1 {-| Implementation of cluster-wide logic.
3 This module holds all pure cluster-logic; I\/O related functionality
4 goes into the /Main/ module for the individual binaries.
10 Copyright (C) 2009, 2010, 2011, 2012 Google Inc.
12 This program is free software; you can redistribute it and/or modify
13 it under the terms of the GNU General Public License as published by
14 the Free Software Foundation; either version 2 of the License, or
15 (at your option) any later version.
17 This program is distributed in the hope that it will be useful, but
18 WITHOUT ANY WARRANTY; without even the implied warranty of
19 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 General Public License for more details.
22 You should have received a copy of the GNU General Public License
23 along with this program; if not, write to the Free Software
24 Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
29 module Ganeti.HTools.Cluster
38 -- * Generic functions
40 , computeAllocationDelta
41 -- * First phase functions
43 -- * Second phase functions
48 -- * Display functions
51 -- * Balacing functions
60 -- * IAllocator functions
67 -- * Allocation functions
70 -- * Node group functions
76 import qualified Data.IntSet as IntSet
78 import Data.Maybe (fromJust, isNothing)
79 import Data.Ord (comparing)
80 import Text.Printf (printf)
82 import qualified Ganeti.HTools.Container as Container
83 import qualified Ganeti.HTools.Instance as Instance
84 import qualified Ganeti.HTools.Node as Node
85 import qualified Ganeti.HTools.Group as Group
86 import Ganeti.HTools.Types
87 import Ganeti.HTools.Utils
88 import Ganeti.HTools.Compat
89 import qualified Ganeti.OpCodes as OpCodes
93 -- | Allocation\/relocation solution.
94 data AllocSolution = AllocSolution
95 { asFailures :: [FailMode] -- ^ Failure counts
96 , asAllocs :: Int -- ^ Good allocation count
97 , asSolution :: Maybe Node.AllocElement -- ^ The actual allocation result
98 , asLog :: [String] -- ^ Informational messages
101 -- | Node evacuation/group change iallocator result type. This result
102 -- type consists of actual opcodes (a restricted subset) that are
103 -- transmitted back to Ganeti.
104 data EvacSolution = EvacSolution
105 { esMoved :: [(Idx, Gdx, [Ndx])] -- ^ Instances moved successfully
106 , esFailed :: [(Idx, String)] -- ^ Instances which were not
108 , esOpCodes :: [[OpCodes.OpCode]] -- ^ List of jobs
111 -- | Allocation results, as used in 'iterateAlloc' and 'tieredAlloc'.
112 type AllocResult = (FailStats, Node.List, Instance.List,
113 [Instance.Instance], [CStats])
115 -- | A type denoting the valid allocation mode/pairs.
117 -- For a one-node allocation, this will be a @Left ['Ndx']@, whereas
118 -- for a two-node allocation, this will be a @Right [('Ndx',
119 -- ['Ndx'])]@. In the latter case, the list is basically an
120 -- association list, grouped by primary node and holding the potential
121 -- secondary nodes in the sub-list.
122 type AllocNodes = Either [Ndx] [(Ndx, [Ndx])]
124 -- | The empty solution we start with when computing allocations.
125 emptyAllocSolution :: AllocSolution
126 emptyAllocSolution = AllocSolution { asFailures = [], asAllocs = 0
127 , asSolution = Nothing, asLog = [] }
129 -- | The empty evac solution.
130 emptyEvacSolution :: EvacSolution
131 emptyEvacSolution = EvacSolution { esMoved = []
136 -- | The complete state for the balancing solution.
137 data Table = Table Node.List Instance.List Score [Placement]
138 deriving (Show, Read)
140 -- | Cluster statistics data type.
142 { csFmem :: Integer -- ^ Cluster free mem
143 , csFdsk :: Integer -- ^ Cluster free disk
144 , csAmem :: Integer -- ^ Cluster allocatable mem
145 , csAdsk :: Integer -- ^ Cluster allocatable disk
146 , csAcpu :: Integer -- ^ Cluster allocatable cpus
147 , csMmem :: Integer -- ^ Max node allocatable mem
148 , csMdsk :: Integer -- ^ Max node allocatable disk
149 , csMcpu :: Integer -- ^ Max node allocatable cpu
150 , csImem :: Integer -- ^ Instance used mem
151 , csIdsk :: Integer -- ^ Instance used disk
152 , csIcpu :: Integer -- ^ Instance used cpu
153 , csTmem :: Double -- ^ Cluster total mem
154 , csTdsk :: Double -- ^ Cluster total disk
155 , csTcpu :: Double -- ^ Cluster total cpus
156 , csVcpu :: Integer -- ^ Cluster total virtual cpus
157 , csNcpu :: Double -- ^ Equivalent to 'csIcpu' but in terms of
158 -- physical CPUs, i.e. normalised used phys CPUs
159 , csXmem :: Integer -- ^ Unnacounted for mem
160 , csNmem :: Integer -- ^ Node own memory
161 , csScore :: Score -- ^ The cluster score
162 , csNinst :: Int -- ^ The total number of instances
163 } deriving (Show, Read)
165 -- | A simple type for allocation functions.
166 type AllocMethod = Node.List -- ^ Node list
167 -> Instance.List -- ^ Instance list
168 -> Maybe Int -- ^ Optional allocation limit
169 -> Instance.Instance -- ^ Instance spec for allocation
170 -> AllocNodes -- ^ Which nodes we should allocate on
171 -> [Instance.Instance] -- ^ Allocated instances
172 -> [CStats] -- ^ Running cluster stats
173 -> Result AllocResult -- ^ Allocation result
175 -- | A simple type for the running solution of evacuations.
