Synnefo » snf-ganeti » ganeti-local

{-| Implementation of cluster-wide logic.

This module holds all pure cluster-logic; I\/O related functionality

goes into the /Main/ module for the individual binaries.

-}

{-

Copyright (C) 2009, 2010, 2011, 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.HTools.Cluster

  (

    -- * Types

    AllocSolution(..)

  , EvacSolution(..)

  , Table(..)

  , CStats(..)

  , AllocResult

  , AllocMethod

  -- * Generic functions

  , totalResources

  , computeAllocationDelta

  -- * First phase functions

  , computeBadItems

  -- * Second phase functions

  , printSolutionLine

  , formatCmds

  , involvedNodes

  , splitJobs

  -- * Display functions

  , printNodes

  , printInsts

  -- * Balacing functions

  , checkMove

  , doNextBalance

  , tryBalance

  , compCV

  , compCVNodes

  , compDetailedCV

  , printStats

  , iMoveToJob

  -- * IAllocator functions

  , genAllocNodes

  , tryAlloc

  , tryMGAlloc

  , tryNodeEvac

  , tryChangeGroup

  , collapseFailures

  -- * Allocation functions

  , iterateAlloc

  , tieredAlloc

  -- * Node group functions

  , instanceGroup

  , findSplitInstances

  , splitCluster

  ) where

import qualified Data.IntSet as IntSet

import Data.List

import Data.Maybe (fromJust, isNothing)

import Data.Ord (comparing)

import Text.Printf (printf)

import qualified Ganeti.HTools.Container as Container

import qualified Ganeti.HTools.Instance as Instance

import qualified Ganeti.HTools.Node as Node

import qualified Ganeti.HTools.Group as Group

import Ganeti.HTools.Types

import Ganeti.HTools.Utils

import Ganeti.HTools.Compat

import qualified Ganeti.OpCodes as OpCodes

-- * Types

-- | Allocation\/relocation solution.

data AllocSolution = AllocSolution

  { asFailures :: [FailMode]              -- ^ Failure counts

  , asAllocs   :: Int                     -- ^ Good allocation count

  , asSolution :: Maybe Node.AllocElement -- ^ The actual allocation result

  , asLog      :: [String]                -- ^ Informational messages

  }

-- | Node evacuation/group change iallocator result type. This result

-- type consists of actual opcodes (a restricted subset) that are

-- transmitted back to Ganeti.

data EvacSolution = EvacSolution

  { esMoved   :: [(Idx, Gdx, [Ndx])]  -- ^ Instances moved successfully

  , esFailed  :: [(Idx, String)]      -- ^ Instances which were not

                                      -- relocated

  , esOpCodes :: [[OpCodes.OpCode]]   -- ^ List of jobs

  } deriving (Show)

-- | Allocation results, as used in 'iterateAlloc' and 'tieredAlloc'.

type AllocResult = (FailStats, Node.List, Instance.List,

                    [Instance.Instance], [CStats])

-- | A type denoting the valid allocation mode/pairs.

--

-- For a one-node allocation, this will be a @Left ['Ndx']@, whereas

-- for a two-node allocation, this will be a @Right [('Ndx',

-- ['Ndx'])]@. In the latter case, the list is basically an

-- association list, grouped by primary node and holding the potential

-- secondary nodes in the sub-list.

type AllocNodes = Either [Ndx] [(Ndx, [Ndx])]

-- | The empty solution we start with when computing allocations.

emptyAllocSolution :: AllocSolution

emptyAllocSolution = AllocSolution { asFailures = [], asAllocs = 0

                                   , asSolution = Nothing, asLog = [] }

-- | The empty evac solution.

emptyEvacSolution :: EvacSolution

emptyEvacSolution = EvacSolution { esMoved = []

                                 , esFailed = []

                                 , esOpCodes = []

                                 }

-- | The complete state for the balancing solution.

data Table = Table Node.List Instance.List Score [Placement]

             deriving (Show, Read)

-- | Cluster statistics data type.

data CStats = CStats

  { csFmem :: Integer -- ^ Cluster free mem

  , csFdsk :: Integer -- ^ Cluster free disk

  , csAmem :: Integer -- ^ Cluster allocatable mem

  , csAdsk :: Integer -- ^ Cluster allocatable disk

  , csAcpu :: Integer -- ^ Cluster allocatable cpus

  , csMmem :: Integer -- ^ Max node allocatable mem

  , csMdsk :: Integer -- ^ Max node allocatable disk

  , csMcpu :: Integer -- ^ Max node allocatable cpu

  , csImem :: Integer -- ^ Instance used mem

  , csIdsk :: Integer -- ^ Instance used disk

  , csIcpu :: Integer -- ^ Instance used cpu

  , csTmem :: Double  -- ^ Cluster total mem

  , csTdsk :: Double  -- ^ Cluster total disk

  , csTcpu :: Double  -- ^ Cluster total cpus

  , csVcpu :: Integer -- ^ Cluster total virtual cpus

  , csNcpu :: Double  -- ^ Equivalent to 'csIcpu' but in terms of

                      -- physical CPUs, i.e. normalised used phys CPUs

  , csXmem :: Integer -- ^ Unnacounted for mem

  , csNmem :: Integer -- ^ Node own memory

  , csScore :: Score  -- ^ The cluster score

  , csNinst :: Int    -- ^ The total number of instances

  } deriving (Show, Read)

-- | A simple type for allocation functions.

type AllocMethod =  Node.List           -- ^ Node list

                 -> Instance.List       -- ^ Instance list

                 -> Maybe Int           -- ^ Optional allocation limit

                 -> Instance.Instance   -- ^ Instance spec for allocation

                 -> AllocNodes          -- ^ Which nodes we should allocate on

                 -> [Instance.Instance] -- ^ Allocated instances

                 -> [CStats]            -- ^ Running cluster stats

                 -> Result AllocResult  -- ^ Allocation result

-- * Utility functions

-- | Verifies the N+1 status and return the affected nodes.

verifyN1 :: [Node.Node] -> [Node.Node]

verifyN1 = filter Node.failN1

{-| Computes the pair of bad nodes and instances.

