Let's immediately start with a scrap of the pre-receive hook that I've already written:
#!/bin/sh
##
format_bold='\033[1m'
format_red='\033[31m'
format_yellow='\033[33m'
format_normal='\033[0m'
##
format_error="${format_bold}${format_red}%s${format_normal}"
format_warning="${format_bold}${format_yellow}%s${format_normal}"
##
stdout() {
format="${1}"
shift
printf "${format}" "${@}"
}
##
stderr() {
stdout "${@}" 1>&2
}
##
output() {
format="${1}"
shift
stdout "${format}\n" "${@}"
}
##
error() {
format="${1}"
shift
stderr "${format_error}: ${format}\n" 'error' "${@}"
}
##
warning() {
format="${1}"
shift
stdout "${format_warning}: ${format}\n" 'warning' "${@}"
}
##
die() {
error "${@}"
exit 1
}
##
git() {
command git --no-pager "${@}"
}
##
list() {
git rev-list "${@}"
}
##
clang_format() {
git clang-format --style='file' "${@}"
}
##
while read sha1_old sha1_new ref; do
case "${ref}" in
refs/heads/*)
branch="$(expr "${ref}" : 'refs/heads/\(.*\)')"
if [ "$(expr "${sha1_new}" : '0*$')" -ne 0 ]; then # delete
unset sha1_new
# ...
else # update
if [ "$(expr "${sha1_old}" : '0*$')" -ne 0 ]; then # create
unset sha1_old
sha1_range="${sha1_new}"
else
sha1_range="${sha1_old}..${sha1_new}"
# ...
fi
fi
# ...
GIT_WORK_TREE="$(mktemp --tmpdir -d 'gitXXXXXX')"
export GIT_WORK_TREE
GIT_DIR="${GIT_WORK_TREE}/.git"
export GIT_DIR
mkdir -p "${GIT_DIR}"
cp -a * "${GIT_DIR}/"
ln -s "${PWD}/../.clang-format" "${GIT_WORK_TREE}/"
error=
for sha1 in $(list "${sha1_range}"); do
git checkout --force "${sha1}" > '/dev/null' 2>&1
if [ "$(list --count "${sha1}")" -eq 1 ]; then
# What should I put here?
else
git reset --soft 'HEAD~1' > '/dev/null' 2>&1
fi
diff="$(clang_format --diff)"
if [ "${diff%% *}" = 'diff' ]; then
error=1
error '%s: %s\n%s' \
'Code style issues detected' \
"${sha1}" \
"${diff}" \
1>&2
fi
done
if [ -n "${error}" ]; then
die '%s' 'Code style issues detected'
fi
fi
;;
refs/tags/*)
tag="$(expr "${ref}" : 'refs/tags/\(.*\)')"
# ...
;;
*)
# ...
;;
esac
done
exit 0
NOTE:
Places with irrelevant code are stubbed with # ....
NOTE:
If you are not familiar with git-clang-format, take a look here.
That hook works as expected, and so far, I didn't notice any bugs, but if you spot any problem or have an improvement suggestion, I'd appreciate any report. Probably, I should give a comment on what's the intention behind this hook. Well, it does check every pushed revision for compliance with code style conventions using git-clang-format, and if any of them does not comply, it will output the relevant diff (the one telling developers what should be fixed) for each of them. Basically, I have two in-depth questions regarding this hook.
First, notice that I perform copy of the remote's (server) bare repository to some temporary directory and check out the code for analysis there. Let me explain the intention of this. Note that I do several git checkouts and git resets (due to for loop) in order to analyze all of the pushed revisions individually with git-clang-format. What I am trying to avoid here, is the (possible) concurrency issue on push access to the remote's (server) bare repository. That is, I'm under impression that if multiple developers will try to push at the same time to a remote with this pre-receive hook installed, that might cause problems if each of these push "sessions" does not do git checkouts and git resets with its private copy of the repository. So, to put it simple, does git-daemon have built-in lock management for concurrent push "sessions"? Will it execute the corresponding pre-receive hook instances strictly sequentially or there is a possibility of interleaving (which can potentially cause undefined behavior)? Something tells me that there should be a built-in solution for this problem with concrete guarantees, otherwise how would remotes work in general (even without complex hooks) being subjected to concurrent pushes? If there is such a built-in solution, then the copy is redundant and simply reusing the bare repository would actually speed up the processing. By the way, any reference to official documentation regarding this question is very welcome.
Second, git-clang-format processes only staged (but not committed) changes vs. specific commit (HEAD by default). Thus, you can easily see where a corner case lies. Yes, it's with the root commits (revisions). In fact, git reset --soft 'HEAD~1' cannot be applied to root commits as they have no parents to reset to. Hence, the following check with my second question is there:
if [ "$(list --count "${sha1}")" -eq 1 ]; then
# What should I put here?
else
git reset --soft 'HEAD~1' > '/dev/null' 2>&1
fi
I've tried git update-ref -d 'HEAD' but this breaks the repository in such a way that git-clang-format is not able to process it anymore. I believe this is related to the fact that all of these pushed revisions that are being analyzed (including this root one) do not really belong to any branch yet. That is, they are in detached HEAD state. It would be perfect to find a solution to this corner case as well, so that initial commits can also undergo the same check by git-clang-format for compliance with code style conventions.
Peace.
