparts, e.g. git_protected_config_check_init() vs git_config_check_init().

In light of constraint 1, this implementation can still be improved since git_protected_config() iterates through every variable in protected_config, which may still be too expensive. There exist constant time lookup functions for non-protected configuration (repo_config_get_*()), but for simplicity, this commit does not implement similar functions for protected configuration.

An alternative that avoids introducing another configset is to continue to read all config using git_config(), but only accept values that have the correct config scope [2]. This technically fulfills constraint 2, because git_config() simply ignores the local and worktree config when the gitdir is not known. However, this would read incomplete config into the_repository->config, which would need to be reset when the gitdir is known and git_config() needs to read the local and worktree config. Resetting the_repository->config might be reasonable while we only have these 'protected configuration only' variables, but it's not clear whether this extends well to future variables.

[1] In this case, we do have a candidate gitdir though, so with a little refactoring, it might be possible to provide a gitdir. [2] This is how `uploadpack.packObjectsHook` was implemented prior to this commit.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 config.c                     | 51 ++++++++++++++++++++++++++++++++++++
 config.h                     | 17 ++++++++++++
 t/t5544-pack-objects-hook.sh |  7 ++++-
 upload-pack.c                | 27 ++++++++++++-------
 4 files changed, 91 insertions(+), 11 deletions(-)

On Thu, Jun 30, 2022 at 06:13:57PM +0000, Glen Choo via GitGitGadget wrote:

I don't quite follow along with this paragraph: it sounds like reading protected configuration is supposed to be as fast as possible. But you note that only the slower variant of reading each configuration variable one at a time is implemented.

If we care about speed (and I think we should here), then would it make more sense to implement only the lookup functions like repo_config_get_*() for protected context? That would encourage usage by providing a more limited set of options to callers.

Should we be using the word "scope" here instead of sources? I think it's clear enough from the context what you're referring to, but in the spirit of being consistent...

This may be copying from an existing pattern, but I think you could avoid the extra function declaration by writing git_protected_config() as:

    if (!protected_config.hash_initialized)
        read_protected_config();
    configset_iter(&protected_config, fn, data);

Thanks, Taylor

Taylor Blau <me@ttaylorr.com> writes:

Right. I should have been clearer that this implementation is "fast enough without introducing too much noise/complexity", and not "as fast as possible".

I held off on implementing these functions because:

- It requires rewriting `safe.directory`, which reads a multivalued
  string using a config iterator. It's not onerous to do (I had a POC
  of this at some point), but it seemed pretty noisy.
- It seems too noisy to implement all of the protected_config_get_*()
  functions, and a little inconsistent to only implement the ones used
  in this series (but maybe a little inconsistency is ok?)

But maybe a little noise and inconsistency is worth the performance improvement, especially since it's been brought up ~1.5 times before this [1] [2]. I'll do this for sure if you feel strongly about it, otherwise I'll just try it out just to see what I think about it.

[1] https://lore.kernel.org/git/802c3541-3301-43fc-c39e-edd44e61a4eb@github.com [2] https://lore.kernel.org/git/xmqqbkv4t7gp.fsf@gitster.g

Good catch.

You're right, I can drop this if I don't implement protected_config_get_*(); this pattern only makes sense for git_config_check_init() because it's called by multiple functions.

From: Glen Choo <chooglen@google.com>

Use git_protected_config() to read `safe.directory` instead of read_very_early_config(), making it 'protected configuration only'.

As a result, `safe.directory` now respects "-c", so update the tests and docs accordingly. It used to ignore "-c" due to how it was implemented, not because of security or correctness concerns [1].

[1] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/safe.txt |  6 +++---
 setup.c                       |  2 +-
 t/t0033-safe-directory.sh     | 24 ++++++++++--------------
 3 files changed, 14 insertions(+), 18 deletions(-)

From: Glen Choo <chooglen@google.com>

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

A simple mitigation is to forbid bare repositories unless specified via `--git-dir` or `GIT_DIR`. In environments that don't use bare repositories, this would be minimally disruptive.

Create a config variable, `discovery.bare`, that tells Git whether or not to die() when it discovers a bare repository. This only affects repository discovery, thus it has no effect if discovery was not done, e.g. if the user passes `--git-dir=my-dir`, discovery will be skipped and my-dir will be used as the repo regardless of the `discovery.bare` value.

This config is an enum of:

- "always": always allow bare repositories (this is the default)
- "never": never allow bare repositories

If we want to protect users from such attacks by default, neither value will suffice - "always" provides no protection, but "never" is impractical for bare repository users. A more usable default would be to allow only non-embedded bare repositories ([2] contains one such proposal), but detecting if a repository is embedded is potentially non-trivial, so this work is not implemented in this series.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  2 ++
 Documentation/config/discovery.txt | 23 ++++++++++++
 setup.c                            | 57 +++++++++++++++++++++++++++++-
 t/t0035-discovery-bare.sh          | 52 +++++++++++++++++++++++++++
 4 files changed, 133 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

On Thu, Jun 30, 2022 at 06:13:59PM +0000, Glen Choo via GitGitGadget wrote:

I think that everything you said in your patch message makes sense, but I appreciate this paragraph in particular. The historical record is definitely important and worth preserving here, and I hope that it'll be helpful to future readers who may wonder why the default wasn't chosen as "never".

Is it clear from the context what "discovery" means here? It's probably easier to describe what it isn't, which you kind of do in the next sentence. But it may be clearer to say something like:

    Specifies whether Git will recognize bare repositories that aren't
    specified via the top-level `--git-dir` command-line option, or the
    `GIT_DIR` environment variable (see linkgit:git[1]).

I think the default being subject to change is par for the course. It's probably easy enough to just say "Defaults to 'always'" and leave it at that.

I think we still don't have a great answer for people who trust some bare repositories (e.g., known-embedded repositories that are used for testing) but not others. To be clear, I think that is a fine point to concede with this direction.

But we should be clear about that limitation by stating that Git does not support the "I trust some bare repositories to be safely discoverable but not others".

Should we have a default case here since the case arms above are exhaustive?

Relying on NEVER being the zero value here seems fragile to me. Should we check that `if (get_discovery_bare() == DISCOVERY_BARE_NEVER)` to be more explicit here?

Thanks, Taylor

Taylor Blau <me@ttaylorr.com> writes:

Hm that's a good point and the suggestion is very well-worded. In addition to what you have, I think we should make reference to "discovery" _somewhere_ in here since the option is named `discovery.bare`, and this seems like a good teaching opportunity.

Makes sense.

Ah, this "default:" was suggested by Stolee in https://lore.kernel.org/git/7b37f3b7-58c5-1ac5-46eb-d995dc3cc33b@github.com

  This case should be a "default:" in case somehow an arbitrary integer
  value was placed in the variable. [...]

I'm not sure where we stand on this kind of defensiveness. It's not really necessary, but I suppose having a "default:" won't hurt here, especially if it BUG()-s instead of silently passing.

This was also originally suggested by Stolee in https://lore.kernel.org/git/7b37f3b7-58c5-1ac5-46eb-d995dc3cc33b@github.com

  With (some changes to return the enum), we can [...] let the caller
  treat the response as a simple boolean.

but.. your suggestion does seem less fragile. It won't really matter when we add a third enum and replace the "if" with a "switch", but it does matter if we ever muck around with the integer values of DISCOVER_BARE_*.

On Thu, Jun 30, 2022 at 06:13:54PM +0000, Glen Choo via GitGitGadget wrote:

Thanks for this summary, and sorry for not taking a look at this series earlier. I spent quite a bit of time with v1 and the RFC patches but haven't had as much time as I would have liked to devote to reviewing the later rounds.

I'm happy with this direction. I think "never" is an understandable direction to go in for compliance reasons, or in deployments where bare repositories are not expected. And I am glad that we are treating it as unlikely to be the default in the future, I agree that it is likely too restrictive for that to make sense.

So leaving the default behavior unchanged, and providing a big hammer to prevent bare repository discovery entirely seems like a good first step for this feature. I'm interested to see a potential future no-embedded implementation, since I think that has a reasonable chance of being a sensible default, depending on how it's implemented.

This concept is new from the earlier rounds, so I'll be curious to see how it plays out as I take a closer look at the patches.

Thanks again for your work on this complicated and tricky-to-get-right feature ;-).

Thanks, Taylor

On Thu, Jun 30 2022, Glen Choo via GitGitGadget wrote:

Thanks!

Just more generally on this series & approach. I know this is a v6 by now, but I haven't kept up with this topic, but to be fair I did mention pretty much this in: https://lore.kernel.org/git/220407.86lewhc6bz.gmgdl@evledraar.gmail.com/

So...

I'm not insisting that the entire approach here should be changed, but in the above exchange you seemed to have performance concerns about the "just walk up in setup.c" approach I mentioned, but it's not clear if that's still the only thing that necessitates taking this approach.

There may be security subtleties that I've missed, but from the description here it seems like that would work equally well, and wouldn't require configuration, except insofar as we'd need to opt-in to reading config from bare repositores *that also exist in a parent tree*.

And it would be a more narrow & more secure solution, since it would e.g. allow you to intentionally navigate to /var/repos/git/git.git in a server setup and read the config there, which it could distinguish from a case of /var/repos/.git existing, and git/git.git being brought in as a part of that "parent" repo.

The "more narrow" and "more secure" go hand-in-hand, since if you work on such servers you'd turn this to "always" because you want to read such config, but then be left vulnerable to the actual (and muche rarer) exploit we're trying to prevent.

Which, it seems...

This series has (since v3?) been noting aspirations to have a "no-embedded" variant of this config, which your 5/5 here notes would be better, but isn't implemented by this series.

But your 5/5 also notes:

    but detecting if a repository is embedded is potentially
    non-trivial, so this work is not implemented in this series.

Hrm, well, the diff-stat isn't quite that trivial either :) :

In threads linked from the above ML link I linked to some POC code showing how to hack a second .git discovery walk into setup.c. This was as part of the "submodule parent dir" proposal, which is a different feature, but also needs such "find the parent" code: https://lore.kernel.org/git/211109.86v912dtfw.gmgdl@evledraar.gmail.com/

Now, obviously that's a dirty hack, but it's not that hard to just change the part of setup.c where we're satisfied that we've found the git dir, then walk up "$THAT_DIR/..", and start our search again.

Then:

	if (first_dir_was_bare() && found_parent_dir())
        	enforce_no_embedded();

Isn't that what your proposed "no embedded" option would need to do? Well, maybe we'd also check if the "first dir" is in the index of the parent, as opposed to just being a bare .git somewhere in ~/Downloads, e.g. if you have a ~/.git and keep your dot-files in git.

But I think for an initial implementation just doing the walk would be good enough, and would have a more narrow scope than this configuration setting.

AFAICT the performance concerns aren't supported by any data, in the case of the "submodule superproject" feature it turned out to not be the directory walk, but us shelling out in a loop in git-submodule.sh.

Well, *maybe* that's not the case, I think I have managed to read between the lines of some of these past exchanges that there's some odd propriterary internal NFS-like setup at Google where *parent dirs* are auto-mounted and searched on access, so a "walk up" pattern would be much more expensive.