The bad node list is computed via a simple 'verifyN1' check, and the

bad instance list is the list of primary and secondary instances of

those nodes.

-}

computeBadItems :: Node.List -> Instance.List ->

                   ([Node.Node], [Instance.Instance])

computeBadItems nl il =

  let bad_nodes = verifyN1 $ getOnline nl

      bad_instances = map (`Container.find` il) .

                      sort . nub $

                      concatMap (\ n -> Node.sList n ++ Node.pList n) bad_nodes

  in

    (bad_nodes, bad_instances)

-- | Extracts the node pairs for an instance. This can fail if the

-- instance is single-homed. FIXME: this needs to be improved,

-- together with the general enhancement for handling non-DRBD moves.

instanceNodes :: Node.List -> Instance.Instance ->

                 (Ndx, Ndx, Node.Node, Node.Node)

instanceNodes nl inst =

  let old_pdx = Instance.pNode inst

      old_sdx = Instance.sNode inst

      old_p = Container.find old_pdx nl

      old_s = Container.find old_sdx nl

  in (old_pdx, old_sdx, old_p, old_s)

-- | Zero-initializer for the CStats type.

emptyCStats :: CStats

emptyCStats = CStats 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

-- | Update stats with data from a new node.

updateCStats :: CStats -> Node.Node -> CStats

updateCStats cs node =

  let CStats { csFmem = x_fmem, csFdsk = x_fdsk,

               csAmem = x_amem, csAcpu = x_acpu, csAdsk = x_adsk,

               csMmem = x_mmem, csMdsk = x_mdsk, csMcpu = x_mcpu,

               csImem = x_imem, csIdsk = x_idsk, csIcpu = x_icpu,

               csTmem = x_tmem, csTdsk = x_tdsk, csTcpu = x_tcpu,

               csVcpu = x_vcpu, csNcpu = x_ncpu,

               csXmem = x_xmem, csNmem = x_nmem, csNinst = x_ninst

             }

        = cs

      inc_amem = Node.fMem node - Node.rMem node

      inc_amem' = if inc_amem > 0 then inc_amem else 0

      inc_adsk = Node.availDisk node

      inc_imem = truncate (Node.tMem node) - Node.nMem node

                 - Node.xMem node - Node.fMem node

      inc_icpu = Node.uCpu node

      inc_idsk = truncate (Node.tDsk node) - Node.fDsk node

      inc_vcpu = Node.hiCpu node

      inc_acpu = Node.availCpu node

      inc_ncpu = fromIntegral (Node.uCpu node) /

                 iPolicyVcpuRatio (Node.iPolicy node)

  in cs { csFmem = x_fmem + fromIntegral (Node.fMem node)

        , csFdsk = x_fdsk + fromIntegral (Node.fDsk node)

        , csAmem = x_amem + fromIntegral inc_amem'

        , csAdsk = x_adsk + fromIntegral inc_adsk

        , csAcpu = x_acpu + fromIntegral inc_acpu

        , csMmem = max x_mmem (fromIntegral inc_amem')

        , csMdsk = max x_mdsk (fromIntegral inc_adsk)

        , csMcpu = max x_mcpu (fromIntegral inc_acpu)

        , csImem = x_imem + fromIntegral inc_imem

        , csIdsk = x_idsk + fromIntegral inc_idsk

        , csIcpu = x_icpu + fromIntegral inc_icpu

        , csTmem = x_tmem + Node.tMem node

        , csTdsk = x_tdsk + Node.tDsk node

        , csTcpu = x_tcpu + Node.tCpu node

        , csVcpu = x_vcpu + fromIntegral inc_vcpu

        , csNcpu = x_ncpu + inc_ncpu

        , csXmem = x_xmem + fromIntegral (Node.xMem node)

        , csNmem = x_nmem + fromIntegral (Node.nMem node)

        , csNinst = x_ninst + length (Node.pList node)

        }

-- | Compute the total free disk and memory in the cluster.

totalResources :: Node.List -> CStats

totalResources nl =

  let cs = foldl' updateCStats emptyCStats . Container.elems $ nl

  in cs { csScore = compCV nl }

-- | Compute the delta between two cluster state.

