remotedaemon: When network connection is lost, close all cached ssh connections.

This commit was sponsored by Cedric Staub.

Annex/Ssh.hs

@@ -11,6 +11,7 @@ module Annex.Ssh (
 	sshCachingOptions,
 	sshCacheDir,
 	sshReadPort,
+	forceSshCleanup,
 	sshCachingEnv,
 	sshCachingTo,
 	inRepoWithSshCachingTo,
@@ -124,21 +125,27 @@ prepSocket socketfile = do
 	liftIO $ createDirectoryIfMissing True $ parentDir socketfile
 	lockFile $ socket2lock socketfile
 
-{- Stop any unused ssh processes. -}
+enumSocketFiles :: Annex [FilePath]
-sshCleanup :: Annex ()
+enumSocketFiles = go =<< sshCacheDir
-sshCleanup = go =<< sshCacheDir
+  where
+	go Nothing = return []
+	go (Just dir) = liftIO $ filter (not . isLock)
+		<$> catchDefaultIO [] (dirContents dir)
+
+{- Stop any unused ssh connection caching processes. -}
+sshCleanup :: Annex ()
+sshCleanup = mapM_ cleanup =<< enumSocketFiles
   where
-	go Nothing = noop
-	go (Just dir) = do
-		sockets <- liftIO $ filter (not . isLock)
-			<$> catchDefaultIO [] (dirContents dir)
-		forM_ sockets cleanup
 	cleanup socketfile = do
 #ifndef mingw32_HOST_OS
 		-- Drop any shared lock we have, and take an
 		-- exclusive lock, without blocking. If the lock
 		-- succeeds, nothing is using this ssh, and it can
 		-- be stopped.
+		--
+		-- After ssh is stopped cannot remove the lock file;
+		-- other processes may be waiting on our exclusive
+		-- lock to use it.
 		let lockfile = socket2lock socketfile
 		unlockFile lockfile
 		mode <- annexFileMode
@@ -148,24 +155,28 @@ sshCleanup = go =<< sshCacheDir
 			setLock fd (WriteLock, AbsoluteSeek, 0, 0)
 		case v of
 			Left _ -> noop
-			Right _ -> stopssh socketfile
+			Right _ -> forceStopSsh socketfile
 		liftIO $ closeFd fd
 #else
-		stopssh socketfile
+		forceStopSsh socketfile
 #endif
-	stopssh socketfile = do
+
-		let (dir, base) = splitFileName socketfile
+{- Stop all ssh connection caching processes, even when they're in use. -}
-		let params = sshConnectionCachingParams base
+forceSshCleanup :: Annex ()
-		-- "ssh -O stop" is noisy on stderr even with -q
+forceSshCleanup = mapM_ forceStopSsh =<< enumSocketFiles
-		void $ liftIO $ catchMaybeIO $
+
-			withQuietOutput createProcessSuccess $
+forceStopSsh :: FilePath -> Annex ()
-				(proc "ssh" $ toCommand $
+forceStopSsh socketfile = do
-					[ Params "-O stop"
+	let (dir, base) = splitFileName socketfile
-					] ++ params ++ [Param "localhost"])
+	let params = sshConnectionCachingParams base
-					{ cwd = Just dir }
+	-- "ssh -O stop" is noisy on stderr even with -q
-		liftIO $ nukeFile socketfile
+	void $ liftIO $ catchMaybeIO $
-		-- Cannot remove the lock file; other processes may
+		withQuietOutput createProcessSuccess $
-		-- be waiting on our exclusive lock to use it.
+			(proc "ssh" $ toCommand $
+				[ Params "-O stop"
+				] ++ params ++ [Param "localhost"])
+				{ cwd = Just dir }
+	liftIO $ nukeFile socketfile
 
 {- This needs to be as short as possible, due to limitations on the length
  - of the path to a socket file. At the same time, it needs to be unique

Assistant/Threads/NetWatcher.hs

@@ -71,7 +71,7 @@ dbusThread = do
 		)
 	handleconn = do
 		debug ["detected network connection"]
-		sendRemoteControl PAUSE
+		sendRemoteControl LOSTNET
 		notifyNetMessagerRestart
 		handleConnection
 		sendRemoteControl RESUME

