path: root/test

#!/bin/sh
# Drives the real image end to end against a local object store: the container
# is destroyed and the forge comes back from the bucket alone.
#
# Needs docker, go, git, curl and openssl here, nothing extra inside the image.
# fake-r2.go stands in for the object handler in worker.js, so no Cloudflare
# account is needed.
set -eu

IMAGE=${IMAGE:-git-forge-cf:smoke}
PORT=${PORT:-18081}
R2PORT=${R2PORT:-19000}
BASE=http://127.0.0.1:$PORT
NAME=git-forge-cf-smoke

WORK=$(mktemp -d)
BUCKET=$WORK/bucket
R2PID=""
cleanup() {
	docker rm -f "$NAME" >/dev/null 2>&1 || true
	[ -n "$R2PID" ] && kill "$R2PID" 2>/dev/null || true
	rm -rf "$WORK"
}
trap cleanup EXIT

ok() { printf 'ok   %s\n' "$1"; }
no() {
	printf 'FAIL %s\n' "$1" >&2
	docker logs "$NAME" 2>&1 | tail -25 >&2 || true
	exit 1
}
want() { grep -q -- "$2" "$3" || no "$1"; ok "$1"; }
code() {
	description=$1 expected=$2
	shift 2
	got=$(curl -s -o "$WORK/body" -w '%{http_code}' "$@")
	[ "$got" = "$expected" ] || no "$description: got $got, wanted $expected"
	ok "$description"
}
csrf() { grep -o 'name="csrf" value="[^"]*"' "$1" | head -1 | cut -d'"' -f4; }

# Same trick git-forge's own test uses: the secret is the base32 of the ascii
# "12345678901234567890", so a code needs openssl and no base32 decoder.
SECRET=GEZDGNBVGY3TQOJQGEZDGNBVGY3TQOJQ
KEYHEX=3132333435363738393031323334353637383930
totp() {
	step=$(( $(date -u +%s) / 30 ))
	# The counter as eight raw bytes. Octal escapes rather than \x, which is not
	# POSIX and which dash does not interpret.
	msg=""
	for byte in $(printf '%016x' "$step" | sed 's/../& /g'); do
		msg="$msg$(printf '\\%03o' "$((0x$byte))")"
	done
	digest=$(printf "$msg" | openssl dgst -sha1 -mac HMAC -macopt hexkey:$KEYHEX | sed 's/.*= *//')
	offset=$(( 0x$(printf %s "$digest" | cut -c40) * 2 + 1 ))
	chunk=$(printf %s "$digest" | cut -c${offset}-$((offset + 7)))
	printf '%06d' $(( (0x$chunk & 0x7fffffff) % 1000000 ))
}

export GIT_TERMINAL_PROMPT=0 GIT_CONFIG_NOSYSTEM=1 GIT_CONFIG_GLOBAL=/dev/null
export GIT_AUTHOR_NAME=Smoke GIT_AUTHOR_EMAIL=smoke@example.com
export GIT_COMMITTER_NAME=Smoke GIT_COMMITTER_EMAIL=smoke@example.com
# alice's cookie jar, as two curl flags. Left unquoted at every use on purpose.
alice="-c $WORK/alice -b $WORK/alice"

# start <description> brings up a container against the same bucket. No volume,
# so the datadir it starts from is whatever it can fetch.
#
# GIT_FORGE_DATADIR is deliberately wrong: syncd mirrors /var/lib/git-forge and
# is what tells git-forge where to write, so a git-forge that took this value
# would serve normally while every flush went nowhere.
start() {
	docker run -d --name "$NAME" \
		-p "$PORT:8081" \
		--add-host=host.docker.internal:host-gateway \
		-e SYNCD_R2="http://host.docker.internal:$R2PORT" \
		-e GIT_FORGE_DATADIR=/tmp/not-mirrored \
		-e GIT_FORGE_BASE_URL="$BASE" \
		-e GIT_FORGE_COOKIE_KEY=0123456789abcdef0123456789abcdef \
		-e SYNCD_LOOSE_LIMIT=1 \
		-e SYNCD_LOGS="http://host.docker.internal:$R2PORT/" \
		"$IMAGE" >/dev/null
	i=0
	while [ $i -lt 100 ]; do
		if curl -fsS "$BASE/healthz" >/dev/null 2>&1; then
			ok "$1"
			return
		fi
		i=$((i + 1))
		sleep 0.2
	done
	docker logs "$NAME" >&2 || true
	no "$1"
}

echo "== build"
./fetch-forge >/dev/null
docker build -q -f Containerfile -t "$IMAGE" . >/dev/null
ok "the image builds"

