Synnefo » snf-ganeti

{-# LANGUAGE TemplateHaskell #-}

{-| Implementation of the job queue.

-}

{-

Copyright (C) 2010, 2012 Google Inc.

This program is free software; you can redistribute it and/or modify

it under the terms of the GNU General Public License as published by

the Free Software Foundation; either version 2 of the License, or

(at your option) any later version.

This program is distributed in the hope that it will be useful, but

WITHOUT ANY WARRANTY; without even the implied warranty of

MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

General Public License for more details.

You should have received a copy of the GNU General Public License

along with this program; if not, write to the Free Software

Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA

02110-1301, USA.

-}

module Ganeti.JQueue

    ( QueuedOpCode(..)

    , QueuedJob(..)

    , InputOpCode(..)

    , queuedOpCodeFromMetaOpCode

    , queuedJobFromOpCodes

    , cancelQueuedJob

    , Timestamp

    , fromClockTime

    , noTimestamp

    , currentTimestamp

    , advanceTimestamp

    , setReceivedTimestamp

    , opStatusFinalized

    , extractOpSummary

    , calcJobStatus

    , jobStarted

    , jobFinalized

    , jobArchivable

    , calcJobPriority

    , jobFileName

    , liveJobFile

    , archivedJobFile

    , determineJobDirectories

    , getJobIDs

    , sortJobIDs

    , loadJobFromDisk

    , noSuchJob

    , readSerialFromDisk

    , allocateJobIds

    , allocateJobId

    , writeJobToDisk

    , replicateManyJobs

    , isQueueOpen

    , startJobs

    , cancelJob

    , queueDirPermissions

    , archiveJobs

    ) where

import Control.Applicative (liftA2, (<|>))

import Control.Arrow (first, second)

import Control.Concurrent (forkIO)

import Control.Concurrent.MVar

import Control.Exception

import Control.Monad

import Control.Monad.IO.Class

import Data.Functor ((<$))

import Data.List

import Data.Maybe

import Data.Ord (comparing)

-- workaround what seems to be a bug in ghc 7.4's TH shadowing code

import Prelude hiding (id, log)

import System.Directory

import System.FilePath

import System.IO.Error (isDoesNotExistError)

import System.Posix.Files

import System.Time

import qualified Text.JSON

import Text.JSON.Types

import Ganeti.BasicTypes

import qualified Ganeti.Config as Config

import qualified Ganeti.Constants as C

import Ganeti.Errors (ErrorResult)

import Ganeti.JSON

import Ganeti.Logging

import Ganeti.Luxi

import Ganeti.Objects (ConfigData, Node)

import Ganeti.OpCodes

import Ganeti.Path

import Ganeti.Rpc (executeRpcCall, ERpcError, logRpcErrors,

                   RpcCallJobqueueUpdate(..), RpcCallJobqueueRename(..))

import Ganeti.THH

import Ganeti.Types

import Ganeti.Utils

import Ganeti.VCluster (makeVirtualPath)

-- * Data types

-- | The ganeti queue timestamp type. It represents the time as the pair

-- of seconds since the epoch and microseconds since the beginning of the

-- second.

type Timestamp = (Int, Int)

-- | Missing timestamp type.

noTimestamp :: Timestamp

noTimestamp = (-1, -1)

-- | Obtain a Timestamp from a given clock time

fromClockTime :: ClockTime -> Timestamp

fromClockTime (TOD ctime pico) =

  (fromIntegral ctime, fromIntegral $ pico `div` 1000000)

-- | Get the current time in the job-queue timestamp format.

currentTimestamp :: IO Timestamp

currentTimestamp = fromClockTime `liftM` getClockTime

-- | From a given timestamp, obtain the timestamp of the

-- time that is the given number of seconds later.

advanceTimestamp :: Int -> Timestamp -> Timestamp

advanceTimestamp = first . (+)

-- | An input opcode.

