FileMonitor.hs

{-# LANGUAGE BangPatterns #-}
{-# LANGUAGE DeriveFunctor #-}
{-# LANGUAGE DeriveGeneric #-}
{-# LANGUAGE GeneralizedNewtypeDeriving #-}
{-# LANGUAGE NamedFieldPuns #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# OPTIONS_GHC -fno-warn-orphans #-}

-- | An abstraction to help with re-running actions when files or other
-- input values they depend on have changed.
module Distribution.Client.FileMonitor
  ( -- * Declaring files to monitor
      module Distribution.Simple.FileMonitor.Types

    -- * Creating and checking sets of monitored files
  , FileMonitor (..)
  , newFileMonitor
  , MonitorChanged (..)
  , MonitorChangedReason (..)
  , checkFileMonitorChanged
  , updateFileMonitor
  , MonitorTimestamp
  , beginUpdateFileMonitor

    -- * Internal
  , MonitorStateFileSet
  , MonitorStateFile
  , MonitorStateGlob
  ) where

import Distribution.Client.Compat.Prelude
import qualified Distribution.Compat.Binary as Binary
import Prelude ()

import Data.Binary.Get (runGetOrFail)
import qualified Data.ByteString.Lazy as BS
import qualified Data.Map.Strict as Map

import Control.Exception
import Control.Monad
import Control.Monad.Except
  ( ExceptT
  , runExceptT
  , throwError
  , withExceptT
  )
import Control.Monad.State (StateT, mapStateT)
import qualified Control.Monad.State as State
import Control.Monad.Trans (MonadIO, liftIO)

import Distribution.Client.Glob
import Distribution.Client.HashValue
import Distribution.Client.Utils (MergeResult (..), mergeBy)
import Distribution.Compat.Time
import Distribution.Simple.FileMonitor.Types
import Distribution.Simple.Utils (handleDoesNotExist, writeFileAtomic)
import Distribution.Utils.Structured (Tag (..), structuredEncode)

import System.Directory
import System.FilePath
import System.IO

------------------------------------------------------------------------------
-- Implementation types, files status
--

-- | The state necessary to determine whether a set of monitored
-- files has changed.  It consists of two parts: a set of specific
-- files to be monitored (index by their path), and a list of
-- globs, which monitor may files at once.
data MonitorStateFileSet
  = MonitorStateFileSet
      ![MonitorStateFile]
      ![MonitorStateGlob]
  -- Morally this is not actually a set but a bag (represented by lists).
  -- There is no principled reason to use a bag here rather than a set, but
  -- there is also no particular gain either. That said, we do preserve the
  -- order of the lists just to reduce confusion (and have predictable I/O
  -- patterns).
  deriving (Show, Generic)

instance Binary MonitorStateFileSet
instance Structured MonitorStateFileSet

-- | The state necessary to determine whether a monitored file has changed.
--
-- This covers all the cases of 'MonitorFilePath' except for globs which is
-- covered separately by 'MonitorStateGlob'.
--
-- The @Maybe ModTime@ is to cover the case where we already consider the
-- file to have changed, either because it had already changed by the time we
-- did the snapshot (i.e. too new, changed since start of update process) or it
-- no longer exists at all.
data MonitorStateFile
  = MonitorStateFile
      !MonitorKindFile
      !MonitorKindDir
      !FilePath
      !MonitorStateFileStatus
  deriving (Show, Generic)

data MonitorStateFileStatus
  = MonitorStateFileExists
  | -- | cached file mtime
    MonitorStateFileModTime !ModTime
  | -- | cached mtime and content hash
    MonitorStateFileHashed !ModTime !HashValue
  | MonitorStateDirExists
  | -- | cached dir mtime
    MonitorStateDirModTime !ModTime
  | MonitorStateNonExistent
  | MonitorStateAlreadyChanged
  deriving (Show, Generic)

instance Binary MonitorStateFile
instance Binary MonitorStateFileStatus
instance Structured MonitorStateFile
instance Structured MonitorStateFileStatus

-- | The state necessary to determine whether the files matched by a globbing
-- match have changed.
data MonitorStateGlob
  = MonitorStateGlob
      !MonitorKindFile
      !MonitorKindDir
      !FilePathRoot
      !MonitorStateGlobRel
  deriving (Show, Generic)

data MonitorStateGlobRel
  = MonitorStateGlobDirs
      !GlobPieces
      !Glob
      !ModTime
      ![(FilePath, MonitorStateGlobRel)] -- invariant: sorted
  | MonitorStateGlobFiles
      !GlobPieces
      !ModTime
      ![(FilePath, MonitorStateFileStatus)] -- invariant: sorted
  | MonitorStateGlobDirTrailing
  deriving (Show, Generic)

instance Binary MonitorStateGlob
instance Binary MonitorStateGlobRel

instance Structured MonitorStateGlob
instance Structured MonitorStateGlobRel

-- | We can build a 'MonitorStateFileSet' from a set of 'MonitorFilePath' by
-- inspecting the state of the file system, and we can go in the reverse
-- direction by just forgetting the extra info.
reconstructMonitorFilePaths :: MonitorStateFileSet -> [MonitorFilePath]
reconstructMonitorFilePaths (MonitorStateFileSet singlePaths globPaths) =
  map getSinglePath singlePaths ++ map getGlobPath globPaths
  where
    getSinglePath :: MonitorStateFile -> MonitorFilePath
    getSinglePath (MonitorStateFile kindfile kinddir filepath _) =
      MonitorFile kindfile kinddir filepath

    getGlobPath :: MonitorStateGlob -> MonitorFilePath
    getGlobPath (MonitorStateGlob kindfile kinddir root gstate) =
      MonitorFileGlob kindfile kinddir $
        RootedGlob root $
          case gstate of
            MonitorStateGlobDirs glob globs _ _ -> GlobDir glob globs
            MonitorStateGlobFiles glob _ _ -> GlobFile glob
            MonitorStateGlobDirTrailing -> GlobDirTrailing