176 type EvacInnerState =
177 Either String (Node.List, Instance.Instance, Score, Ndx)
179 -- * Utility functions
181 -- | Verifies the N+1 status and return the affected nodes.
182 verifyN1 :: [Node.Node] -> [Node.Node]
183 verifyN1 = filter Node.failN1
185 {-| Computes the pair of bad nodes and instances.
187 The bad node list is computed via a simple 'verifyN1' check, and the
188 bad instance list is the list of primary and secondary instances of
192 computeBadItems :: Node.List -> Instance.List ->
193 ([Node.Node], [Instance.Instance])
194 computeBadItems nl il =
195 let bad_nodes = verifyN1 $ getOnline nl
196 bad_instances = map (`Container.find` il) .
198 concatMap (\ n -> Node.sList n ++ Node.pList n) bad_nodes
200 (bad_nodes, bad_instances)
202 -- | Extracts the node pairs for an instance. This can fail if the
203 -- instance is single-homed. FIXME: this needs to be improved,
204 -- together with the general enhancement for handling non-DRBD moves.
205 instanceNodes :: Node.List -> Instance.Instance ->
206 (Ndx, Ndx, Node.Node, Node.Node)
207 instanceNodes nl inst =
208 let old_pdx = Instance.pNode inst
209 old_sdx = Instance.sNode inst
210 old_p = Container.find old_pdx nl
211 old_s = Container.find old_sdx nl
212 in (old_pdx, old_sdx, old_p, old_s)
214 -- | Zero-initializer for the CStats type.
215 emptyCStats :: CStats
216 emptyCStats = CStats 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
218 -- | Update stats with data from a new node.
219 updateCStats :: CStats -> Node.Node -> CStats
220 updateCStats cs node =
221 let CStats { csFmem = x_fmem, csFdsk = x_fdsk,
222 csAmem = x_amem, csAcpu = x_acpu, csAdsk = x_adsk,
223 csMmem = x_mmem, csMdsk = x_mdsk, csMcpu = x_mcpu,
224 csImem = x_imem, csIdsk = x_idsk, csIcpu = x_icpu,
225 csTmem = x_tmem, csTdsk = x_tdsk, csTcpu = x_tcpu,
226 csVcpu = x_vcpu, csNcpu = x_ncpu,
227 csXmem = x_xmem, csNmem = x_nmem, csNinst = x_ninst
230 inc_amem = Node.fMem node - Node.rMem node
231 inc_amem' = if inc_amem > 0 then inc_amem else 0
232 inc_adsk = Node.availDisk node
233 inc_imem = truncate (Node.tMem node) - Node.nMem node
234 - Node.xMem node - Node.fMem node
235 inc_icpu = Node.uCpu node
236 inc_idsk = truncate (Node.tDsk node) - Node.fDsk node
237 inc_vcpu = Node.hiCpu node
238 inc_acpu = Node.availCpu node
239 inc_ncpu = fromIntegral (Node.uCpu node) /
240 iPolicyVcpuRatio (Node.iPolicy node)
241 in cs { csFmem = x_fmem + fromIntegral (Node.fMem node)
242 , csFdsk = x_fdsk + fromIntegral (Node.fDsk node)
243 , csAmem = x_amem + fromIntegral inc_amem'
244 , csAdsk = x_adsk + fromIntegral inc_adsk
245 , csAcpu = x_acpu + fromIntegral inc_acpu
246 , csMmem = max x_mmem (fromIntegral inc_amem')
247 , csMdsk = max x_mdsk (fromIntegral inc_adsk)
248 , csMcpu = max x_mcpu (fromIntegral inc_acpu)
249 , csImem = x_imem + fromIntegral inc_imem
250 , csIdsk = x_idsk + fromIntegral inc_idsk
251 , csIcpu = x_icpu + fromIntegral inc_icpu
252 , csTmem = x_tmem + Node.tMem node
253 , csTdsk = x_tdsk + Node.tDsk node
254 , csTcpu = x_tcpu + Node.tCpu node
255 , csVcpu = x_vcpu + fromIntegral inc_vcpu
256 , csNcpu = x_ncpu + inc_ncpu
257 , csXmem = x_xmem + fromIntegral (Node.xMem node)
258 , csNmem = x_nmem + fromIntegral (Node.nMem node)
259 , csNinst = x_ninst + length (Node.pList node)
262 -- | Compute the total free disk and memory in the cluster.
263 totalResources :: Node.List -> CStats
265 let cs = foldl' updateCStats emptyCStats . Container.elems $ nl
266 in cs { csScore = compCV nl }
268 -- | Compute the delta between two cluster state.
270 -- This is used when doing allocations, to understand better the
271 -- available cluster resources. The return value is a triple of the
272 -- current used values, the delta that was still allocated, and what
273 -- was left unallocated.
274 computeAllocationDelta :: CStats -> CStats -> AllocStats
275 computeAllocationDelta cini cfin =
276 let CStats {csImem = i_imem, csIdsk = i_idsk, csIcpu = i_icpu,
277 csNcpu = i_ncpu } = cini
278 CStats {csImem = f_imem, csIdsk = f_idsk, csIcpu = f_icpu,
279 csTmem = t_mem, csTdsk = t_dsk, csVcpu = f_vcpu,
280 csNcpu = f_ncpu, csTcpu = f_tcpu } = cfin
281 rini = AllocInfo { allocInfoVCpus = fromIntegral i_icpu
282 , allocInfoNCpus = i_ncpu
283 , allocInfoMem = fromIntegral i_imem
284 , allocInfoDisk = fromIntegral i_idsk
286 rfin = AllocInfo { allocInfoVCpus = fromIntegral (f_icpu - i_icpu)
287 , allocInfoNCpus = f_ncpu - i_ncpu
288 , allocInfoMem = fromIntegral (f_imem - i_imem)
289 , allocInfoDisk = fromIntegral (f_idsk - i_idsk)
291 runa = AllocInfo { allocInfoVCpus = fromIntegral (f_vcpu - f_icpu)
292 , allocInfoNCpus = f_tcpu - f_ncpu
293 , allocInfoMem = truncate t_mem - fromIntegral f_imem
294 , allocInfoDisk = truncate t_dsk - fromIntegral f_idsk
296 in (rini, rfin, runa)
298 -- | The names and weights of the individual elements in the CV list.
