Improve robustness of direct mode merge, avoiding a crash if the index file is missing.

I couldn't find a good way to make an *empty* index file (zero byte file
won't do), so I punted and just don't make index.lock when there's no index
yet. This means some other git process could race and write an index file
at the same time as the merge is ongoing, in theory. Only happens in new
repos though.

Annex/Direct.hs

@@ -153,7 +153,7 @@ addDirect file cache = do
  -
  - A lock file is used to avoid races with any other caller of mergeDirect.
  - 
- - To avoid other git processes from making change to the index while our
+ - To avoid other git processes from making changes to the index while our
  - merge is in progress, the index lock file is used as the temp index
  - file. This is the same as what git does when updating the index
  - normally.
@@ -162,7 +162,8 @@ mergeDirect :: Maybe Git.Ref -> Maybe Git.Ref -> Git.Branch -> Annex Bool -> Git
 mergeDirect startbranch oldref branch resolvemerge commitmode = exclusively $ do
 	reali <- liftIO . absPath =<< fromRepo indexFile
 	tmpi <- liftIO . absPath =<< fromRepo indexFileLock
-	liftIO $ copyFile reali tmpi
+	liftIO $ whenM (doesFileExist reali) $
+		copyFile reali tmpi
 
 	d <- fromRepo gitAnnexMergeDir
 	liftIO $ do
@@ -178,7 +179,8 @@ mergeDirect startbranch oldref branch resolvemerge commitmode = exclusively $ do
 		mergeDirectCleanup d (fromMaybe Git.Sha.emptyTree oldref)
 		mergeDirectCommit merged startbranch branch commitmode
 
-		liftIO $ rename tmpi reali
+		liftIO $ whenM (doesFileExist tmpi) $
+			rename tmpi reali
 
 		return r
   where

debian/changelog

@@ -10,6 +10,9 @@ git-annex (5.20150917) UNRELEASED; urgency=medium
     to use Google's NearLine storage.
   * Improve ~/.ssh/config modification code to not add trailing spaces
     to lines it cannot parse.
+  * Fix a crash at direct mode merge time when .git/index doesn't exist
+    yet. Triggered by eg, git-annex sync --no-commit in a fresh clone of
+    a repository.
 
  -- Joey Hess <id@joeyh.name>  Wed, 16 Sep 2015 12:23:33 -0400
 

doc/bugs/Assistant:_synchronisation_between_two_regular_repositories_hangs_.mdwn

@@ -26,3 +26,5 @@ Also note that the second repository is on a FAT32 usb stick
 
 # End of transcript or log.
 """]]
+
+> [[fixed|done]] --[[Joey]]

doc/bugs/Assistant:_synchronisation_between_two_regular_repositories_hangs_/comment_3_db50cbdc302480c4b1fe0f364aa769d9._comment

@@ -0,0 +1,19 @@
+[[!comment format=mdwn
+ username="joey"
+ subject="""comment 3"""
+ date="2015-09-22T16:25:15Z"
+ content="""
+Ok, that seems clear enough, and there's only 1 place where git-annex
+copies .git/index; in `mergeDirect`. Indeed, if .git/index doesn't exist
+yet when that is called, it'll crash. And, a freshly created git repo
+starts off without a .git/index until changes start to be staged.
+
+However, I can't reproduce the crash with a current version of git-annex,
+and this bug report is about a version nearly a year old now. AFAICS,
+the sync (or the assistant) will make a commit before merging, and that
+commit results in the index file being created, as a side effect.
+Also, I can't see anything that VFAT could have to do with this.
+
+Hmm, I did manage to reproduce the crash using the new --no-commit flag to
+git-annex sync. Will fix on that basis.
+"""]]