--

-- This is used when doing allocations, to understand better the

-- available cluster resources. The return value is a triple of the

-- current used values, the delta that was still allocated, and what

-- was left unallocated.

computeAllocationDelta :: CStats -> CStats -> AllocStats

computeAllocationDelta cini cfin =

  let CStats {csImem = i_imem, csIdsk = i_idsk, csIcpu = i_icpu,

              csNcpu = i_ncpu } = cini

      CStats {csImem = f_imem, csIdsk = f_idsk, csIcpu = f_icpu,

              csTmem = t_mem, csTdsk = t_dsk, csVcpu = f_vcpu,

              csNcpu = f_ncpu, csTcpu = f_tcpu } = cfin

      rini = AllocInfo { allocInfoVCpus = fromIntegral i_icpu

                       , allocInfoNCpus = i_ncpu

                       , allocInfoMem   = fromIntegral i_imem

                       , allocInfoDisk  = fromIntegral i_idsk

                       }

      rfin = AllocInfo { allocInfoVCpus = fromIntegral (f_icpu - i_icpu)

                       , allocInfoNCpus = f_ncpu - i_ncpu

                       , allocInfoMem   = fromIntegral (f_imem - i_imem)

                       , allocInfoDisk  = fromIntegral (f_idsk - i_idsk)

                       }

      runa = AllocInfo { allocInfoVCpus = fromIntegral (f_vcpu - f_icpu)

                       , allocInfoNCpus = f_tcpu - f_ncpu

                       , allocInfoMem   = truncate t_mem - fromIntegral f_imem

                       , allocInfoDisk  = truncate t_dsk - fromIntegral f_idsk

                       }

  in (rini, rfin, runa)

-- | The names and weights of the individual elements in the CV list.

detailedCVInfo :: [(Double, String)]

detailedCVInfo = [ (1,  "free_mem_cv")

                 , (1,  "free_disk_cv")

                 , (1,  "n1_cnt")

                 , (1,  "reserved_mem_cv")

                 , (4,  "offline_all_cnt")

                 , (16, "offline_pri_cnt")

                 , (1,  "vcpu_ratio_cv")

                 , (1,  "cpu_load_cv")

                 , (1,  "mem_load_cv")

                 , (1,  "disk_load_cv")

                 , (1,  "net_load_cv")

                 , (2,  "pri_tags_score")

                 , (1,  "spindles_cv")

                 ]

-- | Holds the weights used by 'compCVNodes' for each metric.

detailedCVWeights :: [Double]

detailedCVWeights = map fst detailedCVInfo

-- | Compute the mem and disk covariance.

compDetailedCV :: [Node.Node] -> [Double]

compDetailedCV all_nodes =

  let (offline, nodes) = partition Node.offline all_nodes

      mem_l = map Node.pMem nodes

      dsk_l = map Node.pDsk nodes

      -- metric: memory covariance

      mem_cv = stdDev mem_l

      -- metric: disk covariance

      dsk_cv = stdDev dsk_l

      -- metric: count of instances living on N1 failing nodes

      n1_score = fromIntegral . sum . map (\n -> length (Node.sList n) +

                                                 length (Node.pList n)) .

                 filter Node.failN1 $ nodes :: Double

      res_l = map Node.pRem nodes

      -- metric: reserved memory covariance

      res_cv = stdDev res_l

      -- offline instances metrics

      offline_ipri = sum . map (length . Node.pList) $ offline

      offline_isec = sum . map (length . Node.sList) $ offline

      -- metric: count of instances on offline nodes

      off_score = fromIntegral (offline_ipri + offline_isec)::Double

      -- metric: count of primary instances on offline nodes (this

      -- helps with evacuation/failover of primary instances on

      -- 2-node clusters with one node offline)

      off_pri_score = fromIntegral offline_ipri::Double

      cpu_l = map Node.pCpu nodes

      -- metric: covariance of vcpu/pcpu ratio

      cpu_cv = stdDev cpu_l

      -- metrics: covariance of cpu, memory, disk and network load

      (c_load, m_load, d_load, n_load) =

        unzip4 $ map (\n ->

                      let DynUtil c1 m1 d1 n1 = Node.utilLoad n

                          DynUtil c2 m2 d2 n2 = Node.utilPool n

                      in (c1/c2, m1/m2, d1/d2, n1/n2)) nodes

      -- metric: conflicting instance count

      pri_tags_inst = sum $ map Node.conflictingPrimaries nodes

      pri_tags_score = fromIntegral pri_tags_inst::Double

      -- metric: spindles %

      spindles_cv = map (\n -> Node.instSpindles n / Node.hiSpindles n) nodes

  in [ mem_cv, dsk_cv, n1_score, res_cv, off_score, off_pri_score, cpu_cv

     , stdDev c_load, stdDev m_load , stdDev d_load, stdDev n_load

     , pri_tags_score, stdDev spindles_cv ]

-- | Compute the /total/ variance.

compCVNodes :: [Node.Node] -> Double

compCVNodes = sum . zipWith (*) detailedCVWeights . compDetailedCV

-- | Wrapper over 'compCVNodes' for callers that have a 'Node.List'.

compCV :: Node.List -> Double

compCV = compCVNodes . Container.elems

-- | Compute online nodes from a 'Node.List'.

getOnline :: Node.List -> [Node.Node]

getOnline = filter (not . Node.offline) . Container.elems

-- * Balancing functions

-- | Compute best table. Note that the ordering of the arguments is important.

compareTables :: Table -> Table -> Table

compareTables a@(Table _ _ a_cv _) b@(Table _ _ b_cv _ ) =

  if a_cv > b_cv then b else a

-- | Applies an instance move to a given node list and instance.

applyMove :: Node.List -> Instance.Instance

          -> IMove -> OpResult (Node.List, Instance.Instance, Ndx, Ndx)