299 detailedCVInfo :: [(Double, String)]
300 detailedCVInfo = [ (1, "free_mem_cv")
301 , (1, "free_disk_cv")
303 , (1, "reserved_mem_cv")
304 , (4, "offline_all_cnt")
305 , (16, "offline_pri_cnt")
306 , (1, "vcpu_ratio_cv")
309 , (1, "disk_load_cv")
311 , (2, "pri_tags_score")
315 -- | Holds the weights used by 'compCVNodes' for each metric.
316 detailedCVWeights :: [Double]
317 detailedCVWeights = map fst detailedCVInfo
319 -- | Compute the mem and disk covariance.
320 compDetailedCV :: [Node.Node] -> [Double]
321 compDetailedCV all_nodes =
322 let (offline, nodes) = partition Node.offline all_nodes
323 mem_l = map Node.pMem nodes
324 dsk_l = map Node.pDsk nodes
325 -- metric: memory covariance
326 mem_cv = stdDev mem_l
327 -- metric: disk covariance
328 dsk_cv = stdDev dsk_l
329 -- metric: count of instances living on N1 failing nodes
330 n1_score = fromIntegral . sum . map (\n -> length (Node.sList n) +
331 length (Node.pList n)) .
332 filter Node.failN1 $ nodes :: Double
333 res_l = map Node.pRem nodes
334 -- metric: reserved memory covariance
335 res_cv = stdDev res_l
336 -- offline instances metrics
337 offline_ipri = sum . map (length . Node.pList) $ offline
338 offline_isec = sum . map (length . Node.sList) $ offline
339 -- metric: count of instances on offline nodes
340 off_score = fromIntegral (offline_ipri + offline_isec)::Double
341 -- metric: count of primary instances on offline nodes (this
342 -- helps with evacuation/failover of primary instances on
343 -- 2-node clusters with one node offline)
344 off_pri_score = fromIntegral offline_ipri::Double
345 cpu_l = map Node.pCpu nodes
346 -- metric: covariance of vcpu/pcpu ratio
347 cpu_cv = stdDev cpu_l
348 -- metrics: covariance of cpu, memory, disk and network load
349 (c_load, m_load, d_load, n_load) =
351 let DynUtil c1 m1 d1 n1 = Node.utilLoad n
352 DynUtil c2 m2 d2 n2 = Node.utilPool n
353 in (c1/c2, m1/m2, d1/d2, n1/n2)) nodes
354 -- metric: conflicting instance count
355 pri_tags_inst = sum $ map Node.conflictingPrimaries nodes
356 pri_tags_score = fromIntegral pri_tags_inst::Double
357 -- metric: spindles %
358 spindles_cv = map (\n -> Node.instSpindles n / Node.hiSpindles n) nodes
359 in [ mem_cv, dsk_cv, n1_score, res_cv, off_score, off_pri_score, cpu_cv
360 , stdDev c_load, stdDev m_load , stdDev d_load, stdDev n_load
361 , pri_tags_score, stdDev spindles_cv ]
363 -- | Compute the /total/ variance.
364 compCVNodes :: [Node.Node] -> Double
365 compCVNodes = sum . zipWith (*) detailedCVWeights . compDetailedCV
367 -- | Wrapper over 'compCVNodes' for callers that have a 'Node.List'.
368 compCV :: Node.List -> Double
369 compCV = compCVNodes . Container.elems
371 -- | Compute online nodes from a 'Node.List'.
372 getOnline :: Node.List -> [Node.Node]
373 getOnline = filter (not . Node.offline) . Container.elems
375 -- * Balancing functions
377 -- | Compute best table. Note that the ordering of the arguments is important.
378 compareTables :: Table -> Table -> Table
379 compareTables a@(Table _ _ a_cv _) b@(Table _ _ b_cv _ ) =
380 if a_cv > b_cv then b else a
382 -- | Applies an instance move to a given node list and instance.