I do worry a bit about us ending up with design choices in git that we wouldn't have ended up with, if not to cater to some in-house setup somwhere that 99.99% of git users will never see.

But I don't have the full picture on the "submodule superproject" problem, or this one, and maybe I'm missing something. Just food for thought, and wondering where we're eventually taking this.

Thanks!

Thanks for weighing in :) Despite the different proposed approaches, I think we actually are in broad agreement.

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

Performance is one major concern, yes, and I agree that your findings show that the "just walk up" approach is cheap enough to consider doing. Though in the few cases where it isn't cheap to walk, wouldn't it still be useful to be able to opt out of it?

The other concern is simplicity and correctness. Are we confident that we'll get the design of "just walk up" correct (including edge cases like "bare repo in bare repo in non bare repo")? I'm 100% confident that we'll get it right eventually, and that this approach will be a good default for all users. But in comparison, "never" is so much easier to understand and implement that I don't see why we shouldn't start by presenting this option to the 0.1-1% of users who would find it useful.

And on the topic of simplicity, there's significant interest in maintaining backwards-compatibility with repos with workflows that absolutely depend on embedded bare repos, e.g. libgit2 and Git-LFS. That's yet another special case that we'd have to get right. Stolee's "no-embedded" proposal [1] pretty much covers that, but I don't see the harm in simplifying the design space by making bare repo support a non-goal.

[1] https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

The point that we're not defending bare repo users is fair, but maybe the group we're trying to protect isn't really dedicated Git-serving servers. This exploit requires you to have a bare repo inside the working tree of a non-bare repo. So I think this is less of an issue for a server, and more for "mixed-use" environments with both regular and bare clones.

Well.. a lot of it is refactoring :P

A narrow scope is good, but I don't agree on this definition of "narrow". My preference is to give an obvious solution to a 'narrow' group of users, instead of a more tricky solution that affects all users in a 'narrow' set of cases.

At the very least, I don't think you're saying that it's a bad idea to have "never", just that we might not have come up with it if not for some Google NFS thing.

Another use case I can think of is CI bots, which have no need for bare repos. To some folks (maybe in very security-sensitive environments), "never" might give more peace of mind than "no-embedded".

On Fri, Jul 01 2022, Glen Choo wrote:

Sorry for the late reply, I see there's a v7, but this seems to also apply to it, so...

Maybe, but until we at least have a reason to think that there is a performance concern this all seems rather hypothetical.

In the thread(s) I linked to I noted that I tested my POC implementation on AIX, which has a sloooooooow filesystem.

Why are you concerned about this having a performance impact?

If you run "git status" or whatever in a directory anywhere on your FS now it'll confidently tell you if you're not within a git repository.

We're just talking about re-using that code, why would we be concerned that:

 1. Finding a repo at a given <PATH>
 2. Appending "/.." to that <PATH>
 3. Feeding it to a "given a path, find me the git repo" (code setup.c
    already has)

Is something we'd get wrong?

Sure, and we could have a config know to still use such repos (for config or otherwise) with the "walk up" method, does that address your concern here?

Yes, but this is only something that's even a question because of an artificial limitation your proposal here suffers from.

I.e. in trying to detect nefarious repos where you've got "looks like bare" content *tracked* in another repo you're conflating it with *any bare repo*.

And the only reason we're doing that seems to me to be a premature optimization.

We could still have an option to say "I never want to consider bare repos from config", but if we're able to distinguish *tracked* bare repos from non-tracked ones this is entirely unrelated to the initial stated motivation for this series, isn't it?

I.e. to solve the security issue at hand.

I've got nothing against *also* providing a "I never want config from xyz repos", but that seems to be orthagonal.

I'm just speculating as to why we've ended up with this approach, but maybe I'm wrong.

Sure, but again, I've got nothing against having *more config knobs*, and I've often e.g. wanted some way to tell git to read no config at all (which you can sort of do with the right environment variables, but it's a hassle).

But I don't see how that's not unrelated to addressing the issue this series aims to address, and what I'm pointing out that it's doing so with a method that's less accurate than the "walk up" method, and less secure.

Thanks for following up. I'm a concerned that this thread will be unproductive if all we're doing is reiterating our own opinions. I'm ok if the conclusion is "agree to disagree", but let's not spend too much time talking circles around one another (myself included, of course:)).

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

Right, I hear you. Besides performance, let me offer the perspective that I should have led with in the previous email. In this thread and the original "embedded bare repo" one [1], there is a huge diversity of opinion on what the default behavior should be, e.g.:

- How do we detect an embedded bare repo (fsck check? walk up [and check
  if it's tracked]?)
- What to do when we detect one (ignore the config? block the repo?)
- How to preserve workflows that rely on embedded bare repos (some kind
  of (global|per-repo) exception list? allow the repo but not the
  config?)

And rightfully so! There are a lot of options here, so we want to make sure we get the defaults right. But at the same time that implies a pretty slow, difficult process.

On the other hand, I haven't seen nearly as much disagreement on "just refuse to work with bare repos" because it's so restrictive that it probably won't be the default. So it'll have no effect on most users, but still confers protection for the subset of users who can benefit from it. For those who want the problem fixed _today_ (e.g. my employer), this seems like simple, low-hanging fruit that buys time for us to find good default.

FWIW, when time permits I'd be happy to work on that good default (which will probably be some variant of "walk up"), and to pay off the tech debt introduced by this implementation (I have some ideas about how we could improve the config API to achieve this [2]). Hopefully that helps allay some of your concerns?

[1] https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2] https://lore.kernel.org/git/kl6lr13fi9qn.fsf@chooglen-macbookpro.roam.corp.google.com

On Tue, Jul 12 2022, Glen Choo wrote:

Yes, I have not been following up here to merely repeat what's been said before, but...

...have been following up because it's still genuinely unclear to me what data or design constraints led to this solution. I.e. in [1] you noted ("[...]" interjection is mine):

	"I don't see how we could implement this [the "walk-up" method]
	without imposing a big penalty to all bare repo users[...]."

[Continued below]

I read that thread over again, and some of the highlights were:

 * brian asking if we can't basically do the "walk up" method:
   https://lore.kernel.org/git/Yk9hONuCIVIq6ieV@camp.crustytoothpaste.net/

 * Taylor wondering how much we need to worry about this attack (among
   other things) & worrying about legitimate "bare repo" workflows being
   broken: https://lore.kernel.org/git/YloTQH35r2xVdPm1@nand.local/ &
   https://lore.kernel.org/git/Ylobp7sntKeWTLDX@nand.local/

But most importantly, here's something I hadn't noticed before:

 * Emily talking about the supposed slowness of the "walk up" method:
   https://lore.kernel.org/git/CAJoAoZkgnnvdymuBsM9Ja3+eYSnyohr=FQZMVX_uzZ_pkQhgaw@mail.gmail.com/

I.e.:

	"wantonly scanning up the filesystem for any gitdir above the
	current one is really expensive. When I tried that approach for
	the purposes of including some shared config between
	superproject and submodules, it slowed down the Git test suite
	by something like 3-5x."

Which I'm now 99.99% certain based on past context[2] is a misstatement or misrecollection about an early version of submodule.superprojectGitDir v.s. what setup.c would do.

I.e. that 3-5x slowness referred to git-submodule.sh shelling out to "git rev-parse", it's not a reference to the expense of the few syscalls we'd need to make to discover a parent git directory.

Did you hear about the directory walking being a performance concern from Emily, or was it an independent discovery?

It seems as though this might have come about because of a misrecollection about the git-rev-parse(1)/git-submodule.sh v.s. setup.c performance with reference to submodule.superprojectGitDir, and that we've now got a design that's optimized to avoid a performance problem that doesn't exist, at the cost of accuracy.

And not to reiterate, but I think the performance isn't a concern per-se, but rather that performance concerns seem to have driven one design over another.

I saw some implementation discussion about how we'd do this with fsck, which is one thing, but I don't really see the trickyness or ambiguity on the client side.

I.e. we know when we'd "find a repo", so that's the criteria we'd use to ignore such a contained repo or not. The only trickyness seems to come about if the approach we pick is one where we conflate embedded bare repos v.s. non-embedded bare repos.

It really just seems like a dead end to me, sorry.

I.e. we know what the security problem is, but the side-effects of this approach are such that we'll probably never turn it on by default.

So that'll mean that the vast majority of users who could benefit from the security mitigation won't even know about the config, or if they do might not have it turned on.

And yes, we might end up with a better design later, but then we'll have to still support this config mechanism, potentially deprecate it etc.

1. https://lore.kernel.org/git/kl6lee1z8mcm.fsf@chooglen-macbookpro.roam.corp.google.com/
2. https://lore.kernel.org/git/211109.86v912dtfw.gmgdl@evledraar.gmail.com/

This version incorporates most of Taylor's comments and suggestions. Thanks especially for the wording suggestions, I struggled with those a lot :)

(I believe) I've responded upthread with my intention for each comment. The only differences between that and the actual changes are:

 * In Documentation/git-config.txt, I dropped a suggestion to mention that
   "git config --local" is identical to the default behavior when writing
   options because I found it too hard to fit in.

 * In Documentation/config/discovery.txt, I took Taylor's suggestion, but
   didn't mention "discovery" for the same reasons.

 * I decided to leave out the protected config lookup functions. I made some
   POC patches at:
   
   https://github.com/chooglen/git/compare/setup/disable-bare-repo-config...chooglen:git:config/protected-config-lookup-fns
   
   which show that the code ends up cleaner with the lookup functions (in
   particular, it lets us remove struct safe_directory_data, which we only
   needed to maintain the state of the config iteration). But, I ultimately
   decided to leave them out of this series because the safe.directory
   conversion is pretty noisy and might end up becoming a distraction from
   the discussion here. If there are no strong objections, I'll send them as
   a follow up series instead.

= Description

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

This series implements a simple way of preventing such attacks: create a config option, discovery.bare, that tells Git whether or not to die when it finds a bare repository. discovery.bare has two values:

 * "always": always allow bare repositories (default), identical to current
   behavior
 * "never": never allow bare repositories

and users/system administrators who never expect to work with bare repositories can secure their environments using "never". discovery.bare has no effect if --git-dir or GIT_DIR is passed because we are confident that the user is not confused about which repository is being used.

This series does not change the default behavior, but in the long-run, a "no-embedded" option might be a safe and usable default [2]. "never" is too restrictive and unlikely to be the default.

For security reasons, discovery.bare cannot be read from repository-level config (because we would end up trusting the embedded bare repository that we aren't supposed to trust to begin with). Since this would introduce a 3rd variable that is only read from 'protected/trusted configuration' (the others are safe.directory and uploadpack.packObjectsHook) this series also defines and creates a shared implementation for 'protected configuration'

= Patch organization

 * Patch 1 add a section on configuration scopes to our docs
 * Patches 2-3 define 'protected configuration' and create a shared
   implementation.
 * Patch 4 refactors safe.directory to use protected configuration
 * Patch 5 adds discovery.bare

= Series history

Changes in v7:

 * Numerous docs improvements and code cleanup.
 * In 3/5's commit message, drop "as fast as possible" and allude to lookup
   functions coming in a later series.
 * Remove a comment in 3/5 about repository.protected_config. That was stale
   since v4, but slipped under the radar until now.
 * Fix some s/protected config/protected configuration (leftover from v5).

