Intermittent file locking bug in installPackage.hs
|Reported by:||gw||Owned by:||duncan|
|Type of failure:||Test Case:|
|Related Tickets:||Differential Rev(s):|
After fixing the bug in ticket #1992, I still saw lock occasional lock errors building ghc. Dumping all of the fd/dev/inode info in lockFile and unlockFile showed that the lock hash tables were not inconsistent, as they had been in #1992, but that a write lock was being placed on a file before a read lock was released.
The culprit is the reading of the configuration in libraries/installPackage.hs (lbi <- getConfig verbosity). This ultimately uses tryGetConfigStateFile from libraries/Cabal/Distribution/Simple/Configure.hs. And this function uses readFile to read in the configuration file.
The dump of fd/dev/inode info showed that readFile was acquiring a lock that was not released before the end of the program. Occasionally, there was an attempt to obtain a write lock on the same dev/inode, leading to a lock error. It looks as if this is the laziness problem with readFile noted in the IO library documentation.
Making the read of the configuration file strict and explicitly closing the handle fixes the problem for me. Here is the patch:
--- libraries/Cabal/Distribution/Simple/Configure.hs.sav 2008-03-01 17:01:39.000000000 -0500 +++ libraries/Cabal/Distribution/Simple/Configure.hs 2008-03-01 17:04:25.000000000 -0500 @@ -130,7 +130,7 @@ import qualified System.Info ( os, arch ) import System.IO - ( hPutStrLn, stderr ) + ( hPutStrLn, stderr, openFile, IOMode(..), hGetContents, hClose ) import Text.PrettyPrint.HughesPJ ( comma, punctuate, render, nest, sep ) @@ -146,7 +146,9 @@ let dieMsg = "error reading " ++ filename ++ "; run \"setup configure\" command?\n" if (not e) then return $ Left dieMsg else do - str <- readFile filename + hdl <- openFile filename ReadMode + str <- hGetContents hdl >>= \s -> last s `seq` return s + hClose hdl case reads str of [(bi,_)] -> return $ Right bi _ -> return $ Left dieMsg
I'm still not certain why the bug occurs only some of the time, and whether it is OS dependent. But readFile's holding of the lock for an unpredictable time is just asking for trouble.
Change History (7)
comment:1 Changed 9 years ago by simonmar
- Component changed from Compiler to libraries (other)
- difficulty set to Easy (1 hr)
- Milestone set to 6.8.3