data InputOpCode = ValidOpCode MetaOpCode -- ^ OpCode was parsed successfully

                 | InvalidOpCode JSValue  -- ^ Invalid opcode

                   deriving (Show, Eq)

-- | JSON instance for 'InputOpCode', trying to parse it and if

-- failing, keeping the original JSValue.

instance Text.JSON.JSON InputOpCode where

  showJSON (ValidOpCode mo) = Text.JSON.showJSON mo

  showJSON (InvalidOpCode inv) = inv

  readJSON v = case Text.JSON.readJSON v of

                 Text.JSON.Error _ -> return $ InvalidOpCode v

                 Text.JSON.Ok mo -> return $ ValidOpCode mo

-- | Invalid opcode summary.

invalidOp :: String

invalidOp = "INVALID_OP"

-- | Tries to extract the opcode summary from an 'InputOpCode'. This

-- duplicates some functionality from the 'opSummary' function in

-- "Ganeti.OpCodes".

extractOpSummary :: InputOpCode -> String

extractOpSummary (ValidOpCode metaop) = opSummary $ metaOpCode metaop

extractOpSummary (InvalidOpCode (JSObject o)) =

  case fromObjWithDefault (fromJSObject o) "OP_ID" ("OP_" ++ invalidOp) of

    Just s -> drop 3 s -- drop the OP_ prefix

    Nothing -> invalidOp

extractOpSummary _ = invalidOp

$(buildObject "QueuedOpCode" "qo"

  [ simpleField "input"           [t| InputOpCode |]

  , simpleField "status"          [t| OpStatus    |]

  , simpleField "result"          [t| JSValue     |]

  , defaultField [| [] |] $

    simpleField "log"             [t| [(Int, Timestamp, ELogType, JSValue)] |]

  , simpleField "priority"        [t| Int         |]

  , optionalNullSerField $

    simpleField "start_timestamp" [t| Timestamp   |]

  , optionalNullSerField $

    simpleField "exec_timestamp"  [t| Timestamp   |]

  , optionalNullSerField $

    simpleField "end_timestamp"   [t| Timestamp   |]

  ])

$(buildObject "QueuedJob" "qj"

  [ simpleField "id"                 [t| JobId          |]

  , simpleField "ops"                [t| [QueuedOpCode] |]

  , optionalNullSerField $

    simpleField "received_timestamp" [t| Timestamp      |]

  , optionalNullSerField $

    simpleField "start_timestamp"    [t| Timestamp      |]

  , optionalNullSerField $

    simpleField "end_timestamp"      [t| Timestamp      |]

  ])

-- | Convenience function to obtain a QueuedOpCode from a MetaOpCode

queuedOpCodeFromMetaOpCode :: MetaOpCode -> QueuedOpCode

queuedOpCodeFromMetaOpCode op =

  QueuedOpCode { qoInput = ValidOpCode op

               , qoStatus = OP_STATUS_QUEUED

               , qoPriority = opSubmitPriorityToRaw . opPriority . metaParams

                              $ op

               , qoLog = []

               , qoResult = JSNull

               , qoStartTimestamp = Nothing

               , qoEndTimestamp = Nothing

               , qoExecTimestamp = Nothing

               }

-- | From a job-id and a list of op-codes create a job. This is

-- the pure part of job creation, as allocating a new job id

-- lives in IO.

queuedJobFromOpCodes :: (Monad m) => JobId -> [MetaOpCode] -> m QueuedJob

queuedJobFromOpCodes jobid ops = do

  ops' <- mapM (`resolveDependencies` jobid) ops

  return QueuedJob { qjId = jobid

                   , qjOps = map queuedOpCodeFromMetaOpCode ops'

                   , qjReceivedTimestamp = Nothing 

                   , qjStartTimestamp = Nothing

                   , qjEndTimestamp = Nothing

                   }

-- | Attach a received timestamp to a Queued Job.