-- Failover (f)

applyMove nl inst Failover =

  let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst

      int_p = Node.removePri old_p inst

      int_s = Node.removeSec old_s inst

      new_nl = do -- Maybe monad

        new_p <- Node.addPriEx (Node.offline old_p) int_s inst

        new_s <- Node.addSec int_p inst old_sdx

        let new_inst = Instance.setBoth inst old_sdx old_pdx

        return (Container.addTwo old_pdx new_s old_sdx new_p nl,

                new_inst, old_sdx, old_pdx)

  in new_nl

-- Replace the primary (f:, r:np, f)

applyMove nl inst (ReplacePrimary new_pdx) =

  let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst

      tgt_n = Container.find new_pdx nl

      int_p = Node.removePri old_p inst

      int_s = Node.removeSec old_s inst

      force_p = Node.offline old_p

      new_nl = do -- Maybe monad

                  -- check that the current secondary can host the instance

                  -- during the migration

        tmp_s <- Node.addPriEx force_p int_s inst

        let tmp_s' = Node.removePri tmp_s inst

        new_p <- Node.addPriEx force_p tgt_n inst

        new_s <- Node.addSecEx force_p tmp_s' inst new_pdx

        let new_inst = Instance.setPri inst new_pdx

        return (Container.add new_pdx new_p $

                Container.addTwo old_pdx int_p old_sdx new_s nl,

                new_inst, new_pdx, old_sdx)

  in new_nl

-- Replace the secondary (r:ns)

applyMove nl inst (ReplaceSecondary new_sdx) =

  let old_pdx = Instance.pNode inst

      old_sdx = Instance.sNode inst

      old_s = Container.find old_sdx nl

      tgt_n = Container.find new_sdx nl

      int_s = Node.removeSec old_s inst

      force_s = Node.offline old_s

      new_inst = Instance.setSec inst new_sdx

      new_nl = Node.addSecEx force_s tgt_n inst old_pdx >>=

               \new_s -> return (Container.addTwo new_sdx

                                 new_s old_sdx int_s nl,

                                 new_inst, old_pdx, new_sdx)

  in new_nl

-- Replace the secondary and failover (r:np, f)

applyMove nl inst (ReplaceAndFailover new_pdx) =

  let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst

      tgt_n = Container.find new_pdx nl

      int_p = Node.removePri old_p inst

      int_s = Node.removeSec old_s inst

      force_s = Node.offline old_s

      new_nl = do -- Maybe monad

        new_p <- Node.addPri tgt_n inst

        new_s <- Node.addSecEx force_s int_p inst new_pdx

        let new_inst = Instance.setBoth inst new_pdx old_pdx

        return (Container.add new_pdx new_p $

                Container.addTwo old_pdx new_s old_sdx int_s nl,

                new_inst, new_pdx, old_pdx)

  in new_nl

-- Failver and replace the secondary (f, r:ns)

applyMove nl inst (FailoverAndReplace new_sdx) =

  let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst

      tgt_n = Container.find new_sdx nl

      int_p = Node.removePri old_p inst

      int_s = Node.removeSec old_s inst

      force_p = Node.offline old_p

      new_nl = do -- Maybe monad

        new_p <- Node.addPriEx force_p int_s inst

        new_s <- Node.addSecEx force_p tgt_n inst old_sdx

        let new_inst = Instance.setBoth inst old_sdx new_sdx

        return (Container.add new_sdx new_s $

                Container.addTwo old_sdx new_p old_pdx int_p nl,

                new_inst, old_sdx, new_sdx)

  in new_nl

-- | Tries to allocate an instance on one given node.

allocateOnSingle :: Node.List -> Instance.Instance -> Ndx

                 -> OpResult Node.AllocElement

allocateOnSingle nl inst new_pdx =

  let p = Container.find new_pdx nl

      new_inst = Instance.setBoth inst new_pdx Node.noSecondary

  in do

    Instance.instMatchesPolicy inst (Node.iPolicy p)

    new_p <- Node.addPri p inst

    let new_nl = Container.add new_pdx new_p nl

        new_score = compCV nl

    return (new_nl, new_inst, [new_p], new_score)

-- | Tries to allocate an instance on a given pair of nodes.

allocateOnPair :: Node.List -> Instance.Instance -> Ndx -> Ndx

               -> OpResult Node.AllocElement

allocateOnPair nl inst new_pdx new_sdx =

  let tgt_p = Container.find new_pdx nl

      tgt_s = Container.find new_sdx nl

  in do

    Instance.instMatchesPolicy inst (Node.iPolicy tgt_p)

    new_p <- Node.addPri tgt_p inst

    new_s <- Node.addSec tgt_s inst new_pdx

    let new_inst = Instance.setBoth inst new_pdx new_sdx

        new_nl = Container.addTwo new_pdx new_p new_sdx new_s nl

    return (new_nl, new_inst, [new_p, new_s], compCV new_nl)

-- | Tries to perform an instance move and returns the best table

-- between the original one and the new one.