RemoteDaemon/Core.hs

@@ -18,6 +18,7 @@ import qualified Git.Types as Git
 import qualified Git.CurrentRepo
 import Utility.SimpleProtocol
 import Config
+import Annex.Ssh
 
 import Control.Concurrent.Async
 import Control.Concurrent
@@ -65,12 +66,19 @@ runController ichan ochan = do
 				let common = M.intersection m m'
 				let new = M.difference m' m
 				let old = M.difference m m'
-				stoprunning old
+				broadcast STOP old
 				unless paused $
 					startrunning new
 				go h paused (M.union common new)
+			LOSTNET -> do
+				-- force close all cached ssh connections
+				-- (done here so that if there are multiple
+				-- ssh remotes, it's only done once)
+				liftAnnex h forceSshCleanup
+				broadcast LOSTNET m
+				go h True M.empty
 			PAUSE -> do
-				stoprunning m
+				broadcast STOP m
 				go h True M.empty
 			RESUME -> do
 				when paused $
@@ -89,9 +97,9 @@ runController ichan ochan = do
 	startrunning m = forM_ (M.elems m) startrunning'
 	startrunning' (transport, _) = void $ async transport
 	
-	-- Ask the transport nicely to stop.
+	broadcast msg m = forM_ (M.elems m) send
-	stoprunning m = forM_ (M.elems m) stoprunning'
+	  where
-	stoprunning' (_, c) = writeChan c STOP
+		send (_, c) = writeChan c msg
 
 -- Generates a map with a transport for each supported remote in the git repo,
 -- except those that have annex.sync = false

RemoteDaemon/Transport/Ssh.hs

@@ -84,6 +84,7 @@ transport' r url transporthandle ichan ochan = do
 		msg <- readChan ichan
 		case msg of
 			STOP -> return Stopping
+			LOSTNET -> return Stopping
 			_ -> handlecontrol
 
 	-- Old versions of git-annex-shell that do not support

RemoteDaemon/Types.hs

@@ -42,6 +42,7 @@ data Emitted
 -- Messages that the deamon consumes.
 data Consumed
 	= PAUSE
+	| LOSTNET
 	| RESUME
 	| CHANGED RefList
 	| RELOAD
@@ -63,6 +64,7 @@ instance Proto.Sendable Emitted where
 
 instance Proto.Sendable Consumed where
 	formatMessage PAUSE = ["PAUSE"]
+	formatMessage LOSTNET = ["LOSTNET"]
 	formatMessage RESUME = ["RESUME"]
 	formatMessage (CHANGED refs) =["CHANGED", Proto.serialize refs]
 	formatMessage RELOAD = ["RELOAD"]
@@ -78,6 +80,7 @@ instance Proto.Receivable Emitted where
 
 instance Proto.Receivable Consumed where
 	parseCommand "PAUSE" = Proto.parse0 PAUSE
+	parseCommand "LOSTNET" = Proto.parse0 LOSTNET
 	parseCommand "RESUME" = Proto.parse0 RESUME
 	parseCommand "CHANGED" = Proto.parse1 CHANGED
 	parseCommand "RELOAD" = Proto.parse0 RELOAD

debian/changelog

@@ -10,6 +10,8 @@ git-annex (5.20140413) UNRELEASED; urgency=medium
     set up.
   * sync, assistant, remotedaemon: Use ssh connection caching for git pushes
     and pulls.
+  * remotedaemon: When network connection is lost, close all cached ssh
+    connections.
   * Improve handling on monthly/yearly scheduling.
 
  -- Joey Hess <joeyh@debian.org>  Fri, 11 Apr 2014 21:33:35 -0400

doc/design/git-remote-daemon.mdwn

@@ -95,18 +95,18 @@ the webapp.
 
 * `PAUSE`
 
-  This indicates that the network connection has gone down,
+  The user has requested a pause.  
-  or the user has requested a pause.
   git-remote-daemon should close connections and idle.
 
-  Affects all remotes.
+* `LOSTNET`
+
+  The network connection has been lost.  
+  git-remote-daemon should close connections and idle.
 
 * `RESUME`
 
-  This indicates that the network connection has come back up, or the user
+  Undoes PAUSE or DISCONNECTED.  
-  has asked it to run again. Start back up network connections.
+  Start back up network connections.
-
-  Affects all remotes.
 
 * `CHANGED ref ...`
 
@@ -170,6 +170,21 @@ TODO:
 * Remote system might not be available. Find a smart way to detect it,
   ideally w/o generating network traffic. One way might be to check
   if the ssh connection caching control socket exists, for example.
+* Now that ssh connection caching is enabled for git push/pull in sync,
+  there's the possibility that a stale ssh connection may linger when 
+  changing network connections, and so attempts to use it will stall.  
+  (This was already a potential issue with transfers, which already
+  used the caching.)  
+
+  One option is ssh's ServerAliveCountMax, which will make a dead
+  ssh connection disconnect after approx 45 seconds, per ssh manual.
+  It would need to be enabled by setting ServerAliveInterval=15.
+  And this would add network traffic..
+
+  Another option is to disable all cached connections when the network
+  connection changes. This would handle *most* cases. The case
+  not handled is eg, my dialup ppp box getting a new public IP address,
+  which my laptop won't notice. **done**
 
 ## telehash