383 applyMove :: Node.List -> Instance.Instance
384 -> IMove -> OpResult (Node.List, Instance.Instance, Ndx, Ndx)
386 applyMove nl inst Failover =
387 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
388 int_p = Node.removePri old_p inst
389 int_s = Node.removeSec old_s inst
390 new_nl = do -- Maybe monad
391 new_p <- Node.addPriEx (Node.offline old_p) int_s inst
392 new_s <- Node.addSec int_p inst old_sdx
393 let new_inst = Instance.setBoth inst old_sdx old_pdx
394 return (Container.addTwo old_pdx new_s old_sdx new_p nl,
395 new_inst, old_sdx, old_pdx)
398 -- Failover to any (fa)
399 applyMove nl inst (FailoverToAny new_pdx) = do
400 let (old_pdx, old_sdx, old_pnode, _) = instanceNodes nl inst
401 new_pnode = Container.find new_pdx nl
402 force_failover = Node.offline old_pnode
403 new_pnode' <- Node.addPriEx force_failover new_pnode inst
404 let old_pnode' = Node.removePri old_pnode inst
405 inst' = Instance.setPri inst new_pdx
406 nl' = Container.addTwo old_pdx old_pnode' new_pdx new_pnode' nl
407 return (nl', inst', new_pdx, old_sdx)
409 -- Replace the primary (f:, r:np, f)
410 applyMove nl inst (ReplacePrimary new_pdx) =
411 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
412 tgt_n = Container.find new_pdx nl
413 int_p = Node.removePri old_p inst
414 int_s = Node.removeSec old_s inst
415 force_p = Node.offline old_p
416 new_nl = do -- Maybe monad
417 -- check that the current secondary can host the instance
418 -- during the migration
419 tmp_s <- Node.addPriEx force_p int_s inst
420 let tmp_s' = Node.removePri tmp_s inst
421 new_p <- Node.addPriEx force_p tgt_n inst
422 new_s <- Node.addSecEx force_p tmp_s' inst new_pdx
423 let new_inst = Instance.setPri inst new_pdx
424 return (Container.add new_pdx new_p $
425 Container.addTwo old_pdx int_p old_sdx new_s nl,
426 new_inst, new_pdx, old_sdx)
429 -- Replace the secondary (r:ns)
430 applyMove nl inst (ReplaceSecondary new_sdx) =
431 let old_pdx = Instance.pNode inst
432 old_sdx = Instance.sNode inst
433 old_s = Container.find old_sdx nl
434 tgt_n = Container.find new_sdx nl
435 int_s = Node.removeSec old_s inst
436 force_s = Node.offline old_s
437 new_inst = Instance.setSec inst new_sdx
438 new_nl = Node.addSecEx force_s tgt_n inst old_pdx >>=
439 \new_s -> return (Container.addTwo new_sdx
440 new_s old_sdx int_s nl,
441 new_inst, old_pdx, new_sdx)
444 -- Replace the secondary and failover (r:np, f)
445 applyMove nl inst (ReplaceAndFailover new_pdx) =
446 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
447 tgt_n = Container.find new_pdx nl
448 int_p = Node.removePri old_p inst
449 int_s = Node.removeSec old_s inst
450 force_s = Node.offline old_s
451 new_nl = do -- Maybe monad
452 new_p <- Node.addPri tgt_n inst
453 new_s <- Node.addSecEx force_s int_p inst new_pdx
454 let new_inst = Instance.setBoth inst new_pdx old_pdx
455 return (Container.add new_pdx new_p $
456 Container.addTwo old_pdx new_s old_sdx int_s nl,
457 new_inst, new_pdx, old_pdx)
460 -- Failver and replace the secondary (f, r:ns)
461 applyMove nl inst (FailoverAndReplace new_sdx) =
462 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
463 tgt_n = Container.find new_sdx nl
464 int_p = Node.removePri old_p inst
465 int_s = Node.removeSec old_s inst
466 force_p = Node.offline old_p
467 new_nl = do -- Maybe monad
468 new_p <- Node.addPriEx force_p int_s inst
469 new_s <- Node.addSecEx force_p tgt_n inst old_sdx
470 let new_inst = Instance.setBoth inst old_sdx new_sdx
471 return (Container.add new_sdx new_s $
472 Container.addTwo old_sdx new_p old_pdx int_p nl,
473 new_inst, old_sdx, new_sdx)
476 -- | Tries to allocate an instance on one given node.
477 allocateOnSingle :: Node.List -> Instance.Instance -> Ndx
478 -> OpResult Node.AllocElement
479 allocateOnSingle nl inst new_pdx =
480 let p = Container.find new_pdx nl
481 new_inst = Instance.setBoth inst new_pdx Node.noSecondary
483 Instance.instMatchesPolicy inst (Node.iPolicy p)
484 new_p <- Node.addPri p inst
485 let new_nl = Container.add new_pdx new_p nl
486 new_score = compCV nl
487 return (new_nl, new_inst, [new_p], new_score)
489 -- | Tries to allocate an instance on a given pair of nodes.
490 allocateOnPair :: Node.List -> Instance.Instance -> Ndx -> Ndx
491 -> OpResult Node.AllocElement
492 allocateOnPair nl inst new_pdx new_sdx =
493 let tgt_p = Container.find new_pdx nl
494 tgt_s = Container.find new_sdx nl
496 Instance.instMatchesPolicy inst (Node.iPolicy tgt_p)
497 new_p <- Node.addPri tgt_p inst
498 new_s <- Node.addSec tgt_s inst new_pdx
499 let new_inst = Instance.setBoth inst new_pdx new_sdx
500 new_nl = Container.addTwo new_pdx new_p new_sdx new_s nl
501 return (new_nl, new_inst, [new_p, new_s], compCV new_nl)
503 -- | Tries to perform an instance move and returns the best table
504 -- between the original one and the new one.
505 checkSingleStep :: Table -- ^ The original table
506 -> Instance.Instance -- ^ The instance to move
507 -> Table -- ^ The current best table
508 -> IMove -- ^ The move to apply
509 -> Table -- ^ The final best table
510 checkSingleStep ini_tbl target cur_tbl move =
511 let Table ini_nl ini_il _ ini_plc = ini_tbl
512 tmp_resu = applyMove ini_nl target move
515 OpGood (upd_nl, new_inst, pri_idx, sec_idx) ->
516 let tgt_idx = Instance.idx target
517 upd_cvar = compCV upd_nl
518 upd_il = Container.add tgt_idx new_inst ini_il
519 upd_plc = (tgt_idx, pri_idx, sec_idx, move, upd_cvar):ini_plc
520 upd_tbl = Table upd_nl upd_il upd_cvar upd_plc
521 in compareTables cur_tbl upd_tbl
523 -- | Given the status of the current secondary as a valid new node and
524 -- the current candidate target node, generate the possible moves for
526 possibleMoves :: MirrorType -- ^ The mirroring type of the instance
527 -> Bool -- ^ Whether the secondary node is a valid new node
528 -> Bool -- ^ Whether we can change the primary node
529 -> Ndx -- ^ Target node candidate
530 -> [IMove] -- ^ List of valid result moves
532 possibleMoves MirrorNone _ _ _ = []
534 possibleMoves MirrorExternal _ False _ = []
536 possibleMoves MirrorExternal _ True tdx =
537 [ FailoverToAny tdx ]
539 possibleMoves MirrorInternal _ False tdx =
540 [ ReplaceSecondary tdx ]
542 possibleMoves MirrorInternal True True tdx =
543 [ ReplaceSecondary tdx
544 , ReplaceAndFailover tdx
546 , FailoverAndReplace tdx
549 possibleMoves MirrorInternal False True tdx =
550 [ ReplaceSecondary tdx
551 , ReplaceAndFailover tdx
554 -- | Compute the best move for a given instance.
