shake-0.17: src/General/Pool.hs
-- | Thread pool implementation. The three names correspond to the following
-- priority levels (highest to lowest):
--
-- * 'addPoolException' - things that probably result in a build error,
-- so kick them off quickly.
--
-- * 'addPoolResume' - things that started, blocked, and may have open
-- resources in their closure.
--
-- * 'addPoolStart' - rules that haven't yet started.
--
-- * 'addPoolBatch' - rules that might batch if other rules start first.
module General.Pool(
Pool, runPool,
addPool, PoolPriority(..),
increasePool, keepAlivePool
) where
import Control.Concurrent.Extra
import System.Time.Extra
import Control.Exception
import Control.Monad.Extra
import General.Timing
import General.Extra
import qualified Data.Heap as Heap
import qualified Data.HashSet as Set
import Data.IORef.Extra
import System.Random
---------------------------------------------------------------------
-- THREAD POOL
{-
Must keep a list of active threads, so can raise exceptions in a timely manner
If any worker throws an exception, must signal to all the other workers
-}
data Pool = Pool
!(Var (Maybe S)) -- Current state, 'Nothing' to say we are aborting
!(Barrier (Either SomeException S)) -- Barrier to signal that we are finished
data S = S
{threads :: !(Set.HashSet ThreadId) -- IMPORTANT: Must be strict or we leak thread stacks
,threadsLimit :: {-# UNPACK #-} !Int -- user supplied thread limit, Set.size threads <= threadsLimit
,threadsCount :: {-# UNPACK #-} !Int -- Set.size threads, but in O(1)
,threadsMax :: {-# UNPACK #-} !Int -- high water mark of Set.size threads (accounting only)
,threadsSum :: {-# UNPACK #-} !Int -- number of threads we have been through (accounting only)
,rand :: IO Int -- operation to give us the next random Int
,todo :: !(Heap.Heap (Heap.Entry (PoolPriority, Int) (IO ()))) -- operations waiting a thread
}
emptyS :: Int -> Bool -> IO S
emptyS n deterministic = do
rand <- if not deterministic then return randomIO else do
ref <- newIORef 0
-- no need to be thread-safe - if two threads race they were basically the same time anyway
return $ do i <- readIORef ref; writeIORef' ref (i+1); return i
return $ S Set.empty n 0 0 0 rand Heap.empty
worker :: Pool -> IO ()
worker pool@(Pool var _) = do
let onVar act = modifyVar var $ maybe (return (Nothing, return ())) act
join $ onVar $ \s ->return $ case Heap.uncons $ todo s of
Nothing -> (Just s, return ())
Just (Heap.Entry _ now, todo2) -> (Just s{todo = todo2}, now >> worker pool)
-- | Given a pool, and a function that breaks the S invariants, restore them
-- They are only allowed to touch threadsLimit or todo
step :: Pool -> (S -> IO S) -> IO ()
step pool@(Pool var done) op = do
let onVar act = modifyVar_ var $ maybe (return Nothing) act
onVar $ \s -> do
s <- op s
case Heap.uncons $ todo s of
Just (Heap.Entry _ now, todo2) | threadsCount s < threadsLimit s -> do
-- spawn a new worker
t <- forkFinallyUnmasked (now >> worker pool) $ \res -> case res of
Left e -> onVar $ \s -> do
t <- myThreadId
mapM_ killThread $ Set.toList $ Set.delete t $ threads s
signalBarrier done $ Left e
return Nothing
Right _ -> do
t <- myThreadId
step pool $ \s -> return s{threads = Set.delete t $ threads s, threadsCount = threadsCount s - 1}
return $ Just s{todo = todo2, threads = Set.insert t $ threads s, threadsCount = threadsCount s + 1
,threadsSum = threadsSum s + 1, threadsMax = threadsMax s `max` (threadsCount s + 1)}
Nothing | threadsCount s == 0 -> do
signalBarrier done $ Right s
return Nothing
_ -> return $ Just s
-- | Add a new task to the pool. See the top of the module for the relative ordering
-- and semantics.
addPool :: PoolPriority -> Pool -> IO a -> IO ()
addPool priority pool act = step pool $ \s -> do
i <- rand s
return s{todo = Heap.insert (Heap.Entry (priority, i) $ void act) $ todo s}
data PoolPriority
= PoolException
| PoolResume
| PoolStart
| PoolBatch
| PoolDeprioritize Double
deriving (Eq,Ord)
-- | Temporarily increase the pool by 1 thread. Call the cleanup action to restore the value.
-- After calling cleanup you should requeue onto a new thread.
increasePool :: Pool -> IO (IO ())
increasePool pool = do
step pool $ \s -> return s{threadsLimit = threadsLimit s + 1}
return $ step pool $ \s -> return s{threadsLimit = threadsLimit s - 1}
-- | Make sure the pool cannot run out of tasks (and thus everything finishes) until after the cancel is called.
-- Ensures that a pool that will requeue in time doesn't go idle.
keepAlivePool :: Pool -> IO (IO ())
keepAlivePool pool = do
bar <- newBarrier
addPool PoolResume pool $ do
cancel <- increasePool pool
waitBarrier bar
cancel
return $ signalBarrier bar ()
-- | Run all the tasks in the pool on the given number of works.
-- If any thread throws an exception, the exception will be reraised.
-- When it completes all threads have either finished, or have had 'killThread'
-- called on them (but may not have actually died yet).
runPool :: Bool -> Int -> (Pool -> IO ()) -> IO () -- run all tasks in the pool
runPool deterministic n act = do
s <- newVar . Just =<< emptyS n deterministic
done <- newBarrier
let cleanup = modifyVar_ s $ \s -> do
-- if someone kills our thread, make sure we kill our child threads
case s of
Just s -> mapM_ killThread $ Set.toList $ threads s
Nothing -> return ()
return Nothing
let ghc10793 = do
-- if this thread dies because it is blocked on an MVar there's a chance we have
-- a better error in the done barrier, and GHC raised the exception wrongly, see:
-- https://ghc.haskell.org/trac/ghc/ticket/10793
sleep 1 -- give it a little bit of time for the finally to run
-- no big deal, since the blocked indefinitely takes a while to fire anyway
res <- waitBarrierMaybe done
case res of
Just (Left e) -> throwIO e
_ -> throwIO BlockedIndefinitelyOnMVar
handle (\BlockedIndefinitelyOnMVar -> ghc10793) $ flip onException cleanup $ do
let pool = Pool s done
addPool PoolStart pool $ act pool
res <- waitBarrier done
case res of
Left e -> throwIO e
Right s -> addTiming $ "Pool finished (" ++ show (threadsSum s) ++ " threads, " ++ show (threadsMax s) ++ " max)"