checkSingleStep :: Table -- ^ The original table

                -> Instance.Instance -- ^ The instance to move

                -> Table -- ^ The current best table

                -> IMove -- ^ The move to apply

                -> Table -- ^ The final best table

checkSingleStep ini_tbl target cur_tbl move =

  let Table ini_nl ini_il _ ini_plc = ini_tbl

      tmp_resu = applyMove ini_nl target move

  in case tmp_resu of

       OpFail _ -> cur_tbl

       OpGood (upd_nl, new_inst, pri_idx, sec_idx) ->

         let tgt_idx = Instance.idx target

             upd_cvar = compCV upd_nl

             upd_il = Container.add tgt_idx new_inst ini_il

             upd_plc = (tgt_idx, pri_idx, sec_idx, move, upd_cvar):ini_plc

             upd_tbl = Table upd_nl upd_il upd_cvar upd_plc

         in compareTables cur_tbl upd_tbl

-- | Given the status of the current secondary as a valid new node and

-- the current candidate target node, generate the possible moves for

-- a instance.

possibleMoves :: Bool      -- ^ Whether the secondary node is a valid new node

              -> Bool      -- ^ Whether we can change the primary node

              -> Ndx       -- ^ Target node candidate

              -> [IMove]   -- ^ List of valid result moves

possibleMoves _ False tdx =

  [ReplaceSecondary tdx]

possibleMoves True True tdx =

  [ ReplaceSecondary tdx

  , ReplaceAndFailover tdx

  , ReplacePrimary tdx

  , FailoverAndReplace tdx

  ]

possibleMoves False True tdx =

  [ ReplaceSecondary tdx

  , ReplaceAndFailover tdx

  ]

-- | Compute the best move for a given instance.

checkInstanceMove :: [Ndx]             -- ^ Allowed target node indices

                  -> Bool              -- ^ Whether disk moves are allowed

                  -> Bool              -- ^ Whether instance moves are allowed

                  -> Table             -- ^ Original table

                  -> Instance.Instance -- ^ Instance to move

                  -> Table             -- ^ Best new table for this instance

checkInstanceMove nodes_idx disk_moves inst_moves ini_tbl target =

  let opdx = Instance.pNode target

      osdx = Instance.sNode target

      bad_nodes = [opdx, osdx]

      nodes = filter (`notElem` bad_nodes) nodes_idx

      use_secondary = elem osdx nodes_idx && inst_moves

      aft_failover = if use_secondary -- if allowed to failover

                       then checkSingleStep ini_tbl target ini_tbl Failover

                       else ini_tbl

      all_moves = if disk_moves

                    then concatMap

                           (possibleMoves use_secondary inst_moves) nodes

                    else []

    in

      -- iterate over the possible nodes for this instance

      foldl' (checkSingleStep ini_tbl target) aft_failover all_moves

-- | Compute the best next move.

checkMove :: [Ndx]               -- ^ Allowed target node indices

          -> Bool                -- ^ Whether disk moves are allowed

          -> Bool                -- ^ Whether instance moves are allowed

          -> Table               -- ^ The current solution

          -> [Instance.Instance] -- ^ List of instances still to move

          -> Table               -- ^ The new solution

checkMove nodes_idx disk_moves inst_moves ini_tbl victims =

  let Table _ _ _ ini_plc = ini_tbl

      -- we're using rwhnf from the Control.Parallel.Strategies

      -- package; we don't need to use rnf as that would force too

      -- much evaluation in single-threaded cases, and in

      -- multi-threaded case the weak head normal form is enough to

      -- spark the evaluation

      tables = parMap rwhnf (checkInstanceMove nodes_idx disk_moves

                             inst_moves ini_tbl)

               victims

      -- iterate over all instances, computing the best move

      best_tbl = foldl' compareTables ini_tbl tables

      Table _ _ _ best_plc = best_tbl

  in if length best_plc == length ini_plc

       then ini_tbl -- no advancement

       else best_tbl

-- | Check if we are allowed to go deeper in the balancing.

doNextBalance :: Table     -- ^ The starting table

              -> Int       -- ^ Remaining length

              -> Score     -- ^ Score at which to stop

              -> Bool      -- ^ The resulting table and commands

doNextBalance ini_tbl max_rounds min_score =

  let Table _ _ ini_cv ini_plc = ini_tbl

      ini_plc_len = length ini_plc

  in (max_rounds < 0 || ini_plc_len < max_rounds) && ini_cv > min_score

-- | Run a balance move.

tryBalance :: Table       -- ^ The starting table

           -> Bool        -- ^ Allow disk moves

           -> Bool        -- ^ Allow instance moves

           -> Bool        -- ^ Only evacuate moves

           -> Score       -- ^ Min gain threshold

           -> Score       -- ^ Min gain

           -> Maybe Table -- ^ The resulting table and commands

tryBalance ini_tbl disk_moves inst_moves evac_mode mg_limit min_gain =

    let Table ini_nl ini_il ini_cv _ = ini_tbl

        all_inst = Container.elems ini_il

        all_inst' = if evac_mode

                    then let bad_nodes = map Node.idx . filter Node.offline $

                                         Container.elems ini_nl

                         in filter (any (`elem` bad_nodes) . Instance.allNodes)

                            all_inst

                    else all_inst

        reloc_inst = filter Instance.movable all_inst'

        node_idx = map Node.idx . filter (not . Node.offline) $

                   Container.elems ini_nl

        fin_tbl = checkMove node_idx disk_moves inst_moves ini_tbl reloc_inst

        (Table _ _ fin_cv _) = fin_tbl

    in

      if fin_cv < ini_cv && (ini_cv > mg_limit || ini_cv - fin_cv >= min_gain)

      then Just fin_tbl -- this round made success, return the new table

      else Nothing

-- * Allocation functions

-- | Build failure stats out of a list of failures.