555 checkInstanceMove :: [Ndx] -- ^ Allowed target node indices
556 -> Bool -- ^ Whether disk moves are allowed
557 -> Bool -- ^ Whether instance moves are allowed
558 -> Table -- ^ Original table
559 -> Instance.Instance -- ^ Instance to move
560 -> Table -- ^ Best new table for this instance
561 checkInstanceMove nodes_idx disk_moves inst_moves ini_tbl target =
562 let opdx = Instance.pNode target
563 osdx = Instance.sNode target
564 bad_nodes = [opdx, osdx]
565 nodes = filter (`notElem` bad_nodes) nodes_idx
566 mir_type = Instance.mirrorType target
567 use_secondary = elem osdx nodes_idx && inst_moves
568 aft_failover = if mir_type == MirrorInternal && use_secondary
569 -- if drbd and allowed to failover
570 then checkSingleStep ini_tbl target ini_tbl Failover
574 then concatMap (possibleMoves mir_type use_secondary inst_moves)
578 -- iterate over the possible nodes for this instance
579 foldl' (checkSingleStep ini_tbl target) aft_failover all_moves
581 -- | Compute the best next move.
582 checkMove :: [Ndx] -- ^ Allowed target node indices
583 -> Bool -- ^ Whether disk moves are allowed
584 -> Bool -- ^ Whether instance moves are allowed
585 -> Table -- ^ The current solution
586 -> [Instance.Instance] -- ^ List of instances still to move
587 -> Table -- ^ The new solution
588 checkMove nodes_idx disk_moves inst_moves ini_tbl victims =
589 let Table _ _ _ ini_plc = ini_tbl
590 -- we're using rwhnf from the Control.Parallel.Strategies
591 -- package; we don't need to use rnf as that would force too
592 -- much evaluation in single-threaded cases, and in
593 -- multi-threaded case the weak head normal form is enough to
594 -- spark the evaluation
595 tables = parMap rwhnf (checkInstanceMove nodes_idx disk_moves
598 -- iterate over all instances, computing the best move
599 best_tbl = foldl' compareTables ini_tbl tables
600 Table _ _ _ best_plc = best_tbl
601 in if length best_plc == length ini_plc
602 then ini_tbl -- no advancement
605 -- | Check if we are allowed to go deeper in the balancing.
606 doNextBalance :: Table -- ^ The starting table
607 -> Int -- ^ Remaining length
608 -> Score -- ^ Score at which to stop
609 -> Bool -- ^ The resulting table and commands
610 doNextBalance ini_tbl max_rounds min_score =
611 let Table _ _ ini_cv ini_plc = ini_tbl
612 ini_plc_len = length ini_plc
613 in (max_rounds < 0 || ini_plc_len < max_rounds) && ini_cv > min_score
615 -- | Run a balance move.
616 tryBalance :: Table -- ^ The starting table
617 -> Bool -- ^ Allow disk moves
618 -> Bool -- ^ Allow instance moves
619 -> Bool -- ^ Only evacuate moves
620 -> Score -- ^ Min gain threshold
621 -> Score -- ^ Min gain
622 -> Maybe Table -- ^ The resulting table and commands
623 tryBalance ini_tbl disk_moves inst_moves evac_mode mg_limit min_gain =
624 let Table ini_nl ini_il ini_cv _ = ini_tbl
625 all_inst = Container.elems ini_il
626 all_nodes = Container.elems ini_nl
627 (offline_nodes, online_nodes) = partition Node.offline all_nodes
628 all_inst' = if evac_mode
629 then let bad_nodes = map Node.idx offline_nodes
630 in filter (any (`elem` bad_nodes) .
631 Instance.allNodes) all_inst
633 reloc_inst = filter Instance.movable all_inst'
634 node_idx = map Node.idx online_nodes
635 fin_tbl = checkMove node_idx disk_moves inst_moves ini_tbl reloc_inst
636 (Table _ _ fin_cv _) = fin_tbl
638 if fin_cv < ini_cv && (ini_cv > mg_limit || ini_cv - fin_cv >= min_gain)
639 then Just fin_tbl -- this round made success, return the new table
642 -- * Allocation functions
644 -- | Build failure stats out of a list of failures.
645 collapseFailures :: [FailMode] -> FailStats
646 collapseFailures flst =
647 map (\k -> (k, foldl' (\a e -> if e == k then a + 1 else a) 0 flst))
650 -- | Compares two Maybe AllocElement and chooses the besst score.
