On Sat, Jan 16 2021, Jonathan Tan wrote:

Someone at $DAYJOB noticed that if a .gitmodules-containing tree and the
.gitmodules blob itself are sent in 2 separate packfiles during a fetch
(which can happen when packfile URIs are used), transfer.fsckobjects
causes the fetch to fail. You can reproduce it as follows (as of the
time of writing):

  $ git -c fetch.uriprotocols=https -c transfer.fsckobjects=true clone https://chromium.googlesource.com/chromiumos/codesearch
  Cloning into 'codesearch'...
  remote: Total 2242 (delta 0), reused 2242 (delta 0)
  Receiving objects: 100% (2242/2242), 1.77 MiB | 4.62 MiB/s, done.
  error: object 1f155c20935ee1154a813a814f03ef2b3976680f: gitmodulesMissing: unable to read .gitmodules blob
  fatal: fsck error in pack objects
  fatal: index-pack failed

This happens because the fsck part is currently being done in
index-pack, which operates on one pack at a time. When index-pack sees
the tree, it runs fsck on it (like any other object), and the fsck
subsystem remembers the .gitmodules target (specifically, in
gitmodules_found in fsck.c). Later, index-pack runs fsck_finish() which
checks if the target exists, but it doesn't, so it reports the failure.

One option is for fetch to do its own pass of checking all downloaded
objects once all packfiles have been downloaded, but that seems wasteful
as all trees would have to be re-inflated.

Another option is to do it within the connectivity check instead - so,
update rev-list and the object walking mechanism to be able to detect
.gitmodules in trees and fsck the target blob whenever such an entry
occurs. This has the advantage that there is no extra re-inflation,
although it might be strange to have object walking be able to fsck.

The simplest solution would be to just relax this - check the blob if it
exists, but if it doesn't, it's OK. Some things in favor of this
solution:

 - This is something we already do in the partial clone case (although
   it could be argued that in this case, we're already trusting the
   server for far more than .gitmodules, so just because it's OK in the
   partial clone case doesn't mean that it's OK in the regular case).

 - Also, the commit message for this feature (from ed8b10f631 ("fsck: check
   .gitmodules content", 2018-05-21)) gives a rationale of a newer
   server being able to protect older clients.
    - Servers using receive-pack (instead of fetch-pack) to obtain
      objects would still be protected, since receive-pack still only
      accepts one packfile at a time (and there are currently no plans
      to expand this).
    - Also, malicious .gitobjects files could still be crafted that pass
      fsck checking - for example, by containing a URL (of another
      server) that refers to a repo with a .gitobjects that would fail
      fsck.

So I would rather go with just relaxing the check, but if consensus is
that we should still do it, I'll investigate doing it in the
connectivity check.

Would this still behave if the $DAYJOB's packfile-uri server support was
behaving as documented in packfile-uri.txt, or just because it has
outside-spec behavior?

I.e. the spec[1] says this:

    This is the implementation: a feature, marked experimental, that
    allows the server to be configured by one or more
    `uploadpack.blobPackfileUri=<sha1> <uri>` entries. Whenever the list
    of objects to be sent is assembled, all such blobs are excluded,
    replaced with URIs. The client will download those URIs, expecting
    them to each point to packfiles containing single blobs.

Which I can't see leaving an opening for more than packfile-uri being to
serve up packfiles which each contain a single blob.

In that case it seems to me we'd be OK (but I haven't tested), because
fsck_finish() will call read_object_file() which'll try to read that
"blob from the object store when it encounters the ".gitmodules" tree,
and because we'd have already downloaded the packfile with the blob
before moving onto the main dialog.

But as we discussed on-list before[2] this isn't the way packfile-uri
actually works in the wild. It's really just sending some arbitrary data
in a pack in that URI, with a server that knows what's in that pack and
will send the rest in such a way that everything ends up being
connected.

As far as I can tell the only reason this is called "packfile URI" and
behaves this way in git.git is because of the convenience of
intrumenting pack-objects.c with an "oidset excluded_by_config" to not
stream those blobs in a pack, but it isn't how the only (I'm pretty
sure) production server implementation in the wild behaves at all.

So *poke* about the reply I had in [3] late last year. I think the first
thing worth doing here is fixing the docs so they describe how this
works. You didn't get back on that (and I also forgot about it until
this thread), but it would be nice to know what you think about the
suggested prose there.

Re-reading it I'd add something like this to the spec:

 A. That the config is called "uploadpack.blobPackfileUri" in git.git
    has nothing to do with how this is expected to behave on the
    wire. It's just to serve the narrow support pack-objects.c has for
    crafting such a pack.

 B. It's then called "packfile-uris" on the wire, nothing to do with
    blobs. Just packs with a checksum that we'll validate. An older
    versions of this spec said "[a] packfiles containing single blobs"
    but it can be any combination of blob/tree/commit data.

 C. A client is then expected to deal with any combination of data
    ordered/sliced/split up etc. in any possible way from such a
    combination of "packfile-uris" and PACK dialog, as long as the end
    result is valid.

Except that the result of this discussion will perhaps be a more narrow
definition for "C".

1. https://github.com/git/git/blob/cd8402e0fd8cfc0ec9fb10e22ffb6aabd992eae1/Documentation/technical/packfile-uri.txt#L37-L41
2. https://lore.kernel.org/git/20201125190957.1113461-1-jonathantanmy@google.com/ (local)
3. https://lore.kernel.org/git/87tut5vghw.fsf@evledraar.gmail.com/ (local)