Patch/feature proposal: Provide access to the runStmt sandbox ThreadID
Hi all, This proposal is related to http://hackage.haskell.org/trac/ghc/ticket/1381, which Simon Marlow closed through commit https://github.com/ghc/ghc/commit/02c4ab049adeb77b8ee0e3b98fbf0f3026eee453. The problem, in a nutshell, is "how do we terminate a code snippet started with runStmt"? Before Simon's patch the only way was to disable ghc's sandboxing, so that the snippet would run in the same thread as the thread that called runStmt in the first place, and then send an asynchronous exception to that thread. This is the approach we used to take. It's a little tricky to get right (have to make sure that these exceptions are thrown only at the right times), but we thought we had it working okay. Until, that is, we realized we had a very nasty problem: snippets were being unexpected interrupted. To debug this, we introduced a CustomUserInterruption data type to serve as the exception that we were throwing. This had two purposes: first, we would be use that if we saw a CustomUserInterrupt that it could only have come from one particular throw, and second, the CustomUserInterrupt carried an integer which was incremented on every throw so that we never threw the same exception twice. What we realized is that snippets were being interrupted by *old* exceptions; that is, exceptions that we had thrown to *previous* snippets (and had been caught too). This should obviously never happen. Ian managed to reproduce this behaviour in a completely different setting ( http://hackage.haskell.org/trac/ghc/ticket/5902#comment:5) and we think that something similar (unsafePerformIO related) must be happening inside ghc. Moreover, a further conjecture is that this only happens when runStmt is still compiling the snippet to be executed (as opposed to the snippet actually executing) -- i.e., that the exception gets stuck in the bowels of ghc somewhere. We don't have any hard evidence for this, other than the problem has not appeared again with the proposed patch (but that might be luck, as it depends on timing). The patch as we currently have it is against 7.4.2, so pre Simon's change for the sandbox behaviour -- but I don't think that Simon's changes affect the above problem. The core of our patch is -sandboxIO :: DynFlags -> MVar Status -> IO [HValue] -> IO Status -sandboxIO dflags statusMVar thing = +sandboxIO :: DynFlags -> MVar Status -> MVar (Maybe ThreadId) -> IO [HValue] -> IO Status +sandboxIO dflags statusMVar tidMVar thing = mask $ \restore -> -- fork starts blocked - let runIt = liftM Complete $ try (restore $ rethrow dflags thing) + let thing' = gbracket (myThreadId >>= putMVar tidMVar . Just) + (\() -> modifyMVar_ tidMVar (\_ -> return Nothing)) + (\() -> thing) + runIt = liftM Complete $ try (restore $ rethrow dflags thing') in if dopt Opt_GhciSandbox dflags then do tid <- forkIO $ do res <- runIt putMVar statusMVar res -- empty: can't block That is, sandboxIO takes an additional MVar (Maybe ThreadId): 1. Initially this MVar should be empty. The MVar gets initialized to Just the Thread ID of the sandbox when the user code starts running. This means that if an secondary thread attempts to read the MVar (in order to kill the snippet), that secondary thread will block until the user code starts running -- it will not interrupt ghc when compiling the snippet. 2. When the user code exists the MVar is updated to be Nothing. This means that if the auxiliary thread reads the MVar and finds Nothing it knows that the snippet has already terminated. 3. When the auxiliary thread finds Just a thread ID (it must use withMVar rather than readMVar to avoid a race condition) it can safely throw the asynchronous exception. The remainder of the patch just propagates these changes up so that runStmt takes this MVar as an argument, too. Probably when integrating the patch into ghc it would be better to leave runStmt along and provide a runStmt' that takes the additional argument. Again, any and all feedback on the above would be appreciated. Edsko
On 05/06/13 14:01, Edsko de Vries wrote:
Hi all,
This proposal is related to http://hackage.haskell.org/trac/ghc/ticket/1381, which Simon Marlow closed through commit https://github.com/ghc/ghc/commit/02c4ab049adeb77b8ee0e3b98fbf0f3026eee453.
The problem, in a nutshell, is "how do we terminate a code snippet started with runStmt"? Before Simon's patch the only way was to disable ghc's sandboxing, so that the snippet would run in the same thread as the thread that called runStmt in the first place, and then send an asynchronous exception to that thread. This is the approach we used to take. It's a little tricky to get right (have to make sure that these exceptions are thrown only at the right times), but we thought we had it working okay.