651 bestAllocElement :: Maybe Node.AllocElement
652 -> Maybe Node.AllocElement
653 -> Maybe Node.AllocElement
654 bestAllocElement a Nothing = a
655 bestAllocElement Nothing b = b
656 bestAllocElement a@(Just (_, _, _, ascore)) b@(Just (_, _, _, bscore)) =
657 if ascore < bscore then a else b
659 -- | Update current Allocation solution and failure stats with new
661 concatAllocs :: AllocSolution -> OpResult Node.AllocElement -> AllocSolution
662 concatAllocs as (OpFail reason) = as { asFailures = reason : asFailures as }
664 concatAllocs as (OpGood ns) =
665 let -- Choose the old or new solution, based on the cluster score
667 osols = asSolution as
668 nsols = bestAllocElement osols (Just ns)
670 -- Note: we force evaluation of nsols here in order to keep the
671 -- memory profile low - we know that we will need nsols for sure
672 -- in the next cycle, so we force evaluation of nsols, since the
673 -- foldl' in the caller will only evaluate the tuple, but not the
674 -- elements of the tuple
675 in nsols `seq` nsuc `seq` as { asAllocs = nsuc, asSolution = nsols }
677 -- | Sums two 'AllocSolution' structures.
678 sumAllocs :: AllocSolution -> AllocSolution -> AllocSolution
679 sumAllocs (AllocSolution aFails aAllocs aSols aLog)
680 (AllocSolution bFails bAllocs bSols bLog) =
681 -- note: we add b first, since usually it will be smaller; when
682 -- fold'ing, a will grow and grow whereas b is the per-group
683 -- result, hence smaller
684 let nFails = bFails ++ aFails
685 nAllocs = aAllocs + bAllocs
686 nSols = bestAllocElement aSols bSols
688 in AllocSolution nFails nAllocs nSols nLog
690 -- | Given a solution, generates a reasonable description for it.
691 describeSolution :: AllocSolution -> String
692 describeSolution as =
693 let fcnt = asFailures as
696 intercalate ", " . map (\(a, b) -> printf "%s: %d" (show a) b) .
697 filter ((> 0) . snd) . collapseFailures $ fcnt
699 Nothing -> "No valid allocation solutions, failure reasons: " ++
700 (if null fcnt then "unknown reasons" else freasons)
701 Just (_, _, nodes, cv) ->
702 printf ("score: %.8f, successes %d, failures %d (%s)" ++
703 " for node(s) %s") cv (asAllocs as) (length fcnt) freasons
704 (intercalate "/" . map Node.name $ nodes)
706 -- | Annotates a solution with the appropriate string.
707 annotateSolution :: AllocSolution -> AllocSolution
708 annotateSolution as = as { asLog = describeSolution as : asLog as }
710 -- | Reverses an evacuation solution.
712 -- Rationale: we always concat the results to the top of the lists, so
713 -- for proper jobset execution, we should reverse all lists.
714 reverseEvacSolution :: EvacSolution -> EvacSolution
715 reverseEvacSolution (EvacSolution f m o) =
716 EvacSolution (reverse f) (reverse m) (reverse o)
718 -- | Generate the valid node allocation singles or pairs for a new instance.
719 genAllocNodes :: Group.List -- ^ Group list
720 -> Node.List -- ^ The node map
721 -> Int -- ^ The number of nodes required
722 -> Bool -- ^ Whether to drop or not
724 -> Result AllocNodes -- ^ The (monadic) result
725 genAllocNodes gl nl count drop_unalloc =
726 let filter_fn = if drop_unalloc
727 then filter (Group.isAllocable .
728 flip Container.find gl . Node.group)
730 all_nodes = filter_fn $ getOnline nl
731 all_pairs = [(Node.idx p,
732 [Node.idx s | s <- all_nodes,
733 Node.idx p /= Node.idx s,
734 Node.group p == Node.group s]) |
737 1 -> Ok (Left (map Node.idx all_nodes))
738 2 -> Ok (Right (filter (not . null . snd) all_pairs))
739 _ -> Bad "Unsupported number of nodes, only one or two supported"
741 -- | Try to allocate an instance on the cluster.
742 tryAlloc :: (Monad m) =>
743 Node.List -- ^ The node list
744 -> Instance.List -- ^ The instance list
745 -> Instance.Instance -- ^ The instance to allocate
746 -> AllocNodes -- ^ The allocation targets
747 -> m AllocSolution -- ^ Possible solution list
748 tryAlloc _ _ _ (Right []) = fail "Not enough online nodes"
749 tryAlloc nl _ inst (Right ok_pairs) =
750 let psols = parMap rwhnf (\(p, ss) ->
752 concatAllocs cstate .
753 allocateOnPair nl inst p)
754 emptyAllocSolution ss) ok_pairs
755 sols = foldl' sumAllocs emptyAllocSolution psols
756 in return $ annotateSolution sols
758 tryAlloc _ _ _ (Left []) = fail "No online nodes"
759 tryAlloc nl _ inst (Left all_nodes) =
760 let sols = foldl' (\cstate ->
761 concatAllocs cstate . allocateOnSingle nl inst
762 ) emptyAllocSolution all_nodes
763 in return $ annotateSolution sols
765 -- | Given a group/result, describe it as a nice (list of) messages.
766 solutionDescription :: Group.List -> (Gdx, Result AllocSolution) -> [String]
767 solutionDescription gl (groupId, result) =
769 Ok solution -> map (printf "Group %s (%s): %s" gname pol) (asLog solution)
770 Bad message -> [printf "Group %s: error %s" gname message]
771 where grp = Container.find groupId gl
772 gname = Group.name grp
773 pol = allocPolicyToRaw (Group.allocPolicy grp)
775 -- | From a list of possibly bad and possibly empty solutions, filter
776 -- only the groups with a valid result. Note that the result will be
777 -- reversed compared to the original list.