Changes in v6:

 * Add TEST_PASSES_SANITIZE_LEAK=true
 * Replace all sub-shells with -C and use test_config_global
 * Change the expect_rejected helper to use "grep -F" with a more specific
   message.
   * This reveals that the "-c discovery.bare=" assertion in the last test
     was passing for the wrong reason (because '' is an invalid value for
     "discovery.bare"). I removed it because it wasn't doing anything useful
     anyway - I was trying to make discovery.bare unset in the command line,
     but the whole point of that test is to assert that we respect the CLI
     arg.

Changes in v5:

 * Standardize the usage of "protected configuration" instead of mixing
   "config" and "configuration". This required some unfortunate rewrapping.
 * Remove mentions of "trustworthiness" when discussing protected
   configuration and focus on what Git does instead.
   * The rationale of protected vs non-protected is still kept.
 * Fix the stale documentation entry for discovery.bare.
 * Include a fuller description of how discovery.bare and "--git-dir"
   interact instead of saying "has no effect".

Changes in v4:

 * 2/5's commit message now justifies what scopes are included in protected
   config
 * The global configset is now a file-scope static inside config.c
   (previously it was a member of the_repository).
 * Rename discovery_bare_config to discovery_bare_allowed
 * Make discovery_bare_allowed function-scoped (instead of global).
 * Add an expect_accepted helper to the discovery.bare tests.
 * Add a helper to "upload-pack" that reads the protected and non-protected
   config

Changes in v3:

 * Rebase onto a more recent 'master'
 * Reframe this feature in only in terms of the 'embedded bare repo' attack.
 * Other docs improvements (thanks Stolee in particular!)
 * Protected config no longer uses read_very_early_config() and is only read
   once
 * Protected config now includes "-c"
 * uploadpack.packObjectsHook now uses protected config instead of ignoring
   repo config using config scopes

Changes in v2:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

= Future work

 * This series doesn't implement config lookup functions for protected
   config. This will be done in a follow up series.
 * This series does not implement the "no-embedded" option [2] and I won't
   work on it any time soon, but I'd be more than happy to review if someone
   sends patches.
 * With discovery.bare, if a builtin is marked RUN_SETUP_GENTLY, setup.c
   doesn't die() and we don't tell users why their repository was rejected,
   e.g. "git config" gives an opaque "fatal: not in a git directory". This
   isn't a new problem though, since safe.directory has the same issue.

[1] https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com

[2] This was first suggested in https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Glen Choo (5):
  Documentation/git-config.txt: add SCOPES section
  Documentation: define protected configuration
  config: learn `git_protected_config()`
  safe.directory: use git_protected_config()
  setup.c: create `discovery.bare`

 Documentation/config.txt            |  2 +
 Documentation/config/discovery.txt  | 21 ++++++++
 Documentation/config/safe.txt       |  6 +--
 Documentation/config/uploadpack.txt |  6 +--
 Documentation/git-config.txt        | 78 +++++++++++++++++++++++------
 config.c                            | 43 ++++++++++++++++
 config.h                            | 16 ++++++
 setup.c                             | 59 +++++++++++++++++++++-
 t/t0033-safe-directory.sh           | 24 ++++-----
 t/t0035-discovery-bare.sh           | 52 +++++++++++++++++++
 t/t5544-pack-objects-hook.sh        |  7 ++-
 upload-pack.c                       | 27 ++++++----
 12 files changed, 294 insertions(+), 47 deletions(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

base-commit: f770e9f396d48b567ef7b37d273e91ad570a3522
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v7
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v7
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v6:

 1:  ee9619f6ec0 ! 1:  5c58db3bb21 Documentation/git-config.txt: add SCOPES section
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
      +Options will be read from all of these files that are available. If the
      +global or the system-wide configuration file are not available they will be
      +ignored. If the repository configuration file is not available or readable,
     -+'git config' will exit with a non-zero error code. However, in neither case
     -+will an error message be issued.
     ++'git config' will exit with a non-zero error code. Note that neither case
     ++produces an error message.
       
       The files are read in the order given above, with last value found taking
       precedence over values read earlier.  When multiple values are taken then all
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
      -You can override these rules using the `--global`, `--system`,
      -`--local`, `--worktree`, and `--file` command-line options; see
      -<<OPTIONS>> above.
     -+You can change the way options are read/written by specifying the path to a
     -+file (`--file`), or by specifying a configuration scope (`--system`,
     -+`--global`, `--local`, `--worktree`); see <<OPTIONS>> above.
     ++You can limit which configuration sources are read from or written to by
     ++specifying the path of a file with the `--file` option, or by specifying a
     ++configuration scope with `--system`, `--global`, `--local`, or `--worktree`.
     ++For more, see <<OPTIONS>> above.
      +
      +SCOPES
      +------
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
      +the `-c` option
      +
      +With the exception of 'command', each scope corresponds to a command line
     -+option - `--system`, `--global`, `--local`, `--worktree`.
     ++option: `--system`, `--global`, `--local`, `--worktree`.
      +
      +When reading options, specifying a scope will only read options from the
      +files within that scope. When writing options, specifying a scope will write
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
       
      +Most configuration options are respected regardless of the scope it is
      +defined in, but some options are only respected in certain scopes. See the
     -+option's documentation for the full details.
     ++respective option's documentation for the full details.
       
       ENVIRONMENT
       -----------
 2:  43627c05c0b ! 2:  58f25612aa3 Documentation: define protected configuration
     @@ Commit message
      
          In our documentation, define 'protected configuration' as the system,
          global and command config scopes. As a shorthand, I will refer to
     -    variables that are only respected in protected config as 'protected
     -    configuration only', but this term is not used in the documentation.
     +    variables that are only respected in protected configuration as
     +    'protected configuration only', but this term is not used in the
     +    documentation.
      
     -    This definition of protected config is based on whether or not Git can
     -    reasonably protect the user by ignoring the configuration scope:
     +    This definition of protected configuration is based on whether or not
     +    Git can reasonably protect the user by ignoring the configuration scope:
      
          - System, global and command line config are considered protected
            because an attacker who has control over any of those can do plenty of
     @@ Documentation/config/uploadpack.txt: uploadpack.packObjectsHook::
      -repository-level config (this is a safety measure against fetching from
      -untrusted repositories).
      +Note that this configuration variable is only respected when it is specified
     -+in protected config (see <<SCOPES>>). This is a safety measure against
     -+fetching from untrusted repositories.
     ++in protected configuration (see <<SCOPES>>). This is a safety measure
     ++against fetching from untrusted repositories.
       
       uploadpack.allowFilter::
       	If this option is set, `upload-pack` will support partial
      
       ## Documentation/git-config.txt ##
     -@@ Documentation/git-config.txt: You can change the way options are read/written by specifying the path to a
     - file (`--file`), or by specifying a configuration scope (`--system`,
     - `--global`, `--local`, `--worktree`); see <<OPTIONS>> above.
     +@@ Documentation/git-config.txt: specifying the path of a file with the `--file` option, or by specifying a
     + configuration scope with `--system`, `--global`, `--local`, or `--worktree`.
     + For more, see <<OPTIONS>> above.
       
      +[[SCOPES]]
       SCOPES
     @@ Documentation/git-config.txt: You can change the way options are read/written by
       
      @@ Documentation/git-config.txt: Most configuration options are respected regardless of the scope it is
       defined in, but some options are only respected in certain scopes. See the
     - option's documentation for the full details.
     + respective option's documentation for the full details.
       
      +Protected configuration
      +~~~~~~~~~~~~~~~~~~~~~~~
 3:  3efe282e6b9 ! 3:  3683d20f232 config: learn `git_protected_config()`
     @@ Commit message
          variable today, but we've noted that `safe.directory` and the upcoming
          `discovery.bare` should also be 'protected configuration only'. So, for
          consistency, we'd like to have a single implementation for protected
     -    config.
     +    configuration.
      
          The primary constraints are:
      
     -    1. Reading from protected configuration should be as fast as possible.
     -       Nearly all "git" commands inside a bare repository will read both
     -       `safe.directory` and `discovery.bare`, so we cannot afford to be
     -       slow.
     +    1. Reading from protected configuration should be fast. Nearly all "git"
     +       commands inside a bare repository will read both `safe.directory` and
     +       `discovery.bare`, so we cannot afford to be slow.
      
     -    2. Protected config must be readable when the gitdir is not known.
     -       `safe.directory` and `discovery.bare` both affect repository
     +    2. Protected configuration must be readable when the gitdir is not
     +       known. `safe.directory` and `discovery.bare` both affect repository
             discovery and the gitdir is not known at that point [1].
      
          The chosen implementation in this commit is to read protected
     @@ Commit message
          non-protected counterparts, e.g. git_protected_config_check_init() vs
          git_config_check_init().
      
     -    In light of constraint 1, this implementation can still be improved
     -    since git_protected_config() iterates through every variable in
     -    protected_config, which may still be too expensive. There exist constant
     -    time lookup functions for non-protected configuration
     -    (repo_config_get_*()), but for simplicity, this commit does not
     -    implement similar functions for protected configuration.
     +    In light of constraint 1, this implementation can still be improved.
     +    git_protected_config() iterates through every variable in
     +    protected_config, which is wasteful, but it makes the conversion simple
     +    because it matches existing patterns. We will likely implement constant
     +    time lookup functions for protected configuration in a future series
     +    (such functions already exist for non-protected configuration, i.e.
     +    repo_config_get_*()).
      
