fix memory leak when staging the journal

The list of files had to be retained until the end so it could be deleted.
Also, a list of update-index lines was generated and only then fed into it.
Now everything streams in constant space.

Annex/Branch.hs

@@ -1,6 +1,6 @@
 {- management of the git-annex branch
  -
- - Copyright 2011 Joey Hess <joey@kitenet.net>
+ - Copyright 2011-2012 Joey Hess <joey@kitenet.net>
  -
  - Licensed under the GNU GPL version 3 or higher.
  -}
@@ -32,7 +32,7 @@ import qualified Git.Command
 import qualified Git.Ref
 import qualified Git.Branch
 import qualified Git.UnionMerge
-import qualified Git.HashObject
+import Git.HashObject
 import qualified Git.Index
 import Annex.CatFile
 
@@ -307,13 +307,14 @@ stageJournal = do
 	fs <- getJournalFiles
 	g <- gitRepo
 	withIndex $ liftIO $ do
-		let dir = gitAnnexJournalDir g
+		h <- hashObjectStart g
-		let paths = map (dir </>) fs
+		Git.UnionMerge.stream_update_index g
-		(shas, cleanup) <- Git.HashObject.hashFiles paths g
+			[genstream (gitAnnexJournalDir g) h fs]
-		Git.UnionMerge.update_index g $
+		hashObjectStop h
-			index_lines shas (map fileJournal fs)
-		cleanup
-		mapM_ removeFile paths
 	where
-		index_lines shas = map genline . zip shas
+		genstream dir h fs streamer = forM_ fs $ \file -> do
-		genline (sha, file) = Git.UnionMerge.update_index_line sha file
+			let path = dir </> file
+			sha <- hashFile h path
+			streamer $ Git.UnionMerge.update_index_line
+				sha (fileJournal file)
+			removeFile path

Git/HashObject.hs

@@ -1,6 +1,6 @@
 {- git hash-object interface
  -
- - Copyright 2011 Joey Hess <joey@kitenet.net>
+ - Copyright 2011-2012 Joey Hess <joey@kitenet.net>
  -
  - Licensed under the GNU GPL version 3 or higher.
  -}
@@ -11,23 +11,32 @@ import Common
 import Git
 import Git.Command
 
-{- Injects a set of files into git, returning the shas of the objects
+type HashObjectHandle = (PipeHandle, Handle, Handle)
- - and an IO action to call once the the shas have been used. -}
+
-hashFiles :: [FilePath] -> Repo -> IO ([Sha], IO ())
+{- Starts git hash-object and returns a handle.  -}
-hashFiles paths repo = do
+hashObjectStart :: Repo -> IO HashObjectHandle
-	(pid, fromh, toh) <- hPipeBoth "git" $ toCommand $ git_hash_object repo
+hashObjectStart repo = do
+	r@(_, _, toh) <- hPipeBoth "git" $
+		toCommand $ gitCommandLine params repo
 	fileEncoding toh
-	_ <- forkProcess (feeder toh)
+	return r
-	hClose toh
-	shas <- map Ref . lines <$> hGetContents fromh
-	return (shas, ender fromh pid)
 	where
-		git_hash_object = gitCommandLine
+		params =
-			[Param "hash-object", Param "-w", Param "--stdin-paths"]
+			[ Param "hash-object"
-		feeder toh = do
+			, Param "-w"
-			hPutStr toh $ unlines paths
+			, Param "--stdin-paths"
-			hClose toh
+			]
-			exitSuccess
+
-		ender fromh pid = do
+{- Stops git hash-object. -}
-			hClose fromh
+hashObjectStop :: HashObjectHandle -> IO ()
-			forceSuccess pid
+hashObjectStop (pid, from, to) = do
+	hClose to
+	hClose from
+	forceSuccess pid
+
+{- Injects a file into git, returning the shas of the objects. -}
+hashFile :: HashObjectHandle -> FilePath -> IO Sha
+hashFile (_, from, to) file = do
+	hPutStrLn to file
+	hFlush to
+	Ref <$> hGetLine from

debian/changelog

@@ -24,8 +24,7 @@ git-annex (3.20120124) UNRELEASED; urgency=low
     its head), and records the size in the key.
   * Fixed to use the strict state monad, to avoid leaking all kinds of memory
     due to lazy state update thunks when adding/fixing many files.
-  * Fixed a memory leak due to excessive strictness when committing journal
+  * Fixed some memory leaks that occurred when committing journal files.
-    files.
   * whereis: Prints the urls of files that the web special remote knows about.
 
  -- Joey Hess <joeyh@debian.org>  Tue, 24 Jan 2012 16:21:55 -0400

doc/bugs/git_annex_add_memory_leak.mdwn

@@ -12,26 +12,27 @@ A history of the leaks:
 * Originally, `git annex add` remembered all the files
   it had added, and fed them to git at the end. Of course
   that made its memory use grow, so it was fixed to periodically
-  flush its buffer. Affected versions: before 0.20110417
+  flush its buffer. Fixed in version 0.20110417.
 
 * Something called a "lazy state monad" caused "thunks" to build
   up and memory to leak. Also affected other git annex commands
   than `add`. Adding files using a SHA* backend hit the worst.
   Fixed in versions afer 3.20120123.
 
-* A strange GHC bug seemed to be responsible for another leak.
-  (In particular, a child process was forked. All the child did 
-  was read filenames from one pipe and shove them reformatted out
-  another pipe. For some reason, it steadily grew in size.)
-  Code was rewritten in a way that happens to avoid that leak.
-  Apparently fixed in versions afer 3.20120123, but this one is not
-  well understood.
-
 * Committing journal files turned out to have another memory leak.
   After adding a lot of files ran out of memory, this left the journal
-  behind and could affect other git-anne commands. Fixed in versions afer
+  behind and could affect other git-annex commands. Fixed in versions afer
   3.20120123.
 
+* Something is still causing a slow leak when adding files.
+  I tested by adding many copies of the whole linux kernel
+  tree into the annex using the WORM backend, and once
+  it had added 1 million files, git-annex used ~100 mb of ram.
+  That's 100 bytes leaked per file on average .. roughly the
+  size of a filename? It's worth noting that `git add` uses more memory
+  than that in such a large tree.
+  **not fixed yet**
+
 * (Note that `git ls-files --others`, which is used to find files to add,
   also uses surpsisingly large amounts
   of memory when you have a lot of files. It buffers