setReceivedTimestamp :: Timestamp -> QueuedJob -> QueuedJob

setReceivedTimestamp ts job = job { qjReceivedTimestamp = Just ts }

-- | Set the state of a QueuedOpCode to canceled.

cancelOpCode :: Timestamp -> QueuedOpCode -> QueuedOpCode

cancelOpCode now op =

  op { qoStatus = OP_STATUS_CANCELED, qoEndTimestamp = Just now }

-- | Transform a QueuedJob that has not been started into its canceled form.

cancelQueuedJob :: Timestamp -> QueuedJob -> QueuedJob

cancelQueuedJob now job =

  let ops' = map (cancelOpCode now) $ qjOps job

  in job { qjOps = ops', qjEndTimestamp = Just now}

-- | Job file prefix.

jobFilePrefix :: String

jobFilePrefix = "job-"

-- | Computes the filename for a given job ID.

jobFileName :: JobId -> FilePath

jobFileName jid = jobFilePrefix ++ show (fromJobId jid)

-- | Parses a job ID from a file name.

parseJobFileId :: (Monad m) => FilePath -> m JobId

parseJobFileId path =

  case stripPrefix jobFilePrefix path of

    Nothing -> fail $ "Job file '" ++ path ++

                      "' doesn't have the correct prefix"

    Just suffix -> makeJobIdS suffix

-- | Computes the full path to a live job.

liveJobFile :: FilePath -> JobId -> FilePath

liveJobFile rootdir jid = rootdir </> jobFileName jid

-- | Computes the full path to an archives job. BROKEN.

archivedJobFile :: FilePath -> JobId -> FilePath

archivedJobFile rootdir jid =

  let subdir = show (fromJobId jid `div` C.jstoreJobsPerArchiveDirectory)

  in rootdir </> jobQueueArchiveSubDir </> subdir </> jobFileName jid

-- | Map from opcode status to job status.

opStatusToJob :: OpStatus -> JobStatus

opStatusToJob OP_STATUS_QUEUED    = JOB_STATUS_QUEUED

opStatusToJob OP_STATUS_WAITING   = JOB_STATUS_WAITING

opStatusToJob OP_STATUS_SUCCESS   = JOB_STATUS_SUCCESS

opStatusToJob OP_STATUS_RUNNING   = JOB_STATUS_RUNNING

opStatusToJob OP_STATUS_CANCELING = JOB_STATUS_CANCELING

opStatusToJob OP_STATUS_CANCELED  = JOB_STATUS_CANCELED

opStatusToJob OP_STATUS_ERROR     = JOB_STATUS_ERROR

-- | Computes a queued job's status.

calcJobStatus :: QueuedJob -> JobStatus

calcJobStatus QueuedJob { qjOps = ops } =

  extractOpSt (map qoStatus ops) JOB_STATUS_QUEUED True

    where

      terminalStatus OP_STATUS_ERROR     = True

      terminalStatus OP_STATUS_CANCELING = True

      terminalStatus OP_STATUS_CANCELED  = True

      terminalStatus _                   = False

      softStatus     OP_STATUS_SUCCESS   = True

      softStatus     OP_STATUS_QUEUED    = True

      softStatus     _                   = False

      extractOpSt [] _ True = JOB_STATUS_SUCCESS

      extractOpSt [] d False = d

      extractOpSt (x:xs) d old_all

           | terminalStatus x = opStatusToJob x -- abort recursion

           | softStatus x     = extractOpSt xs d new_all -- continue unchanged

           | otherwise        = extractOpSt xs (opStatusToJob x) new_all

           where new_all = x == OP_STATUS_SUCCESS && old_all

-- | Determine if a job has started

jobStarted :: QueuedJob -> Bool

jobStarted = (> JOB_STATUS_QUEUED) . calcJobStatus

-- | Determine if a job is finalised.

jobFinalized :: QueuedJob -> Bool

jobFinalized = (> JOB_STATUS_RUNNING) . calcJobStatus

-- | Determine if a job is finalized and its timestamp is before

-- a given time.