          An alternative that avoids introducing another configset is to continue
          to read all config using git_config(), but only accept values that have
     @@ config.c: static enum config_scope current_parsing_scope;
       static int zlib_compression_seen;
       
      +/*
     -+ * Config that comes from trusted sources, namely:
     -+ * - system config files (e.g. /etc/gitconfig)
     -+ * - global config files (e.g. $HOME/.gitconfig,
     -+ *   $XDG_CONFIG_HOME/git)
     -+ * - the command line.
     ++ * Config that comes from trusted scopes, namely:
     ++ * - CONFIG_SCOPE_SYSTEM (e.g. /etc/gitconfig)
     ++ * - CONFIG_SCOPE_GLOBAL (e.g. $HOME/.gitconfig, $XDG_CONFIG_HOME/git)
     ++ * - CONFIG_SCOPE_COMMAND (e.g. "-c" option, environment variables)
      + *
      + * This is declared here for code cleanliness, but unlike the other
      + * static variables, this does not hold config parser state.
     @@ config.c: int repo_config_get_pathname(struct repository *repo,
      +	free(user_config);
      +}
      +
     -+/* Ensure that protected_config has been initialized. */
     -+static void git_protected_config_check_init(void)
     -+{
     -+	if (protected_config.hash_initialized)
     -+		return;
     -+	read_protected_config();
     -+}
     -+
      +void git_protected_config(config_fn_t fn, void *data)
      +{
     -+	git_protected_config_check_init();
     ++	if (!protected_config.hash_initialized)
     ++		read_protected_config();
      +	configset_iter(&protected_config, fn, data);
      +}
      +
     @@ config.h: int repo_config_get_maybe_bool(struct repository *repo,
       
      +/*
      + * Functions for reading protected config. By definition, protected
     -+ * config ignores repository config, so it is unnecessary to read
     -+ * protected config from any `struct repository` other than
     -+ * the_repository.
     ++ * config ignores repository config, so these do not take a `struct
     ++ * repository` parameter.
      + */
      +void git_protected_config(config_fn_t fn, void *data);
      +
 4:  ec925823414 = 4:  6394818ffd8 safe.directory: use git_protected_config()
 5:  a1323d963f9 ! 5:  eff4b07480e setup.c: create `discovery.bare`
     @@ Documentation/config.txt: include::config/diff.txt[]
       ## Documentation/config/discovery.txt (new) ##
      @@
      +discovery.bare::
     -+	Specifies whether Git will work with a bare repository that it
     -+	found during repository discovery. If the repository is
     -+	specified directly via the --git-dir command-line option or the
     -+	GIT_DIR environment variable (see linkgit:git[1]), Git will
     -+	always use the specified repository, regardless of this value.
     ++	Specifies whether Git will work with a bare repository that
     ++	wasn't specified via the top-level `--git-dir` command-line
     ++	option, or the `GIT_DIR` environment variable (see
     ++	linkgit:git[1]). If the repository is specified, Git will always
     ++	use the specified repository, regardless of this value.
      ++
      +This config setting is only respected in protected configuration (see
      +<<SCOPES>>). This prevents the untrusted repository from tampering with
     @@ Documentation/config/discovery.txt (new)
      +* `always`: Git always works with bare repositories
      +* `never`: Git never works with bare repositories
      ++
     -+This defaults to `always`, but this default may change in the future.
     -++
      +If you do not use bare repositories in your workflow, then it may be
      +beneficial to set `discovery.bare` to `never` in your global config.
      +This will protect you from attacks that involve cloning a repository
     @@ setup.c: static enum discovery_result setup_git_directory_gently_1(struct strbuf
       		}
       
       		if (is_git_directory(dir->buf)) {
     -+			if (!get_discovery_bare())
     ++			if (get_discovery_bare() == DISCOVERY_BARE_NEVER)
      +				return GIT_DIR_DISALLOWED_BARE;
       			if (!ensure_valid_ownership(dir->buf))
       				return GIT_DIR_INVALID_OWNERSHIP;

From: Glen Choo <chooglen@google.com>

In a subsequent commit, we will introduce "protected configuration", which is easiest to describe in terms of configuration scopes (i.e. it's the union of the 'system', 'global', and 'command' scopes). This description is fine for ML discussions, but it's inadequate for end users because we don't provide a good description of "configuration scopes" in the public docs.

145d59f482 (config: add '--show-scope' to print the scope of a config value, 2020-02-10) introduced the word "scope" to our public docs, but that only enumerates the scopes and assumes the user can figure out those values mean.

Add a SCOPES section to Documentation/git-config.txt that describes the configuration scopes, their corresponding CLI options, and mentions that some configuration options are only respected in certain scopes. Then, use the word "scope" to simplify the FILES section and change some confusing wording.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/git-config.txt | 65 ++++++++++++++++++++++++++++--------
 1 file changed, 51 insertions(+), 14 deletions(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

Probably: "figure out those values mean" -> "figure out what those
values mean"

Listing "-c" as another one in addition to these files is probably a good simplification, instead of "-c override others" in the original, as we would need to say "more specific ones override less specific ones" anyway.

Problem inherited from the original, but I suspect that rephrasing "not available" to "missing" (or "does not exist") may make it easier to follow. Also, "global" in the preceding description is explained as one of the user-specific configuration files, so it may be better to avoid it, e.g.

	If the user-specific or the system-wide configuration files
	are missing, they will be ignored.  If the repository
	configuration file is missing or unreadable, ...

Alternatively, we may want to tighten the description of $XDG_CONFIG_HOME/git/config and ~/.gitconfig a bit better, e.g. something along the lines of ...

	$XDG_CONFIG_HOME/git/config::
	$HOME/.gitconfig::
		User-specific configuration file.  When the
		XDG_CONFIG_HOME environment variable is not set or
		empty, $HOME/.config/ is used as $XDG_CONFIG_HOME.
	+
	These are often called the "global" configuration file.
	When either or both of them exist(s), both files are read.

Note that I deliberately omitted the mehtion that our $XDG support may be too recent and $HOME/.gitconfig may be preferred for portability. It came from 21cf3227 (config: read (but not write) from $XDG_CONFIG_HOME/git/config file, 2012-06-22) but sufficient number of years have passed.

Also note that I originally wrote the following immediately after the above

	+
	When writing to the "--global" scope (see below),
	$XDG_CONFIG_HOME/git/config is used if it exists; otherwise
	$HOME/.gitconfig is used.

but decided to discard it, since the OPTIONS -> "--global" covers it well enough.

With a tightening of the definition of what "global" is,, we can rephrase "If the global or the system configuration files are missing...".

We'd need to tighten this a bit, like:

	GIT_CONFIG_{COUNT,KEY,VALUE} environment varialbes (see below)

GIT_CONFIG_GLOBAL or GIT_CONFIG_SYSTEM environment varialbes and others are listed in the ENVIRONMENT section that comes after this section, but you do not want the readers to be confused into thinking that we'd give them more precedence over others.

Overall it was a pleasant read.  Thanks, will queue.

Junio C Hamano <gitster@pobox.com> writes:

Makes sense. I think this is simpler and more coherent with your suggested changes.

The only change I'd suggest is to expand "missing" -> "missing or unreadable". The original wording is "not available", which could be interpreted to cover both cases. We'd obviously also have to amend "not available or readable" accordingly.

Ah, good point.

Thanks! Shall I apply your suggestions, or were you planning to apply them yourself?

Glen Choo <chooglen@google.com> writes:

Probably. I wonder if we should document that we at least warn when a file we are expected to read exists but is not readable (instead of simply saying "is ignored"), but other than that I agree with you.

Definitely not the latter ;-)

Junio C Hamano <gitster@pobox.com> writes:

;) Ok, I'll give others some time to weigh in before rerolling.

From: Glen Choo <chooglen@google.com>

For security reasons, there are config variables that are only trusted when they are specified in certain configuration scopes, which are sometimes referred to on-list as 'protected configuration' [1]. A future commit will introduce another such variable, so let's define our terms so that we can have consistent documentation and implementation.

In our documentation, define 'protected configuration' as the system, global and command config scopes. As a shorthand, I will refer to variables that are only respected in protected configuration as 'protected configuration only', but this term is not used in the documentation.

This definition of protected configuration is based on whether or not Git can reasonably protect the user by ignoring the configuration scope:

- System, global and command line config are considered protected
  because an attacker who has control over any of those can do plenty of
  harm without Git, so we gain very little by ignoring those scopes.
- On the other hand, local (and similarly, worktree) config are not
  considered protected because it is relatively easy for an attacker to
  control local config, e.g.:
  - On some shared user environments, a non-admin attacker can create a
    repository high up the directory hierarchy (e.g. C:\.git on
    Windows), and a user may accidentally use it when their PS1
    automatically invokes "git" commands.

    `safe.directory` prevents attacks of this form by making sure that
    the user intended to use the shared repository. It obviously
    shouldn't be read from the repository, because that would end up
    trusting the repository that Git was supposed to reject.
  - "git upload-pack" is expected to run in repositories that may not be
    controlled by the user. We cannot ignore all config in that
    repository (because "git upload-pack" would fail), but we can limit
    the risks by ignoring `uploadpack.packObjectsHook`.

Only `uploadpack.packObjectsHook` is 'protected configuration only'. The following variables are intentionally excluded:

- `safe.directory` should be 'protected configuration only', but it does
  not technically fit the definition because it is not respected in the
  "command" scope. A future commit will fix this.

- `trace2.*` happens to read the same scopes as `safe.directory` because
  they share an implementation. However, this is not for security
  reasons; it is because we want to start tracing so early that
  repository-level config and "-c" are not available [2].

  This requirement is unique to `trace2.*`, so it does not makes sense
  for protected configuration to be subject to the same constraints.

[1] For example, https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/ [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/uploadpack.txt |  6 +++---
 Documentation/git-config.txt        | 13 +++++++++++++
 2 files changed, 16 insertions(+), 3 deletions(-)

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

This is only about the formatting, but have a blank line between each bullet-point (e.g. before the line that talks about "On some shared user enviornments, ..." and "git upload-pack"). A paragraph break within a single bullet-point (i.e. the paragraph that talks about `safe.directory` is a second paragraph of hte same bullet point as the paragraph before it) looks like a stronger break than separation between each bullet-point, which you wrote without any blank lines in between.

Very well reasoned.

From: Glen Choo <chooglen@google.com>

`uploadpack.packObjectsHook` is the only 'protected configuration only' variable today, but we've noted that `safe.directory` and the upcoming `discovery.bare` should also be 'protected configuration only'. So, for consistency, we'd like to have a single implementation for protected configuration.

The primary constraints are:

1. Reading from protected configuration should be fast. Nearly all "git"
   commands inside a bare repository will read both `safe.directory` and
   `discovery.bare`, so we cannot afford to be slow.

2. Protected configuration must be readable when the gitdir is not
   known. `safe.directory` and `discovery.bare` both affect repository
   discovery and the gitdir is not known at that point [1].

The chosen implementation in this commit is to read protected configuration and cache the values in a global configset. This is similar to the caching behavior we get with the_repository->config.

Introduce git_protected_config(), which reads protected configuration and caches them in the global configset protected_config. Then, refactor `uploadpack.packObjectsHook` to use git_protected_config().

The protected configuration functions are named similarly to their non-protected counterparts, e.g. git_protected_config_check_init() vs git_config_check_init().

In light of constraint 1, this implementation can still be improved. git_protected_config() iterates through every variable in protected_config, which is wasteful, but it makes the conversion simple because it matches existing patterns. We will likely implement constant time lookup functions for protected configuration in a future series (such functions already exist for non-protected configuration, i.e. repo_config_get_*()).

An alternative that avoids introducing another configset is to continue to read all config using git_config(), but only accept values that have the correct config scope [2]. This technically fulfills constraint 2, because git_config() simply ignores the local and worktree config when the gitdir is not known. However, this would read incomplete config into the_repository->config, which would need to be reset when the gitdir is known and git_config() needs to read the local and worktree config. Resetting the_repository->config might be reasonable while we only have these 'protected configuration only' variables, but it's not clear whether this extends well to future variables.