Until, that is, we realized we had a very nasty problem: snippets were being unexpected interrupted. To debug this, we introduced a CustomUserInterruption data type to serve as the exception that we were throwing. This had two purposes: first, we would be use that if we saw a CustomUserInterrupt that it could only have come from one particular throw, and second, the CustomUserInterrupt carried an integer which was incremented on every throw so that we never threw the same exception twice.
Yup, it's a nasty problem indeed.
What we realized is that snippets were being interrupted by *old* exceptions; that is, exceptions that we had thrown to *previous* snippets (and had been caught too). This should obviously never happen. Ian managed to reproduce this behaviour in a completely different setting (http://hackage.haskell.org/trac/ghc/ticket/5902#comment:5) and we think that something similar (unsafePerformIO related) must be happening inside ghc.
Moreover, a further conjecture is that this only happens when runStmt is still compiling the snippet to be executed (as opposed to the snippet actually executing) -- i.e., that the exception gets stuck in the bowels of ghc somewhere. We don't have any hard evidence for this, other than the problem has not appeared again with the proposed patch (but that might be luck, as it depends on timing).
I suspect that there may be thunks inside the compiled expression containing unsafePerformIOs left over by the typechecker (I'd be interested to know how this happens, though).
The patch as we currently have it is against 7.4.2, so pre Simon's change for the sandbox behaviour -- but I don't think that Simon's changes affect the above problem. The core of our patch is
-sandboxIO :: DynFlags -> MVar Status -> IO [HValue] -> IO Status -sandboxIO dflags statusMVar thing = +sandboxIO :: DynFlags -> MVar Status -> MVar (Maybe ThreadId) -> IO [HValue] -> IO Status +sandboxIO dflags statusMVar tidMVar thing = mask $ \restore -> -- fork starts blocked - let runIt = liftM Complete $ try (restore $ rethrow dflags thing) + let thing' = gbracket (myThreadId >>= putMVar tidMVar . Just) + (\() -> modifyMVar_ tidMVar (\_ -> return Nothing)) + (\() -> thing) + runIt = liftM Complete $ try (restore $ rethrow dflags thing') in if dopt Opt_GhciSandbox dflags then do tid <- forkIO $ do res <- runIt putMVar statusMVar res -- empty: can't block
That is, sandboxIO takes an additional MVar (Maybe ThreadId):
1. Initially this MVar should be empty. The MVar gets initialized to Just the Thread ID of the sandbox when the user code starts running. This means that if an secondary thread attempts to read the MVar (in order to kill the snippet), that secondary thread will block until the user code starts running -- it will not interrupt ghc when compiling the snippet. 2. When the user code exists the MVar is updated to be Nothing. This means that if the auxiliary thread reads the MVar and finds Nothing it knows that the snippet has already terminated. 3. When the auxiliary thread finds Just a thread ID (it must use withMVar rather than readMVar to avoid a race condition) it can safely throw the asynchronous exception.
I'm not sure this will work, because I'm not sure there is a defined boundary between "compiling" and "running", i.e. there might be thunks hiding inside that contain the exception. I believe the right way to fix this is to find all the places in the compiler that catch and rethrow exceptions inside unsafePerformIO, and make them use throwTo when the exception is asynchronous. I've done this in the IO library and it works - see Note [async] in GHC.IO.Handle.Internals in the base package. I don't think there will be many places that need to change - the main one looks to be the tryM in forkM_maybe in TcRnMonad. The main thing to worry about is what happens when the throwTo *returns*, because that's what it does when the thunk is poked on again after receiving the async exception the first time. Cheers, Simon
The remainder of the patch just propagates these changes up so that runStmt takes this MVar as an argument, too. Probably when integrating the patch into ghc it would be better to leave runStmt along and provide a runStmt' that takes the additional argument.
Again, any and all feedback on the above would be appreciated.
Edsko
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs
Hi, I believe the right way to fix this is to find all the places in the
compiler that catch and rethrow exceptions inside unsafePerformIO, and make them use throwTo when the exception is asynchronous.
Ok, I have looked at this in detail and I'm not sure how feasible this is. I have filed a bug, with a test case, and a detailed explanation of my investigation so far: http://hackage.haskell.org/trac/ghc/ticket/8006. Edsko
participants (2)
-
Edsko de Vries -
Simon Marlow