------------------------------------------------------------------------------
-- Checking the status of monitored files
--

-- | A monitor for detecting changes to a set of files. It can be used to
-- efficiently test if any of a set of files (specified individually or by
-- glob patterns) has changed since some snapshot. In addition, it also checks
-- for changes in a value (of type @a@), and when there are no changes in
-- either it returns a saved value (of type @b@).
--
-- The main use case looks like this: suppose we have some expensive action
-- that depends on certain pure inputs and reads some set of files, and
-- produces some pure result. We want to avoid re-running this action when it
-- would produce the same result. So we need to monitor the files the action
-- looked at, the other pure input values, and we need to cache the result.
-- Then at some later point, if the input value didn't change, and none of the
-- files changed, then we can re-use the cached result rather than re-running
-- the action.
--
-- This can be achieved using a 'FileMonitor'. Each 'FileMonitor' instance
-- saves state in a disk file, so the file for that has to be specified,
-- making sure it is unique. The pattern is to use 'checkFileMonitorChanged'
-- to see if there's been any change. If there is, re-run the action, keeping
-- track of the files, then use 'updateFileMonitor' to record the current
-- set of files to monitor, the current input value for the action, and the
-- result of the action.
--
-- The typical occurrence of this pattern is captured by 'rerunIfChanged'
-- and the 'Rebuild' monad. More complicated cases may need to use
-- 'checkFileMonitorChanged' and 'updateFileMonitor' directly.
data FileMonitor a b = FileMonitor
  { fileMonitorCacheFile :: FilePath
  -- ^ The file where this 'FileMonitor' should store its state.
  , fileMonitorKeyValid :: a -> a -> Bool
  -- ^ Compares a new cache key with old one to determine if a
  -- corresponding cached value is still valid.
  --
  -- Typically this is just an equality test, but in some
  -- circumstances it can make sense to do things like subset
  -- comparisons.
  --
  -- The first arg is the new value, the second is the old cached value.
  , fileMonitorCheckIfOnlyValueChanged :: Bool
  -- ^ When this mode is enabled, if 'checkFileMonitorChanged' returns
  -- 'MonitoredValueChanged' then we have the guarantee that no files
  -- changed, that the value change was the only change. In the default
  -- mode no such guarantee is provided which is slightly faster.
  }

-- | Define a new file monitor.
--
-- It's best practice to define file monitor values once, and then use the
-- same value for 'checkFileMonitorChanged' and 'updateFileMonitor' as this
-- ensures you get the same types @a@ and @b@ for reading and writing.
--
-- The path of the file monitor itself must be unique because it keeps state
-- on disk and these would clash.
newFileMonitor
  :: Eq a
  => FilePath
  -- ^ The file to cache the state of the
  -- file monitor. Must be unique.
  -> FileMonitor a b
newFileMonitor path = FileMonitor path (==) False

-- | The result of 'checkFileMonitorChanged': either the monitored files or
-- value changed (and it tells us which it was) or nothing changed and we get
-- the cached result.
data MonitorChanged a b
  = -- | The monitored files and value did not change. The cached result is
    -- @b@.
    --
    -- The set of monitored files is also returned. This is useful
    -- for composing or nesting 'FileMonitor's.
    MonitorUnchanged b [MonitorFilePath]
  | -- | The monitor found that something changed. The reason is given.
    MonitorChanged (MonitorChangedReason a)
  deriving (Show)

-- | What kind of change 'checkFileMonitorChanged' detected.
data MonitorChangedReason a
  = -- | One of the files changed (existence, file type, mtime or file
    -- content, depending on the 'MonitorFilePath' in question)
    MonitoredFileChanged FilePath
  | -- | The pure input value changed.
    --
    -- The previous cached key value is also returned. This is sometimes
    -- useful when using a 'fileMonitorKeyValid' function that is not simply
    -- '(==)', when invalidation can be partial. In such cases it can make
    -- sense to 'updateFileMonitor' with a key value that's a combination of
    -- the new and old (e.g. set union).
    MonitoredValueChanged a
  | -- | There was no saved monitor state, cached value etc. Ie the file
    -- for the 'FileMonitor' does not exist.
    MonitorFirstRun
  | -- | There was existing state, but we could not read it. This typically
    -- happens when the code has changed compared to an existing 'FileMonitor'
    -- cache file and type of the input value or cached value has changed such
    -- that we cannot decode the values. This is completely benign as we can
    -- treat is just as if there were no cache file and re-run.
    MonitorCorruptCache
  deriving (Eq, Show, Functor)