[1] In this case, we do have a candidate gitdir though, so with a little refactoring, it might be possible to provide a gitdir. [2] This is how `uploadpack.packObjectsHook` was implemented prior to this commit.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 config.c                     | 43 ++++++++++++++++++++++++++++++++++++
 config.h                     | 16 ++++++++++++++
 t/t5544-pack-objects-hook.sh |  7 +++++-
 upload-pack.c                | 27 +++++++++++++---------
 4 files changed, 82 insertions(+), 11 deletions(-)

From: Glen Choo <chooglen@google.com>

Use git_protected_config() to read `safe.directory` instead of read_very_early_config(), making it 'protected configuration only'.

As a result, `safe.directory` now respects "-c", so update the tests and docs accordingly. It used to ignore "-c" due to how it was implemented, not because of security or correctness concerns [1].

[1] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/safe.txt |  6 +++---
 setup.c                       |  2 +-
 t/t0033-safe-directory.sh     | 24 ++++++++++--------------
 3 files changed, 14 insertions(+), 18 deletions(-)

From: Glen Choo <chooglen@google.com>

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

A simple mitigation is to forbid bare repositories unless specified via `--git-dir` or `GIT_DIR`. In environments that don't use bare repositories, this would be minimally disruptive.

Create a config variable, `discovery.bare`, that tells Git whether or not to die() when it discovers a bare repository. This only affects repository discovery, thus it has no effect if discovery was not done, e.g. if the user passes `--git-dir=my-dir`, discovery will be skipped and my-dir will be used as the repo regardless of the `discovery.bare` value.

This config is an enum of:

- "always": always allow bare repositories (this is the default)
- "never": never allow bare repositories

If we want to protect users from such attacks by default, neither value will suffice - "always" provides no protection, but "never" is impractical for bare repository users. A more usable default would be to allow only non-embedded bare repositories ([2] contains one such proposal), but detecting if a repository is embedded is potentially non-trivial, so this work is not implemented in this series.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config.txt           |  2 ++
 Documentation/config/discovery.txt | 21 +++++++++++
 setup.c                            | 57 +++++++++++++++++++++++++++++-
 t/t0035-discovery-bare.sh          | 52 +++++++++++++++++++++++++++
 4 files changed, 131 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/config/discovery.txt
 create mode 100755 t/t0035-discovery-bare.sh

"Glen Choo via GitGitGadget" <gitgitgadget@gmail.com> writes:

These patches overall looked ok. I am not very happy to see the proliferation of namespaces like safe.* and discovery.* that would not likely to get the second variable, though.

Junio C Hamano <gitster@pobox.com> writes:

Fair. I think `discovery.bare` is similar enough to `safe.directory` that it could belong in the safe.* namespace if we find a good name for it.

We rejected "safe.bareRepository" earlier because of the insinuation that bare repos are unsafe. Maybe:

- safe.bareDiscovery
- safe.bareRepositoryDiscovery
- safe.unspecifiedBareRepository
- safe.discoveredBareRepository

"safe.unspecifiedBareRepository" is sounding pretty good to me actually.. Any thoughts?

Glen Choo <chooglen@google.com> writes:

(+CC Johannes Schindelin for thoughts on what should go into `safe.*` and/or design considerations that went into it.)

Another thought is that `discovery.bare` and `safe.directory` should both indeed live in the same namespace, but that namespace should be named something other than `safe.*`, e.g. if we had `allowedRepositories.otherOwner` instead of `safe.directory`, it would have been a no-brainer for me to put this in the `allowedRepositories.*` namespace.

So an alternative proposal would be:

- rename this to `allowedRepositories.discoveredBare`
- (possibly not in this series, but at some point) create a
  `safe.directory` alias in that namespace, e.g.
  `allowedRepositories.otherOwner`

*But* I don't see the former making sense without the latter (I really think both should be in the same namespace), so if we think that's unnecessary churn, I'll drop this idea entirely.

Thanks all! This version takes Junio's wording suggestions on the previous round, renames the config variable and, because of conflicts with 2.37.1, is newly rebased onto master.

The config variable is now named "safe.bareRepository"; I discarded this name earlier in this series, but the intent behind the series has come around to the point where "safe.bareRepository" makes sense again. And thanks to Dscho's suggestion to explicitly create an option for "no discovered bare repos" [1], the UI is self-documenting enough that we can write terser docs.

As mentioned upthread, I would love to see "safe.*" namespace renamed (it's an adjective, not a Git command/entity), but I'm doubling down on it anyway. This config option does such a similar thing to "safe.directory" that they really should be siblings, and I don't foresee a world where "safe.directory" gets renamed.

I've triple-checked to make sure I've scrubbed the commits of the old name, but I'd appreciate the extra eyes :)

= Description

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [2] for a fuller description and deeper discussion.

This series implements a simple way of preventing such attacks: create a config option, safe.bareRepository, that tells Git whether or not to die when it finds a bare repository. safe.bareRepository has two values:

 * "all": allow all bare repositories (default), identical to current
   behavior
 * "explicit": only allow bare repositories specified via --git-dir or
   GIT_DIR

and users/system administrators who never expect to work with bare repositories can secure their environments using "explicit". We still trust explicit bare repositories because we are confident that the user is not confused about which repository is being used.

This series does not change the default behavior, but in the long-run, a "no-embedded" option might be a safe and usable default [3]. "never" is too restrictive and unlikely to be the default.

For security reasons, safe.bareRepository cannot be read from repository-level config (because we would end up trusting the embedded bare repository that we aren't supposed to trust to begin with). Since this would introduce a 3rd variable that is only read from 'protected/trusted configuration' (the others are safe.directory and uploadpack.packObjectsHook) this series also defines and creates a shared implementation for 'protected configuration'

= Patch organization

 * Patch 1 add a section on configuration scopes to our docs
 * Patches 2-3 define 'protected configuration' and create a shared
   implementation.
 * Patch 4 refactors safe.directory to use protected configuration
 * Patch 5 adds safe.bareRepository

= Series history

Changes in v8:

 * Rename discovery.bare -> safe.bareRepository, change values from
   "always|never" -> "all|explicit"
 * Numerous docs improvements
 * Rebase onto post-2.37.1 master

Changes in v7:

 * Numerous docs improvements and code cleanup.
 * In 3/5's commit message, drop "as fast as possible" and allude to lookup
   functions coming in a later series.
 * Remove a comment in 3/5 about repository.protected_config. That was stale
   since v4, but slipped under the radar until now.
 * Fix some s/protected config/protected configuration (leftover from v5).

Changes in v6:

 * Add TEST_PASSES_SANITIZE_LEAK=true
 * Replace all sub-shells with -C and use test_config_global
 * Change the expect_rejected helper to use "grep -F" with a more specific
   message.
   * This reveals that the "-c discovery.bare=" assertion in the last test
     was passing for the wrong reason (because '' is an invalid value for
     "discovery.bare"). I removed it because it wasn't doing anything useful
     anyway - I was trying to make discovery.bare unset in the command line,
     but the whole point of that test is to assert that we respect the CLI
     arg.

Changes in v5:

 * Standardize the usage of "protected configuration" instead of mixing
   "config" and "configuration". This required some unfortunate rewrapping.
 * Remove mentions of "trustworthiness" when discussing protected
   configuration and focus on what Git does instead.
   * The rationale of protected vs non-protected is still kept.
 * Fix the stale documentation entry for discovery.bare.
 * Include a fuller description of how discovery.bare and "--git-dir"
   interact instead of saying "has no effect".

Changes in v4:

 * 2/5's commit message now justifies what scopes are included in protected
   config
 * The global configset is now a file-scope static inside config.c
   (previously it was a member of the_repository).
 * Rename discovery_bare_config to discovery_bare_allowed
 * Make discovery_bare_allowed function-scoped (instead of global).
 * Add an expect_accepted helper to the discovery.bare tests.
 * Add a helper to "upload-pack" that reads the protected and non-protected
   config

Changes in v3:

 * Rebase onto a more recent 'master'
 * Reframe this feature in only in terms of the 'embedded bare repo' attack.
 * Other docs improvements (thanks Stolee in particular!)
 * Protected config no longer uses read_very_early_config() and is only read
   once
 * Protected config now includes "-c"
 * uploadpack.packObjectsHook now uses protected config instead of ignoring
   repo config using config scopes

Changes in v2:

 * Rename safe.barerepository to discovery.bare and make it die()
 * Move tests into t/t0034-discovery-bare.sh
 * Avoid unnecessary config reading by using a static variable
 * Add discovery.bare=cwd
 * Fix typos

= Future work

 * This series doesn't implement config lookup functions for protected
   config. This will be done in a follow up series.
 * This series does not implement the "no-embedded" option [3] and I won't
   work on it any time soon, but I'd be more than happy to review if someone
   sends patches.
 * With discovery.bare, if a builtin is marked RUN_SETUP_GENTLY, setup.c
   doesn't die() and we don't tell users why their repository was rejected,
   e.g. "git config" gives an opaque "fatal: not in a git directory". This
   isn't a new problem though, since safe.directory has the same issue.

[1] https://lore.kernel.org/git/5ps2q552-1rr3-7161-4181-31556pp2ns12@tzk.qr

[2] https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com

[3] This was first suggested in https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Glen Choo (5):
  Documentation/git-config.txt: add SCOPES section
  Documentation: define protected configuration
  config: learn `git_protected_config()`
  safe.directory: use git_protected_config()
  setup.c: create `safe.bareRepository`

 Documentation/config/safe.txt       | 25 +++++++-
 Documentation/config/uploadpack.txt |  6 +-
 Documentation/git-config.txt        | 95 ++++++++++++++++++++++-------
 config.c                            | 43 +++++++++++++
 config.h                            | 16 +++++
 setup.c                             | 59 +++++++++++++++++-
 t/t0033-safe-directory.sh           | 24 +++-----
 t/t0035-safe-bare-repository.sh     | 54 ++++++++++++++++
 t/t5544-pack-objects-hook.sh        |  7 ++-
 upload-pack.c                       | 27 +++++---
 10 files changed, 300 insertions(+), 56 deletions(-)
 create mode 100755 t/t0035-safe-bare-repository.sh

base-commit: 4e2a4d1dd44367d7783f33b169698f2930ff13c0
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1261%2Fchooglen%2Fsetup%2Fdisable-bare-repo-config-v8
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1261/chooglen/setup/disable-bare-repo-config-v8
Pull-Request: https://github.com/git/git/pull/1261

Range-diff vs v7:

 1:  5c58db3bb21 ! 1:  6147751c9c1 Documentation/git-config.txt: add SCOPES section
     @@ Commit message
          145d59f482 (config: add '--show-scope' to print the scope of a config
          value, 2020-02-10) introduced the word "scope" to our public docs, but
          that only enumerates the scopes and assumes the user can figure out
     -    those values mean.
     +    what those values mean.
      
          Add a SCOPES section to Documentation/git-config.txt that describes the
          configuration scopes, their corresponding CLI options, and mentions that
     @@ Documentation/git-config.txt: The default is to use a pager.
       