778 filterMGResults :: Group.List
779 -> [(Gdx, Result AllocSolution)]
780 -> [(Gdx, AllocSolution)]
781 filterMGResults gl = foldl' fn []
782 where unallocable = not . Group.isAllocable . flip Container.find gl
783 fn accu (gdx, rasol) =
786 Ok sol | isNothing (asSolution sol) -> accu
787 | unallocable gdx -> accu
788 | otherwise -> (gdx, sol):accu
790 -- | Sort multigroup results based on policy and score.
791 sortMGResults :: Group.List
792 -> [(Gdx, AllocSolution)]
793 -> [(Gdx, AllocSolution)]
794 sortMGResults gl sols =
795 let extractScore (_, _, _, x) = x
796 solScore (gdx, sol) = (Group.allocPolicy (Container.find gdx gl),
797 (extractScore . fromJust . asSolution) sol)
798 in sortBy (comparing solScore) sols
800 -- | Finds the best group for an instance on a multi-group cluster.
802 -- Only solutions in @preferred@ and @last_resort@ groups will be
803 -- accepted as valid, and additionally if the allowed groups parameter
804 -- is not null then allocation will only be run for those group
806 findBestAllocGroup :: Group.List -- ^ The group list
807 -> Node.List -- ^ The node list
808 -> Instance.List -- ^ The instance list
809 -> Maybe [Gdx] -- ^ The allowed groups
810 -> Instance.Instance -- ^ The instance to allocate
811 -> Int -- ^ Required number of nodes
812 -> Result (Gdx, AllocSolution, [String])
813 findBestAllocGroup mggl mgnl mgil allowed_gdxs inst cnt =
814 let groups = splitCluster mgnl mgil
815 groups' = maybe groups (\gs -> filter ((`elem` gs) . fst) groups)
817 sols = map (\(gid, (nl, il)) ->
818 (gid, genAllocNodes mggl nl cnt False >>=
819 tryAlloc nl il inst))
820 groups'::[(Gdx, Result AllocSolution)]
821 all_msgs = concatMap (solutionDescription mggl) sols
822 goodSols = filterMGResults mggl sols
823 sortedSols = sortMGResults mggl goodSols
824 in if null sortedSols
826 then Bad $ "no groups for evacuation: allowed groups was" ++
827 show allowed_gdxs ++ ", all groups: " ++
828 show (map fst groups)
829 else Bad $ intercalate ", " all_msgs
830 else let (final_group, final_sol) = head sortedSols
831 in return (final_group, final_sol, all_msgs)
833 -- | Try to allocate an instance on a multi-group cluster.
834 tryMGAlloc :: Group.List -- ^ The group list
835 -> Node.List -- ^ The node list
836 -> Instance.List -- ^ The instance list
837 -> Instance.Instance -- ^ The instance to allocate
838 -> Int -- ^ Required number of nodes
839 -> Result AllocSolution -- ^ Possible solution list
840 tryMGAlloc mggl mgnl mgil inst cnt = do
841 (best_group, solution, all_msgs) <-
842 findBestAllocGroup mggl mgnl mgil Nothing inst cnt
843 let group_name = Group.name $ Container.find best_group mggl
844 selmsg = "Selected group: " ++ group_name
845 return $ solution { asLog = selmsg:all_msgs }
847 -- | Function which fails if the requested mode is change secondary.
849 -- This is useful since except DRBD, no other disk template can
850 -- execute change secondary; thus, we can just call this function
851 -- instead of always checking for secondary mode. After the call to
852 -- this function, whatever mode we have is just a primary change.
853 failOnSecondaryChange :: (Monad m) => EvacMode -> DiskTemplate -> m ()
854 failOnSecondaryChange ChangeSecondary dt =
855 fail $ "Instances with disk template '" ++ diskTemplateToRaw dt ++
856 "' can't execute change secondary"
857 failOnSecondaryChange _ _ = return ()
859 -- | Run evacuation for a single instance.
861 -- /Note:/ this function should correctly execute both intra-group
862 -- evacuations (in all modes) and inter-group evacuations (in the
863 -- 'ChangeAll' mode). Of course, this requires that the correct list
864 -- of target nodes is passed.
