fsck: remove overlay for directories that aren't on disk

Summary:
During startup, EdenFS loads inodes for all the directories present in the
overlay, with the assumption that the state of the overlay represents the state
of the filesystem. Today, this is mostly true as the working copy only grows
and thus the only way for the working copy and overlay to shrink is to update
past commits that delete a directory.

In the future however, EdenFS will gain the ability to shrink its working copy,
either manually, or automatically. When this happens, the overlay will still
contain data for directories that are no longer in the working copy, thus
invalidating the assumption that working copy and overlay stay in sync.

To fix this, we simply need to teach the fsck code to remove overlay
directories when a directory isn't present on disk.

Reviewed By: kmancini

Differential Revision: D41007167

fbshipit-source-id: 2a6d93417995b9bb8f96c757d47b444c17ab82f8

Author: xavierd

eden/fs/inodes/sqlitecatalog/WindowsFsck.cpp

@@ -491,7 +491,14 @@ std::optional<InodeNumber> fixup(
           state.desiredHash,
           insensitiveOverlayDir);
     } else {
-      return InodeNumber(*state.overlayEntry->inodeNumber());
+      auto inodeNumber = InodeNumber(*state.overlayEntry->inodeNumber());
+      if (!state.onDisk && state.overlayDtype == dtype_t::Dir) {
+        auto overlayDir = inodeCatalog.loadAndRemoveOverlayDir(inodeNumber);
+        if (overlayDir) {
+          XLOGF(DBG9, "Removed overlay directory for: {}", path);
+        }
+      }
+      return inodeNumber;
     }
   } else {
     if (state.inOverlay) {

eden/integration/windows_fsck_test.py

@@ -9,6 +9,7 @@
 import subprocess
 import sys
 import unittest
+from pathlib import Path
 from typing import Dict, List, Optional, Set, Tuple, Union
 
 from facebook.eden.ttypes import (
@@ -247,6 +248,36 @@ def test_fsck_rename_with_different_case_and_modify_while_stopped(self) -> None:
         self._update_clean()
         self.assertEqual(self._eden_status(), {})
 
+    def test_loaded_inodes_not_loaded_on_restart(self) -> None:
+        """Verifies that a loaded inode not present on disk doesn't get loaded
+        with a positive refcount on restart.
+        """
+
+        def get_all_loaded_under(path: str) -> List[Tuple[Path, int]]:
+            with self.eden.get_thrift_client_legacy() as client:
+                all_loaded = client.debugInodeStatus(
+                    self.mount_path_bytes,
+                    path.encode(),
+                    0,
+                    sync=SyncBehavior(),
+                )
+
+            ret: List[Tuple[Path, int]] = []
+            for loaded in all_loaded:
+                ret += [(Path(loaded.path.decode()), loaded.refcount)]
+
+            return ret
+
+        # This relies on debugInodeStatus to load the inode for the directory.
+        loaded = get_all_loaded_under("subdir/bdir")
+        self.assertEqual(loaded, [(Path("subdir/bdir"), 0)])
+
+        self.eden.shutdown()
+        self.eden.start()
+
+        loaded = get_all_loaded_under("subdir/bdir")
+        self.assertEqual(loaded, [(Path("subdir/bdir"), 0)])
+
 
 MATERIALIZED = True
 UNMATERIALIZED = False