-- | Test if the input value or files monitored by the 'FileMonitor' have
-- changed. If not, return the cached value.
--
-- See 'FileMonitor' for a full explanation.
checkFileMonitorChanged
  :: forall a b
   . (Binary a, Structured a, Binary b, Structured b)
  => FileMonitor a b
  -- ^ cache file path
  -> FilePath
  -- ^ root directory
  -> a
  -- ^ guard or key value
  -> IO (MonitorChanged a b)
  -- ^ did the key or any paths change?
checkFileMonitorChanged
  monitor@FileMonitor
    { fileMonitorKeyValid
    , fileMonitorCheckIfOnlyValueChanged
    }
  root
  currentKey =
    -- Consider it a change if the cache file does not exist,
    -- or we cannot decode it. Sadly ErrorCall can still happen, despite
    -- using decodeFileOrFail, e.g. Data.Char.chr errors

    handleDoesNotExist (MonitorChanged MonitorFirstRun) $
      handleErrorCall (MonitorChanged MonitorCorruptCache) $
        withCacheFile monitor $
          either
            (\_ -> return (MonitorChanged MonitorCorruptCache))
            checkStatusCache
    where
      checkStatusCache :: (MonitorStateFileSet, a, Either String b) -> IO (MonitorChanged a b)
      checkStatusCache (cachedFileStatus, cachedKey, cachedResult) = do
        change <- checkForChanges
        case change of
          Just reason -> return (MonitorChanged reason)
          Nothing -> case cachedResult of
            Left _ -> pure (MonitorChanged MonitorCorruptCache)
            Right cr -> return (MonitorUnchanged cr monitorFiles)
            where
              monitorFiles = reconstructMonitorFilePaths cachedFileStatus
        where
          -- In fileMonitorCheckIfOnlyValueChanged mode we want to guarantee that
          -- if we return MonitoredValueChanged that only the value changed.
          -- We do that by checking for file changes first. Otherwise it makes
          -- more sense to do the cheaper test first.
          checkForChanges :: IO (Maybe (MonitorChangedReason a))
          checkForChanges
            | fileMonitorCheckIfOnlyValueChanged =
                checkFileChange cachedFileStatus cachedKey cachedResult
                  `mplusMaybeT` checkValueChange cachedKey
            | otherwise =
                checkValueChange cachedKey
                  `mplusMaybeT` checkFileChange cachedFileStatus cachedKey cachedResult

      mplusMaybeT :: Monad m => m (Maybe a1) -> m (Maybe a1) -> m (Maybe a1)
      mplusMaybeT ma mb = do
        mx <- ma
        case mx of
          Nothing -> mb
          Just x -> return (Just x)

      -- Check if the guard value has changed
      checkValueChange :: a -> IO (Maybe (MonitorChangedReason a))
      checkValueChange cachedKey
        | not (fileMonitorKeyValid currentKey cachedKey) =
            return (Just (MonitoredValueChanged cachedKey))
        | otherwise =
            return Nothing

      -- Check if any file has changed
      checkFileChange :: MonitorStateFileSet -> a -> Either String b -> IO (Maybe (MonitorChangedReason a))
      checkFileChange cachedFileStatus cachedKey cachedResult = do
        res <- probeFileSystem root cachedFileStatus
        case res of
          -- Some monitored file has changed
          Left changedPath ->
            return (Just (MonitoredFileChanged (normalise changedPath)))
          -- No monitored file has changed
          Right (cachedFileStatus', cacheStatus) -> do
            -- But we might still want to update the cache
            whenCacheChanged cacheStatus $
              case cachedResult of
                Left _ -> pure ()
                Right cr -> rewriteCacheFile monitor cachedFileStatus' cachedKey cr

            return Nothing

-- | Lazily decode a triple, parsing the first two fields strictly and
-- returning a lazy value containing either the last one or an error.
-- This is helpful for cabal cache files where the first two components
-- contain header data that lets one test if the cache is still valid,
-- and the last (potentially large) component is the cached value itself.
-- This way we can test for cache validity without needing to pay the
-- cost of the decode of stale cache data. This lives here rather than
-- Distribution.Utils.Structured because it depends on a newer version of
-- binary than supported in the Cabal library proper.
structuredDecodeTriple
  :: forall a b c
   . (Structured a, Structured b, Structured c, Binary.Binary a, Binary.Binary b, Binary.Binary c)
  => BS.ByteString
  -> Either String (a, b, Either String c)
structuredDecodeTriple lbs =
  let partialDecode =
        (`runGetOrFail` lbs) $ do
          (_ :: Tag (a, b, c)) <- Binary.get
          (a :: a) <- Binary.get
          (b :: b) <- Binary.get
          pure (a, b)
      cleanEither (Left (_, pos, msg)) = Left ("Data.Binary.Get.runGet at position " ++ show pos ++ ": " ++ msg)
      cleanEither (Right (_, _, v)) = Right v
   in case partialDecode of
        Left (_, pos, msg) -> Left ("Data.Binary.Get.runGet at position " ++ show pos ++ ": " ++ msg)
        Right (lbs', _, (x, y)) -> Right (x, y, cleanEither $ runGetOrFail (Binary.get :: Binary.Get c) lbs')

-- | Helper for reading the cache file.
--
-- This determines the type and format of the binary cache file.
withCacheFile
  :: (Binary a, Structured a, Binary b, Structured b)
  => FileMonitor a b
  -> (Either String (MonitorStateFileSet, a, Either String b) -> IO r)
  -> IO r
withCacheFile (FileMonitor{fileMonitorCacheFile}) k =
  withBinaryFile fileMonitorCacheFile ReadMode $ \hnd -> do
    contents <- structuredDecodeTriple <$> BS.hGetContents hnd
    k contents