collapseFailures :: [FailMode] -> FailStats

collapseFailures flst =

    map (\k -> (k, foldl' (\a e -> if e == k then a + 1 else a) 0 flst))

            [minBound..maxBound]

-- | Compares two Maybe AllocElement and chooses the besst score.

bestAllocElement :: Maybe Node.AllocElement

                 -> Maybe Node.AllocElement

                 -> Maybe Node.AllocElement

bestAllocElement a Nothing = a

bestAllocElement Nothing b = b

bestAllocElement a@(Just (_, _, _, ascore)) b@(Just (_, _, _, bscore)) =

  if ascore < bscore then a else b

-- | Update current Allocation solution and failure stats with new

-- elements.

concatAllocs :: AllocSolution -> OpResult Node.AllocElement -> AllocSolution

concatAllocs as (OpFail reason) = as { asFailures = reason : asFailures as }

concatAllocs as (OpGood ns) =

  let -- Choose the old or new solution, based on the cluster score

    cntok = asAllocs as

    osols = asSolution as

    nsols = bestAllocElement osols (Just ns)

    nsuc = cntok + 1

    -- Note: we force evaluation of nsols here in order to keep the

    -- memory profile low - we know that we will need nsols for sure

    -- in the next cycle, so we force evaluation of nsols, since the

    -- foldl' in the caller will only evaluate the tuple, but not the

    -- elements of the tuple

  in nsols `seq` nsuc `seq` as { asAllocs = nsuc, asSolution = nsols }

-- | Sums two 'AllocSolution' structures.

sumAllocs :: AllocSolution -> AllocSolution -> AllocSolution

sumAllocs (AllocSolution aFails aAllocs aSols aLog)

          (AllocSolution bFails bAllocs bSols bLog) =

  -- note: we add b first, since usually it will be smaller; when

  -- fold'ing, a will grow and grow whereas b is the per-group

  -- result, hence smaller

  let nFails  = bFails ++ aFails

      nAllocs = aAllocs + bAllocs

      nSols   = bestAllocElement aSols bSols

      nLog    = bLog ++ aLog

  in AllocSolution nFails nAllocs nSols nLog

-- | Given a solution, generates a reasonable description for it.

describeSolution :: AllocSolution -> String

describeSolution as =

  let fcnt = asFailures as

      sols = asSolution as

      freasons =

        intercalate ", " . map (\(a, b) -> printf "%s: %d" (show a) b) .

        filter ((> 0) . snd) . collapseFailures $ fcnt

  in case sols of

     Nothing -> "No valid allocation solutions, failure reasons: " ++

                (if null fcnt then "unknown reasons" else freasons)

     Just (_, _, nodes, cv) ->

         printf ("score: %.8f, successes %d, failures %d (%s)" ++

                 " for node(s) %s") cv (asAllocs as) (length fcnt) freasons

               (intercalate "/" . map Node.name $ nodes)

-- | Annotates a solution with the appropriate string.

annotateSolution :: AllocSolution -> AllocSolution

annotateSolution as = as { asLog = describeSolution as : asLog as }

-- | Reverses an evacuation solution.

--

-- Rationale: we always concat the results to the top of the lists, so

-- for proper jobset execution, we should reverse all lists.

reverseEvacSolution :: EvacSolution -> EvacSolution

reverseEvacSolution (EvacSolution f m o) =

  EvacSolution (reverse f) (reverse m) (reverse o)

-- | Generate the valid node allocation singles or pairs for a new instance.

genAllocNodes :: Group.List        -- ^ Group list

              -> Node.List         -- ^ The node map

              -> Int               -- ^ The number of nodes required

              -> Bool              -- ^ Whether to drop or not

                                   -- unallocable nodes

              -> Result AllocNodes -- ^ The (monadic) result

genAllocNodes gl nl count drop_unalloc =

  let filter_fn = if drop_unalloc

                    then filter (Group.isAllocable .

                                 flip Container.find gl . Node.group)

                    else id

      all_nodes = filter_fn $ getOnline nl

      all_pairs = [(Node.idx p,

                    [Node.idx s | s <- all_nodes,

                                       Node.idx p /= Node.idx s,

                                       Node.group p == Node.group s]) |

                   p <- all_nodes]

  in case count of

       1 -> Ok (Left (map Node.idx all_nodes))

       2 -> Ok (Right (filter (not . null . snd) all_pairs))

       _ -> Bad "Unsupported number of nodes, only one or two  supported"

-- | Try to allocate an instance on the cluster.

tryAlloc :: (Monad m) =>

            Node.List         -- ^ The node list

         -> Instance.List     -- ^ The instance list

         -> Instance.Instance -- ^ The instance to allocate

         -> AllocNodes        -- ^ The allocation targets

         -> m AllocSolution   -- ^ Possible solution list

tryAlloc _  _ _    (Right []) = fail "Not enough online nodes"

tryAlloc nl _ inst (Right ok_pairs) =

  let psols = parMap rwhnf (\(p, ss) ->

                              foldl' (\cstate ->

                                        concatAllocs cstate .