jobArchivable :: Timestamp -> QueuedJob -> Bool

jobArchivable ts = liftA2 (&&) jobFinalized 

  $ maybe False (< ts)

    .  liftA2 (<|>) qjEndTimestamp qjStartTimestamp

-- | Determine whether an opcode status is finalized.

opStatusFinalized :: OpStatus -> Bool

opStatusFinalized = (> OP_STATUS_RUNNING)

-- | Compute a job's priority.

calcJobPriority :: QueuedJob -> Int

calcJobPriority QueuedJob { qjOps = ops } =

  helper . map qoPriority $ filter (not . opStatusFinalized . qoStatus) ops

    where helper [] = C.opPrioDefault

          helper ps = minimum ps

-- | Log but ignore an 'IOError'.

ignoreIOError :: a -> Bool -> String -> IOError -> IO a

ignoreIOError a ignore_noent msg e = do

  unless (isDoesNotExistError e && ignore_noent) .

    logWarning $ msg ++ ": " ++ show e

  return a

-- | Compute the list of existing archive directories. Note that I/O

-- exceptions are swallowed and ignored.

allArchiveDirs :: FilePath -> IO [FilePath]

allArchiveDirs rootdir = do

  let adir = rootdir </> jobQueueArchiveSubDir

  contents <- getDirectoryContents adir `Control.Exception.catch`

               ignoreIOError [] False

                 ("Failed to list queue directory " ++ adir)

  let fpaths = map (adir </>) $ filter (not . ("." `isPrefixOf`)) contents

  filterM (\path ->

             liftM isDirectory (getFileStatus (adir </> path))

               `Control.Exception.catch`

               ignoreIOError False True

                 ("Failed to stat archive path " ++ path)) fpaths

-- | Build list of directories containing job files. Note: compared to

-- the Python version, this doesn't ignore a potential lost+found

-- file.

determineJobDirectories :: FilePath -> Bool -> IO [FilePath]

determineJobDirectories rootdir archived = do

  other <- if archived

             then allArchiveDirs rootdir

             else return []

  return $ rootdir:other

-- | Computes the list of all jobs in the given directories.

getJobIDs :: [FilePath] -> IO (GenericResult IOError [JobId])

getJobIDs = runResultT . liftM concat . mapM getDirJobIDs

-- | Sorts the a list of job IDs.

sortJobIDs :: [JobId] -> [JobId]

sortJobIDs = sortBy (comparing fromJobId)

-- | Computes the list of jobs in a given directory.

getDirJobIDs :: FilePath -> ResultT IOError IO [JobId]

getDirJobIDs path =

  withErrorLogAt WARNING ("Failed to list job directory " ++ path) .

    liftM (mapMaybe parseJobFileId) $ liftIO (getDirectoryContents path)

-- | Reads the job data from disk.

readJobDataFromDisk :: FilePath -> Bool -> JobId -> IO (Maybe (String, Bool))

readJobDataFromDisk rootdir archived jid = do

  let live_path = liveJobFile rootdir jid

      archived_path = archivedJobFile rootdir jid

      all_paths = if archived

                    then [(live_path, False), (archived_path, True)]

                    else [(live_path, False)]

  foldM (\state (path, isarchived) ->

           liftM (\r -> Just (r, isarchived)) (readFile path)

             `Control.Exception.catch`

             ignoreIOError state True

               ("Failed to read job file " ++ path)) Nothing all_paths

-- | Failed to load job error.

noSuchJob :: Result (QueuedJob, Bool)

noSuchJob = Bad "Can't load job file"

-- | Loads a job from disk.