-- | Helper for writing the cache file.
--
-- This determines the type and format of the binary cache file.
rewriteCacheFile
  :: (Binary a, Structured a, Binary b, Structured b)
  => FileMonitor a b
  -> MonitorStateFileSet
  -> a
  -> b
  -> IO ()
rewriteCacheFile FileMonitor{fileMonitorCacheFile} fileset key result =
  writeFileAtomic fileMonitorCacheFile $
    structuredEncode (fileset, key, result)

-- | Probe the file system to see if any of the monitored files have changed.
--
-- It returns Nothing if any file changed, or returns a possibly updated
-- file 'MonitorStateFileSet' plus an indicator of whether it actually changed.
--
-- We may need to update the cache since there may be changes in the filesystem
-- state which don't change any of our affected files.
--
-- Consider the glob @{proj1,proj2}\/\*.cabal@. Say we first run and find a
-- @proj1@ directory containing @proj1.cabal@ yet no @proj2@. If we later run
-- and find @proj2@ was created, yet contains no files matching @*.cabal@ then
-- we want to update the cache despite no changes in our relevant file set.
-- Specifically, we should add an mtime for this directory so we can avoid
-- re-traversing the directory in future runs.
probeFileSystem
  :: FilePath
  -> MonitorStateFileSet
  -> IO (Either FilePath (MonitorStateFileSet, CacheChanged))
probeFileSystem root (MonitorStateFileSet singlePaths globPaths) =
  runChangedM $ do
    sequence_
      [ probeMonitorStateFileStatus root file status
      | MonitorStateFile _ _ file status <- singlePaths
      ]
    -- The glob monitors can require state changes
    globPaths' <-
      sequence
        [ probeMonitorStateGlob root globPath
        | globPath <- globPaths
        ]
    return (MonitorStateFileSet singlePaths globPaths')

-----------------------------------------------
-- Monad for checking for file system changes
--
-- We need to be able to bail out if we detect a change (using ExceptT),
-- but if there's no change we need to be able to rebuild the monitor
-- state. And we want to optimise that rebuilding by keeping track if
-- anything actually changed (using StateT), so that in the typical case
-- we can avoid rewriting the state file.

newtype ChangedM a = ChangedM (StateT CacheChanged (ExceptT FilePath IO) a)
  deriving (Functor, Applicative, Monad, MonadIO)

runChangedM :: ChangedM a -> IO (Either FilePath (a, CacheChanged))
runChangedM (ChangedM action) =
  runExceptT $ State.runStateT action CacheUnchanged

somethingChanged :: FilePath -> ChangedM a
somethingChanged path = ChangedM $ throwError path

cacheChanged :: ChangedM ()
cacheChanged = ChangedM $ State.put CacheChanged

mapChangedFile :: (FilePath -> FilePath) -> ChangedM a -> ChangedM a
mapChangedFile adjust (ChangedM a) =
  ChangedM (mapStateT (withExceptT adjust) a)

data CacheChanged = CacheChanged | CacheUnchanged

whenCacheChanged :: Monad m => CacheChanged -> m () -> m ()
whenCacheChanged CacheChanged action = action
whenCacheChanged CacheUnchanged _ = return ()

----------------------

-- | Probe the file system to see if a single monitored file has changed.
probeMonitorStateFileStatus
  :: FilePath
  -> FilePath
  -> MonitorStateFileStatus
  -> ChangedM ()
probeMonitorStateFileStatus root file status =
  case status of
    MonitorStateFileExists ->
      probeFileExistence root file
    MonitorStateFileModTime mtime ->
      probeFileModificationTime root file mtime
    MonitorStateFileHashed mtime hash ->
      probeFileModificationTimeAndHash root file mtime hash
    MonitorStateDirExists ->
      probeDirExistence root file
    MonitorStateDirModTime mtime ->
      probeFileModificationTime root file mtime
    MonitorStateNonExistent ->
      probeFileNonExistence root file
    MonitorStateAlreadyChanged ->
      somethingChanged file

-- | Probe the file system to see if a monitored file glob has changed.
probeMonitorStateGlob
  :: FilePath
  -- ^ root path
  -> MonitorStateGlob
  -> ChangedM MonitorStateGlob
probeMonitorStateGlob
  relroot
  (MonitorStateGlob kindfile kinddir globroot glob) = do
    root <- liftIO $ getFilePathRootDirectory globroot relroot
    case globroot of
      FilePathRelative ->
        MonitorStateGlob kindfile kinddir globroot
          <$> probeMonitorStateGlobRel kindfile kinddir root "." glob
      -- for absolute cases, make the changed file we report absolute too
      _ ->
        mapChangedFile (root </>) $
          MonitorStateGlob kindfile kinddir globroot
            <$> probeMonitorStateGlobRel kindfile kinddir root "" glob

probeMonitorStateGlobRel
  :: MonitorKindFile
  -> MonitorKindDir
  -> FilePath
  -- ^ root path
  -> FilePath
  -- ^ path of the directory we are
  --   looking in relative to @root@
  -> MonitorStateGlobRel
  -> ChangedM MonitorStateGlobRel
probeMonitorStateGlobRel
  kindfile
  kinddir
  root
  dirName
  (MonitorStateGlobDirs glob globPath mtime children) = do
    change <- liftIO $ checkDirectoryModificationTime (root </> dirName) mtime
    case change of
      Nothing -> do
        children' <-
          sequence
            [ do
              fstate' <-
                probeMonitorStateGlobRel
                  kindfile
                  kinddir
                  root
                  (dirName </> fname)
                  fstate
              return (fname, fstate')
            | (fname, fstate) <- children
            ]
        return $! MonitorStateGlobDirs glob globPath mtime children'
      Just mtime' -> do
        -- directory modification time changed:
        -- a matching subdir may have been added or deleted
        matches <-
          filterM
            ( \entry ->
                let subdir = root </> dirName </> entry
                 in liftIO $ doesDirectoryExist subdir
            )
            . filter (matchGlobPieces glob)
            =<< liftIO (getDirectoryContents (root </> dirName))

