Take an exclusive lock on DATA_DIR at startup (closes #201) (#220)

internal/datadir/lock.go

@@ -0,0 +1,126 @@
+// Package datadir guards exclusive access to the directory holding
+// every SQLite database webhooker writes.
+//
+// Two processes sharing a DATA_DIR each open the same per-webhook
+// event databases and each run delivery recovery over the same rows,
+// so every pending delivery goes out twice. SQLite's own locking does
+// not prevent that: both writers are serialised correctly and both
+// deliver. The only thing that prevents it is refusing to be the
+// second process.
+//
+// The lock lives here rather than in the server's fx graph so that any
+// entry point which touches DATA_DIR — the server, or a CLI
+// subcommand that must not operate on a live deployment's data — takes
+// it the same way.
+package datadir
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+
+	"github.com/gofrs/flock"
+)
+
+// LockFileName is the advisory lock file created inside DATA_DIR. Its
+// contents are never read: the lock is the flock(2) held on the open
+// descriptor, not the file's existence, so a leftover file from a
+// process that was killed with SIGKILL blocks nothing.
+const LockFileName = "webhooker.lock"
+
+// dirPerm is the mode Acquire creates DATA_DIR with. It matches what
+// internal/database uses, since whichever runs first creates it.
+const dirPerm = 0o750
+
+// ErrLocked reports that another live process holds the data
+// directory. Callers that need to know whether a deployment is running
+// — rather than merely failing to start — test for this with
+// errors.Is.
+var ErrLocked = errors.New(
+	"data directory is already in use by another instance",
+)
+
+// ErrNoDir reports that Acquire was given an empty directory.
+var ErrNoDir = errors.New("no data directory given")
+
+// Lock is a held exclusive advisory lock on a data directory. It is
+// valid only while the process that took it lives: the kernel drops it
+// when the descriptor closes, whether that is Release, a normal exit,
+// or a SIGKILL.
+type Lock struct {
+	dir  string
+	file *flock.Flock
+}
+
+// Acquire takes the exclusive advisory lock on dir, creating dir if it
+// does not exist. It never waits: if another process holds the lock it
+// returns an error wrapping ErrLocked and naming dir.
+//
+// The returned Lock must be held for as long as the caller intends to
+// use dir.
+func Acquire(dir string) (*Lock, error) {
+	if dir == "" {
+		return nil, ErrNoDir
+	}
+
+	err := os.MkdirAll(dir, dirPerm)
+	if err != nil {
+		return nil, fmt.Errorf(
+			"creating data directory %s: %w", dir, err,
+		)
+	}
+
+	path := filepath.Join(dir, LockFileName)
+	fl := flock.New(path)
+
+	held, err := fl.TryLock()
+	if err != nil {
+		return nil, fmt.Errorf(
+			"locking data directory %s: %w", dir, err,
+		)
+	}
+
+	if !held {
+		// A no-op on flock v0.13.0, which closes its own descriptor on
+		// a failed TryLock; kept so no version can leak one.
+		_ = fl.Close()
+
+		return nil, fmt.Errorf(
+			"%w: %s (%s). Only one webhooker may use a data "+
+				"directory: two both run delivery recovery over the "+
+				"same rows and both deliver",
+			ErrLocked, dir, path,
+		)
+	}
+
+	return &Lock{dir: dir, file: fl}, nil
+}
+
+// Dir returns the locked directory.
+func (l *Lock) Dir() string {
+	return l.dir
+}
+
+// Path returns the lock file backing the lock.
+func (l *Lock) Path() string {
+	return l.file.Path()
+}
+
+// Release drops the lock and closes the descriptor. It is safe to call
+// more than once.
+//
+// The lock file is deliberately left on disk. Unlinking it would let
+// the next process create and lock a fresh inode while a third still
+// holds the old one, which is the one outcome this package exists to
+// prevent.
+func (l *Lock) Release() error {
+	err := l.file.Unlock()
+	if err != nil {
+		return fmt.Errorf(
+			"releasing lock on data directory %s: %w", l.dir, err,
+		)
+	}
+
+	return nil
+}