loadJobFromDisk :: FilePath -> Bool -> JobId -> IO (Result (QueuedJob, Bool))

loadJobFromDisk rootdir archived jid = do

  raw <- readJobDataFromDisk rootdir archived jid

  -- note: we need some stricness below, otherwise the wrapping in a

  -- Result will create too much lazyness, and not close the file

  -- descriptors for the individual jobs

  return $! case raw of

             Nothing -> noSuchJob

             Just (str, arch) ->

               liftM (\qj -> (qj, arch)) .

               fromJResult "Parsing job file" $ Text.JSON.decode str

-- | Write a job to disk.

writeJobToDisk :: FilePath -> QueuedJob -> IO (Result ())

writeJobToDisk rootdir job = do

  let filename = liveJobFile rootdir . qjId $ job

      content = Text.JSON.encode . Text.JSON.showJSON $ job

  tryAndLogIOError (atomicWriteFile filename content)

                   ("Failed to write " ++ filename) Ok

-- | Replicate a job to all master candidates.

replicateJob :: FilePath -> [Node] -> QueuedJob -> IO [(Node, ERpcError ())]

replicateJob rootdir mastercandidates job = do

  let filename = liveJobFile rootdir . qjId $ job

      content = Text.JSON.encode . Text.JSON.showJSON $ job

  filename' <- makeVirtualPath filename

  callresult <- executeRpcCall mastercandidates

                  $ RpcCallJobqueueUpdate filename' content

  let result = map (second (() <$)) callresult

  logRpcErrors result

  return result

-- | Replicate many jobs to all master candidates.

replicateManyJobs :: FilePath -> [Node] -> [QueuedJob] -> IO ()

replicateManyJobs rootdir mastercandidates =

  mapM_ (replicateJob rootdir mastercandidates)

-- | Read the job serial number from disk.

readSerialFromDisk :: IO (Result JobId)

readSerialFromDisk = do

  filename <- jobQueueSerialFile

  tryAndLogIOError (readFile filename) "Failed to read serial file"

                   (makeJobIdS . rStripSpace)

-- | Allocate new job ids.

-- To avoid races while accessing the serial file, the threads synchronize

-- over a lock, as usual provided by an MVar.

allocateJobIds :: [Node] -> MVar () -> Int -> IO (Result [JobId])

allocateJobIds mastercandidates lock n =

  if n <= 0

    then return . Bad $ "Can only allocate positive number of job ids"

    else do

      takeMVar lock

      rjobid <- readSerialFromDisk

      case rjobid of

        Bad s -> do

          putMVar lock ()

          return . Bad $ s

        Ok jid -> do

          let current = fromJobId jid

              serial_content = show (current + n) ++  "\n"

          serial <- jobQueueSerialFile

          write_result <- try $ atomicWriteFile serial serial_content

                          :: IO (Either IOError ())

          case write_result of

            Left e -> do

              putMVar lock ()

              let msg = "Failed to write serial file: " ++ show e

              logError msg

              return . Bad $ msg 

            Right () -> do

              serial' <- makeVirtualPath serial

              _ <- executeRpcCall mastercandidates

                     $ RpcCallJobqueueUpdate serial' serial_content

              putMVar lock ()

              return $ mapM makeJobId [(current+1)..(current+n)]

-- | Allocate one new job id.

allocateJobId :: [Node] -> MVar () -> IO (Result JobId)

allocateJobId mastercandidates lock = do

  jids <- allocateJobIds mastercandidates lock 1

  return (jids >>= monadicThe "Failed to allocate precisely one Job ID")

-- | Decide if job queue is open

isQueueOpen :: IO Bool

isQueueOpen = liftM not (jobQueueDrainFile >>= doesFileExist)

-- | Start enqueued jobs, currently by handing them over to masterd.

startJobs :: [QueuedJob] -> IO ()

startJobs jobs = do

  socketpath <- defaultMasterSocket

  client <- getLuxiClient socketpath

  pickupResults <- mapM (flip callMethod client . PickupJob . qjId) jobs

  let failures = map show $ justBad pickupResults

  unless (null failures)