# Built rather than `go run`, so that R2PID is the server and not a wrapper
# whose child would outlive the kill and hold the port.
go build -o "$WORK/fake-r2" fake-r2.go
"$WORK/fake-r2" -dir "$BUCKET" -listen ":$R2PORT" >"$WORK/r2.log" 2>&1 &
R2PID=$!
sleep 1
[ -d "$BUCKET" ] || no "the object store did not start"
ok "the object store is up"

echo "== first start, on an empty bucket"
start "the container starts with nothing to restore"

docker logs "$NAME" 2>&1 | grep -o 'token: .*' | cut -d' ' -f2 >"$WORK/tok"
TOKEN=$(cat "$WORK/tok")
[ -n "$TOKEN" ] || no "no setup token in the log"
curl -s $alice "$BASE/setup" >"$WORK/p"
code "the setup form makes the first admin" 303 $alice -o /dev/null \
	-d "csrf=$(csrf "$WORK/p")&token=$TOKEN&name=alice&password=hunter2hunter2&again=hunter2hunter2" \
	"$BASE/setup"

curl -s $alice "$BASE/settings/totp" >"$WORK/p"
curl -s $alice -o /dev/null \
	-d "csrf=$(csrf "$WORK/p")&secret=$SECRET&code=$(totp)&current=hunter2hunter2" \
	"$BASE/settings/totp"
code "the admin holds a second factor" 200 $alice "$BASE/admin"

curl -s $alice "$BASE/admin" >"$WORK/p"
code "a repository is created" 303 $alice -o /dev/null \
	-d "csrf=$(csrf "$WORK/p")&name=notes&desc=Notes&visibility=public&issues=members&wiki=members" \
	"$BASE/admin/repo/create"

curl -s $alice "$BASE/settings" >"$WORK/p"
curl -s $alice -d "csrf=$(csrf "$WORK/p")&repo=notes" "$BASE/settings/tokens" >"$WORK/minted"
PUSH=$(sed -n 's/.*<pre id="token">\(.*\)<\/pre>.*/\1/p' "$WORK/minted")
[ -n "$PUSH" ] || no "no push token was minted"
ok "a push token is minted"

echo "== push"
git init -q "$WORK/src"
echo "first" >"$WORK/src/a.txt"
git -C "$WORK/src" add a.txt
git -C "$WORK/src" -c commit.gpgsign=false commit -qm "first"
git -C "$WORK/src" push -q "http://alice:$PUSH@127.0.0.1:$PORT/notes.git" master:master
ok "push over http succeeds"
HEAD=$(git -C "$WORK/src" rev-parse HEAD)

# The push returned, so it is already in the bucket. Nothing is waited for: a
# sleep would hide exactly the bug this checks for.
[ -f "$BUCKET/repos/notes.git/REFS" ] || no "the refs are not in the bucket"
grep -q "$HEAD" "$BUCKET/repos/notes.git/REFS" || no "the bucket has the wrong head"
ok "the push was durable before the response returned"
find "$BUCKET/repos/notes.git/objects" -type f >"$WORK/objs" 2>/dev/null || true
[ -s "$WORK/objs" ] || no "no objects in the bucket"
ok "objects are in the bucket"
[ -f "$BUCKET/config/users.json" ] || no "the account table is not in the bucket"
ok "the config is in the bucket"

# git sends anything past http.postBuffer, a megabyte by default, with chunked
# transfer encoding, and that body crosses syncd's proxy. Every other push here
# is a few hundred bytes. Random, so the pack is still over the buffer after
# compression.
head -c 3000000 /dev/urandom >"$WORK/src/big.bin"
git -C "$WORK/src" add big.bin
git -C "$WORK/src" -c commit.gpgsign=false commit -qm "big"
git -C "$WORK/src" push -q "http://alice:$PUSH@127.0.0.1:$PORT/notes.git" master:master
ok "a push past git's post buffer succeeds"
# Reassigned, so everything after the restart is checked against the large push
# rather than the small one.
HEAD=$(git -C "$WORK/src" rev-parse HEAD)
grep -q "$HEAD" "$BUCKET/repos/notes.git/REFS" || no "the chunked push is not in the bucket"
ok "and it was durable before the response returned too"

echo "== destroy the container, keep only the bucket"
docker rm -f "$NAME" >/dev/null
ok "the container and its disk are gone"

start "a fresh container starts from the bucket alone"

# Lazy restore: the directory has to exist so git-forge's startup scan finds it,
# but nothing has asked for the repository yet, so it must still be empty.
docker exec "$NAME" ls /var/lib/git-forge/repos/notes.git >"$WORK/before" 2>&1 || true
[ -s "$WORK/before" ] && no "the repository was fetched before anything asked for it"
ok "the repository is a placeholder until it is asked for"

curl -s "$BASE/" >"$WORK/home"
want "the index lists it anyway" "notes" "$WORK/home"

echo "== hydrate on first touch"
git clone -q "http://127.0.0.1:$PORT/notes.git" "$WORK/clone" 2>"$WORK/cloneerr" \
	|| { cat "$WORK/cloneerr" >&2; no "the clone failed"; }
[ "$(git -C "$WORK/clone" rev-parse HEAD)" = "$HEAD" ] || no "the clone is at the wrong commit"
ok "the clone comes back at the same commit"
want "and the content survived" "first" "$WORK/clone/a.txt"

