Fix the IPC lock used by custom shader compilation

docs/changelog.rst

@@ -206,6 +206,11 @@ Unreleased
   consistently after rearranging windows. Fix nested dividers not tracking
   mouse movement correctly, including when returning from a minimum size.
 
+- Custom shaders: Fix starting a second kitty instance failing to build
+  shaders while another instance is building them, or after a previous
+  instance was killed while building them. Now instances wait for each other
+  to finish instead.
+
 0.49.0 [2026-09-21]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

kitty/utils.py

@@ -1239,11 +1239,20 @@ def unlock_file(f: IO[bytes] | IO[str]) -> None:
 
 @contextmanager
 def lock_with_file(path: str) -> Iterator[None]:
-    os.close(os.open(path, os.O_CREAT | os.O_WRONLY | os.O_EXCL | os.O_CLOEXEC))
+    """Take an exclusive advisory lock on path, waiting for it to become available.
+
+    The lock file is created if needed and left in place, since unlinking it
+    would allow a second process to create and lock a different file with the
+    same name while the lock is still held. The kernel drops the lock when the
+    file descriptor is closed, including on abnormal process exit, so stale lock
+    files are harmless.
+    """
+    fd = os.open(path, os.O_CREAT | os.O_WRONLY | os.O_CLOEXEC, 0o600)
     try:
+        fcntl.flock(fd, fcntl.LOCK_EX)
         yield
     finally:
-        os.remove(path)
+        os.close(fd)
 
 
 def rmtree_best_effort(relpath: str, dir_fd: int) -> None:

kitty_tests/datatypes.py

@@ -758,34 +758,48 @@ class TestDataTypes(BaseTest):
                     os.environ[k] = saved[k]
 
     def test_lock_with_file(self):
+        def is_locked_by_another_process(path):
+            # flock() locks are per file descriptor, so a non-blocking
+            # acquisition in this process would succeed even while the lock is
+            # held here, hence the check has to happen in a child process.
+            code = f'import fcntl, os;fd = os.open({path!r}, os.O_CREAT | os.O_WRONLY);fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB)'
+            return subprocess.run([kitty_exe(), '+runpy', code], capture_output=True).returncode != 0
+
         with tempfile.TemporaryDirectory() as tdir:
             lock_path = os.path.join(tdir, 'test.lock')
 
-            # normal usage: file exists during context and is removed after
+            # normal usage: the lock is held only inside the context
             with lock_with_file(lock_path):
                 self.assertTrue(os.path.exists(lock_path))
-            self.assertFalse(os.path.exists(lock_path))
+                self.assertTrue(is_locked_by_another_process(lock_path))
+            self.assertFalse(is_locked_by_another_process(lock_path))
 
-            # second acquisition of the same path succeeds after the first released it
-            with lock_with_file(lock_path):
-                self.assertTrue(os.path.exists(lock_path))
-            self.assertFalse(os.path.exists(lock_path))
-
-            # file already exists: raises FileExistsError before yielding
-            open(lock_path, 'w').close()
-            with self.assertRaises(FileExistsError):
-                with lock_with_file(lock_path):
-                    pass  # should not be reached
-            # pre-existing file must not be deleted
+            # a pre-existing, unlocked lock file must not prevent acquisition
             self.assertTrue(os.path.exists(lock_path))
-            os.remove(lock_path)
+            with lock_with_file(lock_path):
+                self.assertTrue(is_locked_by_another_process(lock_path))
+            self.assertFalse(is_locked_by_another_process(lock_path))
 
-            # lock file is removed even when the body raises
+            # the lock is released even when the body raises
             with self.assertRaises(RuntimeError):
                 with lock_with_file(lock_path):
-                    self.assertTrue(os.path.exists(lock_path))
                     raise RuntimeError('body error')
-            self.assertFalse(os.path.exists(lock_path))
+            self.assertFalse(is_locked_by_another_process(lock_path))
+
+            # a waiting process acquires the lock once it is released
+            code = f'from kitty.utils import lock_with_file;lock_with_file({lock_path!r}).__enter__();print("acquired", flush=True)'
+            with lock_with_file(lock_path):
+                p = subprocess.Popen([kitty_exe(), '+runpy', code], stdout=subprocess.PIPE)
+                try:
+                    # the child must block while we hold the lock
+                    with self.assertRaises(subprocess.TimeoutExpired):
+                        p.wait(timeout=1)
+                except Exception:
+                    p.kill()
+                    raise
+            self.assertEqual(p.stdout.readline(), b'acquired\n')
+            self.assertEqual(p.wait(timeout=30), 0)
+            p.stdout.close()
 
     def test_historybuf(self):
         lb = filled_line_buf()