   . logWarning . (++) "Failed to notify masterd: " . commaJoin $ failures

-- | Try to cancel a job that has already been handed over to execution,

-- currently by asking masterd to cancel it.

cancelJob :: JobId -> IO (ErrorResult JSValue)

cancelJob jid = do

  socketpath <- defaultMasterSocket

  client <- getLuxiClient socketpath

  callMethod (CancelJob jid) client

-- | Permissions for the archive directories.

queueDirPermissions :: FilePermissions

queueDirPermissions = FilePermissions { fpOwner = Just C.masterdUser

                                      , fpGroup = Just C.daemonsGroup

                                      , fpPermissions = 0o0750

                                      }

-- | Try, at most until the given endtime, to archive some of the given

-- jobs, if they are older than the specified cut-off time; also replicate

-- archival of the additional jobs. Return the pair of the number of jobs

-- archived, and the number of jobs remaining int he queue, asuming the

-- given numbers about the not considered jobs.

archiveSomeJobsUntil :: ([JobId] -> IO ()) -- ^ replication function

                        -> FilePath -- ^ queue root directory

                        -> ClockTime -- ^ Endtime

                        -> Timestamp -- ^ cut-off time for archiving jobs

                        -> Int -- ^ number of jobs alread archived

                        -> [JobId] -- ^ Additional jobs to replicate

                        -> [JobId] -- ^ List of job-ids still to consider

                        -> IO (Int, Int)

archiveSomeJobsUntil replicateFn _ _ _ arch torepl [] = do

  unless (null torepl) . (>> return ())

   . forkIO $ replicateFn torepl

  return (arch, 0)

archiveSomeJobsUntil replicateFn qDir endt cutt arch torepl (jid:jids) = do

  let archiveMore = archiveSomeJobsUntil replicateFn qDir endt cutt

      continue = archiveMore arch torepl jids

      jidname = show $ fromJobId jid

  time <- getClockTime

  if time >= endt

    then do

      _ <- forkIO $ replicateFn torepl

      return (arch, length (jid:jids))

    else do

      logDebug $ "Inspecting job " ++ jidname ++ " for archival"

      loadResult <- loadJobFromDisk qDir False jid

      case loadResult of

        Bad _ -> continue

        Ok (job, _) -> 

          if jobArchivable cutt job

            then do

              let live = liveJobFile qDir jid

                  archive = archivedJobFile qDir jid

              renameResult <- safeRenameFile queueDirPermissions

                                live archive

              case renameResult of                   

                Bad s -> do

                  logWarning $ "Renaming " ++ live ++ " to " ++ archive

                                 ++ " failed unexpectedly: " ++ s

                  continue

                Ok () -> do

                  let torepl' = jid:torepl

                  if length torepl' >= 10

                    then do

                      _ <- forkIO $ replicateFn torepl'

                      archiveMore (arch + 1) [] jids

                    else archiveMore (arch + 1) torepl' jids

            else continue

-- | Archive jobs older than the given time, but do not exceed the timeout for

-- carrying out this task.

archiveJobs :: ConfigData -- ^ cluster configuration

               -> Int  -- ^ time the job has to be in the past in order

                       -- to be archived

               -> Int -- ^ timeout

               -> [JobId] -- ^ jobs to consider

               -> IO (Int, Int)

archiveJobs cfg age timeout jids = do

  now <- getClockTime

  qDir <- queueDir

  let endtime = addToClockTime (noTimeDiff { tdSec = timeout }) now

      cuttime = if age < 0 then noTimestamp

                           else advanceTimestamp (- age) (fromClockTime now)

      mcs = Config.getMasterCandidates cfg

      replicateFn jobs = do

        let olds = map (liveJobFile qDir) jobs

            news = map (archivedJobFile qDir) jobs

        _ <- executeRpcCall mcs . RpcCallJobqueueRename $ zip olds news

        return ()

  archiveSomeJobsUntil replicateFn qDir endtime cuttime 0 [] jids