                                        allocateOnPair nl inst p)

                              emptyAllocSolution ss) ok_pairs

      sols = foldl' sumAllocs emptyAllocSolution psols

  in return $ annotateSolution sols

tryAlloc _  _ _    (Left []) = fail "No online nodes"

tryAlloc nl _ inst (Left all_nodes) =

  let sols = foldl' (\cstate ->

                       concatAllocs cstate . allocateOnSingle nl inst

                    ) emptyAllocSolution all_nodes

  in return $ annotateSolution sols

-- | Given a group/result, describe it as a nice (list of) messages.

solutionDescription :: Group.List -> (Gdx, Result AllocSolution) -> [String]

solutionDescription gl (groupId, result) =

  case result of

    Ok solution -> map (printf "Group %s (%s): %s" gname pol) (asLog solution)

    Bad message -> [printf "Group %s: error %s" gname message]

  where grp = Container.find groupId gl

        gname = Group.name grp

        pol = allocPolicyToRaw (Group.allocPolicy grp)

-- | From a list of possibly bad and possibly empty solutions, filter

-- only the groups with a valid result. Note that the result will be

-- reversed compared to the original list.

filterMGResults :: Group.List

                -> [(Gdx, Result AllocSolution)]

                -> [(Gdx, AllocSolution)]

filterMGResults gl = foldl' fn []

  where unallocable = not . Group.isAllocable . flip Container.find gl

        fn accu (gdx, rasol) =

          case rasol of

            Bad _ -> accu

            Ok sol | isNothing (asSolution sol) -> accu

                   | unallocable gdx -> accu

                   | otherwise -> (gdx, sol):accu

-- | Sort multigroup results based on policy and score.

sortMGResults :: Group.List

             -> [(Gdx, AllocSolution)]

             -> [(Gdx, AllocSolution)]

sortMGResults gl sols =

  let extractScore (_, _, _, x) = x

      solScore (gdx, sol) = (Group.allocPolicy (Container.find gdx gl),

                             (extractScore . fromJust . asSolution) sol)

  in sortBy (comparing solScore) sols

-- | Finds the best group for an instance on a multi-group cluster.

--

-- Only solutions in @preferred@ and @last_resort@ groups will be

-- accepted as valid, and additionally if the allowed groups parameter

-- is not null then allocation will only be run for those group

-- indices.

findBestAllocGroup :: Group.List           -- ^ The group list

                   -> Node.List            -- ^ The node list

                   -> Instance.List        -- ^ The instance list

                   -> Maybe [Gdx]          -- ^ The allowed groups

                   -> Instance.Instance    -- ^ The instance to allocate

                   -> Int                  -- ^ Required number of nodes

                   -> Result (Gdx, AllocSolution, [String])

findBestAllocGroup mggl mgnl mgil allowed_gdxs inst cnt =

  let groups = splitCluster mgnl mgil

      groups' = maybe groups (\gs -> filter ((`elem` gs) . fst) groups)

                allowed_gdxs

      sols = map (\(gid, (nl, il)) ->

                   (gid, genAllocNodes mggl nl cnt False >>=

                       tryAlloc nl il inst))

             groups'::[(Gdx, Result AllocSolution)]

      all_msgs = concatMap (solutionDescription mggl) sols

      goodSols = filterMGResults mggl sols

      sortedSols = sortMGResults mggl goodSols

  in if null sortedSols

       then if null groups'

              then Bad $ "no groups for evacuation: allowed groups was" ++

                     show allowed_gdxs ++ ", all groups: " ++

                     show (map fst groups)

              else Bad $ intercalate ", " all_msgs

       else let (final_group, final_sol) = head sortedSols

            in return (final_group, final_sol, all_msgs)

-- | Try to allocate an instance on a multi-group cluster.

tryMGAlloc :: Group.List           -- ^ The group list

           -> Node.List            -- ^ The node list

           -> Instance.List        -- ^ The instance list

           -> Instance.Instance    -- ^ The instance to allocate

           -> Int                  -- ^ Required number of nodes

           -> Result AllocSolution -- ^ Possible solution list

tryMGAlloc mggl mgnl mgil inst cnt = do

  (best_group, solution, all_msgs) <-

      findBestAllocGroup mggl mgnl mgil Nothing inst cnt

  let group_name = Group.name $ Container.find best_group mggl

      selmsg = "Selected group: " ++ group_name

  return $ solution { asLog = selmsg:all_msgs }

-- | Function which fails if the requested mode is change secondary.

--

-- This is useful since except DRBD, no other disk template can

-- execute change secondary; thus, we can just call this function

-- instead of always checking for secondary mode. After the call to

-- this function, whatever mode we have is just a primary change.

failOnSecondaryChange :: (Monad m) => EvacMode -> DiskTemplate -> m ()

failOnSecondaryChange ChangeSecondary dt =

  fail $ "Instances with disk template '" ++ diskTemplateToRaw dt ++

         "' can't execute change secondary"

failOnSecondaryChange _ _ = return ()

-- | Run evacuation for a single instance.