       $(prefix)/etc/gitconfig::
       	System-wide configuration file.
     + 
     + $XDG_CONFIG_HOME/git/config::
     +-	Second user-specific configuration file. If $XDG_CONFIG_HOME is not set
     +-	or empty, `$HOME/.config/git/config` will be used. Any single-valued
     +-	variable set in this file will be overwritten by whatever is in
     +-	`~/.gitconfig`.  It is a good idea not to create this file if
     +-	you sometimes use older versions of Git, as support for this
     +-	file was added fairly recently.
     +-
     + ~/.gitconfig::
     +-	User-specific configuration file. Also called "global"
     +-	configuration file.
     ++	User-specific configuration files. When the XDG_CONFIG_HOME environment
     ++	variable is not set or empty, $HOME/.config/ is used as
     ++	$XDG_CONFIG_HOME.
     +++
     ++These are also called "global" configuration files. If both files exist, both
     ++files are read in the order given above.
     + 
     + $GIT_DIR/config::
     + 	Repository specific configuration file.
      @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
       	This is optional and is only searched when
       	`extensions.worktreeConfig` is present in $GIT_DIR/config.
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
      +git command by using the `-c` option. See linkgit:git[1] for details.
      +
      +Options will be read from all of these files that are available. If the
     -+global or the system-wide configuration file are not available they will be
     -+ignored. If the repository configuration file is not available or readable,
     -+'git config' will exit with a non-zero error code. Note that neither case
     -+produces an error message.
     ++global or the system-wide configuration files are missing or unreadable they
     ++will be ignored. If the repository configuration file is missing or unreadable,
     ++'git config' will exit with a non-zero error code. An error message is produced
     ++if the file is unreadable, but not if it is missing.
       
       The files are read in the order given above, with last value found taking
       precedence over values read earlier.  When multiple values are taken then all
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
      +	$GIT_DIR/config.worktree
      +
      +command::
     -+	environment variables
     ++	GIT_CONFIG_{COUNT,KEY,VALUE} environment variables (see <<ENVIRONMENT>>
     ++	below)
      ++
      +the `-c` option
      +
     @@ Documentation/git-config.txt: $GIT_DIR/config.worktree::
      +defined in, but some options are only respected in certain scopes. See the
      +respective option's documentation for the full details.
       
     ++[[ENVIRONMENT]]
       ENVIRONMENT
       -----------
     + 
 2:  58f25612aa3 ! 2:  df8a1a78d53 Documentation: define protected configuration
     @@ Commit message
          - System, global and command line config are considered protected
            because an attacker who has control over any of those can do plenty of
            harm without Git, so we gain very little by ignoring those scopes.
     +
          - On the other hand, local (and similarly, worktree) config are not
            considered protected because it is relatively easy for an attacker to
            control local config, e.g.:
     +
            - On some shared user environments, a non-admin attacker can create a
              repository high up the directory hierarchy (e.g. C:\.git on
              Windows), and a user may accidentally use it when their PS1
     @@ Commit message
              the user intended to use the shared repository. It obviously
              shouldn't be read from the repository, because that would end up
              trusting the repository that Git was supposed to reject.
     +
            - "git upload-pack" is expected to run in repositories that may not be
              controlled by the user. We cannot ignore all config in that
              repository (because "git upload-pack" would fail), but we can limit
     @@ Documentation/git-config.txt: Most configuration options are respected regardles
      +substantial harm without using Git, so it is assumed that the user's environment
      +protects these scopes against attackers.
      +
     + [[ENVIRONMENT]]
       ENVIRONMENT
       -----------
     - 
 3:  3683d20f232 ! 3:  30ac73716cb config: learn `git_protected_config()`
     @@ Commit message
      
          `uploadpack.packObjectsHook` is the only 'protected configuration only'
          variable today, but we've noted that `safe.directory` and the upcoming
     -    `discovery.bare` should also be 'protected configuration only'. So, for
     -    consistency, we'd like to have a single implementation for protected
     +    `safe.bareRepository` should also be 'protected configuration only'. So,
     +    for consistency, we'd like to have a single implementation for protected
          configuration.
      
          The primary constraints are:
      
          1. Reading from protected configuration should be fast. Nearly all "git"
             commands inside a bare repository will read both `safe.directory` and
     -       `discovery.bare`, so we cannot afford to be slow.
     +       `safe.bareRepository`, so we cannot afford to be slow.
      
          2. Protected configuration must be readable when the gitdir is not
     -       known. `safe.directory` and `discovery.bare` both affect repository
     -       discovery and the gitdir is not known at that point [1].
     +       known. `safe.directory` and `safe.bareRepository` both affect
     +       repository discovery and the gitdir is not known at that point [1].
      
          The chosen implementation in this commit is to read protected
          configuration and cache the values in a global configset. This is
 4:  6394818ffd8 ! 4:  b3256d68f84 safe.directory: use git_protected_config()
     @@ Documentation/config/safe.txt: via `git config --add`. To reset the list of safe
       path relative to the home directory and `%(prefix)/<path>` expands to a
      
       ## setup.c ##
     -@@ setup.c: static int ensure_valid_ownership(const char *path)
     - 	    is_path_owned_by_current_user(path))
     - 		return 1;
     - 
     +@@ setup.c: static int ensure_valid_ownership(const char *gitfile,
     + 	 * constant regardless of what failed above. data.is_safe should be
     + 	 * initialized to false, and might be changed by the callback.
     + 	 */
      -	read_very_early_config(safe_directory_cb, &data);
      +	git_protected_config(safe_directory_cb, &data);
       
     @@ t/t0033-safe-directory.sh: test_expect_success 'safe.directory is not set' '
       
      -test_expect_success 'ignoring safe.directory on the command line' '
      -	test_must_fail git -c safe.directory="$(pwd)" status 2>err &&
     --	grep "unsafe repository" err
     +-	grep "dubious ownership" err
      +test_expect_success 'safe.directory on the command line' '
      +	git -c safe.directory="$(pwd)" status
       '
     @@ t/t0033-safe-directory.sh: test_expect_success 'safe.directory is not set' '
      -		GIT_CONFIG_KEY_0="safe.directory" \
      -		GIT_CONFIG_VALUE_0="$(pwd)" \
      -		git status 2>err &&
     --	grep "unsafe repository" err
     +-	grep "dubious ownership" err
      +test_expect_success 'safe.directory in the environment' '
      +	env GIT_CONFIG_COUNT=1 \
      +	    GIT_CONFIG_KEY_0="safe.directory" \
     @@ t/t0033-safe-directory.sh: test_expect_success 'safe.directory is not set' '
      -	test_must_fail env \
      -		GIT_CONFIG_PARAMETERS="${SQ}safe.directory${SQ}=${SQ}$(pwd)${SQ}" \
      -		git status 2>err &&
     --	grep "unsafe repository" err
     +-	grep "dubious ownership" err
      +test_expect_success 'safe.directory in GIT_CONFIG_PARAMETERS' '
      +	env GIT_CONFIG_PARAMETERS="${SQ}safe.directory${SQ}=${SQ}$(pwd)${SQ}" \
      +	    git status
 5:  eff4b07480e ! 5:  50069bba9a5 setup.c: create `discovery.bare`
     @@ Metadata
      Author: Glen Choo <chooglen@google.com>
      
       ## Commit message ##
     -    setup.c: create `discovery.bare`
     +    setup.c: create `safe.bareRepository`
      
          There is a known social engineering attack that takes advantage of the
          fact that a working tree can include an entire bare repository,
     @@ Commit message
          `--git-dir` or `GIT_DIR`. In environments that don't use bare
          repositories, this would be minimally disruptive.
      
     -    Create a config variable, `discovery.bare`, that tells Git whether or
     -    not to die() when it discovers a bare repository. This only affects
     -    repository discovery, thus it has no effect if discovery was not
     -    done, e.g. if the user passes `--git-dir=my-dir`, discovery will be
     -    skipped and my-dir will be used as the repo regardless of the
     -    `discovery.bare` value.
     +    Create a config variable, `safe.bareRepository`, that tells Git whether
     +    or not to die() when working with a bare repository. This config is an
     +    enum of:
      
     -    This config is an enum of:
     -
     -    - "always": always allow bare repositories (this is the default)
     -    - "never": never allow bare repositories
     +    - "all": allow all bare repositories (this is the default)
     +    - "explicit": only allow bare repositories specified via --git-dir
     +      or GIT_DIR.
      
          If we want to protect users from such attacks by default, neither value
     -    will suffice - "always" provides no protection, but "never" is
     +    will suffice - "all" provides no protection, but "explicit" is
          impractical for bare repository users. A more usable default would be to
          allow only non-embedded bare repositories ([2] contains one such
          proposal), but detecting if a repository is embedded is potentially
     @@ Commit message
      
          Signed-off-by: Glen Choo <chooglen@google.com>
      
     - ## Documentation/config.txt ##
     -@@ Documentation/config.txt: include::config/diff.txt[]
     - 
     - include::config/difftool.txt[]
     - 
     -+include::config/discovery.txt[]
     -+
     - include::config/extensions.txt[]
     - 
     - include::config/fastimport.txt[]
     -
     - ## Documentation/config/discovery.txt (new) ##
     + ## Documentation/config/safe.txt ##
      @@
     -+discovery.bare::
     -+	Specifies whether Git will work with a bare repository that
     -+	wasn't specified via the top-level `--git-dir` command-line
     -+	option, or the `GIT_DIR` environment variable (see
     -+	linkgit:git[1]). If the repository is specified, Git will always
     -+	use the specified repository, regardless of this value.
     ++safe.bareRepository::
     ++	Specifies which bare repositories Git will work with. The currently
     ++	supported values are:
     +++
     ++* `all`: Git works with all bare repositories. This is the default.
     ++* `explicit`: Git only works with bare repositories specified via
     ++  the top-level `--git-dir` command-line option, or the `GIT_DIR`
     ++  environment variable (see linkgit:git[1]).
     +++
     ++If you do not use bare repositories in your workflow, then it may be
     ++beneficial to set `safe.bareRepository` to `explicit` in your global
     ++config. This will protect you from attacks that involve cloning a
     ++repository that contains a bare repository and running a Git command
     ++within that directory.
      ++
      +This config setting is only respected in protected configuration (see
      +<<SCOPES>>). This prevents the untrusted repository from tampering with
      +this value.
     -++
     -+The currently supported values are:
     -++
     -+* `always`: Git always works with bare repositories
     -+* `never`: Git never works with bare repositories
     -++
     -+If you do not use bare repositories in your workflow, then it may be
     -+beneficial to set `discovery.bare` to `never` in your global config.
     -+This will protect you from attacks that involve cloning a repository
     -+that contains a bare repository and running a Git command within that
     -+directory.
     ++
     + safe.directory::
     + 	These config entries specify Git-tracked directories that are
     + 	considered safe even if they are owned by someone other than the
      