        children' <-
          traverse probeMergeResult $
            mergeBy
              (\(path1, _) path2 -> compare path1 path2)
              children
              (sort matches)
        return $! MonitorStateGlobDirs glob globPath mtime' children'
    where
      -- Note that just because the directory has changed, we don't force
      -- a cache rewrite with 'cacheChanged' since that has some cost, and
      -- all we're saving is scanning the directory. But we do rebuild the
      -- cache with the new mtime', so that if the cache is rewritten for
      -- some other reason, we'll take advantage of that.

      probeMergeResult
        :: MergeResult (FilePath, MonitorStateGlobRel) FilePath
        -> ChangedM (FilePath, MonitorStateGlobRel)

      -- Only in cached (directory deleted)
      probeMergeResult (OnlyInLeft (path, fstate)) = do
        case allMatchingFiles (dirName </> path) fstate of
          [] -> return (path, fstate)
          -- Strictly speaking we should be returning 'CacheChanged' above
          -- as we should prune the now-missing 'MonitorStateGlobRel'. However
          -- we currently just leave these now-redundant entries in the
          -- cache as they cost no IO and keeping them allows us to avoid
          -- rewriting the cache.
          (file : _) -> somethingChanged file

      -- Only in current filesystem state (directory added)
      probeMergeResult (OnlyInRight path) = do
        fstate <-
          liftIO $
            buildMonitorStateGlobRel
              Nothing
              Map.empty
              kindfile
              kinddir
              root
              (dirName </> path)
              globPath
        case allMatchingFiles (dirName </> path) fstate of
          (file : _) -> somethingChanged file
          -- This is the only case where we use 'cacheChanged' because we can
          -- have a whole new dir subtree (of unbounded size and cost), so we
          -- need to save the state of that new subtree in the cache.
          [] -> cacheChanged >> return (path, fstate)

      -- Found in path
      probeMergeResult (InBoth (path, fstate) _) = do
        fstate' <-
          probeMonitorStateGlobRel
            kindfile
            kinddir
            root
            (dirName </> path)
            fstate
        return (path, fstate')

      -- \| Does a 'MonitorStateGlob' have any relevant files within it?
      allMatchingFiles :: FilePath -> MonitorStateGlobRel -> [FilePath]
      allMatchingFiles dir (MonitorStateGlobFiles _ _ entries) =
        [dir </> fname | (fname, _) <- entries]
      allMatchingFiles dir (MonitorStateGlobDirs _ _ _ entries) =
        [ res
        | (subdir, fstate) <- entries
        , res <- allMatchingFiles (dir </> subdir) fstate
        ]
      allMatchingFiles dir MonitorStateGlobDirTrailing =
        [dir]
probeMonitorStateGlobRel
  _
  _
  root
  dirName
  (MonitorStateGlobFiles glob mtime children) = do
    change <- liftIO $ checkDirectoryModificationTime (root </> dirName) mtime
    mtime' <- case change of
      Nothing -> return mtime
      Just mtime' -> do
        -- directory modification time changed:
        -- a matching file may have been added or deleted
        matches <-
          return . filter (matchGlobPieces glob)
            =<< liftIO (getDirectoryContents (root </> dirName))

        traverse_ probeMergeResult $
          mergeBy
            (\(path1, _) path2 -> compare path1 path2)
            children
            (sort matches)
        return mtime'

    -- Check that none of the children have changed
    for_ children $ \(file, status) ->
      probeMonitorStateFileStatus root (dirName </> file) status

    return (MonitorStateGlobFiles glob mtime' children)
    where
      -- Again, we don't force a cache rewrite with 'cacheChanged', but we do use
      -- the new mtime' if any.

      probeMergeResult
        :: MergeResult (FilePath, MonitorStateFileStatus) FilePath
        -> ChangedM ()
      probeMergeResult mr = case mr of
        InBoth _ _ -> return ()
        -- this is just to be able to accurately report which file changed:
        OnlyInLeft (path, _) -> somethingChanged (dirName </> path)
        OnlyInRight path -> somethingChanged (dirName </> path)
probeMonitorStateGlobRel _ _ _ _ MonitorStateGlobDirTrailing =
  return MonitorStateGlobDirTrailing