--

-- /Note:/ this function should correctly execute both intra-group

-- evacuations (in all modes) and inter-group evacuations (in the

-- 'ChangeAll' mode). Of course, this requires that the correct list

-- of target nodes is passed.

nodeEvacInstance :: Node.List         -- ^ The node list (cluster-wide)

                 -> Instance.List     -- ^ Instance list (cluster-wide)

                 -> EvacMode          -- ^ The evacuation mode

                 -> Instance.Instance -- ^ The instance to be evacuated

                 -> Gdx               -- ^ The group we're targetting

                 -> [Ndx]             -- ^ The list of available nodes

                                      -- for allocation

                 -> Result (Node.List, Instance.List, [OpCodes.OpCode])

nodeEvacInstance _ _ mode (Instance.Instance

                           {Instance.diskTemplate = dt@DTDiskless}) _ _ =

                  failOnSecondaryChange mode dt >>

                  fail "Diskless relocations not implemented yet"

nodeEvacInstance _ _ _ (Instance.Instance

                        {Instance.diskTemplate = DTPlain}) _ _ =

                  fail "Instances of type plain cannot be relocated"

nodeEvacInstance _ _ _ (Instance.Instance

                        {Instance.diskTemplate = DTFile}) _ _ =

                  fail "Instances of type file cannot be relocated"

nodeEvacInstance _ _ mode  (Instance.Instance

                            {Instance.diskTemplate = dt@DTSharedFile}) _ _ =

                  failOnSecondaryChange mode dt >>

                  fail "Shared file relocations not implemented yet"

nodeEvacInstance _ _ mode (Instance.Instance

                           {Instance.diskTemplate = dt@DTBlock}) _ _ =

                  failOnSecondaryChange mode dt >>

                  fail "Block device relocations not implemented yet"

nodeEvacInstance _ _ mode  (Instance.Instance

                            {Instance.diskTemplate = dt@DTRbd}) _ _ =

                  failOnSecondaryChange mode dt >>

                  fail "Rbd relocations not implemented yet"

nodeEvacInstance nl il ChangePrimary

                 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})

                 _ _ =

  do

    (nl', inst', _, _) <- opToResult $ applyMove nl inst Failover

    let idx = Instance.idx inst

        il' = Container.add idx inst' il

        ops = iMoveToJob nl' il' idx Failover

    return (nl', il', ops)

nodeEvacInstance nl il ChangeSecondary

                 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})

                 gdx avail_nodes =

  do

    (nl', inst', _, ndx) <- annotateResult "Can't find any good node" $

                            eitherToResult $

                            foldl' (evacDrbdSecondaryInner nl inst gdx)

                            (Left "no nodes available") avail_nodes

    let idx = Instance.idx inst

        il' = Container.add idx inst' il

        ops = iMoveToJob nl' il' idx (ReplaceSecondary ndx)

    return (nl', il', ops)

-- The algorithm for ChangeAll is as follows:

--

-- * generate all (primary, secondary) node pairs for the target groups

-- * for each pair, execute the needed moves (r:s, f, r:s) and compute

--   the final node list state and group score

-- * select the best choice via a foldl that uses the same Either

--   String solution as the ChangeSecondary mode

nodeEvacInstance nl il ChangeAll

                 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})

                 gdx avail_nodes =

  do

    let no_nodes = Left "no nodes available"

        node_pairs = [(p,s) | p <- avail_nodes, s <- avail_nodes, p /= s]

    (nl', il', ops, _) <-

        annotateResult "Can't find any good nodes for relocation" $

        eitherToResult $

        foldl'

        (\accu nodes -> case evacDrbdAllInner nl il inst gdx nodes of

                          Bad msg ->

                              case accu of

                                Right _ -> accu

                                -- we don't need more details (which

                                -- nodes, etc.) as we only selected

                                -- this group if we can allocate on

                                -- it, hence failures will not

                                -- propagate out of this fold loop

                                Left _ -> Left $ "Allocation failed: " ++ msg

                          Ok result@(_, _, _, new_cv) ->

                              let new_accu = Right result in

                              case accu of

                                Left _ -> new_accu

                                Right (_, _, _, old_cv) ->

                                    if old_cv < new_cv

                                    then accu

                                    else new_accu

        ) no_nodes node_pairs

    return (nl', il', ops)

-- | Inner fold function for changing secondary of a DRBD instance.

--

-- The running solution is either a @Left String@, which means we

-- don't have yet a working solution, or a @Right (...)@, which

-- represents a valid solution; it holds the modified node list, the

-- modified instance (after evacuation), the score of that solution,

-- and the new secondary node index.

evacDrbdSecondaryInner :: Node.List -- ^ Cluster node list

                       -> Instance.Instance -- ^ Instance being evacuated

                       -> Gdx -- ^ The group index of the instance

                       -> Either String ( Node.List

                                        , Instance.Instance

                                        , Score

                                        , Ndx)  -- ^ Current best solution

                       -> Ndx  -- ^ Node we're evaluating as new secondary

                       -> Either String ( Node.List

                                        , Instance.Instance

                                        , Score

                                        , Ndx) -- ^ New best solution

evacDrbdSecondaryInner nl inst gdx accu ndx =

  case applyMove nl inst (ReplaceSecondary ndx) of

    OpFail fm ->

      case accu of

        Right _ -> accu

        Left _ -> Left $ "Node " ++ Container.nameOf nl ndx ++

                  " failed: " ++ show fm

    OpGood (nl', inst', _, _) ->

      let nodes = Container.elems nl'

          -- The fromJust below is ugly (it can fail nastily), but

          -- at this point we should have any internal mismatches,

          -- and adding a monad here would be quite involved

          grpnodes = fromJust (gdx `lookup` Node.computeGroups nodes)

          new_cv = compCVNodes grpnodes

          new_accu = Right (nl', inst', new_cv, ndx)