threads / bug / 66860

[BUG] push: a remote group that lists a remote of its own name makes "git push <name>" recurse

Subject: [BUG] push: a remote group that lists a remote of its own name makes "git push <name>" recurse

## tl;dr

A report shows that a group named after one of its own members starts thousands of nested push processes instead of failing. Read the story.

replies: 0people: 1as markdown or json

wheakerd· Nov 17, 2026, 01:07 UTC · lore
What did you do before the bug happened? (Steps to reproduce your issue)
  git init -q -b main client
  git init -q --bare -b main a.git
  git init -q --bare -b main b.git
  git -C client commit -q --allow-empty -m base
  git -C client remote add origin ../a.git
  git -C client remote add member ../b.git
  git -C client config remotes.origin "origin member"
  timeout -s KILL 12 git -C client push origin refs/heads/main:refs/heads/main
  Caution: the last command starts several thousand nested "git push"
  processes. Please keep the time limit.
What did you expect to happen? (Expected behavior)
  Either one push to the remote "origin" and one to "member", or an error
  that says a group must not contain its own name.
What happened instead? (Actual behavior)
  The command prints "Pushing to origin" again and again and starts one
  nested "git push origin" per line. In a 12-second run it printed that
  line 5,945 times, and I sampled up to 5,898 "git" processes alive at the
  same time. Once the nesting stops growing, each level reports "error:
  could not push to origin", then prints "Pushing to member" and pushes to
  member. When the 12 seconds were over, b.git had received main, and
  a.git, the repository that the remote "origin" names, had no ref.
  An earlier run without a time limit ended by itself with exit status 1.
  I did not measure how long it took. There, too, the other member had
  received the push and the repository of the remote "origin" had not.
What's different between what you expected and what actually happened?
  The remote that has the same name as the group is never pushed to, the
  other member is pushed to once per nesting level, and the number of
  processes is limited only by whatever makes the deepest child fail.
Anything else you want to add:
  Reading v2.56.0: push_multiple() in builtin/push.c runs
  "git push <member>" as a child process for every member of the group,
  and add_remote_or_group() in remote.c resolves a name as a group first
  when remotes.<name> exists. The child for the member "origin" therefore
  expands the same group again. On 2026-11-15 the two functions read
  the same way on master, and push_multiple() on next and seen still
  runs one child per member without a guard. I did not run the
  reproduction on anything newer than 2.56.0.
  "git push <group>" is new in 2.55.0. t/t5566-push-group.sh has no case
  for a group that contains its own name, and git-push(1) does not
  mention one.
  The fetch side has an older report of the same cycle:
    Fwd: Runaway "git remote" if group definition contains a remote by
    the same name
    https://mail-archive.com/git@vger.kernel.org/msg41422.html
  Alex Riesen first sent it on 2010-11-17 and re-posted it on
  2013-12-28, with
    git -c remotes.origin='origin other' remote update origin
  as the reproduction. Jeff King's reply names the reference cycle and
  sketches a fix: expand groups in one process, tell the child
  processes not to expand again, and detect cycles. I found no patch in
  that thread, and fetch_multiple() in builtin/fetch.c on master passes
  no such option to its children.
  A group named like one of its members is easy to write by accident, for
  example remotes.origin = "origin backup" to mean "origin and its
  backup".

[System Info] git version: git version 2.56.0 cpu: x86_64 no commit associated with this build sizeof-long: 8 sizeof-size_t: 8 shell-path: /bin/sh rust: disabled feature: fsmonitor--daemon uname: Linux 7.2.11-200.fc44.x86_64 x86_64

[Enabled Hooks] none in the reproduction repository

← back to recent threads