docker exec "$NAME" ls /var/lib/git-forge/repos/notes.git >"$WORK/after" 2>&1 || true
grep -q objects "$WORK/after" || no "the repository was not filled in"
ok "the repository is filled in only after it was asked for"

echo "== the accounts survived too"
# The push token is a digest in users.json, so a push that authenticates proves
# the account table came back, and proves a restored repository takes writes.
echo "second" >"$WORK/clone/b.txt"
git -C "$WORK/clone" add b.txt
git -C "$WORK/clone" -c commit.gpgsign=false commit -qm "second"
git -C "$WORK/clone" push -q "http://alice:$PUSH@127.0.0.1:$PORT/notes.git" master:master
ok "the push token still authenticates after the restore"
SECOND=$(git -C "$WORK/clone" rev-parse HEAD)
grep -q "$SECOND" "$BUCKET/repos/notes.git/REFS" || no "the second push is not in the bucket"
ok "and the second push is durable too"

# alice enrolled a second factor before the restart, so a password alone must
# now be answered with a demand for a code rather than a session.
curl -s -o "$WORK/login" -w '%{http_code}' \
	-d "name=alice&password=hunter2hunter2" "$BASE/login" >"$WORK/lc"
grep -qx 200 "$WORK/lc" || no "signing in did not reach the second factor"
want "the second factor survived the restore" 'name="code"' "$WORK/login"

echo "== compaction packs the audit log and prunes what it replaced"
# SYNCD_LOOSE_LIMIT is 1 here, so a sweep packs rather than waiting for the five
# hundred a deployment would. Long enough for two of them, because pruning waits
# for the pass after the one that finds an object gone.
loose() { find "$BUCKET/config/git/objects" -type f -not -path '*/pack/*' 2>/dev/null | wc -l; }
before=$(loose)
[ "$before" -gt 0 ] || no "there were no loose objects to pack"
sleep 35
ls "$BUCKET"/config/git/objects/pack/*.pack >/dev/null 2>&1 || no "the audit log was not packed"
ok "the audit log is packed"
after=$(loose)
[ "$after" -lt "$before" ] || no "nothing was pruned: $before loose objects before, $after after"
ok "what packing replaced was pruned from the bucket ($before down to $after)"

# Pruning deletes from object storage, so the question is whether it deleted
# something still reachable.
docker rm -f "$NAME" >/dev/null
start "a third container starts from the compacted bucket"
curl -s "$BASE/" >"$WORK/home3"
want "the index still lists the repository" "notes" "$WORK/home3"
git clone -q "http://127.0.0.1:$PORT/notes.git" "$WORK/clone3" 2>"$WORK/err3" \
	|| { cat "$WORK/err3" >&2; no "the clone failed after compaction"; }
[ "$(git -C "$WORK/clone3" rev-parse HEAD)" = "$SECOND" ] || no "the clone is at the wrong commit after compaction"
ok "and a clone after compaction is still at the right commit"

echo "== an idle forge writes nothing"
# The sweep runs every fifteen seconds and must find nothing to do.
before=$(grep -c 'PUT ' "$WORK/r2.log" || true)
sleep 20
after=$(grep -c 'PUT ' "$WORK/r2.log" || true)
[ "$before" = "$after" ] || no "an idle sweep wrote $((after - before)) objects, it should write none"
ok "a sweep over unchanged state writes nothing"

echo "== a graceful stop is durable and says so"
# SIGTERM, not a kill: syncd drains git-forge, flushes once more, and ships
# those last lines rather than losing them with the process.
grep -q 'container: syncd:' "$WORK/r2.log" || no "no container logs were shipped at all"
ok "container logs are shipped"
docker stop -t 30 "$NAME" >/dev/null
grep -q 'container: syncd: draining' "$WORK/r2.log" || no "the drain was never reported"
ok "SIGTERM drains git-forge"
grep -q 'container: syncd: flushed and stopping' "$WORK/r2.log" || no "the final flush was never reported"
ok "and the final flush is reported before the process exits"

echo "== a container that refuses to start says why"
# Container stdout is forwarded nowhere, so a fatal must ship its line before it
# goes. No cookie key is the cheapest way to make it refuse.
docker run --rm --name "$NAME-fatal" \
	--add-host=host.docker.internal:host-gateway \
	-e SYNCD_R2="http://host.docker.internal:$R2PORT" \
	-e SYNCD_LOGS="http://host.docker.internal:$R2PORT/" \
	"$IMAGE" >/dev/null 2>&1 || true
sleep 1
want "the reason it refused to start reaches the log" \
	'container: syncd: GIT_FORGE_COOKIE_KEY must be set' "$WORK/r2.log"

printf '\nall good\n'