       ## setup.c ##
      @@
       static int inside_git_dir = -1;
       static int inside_work_tree = -1;
       static int work_tree_config_is_bogus;
     -+enum discovery_bare_allowed {
     -+	DISCOVERY_BARE_NEVER = 0,
     -+	DISCOVERY_BARE_ALWAYS,
     ++enum allowed_bare_repo {
     ++	ALLOWED_BARE_REPO_EXPLICIT = 0,
     ++	ALLOWED_BARE_REPO_ALL,
      +};
       
       static struct startup_info the_startup_info;
       struct startup_info *startup_info = &the_startup_info;
     -@@ setup.c: static int ensure_valid_ownership(const char *path)
     +@@ setup.c: static int ensure_valid_ownership(const char *gitfile,
       	return data.is_safe;
       }
       
     -+static int discovery_bare_cb(const char *key, const char *value, void *d)
     ++static int allowed_bare_repo_cb(const char *key, const char *value, void *d)
      +{
     -+	enum discovery_bare_allowed *discovery_bare_allowed = d;
     ++	enum allowed_bare_repo *allowed_bare_repo = d;
      +
     -+	if (strcmp(key, "discovery.bare"))
     ++	if (strcasecmp(key, "safe.bareRepository"))
      +		return 0;
      +
     -+	if (!strcmp(value, "never")) {
     -+		*discovery_bare_allowed = DISCOVERY_BARE_NEVER;
     ++	if (!strcmp(value, "explicit")) {
     ++		*allowed_bare_repo = ALLOWED_BARE_REPO_EXPLICIT;
      +		return 0;
      +	}
     -+	if (!strcmp(value, "always")) {
     -+		*discovery_bare_allowed = DISCOVERY_BARE_ALWAYS;
     ++	if (!strcmp(value, "all")) {
     ++		*allowed_bare_repo = ALLOWED_BARE_REPO_ALL;
      +		return 0;
      +	}
      +	return -1;
      +}
      +
     -+static enum discovery_bare_allowed get_discovery_bare(void)
     ++static enum allowed_bare_repo get_allowed_bare_repo(void)
      +{
     -+	enum discovery_bare_allowed result = DISCOVERY_BARE_ALWAYS;
     -+	git_protected_config(discovery_bare_cb, &result);
     ++	enum allowed_bare_repo result = ALLOWED_BARE_REPO_ALL;
     ++	git_protected_config(allowed_bare_repo_cb, &result);
      +	return result;
      +}
      +
     -+static const char *discovery_bare_allowed_to_string(
     -+	enum discovery_bare_allowed discovery_bare_allowed)
     ++static const char *allowed_bare_repo_to_string(
     ++	enum allowed_bare_repo allowed_bare_repo)
      +{
     -+	switch (discovery_bare_allowed) {
     -+	case DISCOVERY_BARE_NEVER:
     -+		return "never";
     -+	case DISCOVERY_BARE_ALWAYS:
     -+		return "always";
     ++	switch (allowed_bare_repo) {
     ++	case ALLOWED_BARE_REPO_EXPLICIT:
     ++		return "explicit";
     ++	case ALLOWED_BARE_REPO_ALL:
     ++		return "all";
      +	default:
     -+		BUG("invalid discovery_bare_allowed %d",
     -+		    discovery_bare_allowed);
     ++		BUG("invalid allowed_bare_repo %d",
     ++		    allowed_bare_repo);
      +	}
      +	return NULL;
      +}
     @@ setup.c: static enum discovery_result setup_git_directory_gently_1(struct strbuf
       		}
       
       		if (is_git_directory(dir->buf)) {
     -+			if (get_discovery_bare() == DISCOVERY_BARE_NEVER)
     ++			if (get_allowed_bare_repo() == ALLOWED_BARE_REPO_EXPLICIT)
      +				return GIT_DIR_DISALLOWED_BARE;
     - 			if (!ensure_valid_ownership(dir->buf))
     + 			if (!ensure_valid_ownership(NULL, NULL, dir->buf))
       				return GIT_DIR_INVALID_OWNERSHIP;
       			strbuf_addstr(gitdir, ".");
      @@ setup.c: const char *setup_git_directory_gently(int *nongit_ok)
     @@ setup.c: const char *setup_git_directory_gently(int *nongit_ok)
       		break;
      +	case GIT_DIR_DISALLOWED_BARE:
      +		if (!nongit_ok) {
     -+			die(_("cannot use bare repository '%s' (discovery.bare is '%s')"),
     ++			die(_("cannot use bare repository '%s' (safe.bareRepository is '%s')"),
      +			    dir.buf,
     -+			    discovery_bare_allowed_to_string(get_discovery_bare()));
     ++			    allowed_bare_repo_to_string(get_allowed_bare_repo()));
      +		}
      +		*nongit_ok = 1;
      +		break;
     @@ setup.c: const char *setup_git_directory_gently(int *nongit_ok)
       		/*
       		 * As a safeguard against setup_git_directory_gently_1 returning
      
     - ## t/t0035-discovery-bare.sh (new) ##
     + ## t/t0035-safe-bare-repository.sh (new) ##
      @@
      +#!/bin/sh
      +
     -+test_description='verify discovery.bare checks'
     ++test_description='verify safe.bareRepository checks'
      +
      +TEST_PASSES_SANITIZE_LEAK=true
      +. ./test-lib.sh
     @@ t/t0035-discovery-bare.sh (new)
      +	git init --bare outer-repo/bare-repo
      +'
      +
     -+test_expect_success 'discovery.bare unset' '
     ++test_expect_success 'safe.bareRepository unset' '
      +	expect_accepted -C outer-repo/bare-repo
      +'
      +
     -+test_expect_success 'discovery.bare=always' '
     -+	test_config_global discovery.bare always &&
     ++test_expect_success 'safe.bareRepository=all' '
     ++	test_config_global safe.bareRepository all &&
      +	expect_accepted -C outer-repo/bare-repo
      +'
      +
     -+test_expect_success 'discovery.bare=never' '
     -+	test_config_global discovery.bare never &&
     ++test_expect_success 'safe.bareRepository=explicit' '
     ++	test_config_global safe.bareRepository explicit &&
      +	expect_rejected -C outer-repo/bare-repo
      +'
      +
     -+test_expect_success 'discovery.bare in the repository' '
     -+	# discovery.bare must not be "never", otherwise git config fails
     -+	# with "fatal: not in a git directory" (like safe.directory)
     -+	test_config -C outer-repo/bare-repo discovery.bare always &&
     -+	test_config_global discovery.bare never &&
     ++test_expect_success 'safe.bareRepository in the repository' '
     ++	# safe.bareRepository must not be "explicit", otherwise
     ++	# git config fails with "fatal: not in a git directory" (like
     ++	# safe.directory)
     ++	test_config -C outer-repo/bare-repo safe.bareRepository \
     ++		all &&
     ++	test_config_global safe.bareRepository explicit &&
      +	expect_rejected -C outer-repo/bare-repo
      +'
      +
     -+test_expect_success 'discovery.bare on the command line' '
     -+	test_config_global discovery.bare never &&
     ++test_expect_success 'safe.bareRepository on the command line' '
     ++	test_config_global safe.bareRepository explicit &&
      +	expect_accepted -C outer-repo/bare-repo \
     -+		-c discovery.bare=always
     ++		-c safe.bareRepository=all
      +'
      +
      +test_done

From: Glen Choo <chooglen@google.com>

In a subsequent commit, we will introduce "protected configuration", which is easiest to describe in terms of configuration scopes (i.e. it's the union of the 'system', 'global', and 'command' scopes). This description is fine for ML discussions, but it's inadequate for end users because we don't provide a good description of "configuration scopes" in the public docs.

145d59f482 (config: add '--show-scope' to print the scope of a config value, 2020-02-10) introduced the word "scope" to our public docs, but that only enumerates the scopes and assumes the user can figure out what those values mean.

Add a SCOPES section to Documentation/git-config.txt that describes the configuration scopes, their corresponding CLI options, and mentions that some configuration options are only respected in certain scopes. Then, use the word "scope" to simplify the FILES section and change some confusing wording.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/git-config.txt | 82 ++++++++++++++++++++++++++----------
 1 file changed, 59 insertions(+), 23 deletions(-)

From: Glen Choo <chooglen@google.com>

For security reasons, there are config variables that are only trusted when they are specified in certain configuration scopes, which are sometimes referred to on-list as 'protected configuration' [1]. A future commit will introduce another such variable, so let's define our terms so that we can have consistent documentation and implementation.

In our documentation, define 'protected configuration' as the system, global and command config scopes. As a shorthand, I will refer to variables that are only respected in protected configuration as 'protected configuration only', but this term is not used in the documentation.

This definition of protected configuration is based on whether or not Git can reasonably protect the user by ignoring the configuration scope:

- System, global and command line config are considered protected
  because an attacker who has control over any of those can do plenty of
  harm without Git, so we gain very little by ignoring those scopes.

- On the other hand, local (and similarly, worktree) config are not
  considered protected because it is relatively easy for an attacker to
  control local config, e.g.:

  - On some shared user environments, a non-admin attacker can create a
    repository high up the directory hierarchy (e.g. C:\.git on
    Windows), and a user may accidentally use it when their PS1
    automatically invokes "git" commands.

    `safe.directory` prevents attacks of this form by making sure that
    the user intended to use the shared repository. It obviously
    shouldn't be read from the repository, because that would end up
    trusting the repository that Git was supposed to reject.

  - "git upload-pack" is expected to run in repositories that may not be
    controlled by the user. We cannot ignore all config in that
    repository (because "git upload-pack" would fail), but we can limit
    the risks by ignoring `uploadpack.packObjectsHook`.

Only `uploadpack.packObjectsHook` is 'protected configuration only'. The following variables are intentionally excluded:

- `safe.directory` should be 'protected configuration only', but it does
  not technically fit the definition because it is not respected in the
  "command" scope. A future commit will fix this.

- `trace2.*` happens to read the same scopes as `safe.directory` because
  they share an implementation. However, this is not for security
  reasons; it is because we want to start tracing so early that
  repository-level config and "-c" are not available [2].

  This requirement is unique to `trace2.*`, so it does not makes sense
  for protected configuration to be subject to the same constraints.

[1] For example, https://lore.kernel.org/git/6af83767-576b-75c4-c778-0284344a8fe7@github.com/ [2] https://lore.kernel.org/git/a0c89d0d-669e-bf56-25d2-cbb09b012e70@jeffhostetler.com/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/uploadpack.txt |  6 +++---
 Documentation/git-config.txt        | 13 +++++++++++++
 2 files changed, 16 insertions(+), 3 deletions(-)

From: Glen Choo <chooglen@google.com>

`uploadpack.packObjectsHook` is the only 'protected configuration only' variable today, but we've noted that `safe.directory` and the upcoming `safe.bareRepository` should also be 'protected configuration only'. So, for consistency, we'd like to have a single implementation for protected configuration.

The primary constraints are:

1. Reading from protected configuration should be fast. Nearly all "git"
   commands inside a bare repository will read both `safe.directory` and
   `safe.bareRepository`, so we cannot afford to be slow.

2. Protected configuration must be readable when the gitdir is not
   known. `safe.directory` and `safe.bareRepository` both affect
   repository discovery and the gitdir is not known at that point [1].

The chosen implementation in this commit is to read protected configuration and cache the values in a global configset. This is similar to the caching behavior we get with the_repository->config.