------------------------------------------------------------------------------

-- | Update the input value and the set of files monitored by the
-- 'FileMonitor', plus the cached value that may be returned in future.
--
-- This takes a snapshot of the state of the monitored files right now, so
-- 'checkFileMonitorChanged' will look for file system changes relative to
-- this snapshot.
--
-- This is typically done once the action has been completed successfully and
-- we have the action's result and we know what files it looked at. See
-- 'FileMonitor' for a full explanation.
--
-- If we do take the snapshot after the action has completed then we have a
-- problem. The problem is that files might have changed /while/ the action was
-- running but /after/ the action read them. If we take the snapshot after the
-- action completes then we will miss these changes. The solution is to record
-- a timestamp before beginning execution of the action and then we make the
-- conservative assumption that any file that has changed since then has
-- already changed, ie the file monitor state for these files will be such that
-- 'checkFileMonitorChanged' will report that they have changed.
--
-- So if you do use 'updateFileMonitor' after the action (so you can discover
-- the files used rather than predicting them in advance) then use
-- 'beginUpdateFileMonitor' to get a timestamp and pass that. Alternatively,
-- if you take the snapshot in advance of the action, or you're not monitoring
-- any files then you can use @Nothing@ for the timestamp parameter.
updateFileMonitor
  :: (Binary a, Structured a, Binary b, Structured b)
  => FileMonitor a b
  -- ^ cache file path
  -> FilePath
  -- ^ root directory
  -> Maybe MonitorTimestamp
  -- ^ timestamp when the update action started
  -> [MonitorFilePath]
  -- ^ files of interest relative to root
  -> a
  -- ^ the current key value
  -> b
  -- ^ the current result value
  -> IO ()
updateFileMonitor
  monitor
  root
  startTime
  monitorFiles
  cachedKey
  cachedResult = do
    hashcache <- readCacheFileHashes monitor
    msfs <- buildMonitorStateFileSet startTime hashcache root monitorFiles
    rewriteCacheFile monitor msfs cachedKey cachedResult

-- | A timestamp to help with the problem of file changes during actions.
-- See 'updateFileMonitor' for details.
newtype MonitorTimestamp = MonitorTimestamp ModTime

-- | Record a timestamp at the beginning of an action, and when the action
-- completes call 'updateFileMonitor' passing it the timestamp.
-- See 'updateFileMonitor' for details.
beginUpdateFileMonitor :: IO MonitorTimestamp
beginUpdateFileMonitor = MonitorTimestamp <$> getCurTime

-- | Take the snapshot of the monitored files. That is, given the
-- specification of the set of files we need to monitor, inspect the state
-- of the file system now and collect the information we'll need later to
-- determine if anything has changed.
buildMonitorStateFileSet
  :: Maybe MonitorTimestamp
  -- ^ optional: timestamp
  -- of the start of the action
  -> FileHashCache
  -- ^ existing file hashes
  -> FilePath
  -- ^ root directory
  -> [MonitorFilePath]
  -- ^ patterns of interest
  --   relative to root
  -> IO MonitorStateFileSet
buildMonitorStateFileSet mstartTime hashcache root =
  go [] []
  where
    go
      :: [MonitorStateFile]
      -> [MonitorStateGlob]
      -> [MonitorFilePath]
      -> IO MonitorStateFileSet
    go !singlePaths !globPaths [] =
      return (MonitorStateFileSet (reverse singlePaths) (reverse globPaths))
    go
      !singlePaths
      !globPaths
      (MonitorFile kindfile kinddir path : monitors) = do
        monitorState <-
          MonitorStateFile kindfile kinddir path
            <$> buildMonitorStateFile
              mstartTime
              hashcache
              kindfile
              kinddir
              root
              path
        go (monitorState : singlePaths) globPaths monitors
    go
      !singlePaths
      !globPaths
      (MonitorFileGlob kindfile kinddir globPath : monitors) = do
        monitorState <-
          buildMonitorStateGlob
            mstartTime
            hashcache
            kindfile
            kinddir
            root
            globPath
        go singlePaths (monitorState : globPaths) monitors

buildMonitorStateFile
  :: Maybe MonitorTimestamp
  -- ^ start time of update
  -> FileHashCache
  -- ^ existing file hashes
  -> MonitorKindFile
  -> MonitorKindDir
  -> FilePath
  -- ^ the root directory
  -> FilePath
  -> IO MonitorStateFileStatus
buildMonitorStateFile mstartTime hashcache kindfile kinddir root path = do
  let abspath = root </> path
  isFile <- doesFileExist abspath
  isDir <- doesDirectoryExist abspath
  case (isFile, kindfile, isDir, kinddir) of
    (_, FileNotExists, _, DirNotExists) ->
      -- we don't need to care if it exists now, since we check at probe time
      return MonitorStateNonExistent
    (False, _, False, _) ->
      return MonitorStateAlreadyChanged
    (True, FileExists, _, _) ->
      return MonitorStateFileExists
    (True, FileModTime, _, _) ->
      handleIOException MonitorStateAlreadyChanged $ do
        mtime <- getModTime abspath
        if changedDuringUpdate mstartTime mtime
          then return MonitorStateAlreadyChanged
          else return (MonitorStateFileModTime mtime)
    (True, FileHashed, _, _) ->
      handleIOException MonitorStateAlreadyChanged $ do
        mtime <- getModTime abspath
        if changedDuringUpdate mstartTime mtime
          then return MonitorStateAlreadyChanged
          else do
            hash <- getFileHash hashcache abspath abspath mtime
            return (MonitorStateFileHashed mtime hash)
    (_, _, True, DirExists) ->
      return MonitorStateDirExists
    (_, _, True, DirModTime) ->
      handleIOException MonitorStateAlreadyChanged $ do
        mtime <- getModTime abspath
        if changedDuringUpdate mstartTime mtime
          then return MonitorStateAlreadyChanged
          else return (MonitorStateDirModTime mtime)
    (False, _, True, DirNotExists) -> return MonitorStateAlreadyChanged
    (True, FileNotExists, False, _) -> return MonitorStateAlreadyChanged

-- | If we have a timestamp for the beginning of the update, then any file
-- mtime later than this means that it changed during the update and we ought
-- to consider the file as already changed.
changedDuringUpdate :: Maybe MonitorTimestamp -> ModTime -> Bool
changedDuringUpdate (Just (MonitorTimestamp startTime)) mtime =
  mtime > startTime
changedDuringUpdate _ _ = False

-- | Much like 'buildMonitorStateFileSet' but for the somewhat complicated case
-- of a file glob.
--
-- This gets used both by 'buildMonitorStateFileSet' when we're taking the
-- file system snapshot, but also by 'probeGlobStatus' as part of checking
-- the monitored (globed) files for changes when we find a whole new subtree.
buildMonitorStateGlob
  :: Maybe MonitorTimestamp
  -- ^ start time of update
  -> FileHashCache
  -- ^ existing file hashes
  -> MonitorKindFile
  -> MonitorKindDir
  -> FilePath
  -- ^ the root directory
  -> RootedGlob
  -- ^ the matching glob
  -> IO MonitorStateGlob
buildMonitorStateGlob
  mstartTime
  hashcache
  kindfile
  kinddir
  relroot
  (RootedGlob globroot globPath) = do
    root <- liftIO $ getFilePathRootDirectory globroot relroot
    MonitorStateGlob kindfile kinddir globroot
      <$> buildMonitorStateGlobRel
        mstartTime
        hashcache
        kindfile
        kinddir
        root
        "."
        globPath

buildMonitorStateGlobRel
  :: Maybe MonitorTimestamp
  -- ^ start time of update
  -> FileHashCache
  -- ^ existing file hashes
  -> MonitorKindFile
  -> MonitorKindDir
  -> FilePath
  -- ^ the root directory
  -> FilePath
  -- ^ directory we are examining
  --   relative to the root
  -> Glob
  -- ^ the matching glob
  -> IO MonitorStateGlobRel
buildMonitorStateGlobRel
  mstartTime
  hashcache
  kindfile
  kinddir
  root
  dir
  globPath = do
    let absdir = root </> dir
    dirEntries <- getDirectoryContents absdir
    dirMTime <- getModTime absdir
    case globPath of
      GlobDirRecursive{} -> error "Monitoring directory-recursive globs (i.e. ../**/...) is currently unsupported"
      GlobDir glob globPath' -> do
        subdirs <-
          filterM (\subdir -> doesDirectoryExist (absdir </> subdir)) $
            filter (matchGlobPieces glob) dirEntries
        subdirStates <-
          for (sort subdirs) $ \subdir -> do
            fstate <-
              buildMonitorStateGlobRel
                mstartTime
                hashcache
                kindfile
                kinddir
                root
                (dir </> subdir)
                globPath'
            return (subdir, fstate)
        return $! MonitorStateGlobDirs glob globPath' dirMTime subdirStates
      GlobFile glob -> do
        let files = filter (matchGlobPieces glob) dirEntries
        filesStates <-
          for (sort files) $ \file -> do
            fstate <-
              buildMonitorStateFile
                mstartTime
                hashcache
                kindfile
                kinddir
                root
                (dir </> file)
            return (file, fstate)
        return $! MonitorStateGlobFiles glob dirMTime filesStates
      GlobDirTrailing ->
        return MonitorStateGlobDirTrailing

-- | We really want to avoid re-hashing files all the time. We already make
-- the assumption that if a file mtime has not changed then we don't need to
-- bother checking if the content hash has changed. We can apply the same
-- assumption when updating the file monitor state. In the typical case of
-- updating a file monitor the set of files is the same or largely the same so
-- we can grab the previously known content hashes with their corresponding
-- mtimes.
type FileHashCache = Map FilePath (ModTime, HashValue)

-- | We declare it a cache hit if the mtime of a file is the same as before.
lookupFileHashCache :: FileHashCache -> FilePath -> ModTime -> Maybe HashValue
lookupFileHashCache hashcache file mtime = do
  (mtime', hash) <- Map.lookup file hashcache
  guard (mtime' == mtime)
  return hash

-- | Either get it from the cache or go read the file
getFileHash :: FileHashCache -> FilePath -> FilePath -> ModTime -> IO HashValue
getFileHash hashcache relfile absfile mtime =
  case lookupFileHashCache hashcache relfile mtime of
    Just hash -> return hash
    Nothing -> readFileHashValue absfile

-- | Build a 'FileHashCache' from the previous 'MonitorStateFileSet'. While
-- in principle we could preserve the structure of the previous state, given
-- that the set of files to monitor can change then it's simpler just to throw
-- away the structure and use a finite map.
readCacheFileHashes
  :: (Binary a, Structured a, Binary b, Structured b)
  => FileMonitor a b
  -> IO FileHashCache
readCacheFileHashes monitor =
  handleDoesNotExist Map.empty $
    handleErrorCall Map.empty $
      withCacheFile monitor $ \res ->
        case res of
          Left _ -> return Map.empty
          Right (msfs, _, _) -> return (mkFileHashCache msfs)
  where
    mkFileHashCache :: MonitorStateFileSet -> FileHashCache
    mkFileHashCache (MonitorStateFileSet singlePaths globPaths) =
      collectAllFileHashes singlePaths
        `Map.union` collectAllGlobHashes globPaths