Introduce git_protected_config(), which reads protected configuration and caches them in the global configset protected_config. Then, refactor `uploadpack.packObjectsHook` to use git_protected_config().

The protected configuration functions are named similarly to their non-protected counterparts, e.g. git_protected_config_check_init() vs git_config_check_init().

In light of constraint 1, this implementation can still be improved. git_protected_config() iterates through every variable in protected_config, which is wasteful, but it makes the conversion simple because it matches existing patterns. We will likely implement constant time lookup functions for protected configuration in a future series (such functions already exist for non-protected configuration, i.e. repo_config_get_*()).

An alternative that avoids introducing another configset is to continue to read all config using git_config(), but only accept values that have the correct config scope [2]. This technically fulfills constraint 2, because git_config() simply ignores the local and worktree config when the gitdir is not known. However, this would read incomplete config into the_repository->config, which would need to be reset when the gitdir is known and git_config() needs to read the local and worktree config. Resetting the_repository->config might be reasonable while we only have these 'protected configuration only' variables, but it's not clear whether this extends well to future variables.

[1] In this case, we do have a candidate gitdir though, so with a little refactoring, it might be possible to provide a gitdir. [2] This is how `uploadpack.packObjectsHook` was implemented prior to this commit.

Signed-off-by: Glen Choo <chooglen@google.com>
---
 config.c                     | 43 ++++++++++++++++++++++++++++++++++++
 config.h                     | 16 ++++++++++++++
 t/t5544-pack-objects-hook.sh |  7 +++++-
 upload-pack.c                | 27 +++++++++++++---------
 4 files changed, 82 insertions(+), 11 deletions(-)

On Thu, Jul 14 2022, Glen Choo via GitGitGadget wrote:

Noticed after it landed on master: This change fails with:

	make SANITIZE=address test T=t0410*.sh

Running that manually shows that we fail like this:
	
	$ cat trash\ directory.t0410-partial-clone/httpd/error.log | grep -o AH0.*
	AH00163: Apache/2.4.54 (Debian) configured -- resuming normal operations
	AH00094: Command line: '/usr/sbin/apache2 -d /home/avar/g/git/t/trash directory.t0410-partial-clone/httpd -f /home/avar/g/git/t/lib-httpd/apache.conf -c Listen 127.0.0.1:10410'
	AH01215: AddressSanitizer:DEADLYSIGNAL: /home/avar/g/git/git-http-backend
	AH01215: =================================================================: /home/avar/g/git/git-http-backend
	AH01215: ==27820==ERROR: AddressSanitizer: SEGV on unknown address 0x000000000000 (pc 0x7f7af5dc0d66 bp 0x7fff11964450 sp 0x7fff11963be8 T0): /home/avar/g/git/git-http-backend
	AH01215: ==27820==The signal is caused by a READ memory access.: /home/avar/g/git/git-http-backend
	AH01215: ==27820==Hint: address points to the zero page.: /home/avar/g/git/git-http-backend
	AH01215:     #0 0x7f7af5dc0d66 in __sanitizer::internal_strlen(char const*) ../../../../src/libsanitizer/sanitizer_common/sanitizer_libc.cpp:167: /home/avar/g/git/git-http-backend
	AH01215:     #1 0x7f7af5d512f2 in __interceptor_fopen64 ../../../../src/libsanitizer/sanitizer_common/sanitizer_common_interceptors.inc:6220: /home/avar/g/git/git-http-backend
	AH01215:     #2 0x562a65e37cc8 in git_fopen compat/fopen.c:22: /home/avar/g/git/git-http-backend
	AH01215:     #3 0x562a65df3879 in fopen_or_warn wrapper.c:431: /home/avar/g/git/git-http-backend
	AH01215:     #4 0x562a65a12476 in git_config_from_file_with_options config.c:1982: /home/avar/g/git/git-http-backend
	AH01215:     #5 0x562a65a124f4 in git_config_from_file config.c:1993: /home/avar/g/git/git-http-backend
	AH01215:     #6 0x562a65a15288 in git_configset_add_file config.c:2389: /home/avar/g/git/git-http-backend
	AH01215:     #7 0x562a65a16a37 in read_protected_config config.c:2649: /home/avar/g/git/git-http-backend
	AH01215:     #8 0x562a65a16b5c in git_protected_config config.c:2661: /home/avar/g/git/git-http-backend
	AH01215:     #9 0x562a65dd9f9a in get_upload_pack_config upload-pack.c:1342: /home/avar/g/git/git-http-backend
	AH01215:     #10 0x562a65ddc1cb in upload_pack_v2 upload-pack.c:1706: /home/avar/g/git/git-http-backend
	AH01215:     #11 0x562a65d2eb8a in process_request serve.c:308: /home/avar/g/git/git-http-backend
	AH01215:     #12 0x562a65d2ec18 in protocol_v2_serve_loop serve.c:323: /home/avar/g/git/git-http-backend
	AH01215:     #13 0x562a6593c5ae in cmd_upload_pack builtin/upload-pack.c:55: /home/avar/g/git/git-http-backend
	AH01215:     #14 0x562a656cf8ff in run_builtin git.c:466: /home/avar/g/git/git-http-backend
	AH01215:     #15 0x562a656d02ab in handle_builtin git.c:720: /home/avar/g/git/git-http-backend
	AH01215:     #16 0x562a656d09d5 in run_argv git.c:787: /home/avar/g/git/git-http-backend
	AH01215:     #17 0x562a656d174f in cmd_main git.c:920: /home/avar/g/git/git-http-backend
	AH01215:     #18 0x562a6594b0b9 in main common-main.c:56: /home/avar/g/git/git-http-backend
	AH01215:     #19 0x7f7af5a5681c in __libc_start_main ../csu/libc-start.c:332: /home/avar/g/git/git-http-backend
	AH01215:     #20 0x562a656cb209 in _start (git+0x1d1209): /home/avar/g/git/git-http-backend
	AH01215: : /home/avar/g/git/git-http-backend
	AH01215: AddressSanitizer can not provide additional info.: /home/avar/g/git/git-http-backend
	AH01215: SUMMARY: AddressSanitizer: SEGV ../../../../src/libsanitizer/sanitizer_common/sanitizer_libc.cpp:167 in __sanitizer::internal_strlen(char const*): /home/avar/g/git/git-http-backend
	AH01215: ==27820==ABORTING: /home/avar/g/git/git-http-backend
	AH01215: error: upload-pack died of signal 6: /home/avar/g/git/git-http-backend

(We really should have a SANITIZE=address in CI, but it takes a while...)

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

Thanks. I narrowed the failure down to the hunk above, specifically this line:

  git_configset_add_file(&protected_config, xdg_config);

Since xdg_config can be NULL, this results in the failing call fopen_or_warn(NULL, "r").

This logic was lifted from do_git_config_sequence(), which checks that each of the paths are not NULL. So a fix might be something like:

----- >8 --------- >8 --------- >8 --------- >8 --------- >8 ----

  diff --git a/config.c b/config.c
  index 015bec360f..208a3dd7a7 100644
  --- a/config.c
  +++ b/config.c
  @@ -2645,9 +2645,13 @@ static void read_protected_config(void)
    system_config = git_system_config();
    git_global_config(&user_config, &xdg_config);

  -	git_configset_add_file(&protected_config, system_config);
  -	git_configset_add_file(&protected_config, xdg_config);
  -	git_configset_add_file(&protected_config, user_config);
  +
  +	if (system_config)
  +		git_configset_add_file(&protected_config, system_config);
  +	if (xdg_config)
  +		git_configset_add_file(&protected_config, xdg_config);
  +	if (user_config)
  +		git_configset_add_file(&protected_config, user_config);
    git_configset_add_parameters(&protected_config);

    free(system_config);

----- >8 --------- >8 --------- >8 --------- >8 --------- >8 ----

I'm not sure if system_config can ever be NULL, but (xdg|user)_config is NULL when $HOME is unset, and xdg_config is also unset if $GIT_CONFIG_GLOBAL is set.

On Mon, Jul 25 2022, Glen Choo wrote:

Not having looked into it much at all: Doesn't this then introduce another logic error where git_protected_config() is now buggy, i.e. it's a "lazy load" method where we'll expect to read_protected_config() first.

The assumption with that seems to have been that it's invariant within a single process, is that still the case, or can e.g. HOME be set during our runtime when we rely on these functions?

(I don't know)

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

I don't think this introduces an error, or at least, not one that we don't already have. This mimics do_git_config_sequence() (which also assumes this invariant), which is used under the hood by (git|repo)_read_config(),

In retrospect, it might have been a good idea to implement read_protected_config() using do_git_config_sequence() / config_with_options(); those functions are a bit bloated, but at least we'd only have one implementation.

From: Glen Choo <chooglen@google.com>

Use git_protected_config() to read `safe.directory` instead of read_very_early_config(), making it 'protected configuration only'.

As a result, `safe.directory` now respects "-c", so update the tests and docs accordingly. It used to ignore "-c" due to how it was implemented, not because of security or correctness concerns [1].

[1] https://lore.kernel.org/git/xmqqlevabcsu.fsf@gitster.g/

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/safe.txt |  6 +++---
 setup.c                       |  2 +-
 t/t0033-safe-directory.sh     | 24 ++++++++++--------------
 3 files changed, 14 insertions(+), 18 deletions(-)

From: Glen Choo <chooglen@google.com>

There is a known social engineering attack that takes advantage of the fact that a working tree can include an entire bare repository, including a config file. A user could run a Git command inside the bare repository thinking that the config file of the 'outer' repository would be used, but in reality, the bare repository's config file (which is attacker-controlled) is used, which may result in arbitrary code execution. See [1] for a fuller description and deeper discussion.

A simple mitigation is to forbid bare repositories unless specified via `--git-dir` or `GIT_DIR`. In environments that don't use bare repositories, this would be minimally disruptive.

Create a config variable, `safe.bareRepository`, that tells Git whether or not to die() when working with a bare repository. This config is an enum of:

- "all": allow all bare repositories (this is the default)
- "explicit": only allow bare repositories specified via --git-dir
  or GIT_DIR.

If we want to protect users from such attacks by default, neither value will suffice - "all" provides no protection, but "explicit" is impractical for bare repository users. A more usable default would be to allow only non-embedded bare repositories ([2] contains one such proposal), but detecting if a repository is embedded is potentially non-trivial, so this work is not implemented in this series.

[1]: https://lore.kernel.org/git/kl6lsfqpygsj.fsf@chooglen-macbookpro.roam.corp.google.com [2]: https://lore.kernel.org/git/5b969c5e-e802-c447-ad25-6acc0b784582@github.com

Signed-off-by: Glen Choo <chooglen@google.com>
---
 Documentation/config/safe.txt   | 19 +++++++++++
 setup.c                         | 57 ++++++++++++++++++++++++++++++++-
 t/t0035-safe-bare-repository.sh | 54 +++++++++++++++++++++++++++++++
 3 files changed, 129 insertions(+), 1 deletion(-)
 create mode 100755 t/t0035-safe-bare-repository.sh