    collectAllFileHashes :: [MonitorStateFile] -> Map FilePath (ModTime, HashValue)
    collectAllFileHashes singlePaths =
      Map.fromList
        [ (fpath, (mtime, hash))
        | MonitorStateFile
            _
            _
            fpath
            (MonitorStateFileHashed mtime hash) <-
            singlePaths
        ]

    collectAllGlobHashes :: [MonitorStateGlob] -> Map FilePath (ModTime, HashValue)
    collectAllGlobHashes globPaths =
      Map.fromList
        [ (fpath, (mtime, hash))
        | MonitorStateGlob _ _ _ gstate <- globPaths
        , (fpath, (mtime, hash)) <- collectGlobHashes "" gstate
        ]

    collectGlobHashes :: FilePath -> MonitorStateGlobRel -> [(FilePath, (ModTime, HashValue))]
    collectGlobHashes dir (MonitorStateGlobDirs _ _ _ entries) =
      [ res
      | (subdir, fstate) <- entries
      , res <- collectGlobHashes (dir </> subdir) fstate
      ]
    collectGlobHashes dir (MonitorStateGlobFiles _ _ entries) =
      [ (dir </> fname, (mtime, hash))
      | (fname, MonitorStateFileHashed mtime hash) <- entries
      ]
    collectGlobHashes _dir MonitorStateGlobDirTrailing =
      []

------------------------------------------------------------------------------
-- Utils
--

-- | Within the @root@ directory, check if @file@ has its 'ModTime' is
-- the same as @mtime@, short-circuiting if it is different.
probeFileModificationTime :: FilePath -> FilePath -> ModTime -> ChangedM ()
probeFileModificationTime root file mtime = do
  unchanged <- liftIO $ checkModificationTimeUnchanged root file mtime
  unless unchanged (somethingChanged file)

-- | Within the @root@ directory, check if @file@ has its 'ModTime' and
-- 'HashValue' is the same as @mtime@ and @hash@, short-circuiting if it is
-- different.
probeFileModificationTimeAndHash
  :: FilePath
  -> FilePath
  -> ModTime
  -> HashValue
  -> ChangedM ()
probeFileModificationTimeAndHash root file mtime hash = do
  unchanged <-
    liftIO $
      checkFileModificationTimeAndHashUnchanged root file mtime hash
  unless unchanged (somethingChanged file)

-- | Within the @root@ directory, check if @file@ still exists as a file.
-- If it *does not* exist, short-circuit.
probeFileExistence :: FilePath -> FilePath -> ChangedM ()
probeFileExistence root file = do
  existsFile <- liftIO $ doesFileExist (root </> file)
  unless existsFile (somethingChanged file)

-- | Within the @root@ directory, check if @dir@ still exists.
-- If it *does not* exist, short-circuit.
probeDirExistence :: FilePath -> FilePath -> ChangedM ()
probeDirExistence root dir = do
  existsDir <- liftIO $ doesDirectoryExist (root </> dir)
  unless existsDir (somethingChanged dir)

-- | Within the @root@ directory, check if @file@ still does not exist.
-- If it *does* exist, short-circuit.
probeFileNonExistence :: FilePath -> FilePath -> ChangedM ()
probeFileNonExistence root file = do
  existsFile <- liftIO $ doesFileExist (root </> file)
  existsDir <- liftIO $ doesDirectoryExist (root </> file)
  when (existsFile || existsDir) (somethingChanged file)

-- | Returns @True@ if, inside the @root@ directory, @file@ has the same
-- 'ModTime' as @mtime@.
checkModificationTimeUnchanged
  :: FilePath
  -> FilePath
  -> ModTime
  -> IO Bool
checkModificationTimeUnchanged root file mtime =
  handleIOException False $ do
    mtime' <- getModTime (root </> file)
    return (mtime == mtime')

-- | Returns @True@ if, inside the @root@ directory, @file@ has the
-- same 'ModTime' and 'HashValue' as @mtime and @chash@.
checkFileModificationTimeAndHashUnchanged
  :: FilePath
  -> FilePath
  -> ModTime
  -> HashValue
  -> IO Bool
checkFileModificationTimeAndHashUnchanged root file mtime chash =
  handleIOException False $ do
    mtime' <- getModTime (root </> file)
    if mtime == mtime'
      then return True
      else do
        chash' <- readFileHashValue (root </> file)
        return (chash == chash')

-- | Given a directory @dir@, return @Nothing@ if its 'ModTime'
-- is the same as @mtime@, and the new 'ModTime' if it is not.
checkDirectoryModificationTime :: FilePath -> ModTime -> IO (Maybe ModTime)
checkDirectoryModificationTime dir mtime =
  handleIOException Nothing $ do
    mtime' <- getModTime dir
    if mtime == mtime'
      then return Nothing
      else return (Just mtime')

-- | Run an IO computation, returning the first argument @e@ if there is an 'error'
-- call. ('ErrorCall')
handleErrorCall :: a -> IO a -> IO a
handleErrorCall e = handle handler
  where
    handler (ErrorCall _) = return e

-- | Run an IO computation, returning @e@ if there is any 'IOException'.
--
-- This policy is OK in the file monitor code because it just causes the
-- monitor to report that something changed, and then code reacting to that
-- will normally encounter the same IO exception when it re-runs the action
-- that uses the file.
handleIOException :: a -> IO a -> IO a
handleIOException e =
  handle (anyIOException e)
  where
    anyIOException :: a -> IOException -> IO a
    anyIOException x _ = return x

------------------------------------------------------------------------------
-- Instances
--