865 nodeEvacInstance :: Node.List -- ^ The node list (cluster-wide)
866 -> Instance.List -- ^ Instance list (cluster-wide)
867 -> EvacMode -- ^ The evacuation mode
868 -> Instance.Instance -- ^ The instance to be evacuated
869 -> Gdx -- ^ The group we're targetting
870 -> [Ndx] -- ^ The list of available nodes
872 -> Result (Node.List, Instance.List, [OpCodes.OpCode])
873 nodeEvacInstance nl il mode inst@(Instance.Instance
874 {Instance.diskTemplate = dt@DTDiskless})
876 failOnSecondaryChange mode dt >>
877 evacOneNodeOnly nl il inst gdx avail_nodes
879 nodeEvacInstance _ _ _ (Instance.Instance
880 {Instance.diskTemplate = DTPlain}) _ _ =
881 fail "Instances of type plain cannot be relocated"
883 nodeEvacInstance _ _ _ (Instance.Instance
884 {Instance.diskTemplate = DTFile}) _ _ =
885 fail "Instances of type file cannot be relocated"
887 nodeEvacInstance nl il mode inst@(Instance.Instance
888 {Instance.diskTemplate = dt@DTSharedFile})
890 failOnSecondaryChange mode dt >>
891 evacOneNodeOnly nl il inst gdx avail_nodes
893 nodeEvacInstance nl il mode inst@(Instance.Instance
894 {Instance.diskTemplate = dt@DTBlock})
896 failOnSecondaryChange mode dt >>
897 evacOneNodeOnly nl il inst gdx avail_nodes
899 nodeEvacInstance nl il mode inst@(Instance.Instance
900 {Instance.diskTemplate = dt@DTRbd})
902 failOnSecondaryChange mode dt >>
903 evacOneNodeOnly nl il inst gdx avail_nodes
905 nodeEvacInstance nl il ChangePrimary
906 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})
909 (nl', inst', _, _) <- opToResult $ applyMove nl inst Failover
910 let idx = Instance.idx inst
911 il' = Container.add idx inst' il
912 ops = iMoveToJob nl' il' idx Failover
913 return (nl', il', ops)
915 nodeEvacInstance nl il ChangeSecondary
916 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})
918 evacOneNodeOnly nl il inst gdx avail_nodes
920 -- The algorithm for ChangeAll is as follows:
922 -- * generate all (primary, secondary) node pairs for the target groups
923 -- * for each pair, execute the needed moves (r:s, f, r:s) and compute
924 -- the final node list state and group score
925 -- * select the best choice via a foldl that uses the same Either
926 -- String solution as the ChangeSecondary mode
927 nodeEvacInstance nl il ChangeAll
928 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})
931 let no_nodes = Left "no nodes available"
932 node_pairs = [(p,s) | p <- avail_nodes, s <- avail_nodes, p /= s]
933 (nl', il', ops, _) <-
934 annotateResult "Can't find any good nodes for relocation" $
937 (\accu nodes -> case evacDrbdAllInner nl il inst gdx nodes of
941 -- we don't need more details (which
942 -- nodes, etc.) as we only selected
943 -- this group if we can allocate on
944 -- it, hence failures will not
945 -- propagate out of this fold loop
946 Left _ -> Left $ "Allocation failed: " ++ msg
947 Ok result@(_, _, _, new_cv) ->
948 let new_accu = Right result in
951 Right (_, _, _, old_cv) ->
955 ) no_nodes node_pairs
957 return (nl', il', ops)
959 -- | Generic function for changing one node of an instance.
961 -- This is similar to 'nodeEvacInstance' but will be used in a few of
962 -- its sub-patterns. It folds the inner function 'evacOneNodeInner'
963 -- over the list of available nodes, which results in the best choice
965 evacOneNodeOnly :: Node.List -- ^ The node list (cluster-wide)
966 -> Instance.List -- ^ Instance list (cluster-wide)
967 -> Instance.Instance -- ^ The instance to be evacuated
968 -> Gdx -- ^ The group we're targetting
969 -> [Ndx] -- ^ The list of available nodes
971 -> Result (Node.List, Instance.List, [OpCodes.OpCode])
972 evacOneNodeOnly nl il inst gdx avail_nodes = do
973 op_fn <- case Instance.mirrorType inst of
974 MirrorNone -> Bad "Can't relocate/evacuate non-mirrored instances"
975 MirrorInternal -> Ok ReplaceSecondary
976 MirrorExternal -> Ok FailoverToAny
977 (nl', inst', _, ndx) <- annotateResult "Can't find any good node" $
979 foldl' (evacOneNodeInner nl inst gdx op_fn)
980 (Left "no nodes available") avail_nodes
981 let idx = Instance.idx inst
982 il' = Container.add idx inst' il
983 ops = iMoveToJob nl' il' idx (op_fn ndx)
984 return (nl', il', ops)
986 -- | Inner fold function for changing one node of an instance.
988 -- Depending on the instance disk template, this will either change
989 -- the secondary (for DRBD) or the primary node (for shared
990 -- storage). However, the operation is generic otherwise.
992 -- The running solution is either a @Left String@, which means we
993 -- don't have yet a working solution, or a @Right (...)@, which
994 -- represents a valid solution; it holds the modified node list, the
995 -- modified instance (after evacuation), the score of that solution,
996 -- and the new secondary node index.
997 evacOneNodeInner :: Node.List -- ^ Cluster node list
998 -> Instance.Instance -- ^ Instance being evacuated
999 -> Gdx -- ^ The group index of the instance
1000 -> (Ndx -> IMove) -- ^ Operation constructor
1001 -> EvacInnerState -- ^ Current best solution
1002 -> Ndx -- ^ Node we're evaluating as target
1003 -> EvacInnerState -- ^ New best solution
1004 evacOneNodeInner nl inst gdx op_fn accu ndx =
1005 case applyMove nl inst (op_fn ndx) of
1006 OpFail fm -> let fail_msg = "Node " ++ Container.nameOf nl ndx ++
1007 " failed: " ++ show fm
1008 in either (const $ Left fail_msg) (const accu) accu
1009 OpGood (nl', inst', _, _) ->
1010 let nodes = Container.elems nl'
1011 -- The fromJust below is ugly (it can fail nastily), but
1012 -- at this point we should have any internal mismatches,
1013 -- and adding a monad here would be quite involved
1014 grpnodes = fromJust (gdx `lookup` Node.computeGroups nodes)
1015 new_cv = compCVNodes grpnodes
1016 new_accu = Right (nl', inst', new_cv, ndx)
1019 Right (_, _, old_cv, _) ->
1024 -- | Compute result of changing all nodes of a DRBD instance.
1026 -- Given the target primary and secondary node (which might be in a
1027 -- different group or not), this function will 'execute' all the
1028 -- required steps and assuming all operations succceed, will return
1029 -- the modified node and instance lists, the opcodes needed for this
1030 -- and the new group score.
1031 evacDrbdAllInner :: Node.List -- ^ Cluster node list
1032 -> Instance.List -- ^ Cluster instance list
1033 -> Instance.Instance -- ^ The instance to be moved
1034 -> Gdx -- ^ The target group index
1035 -- (which can differ from the
1036 -- current group of the
1038 -> (Ndx, Ndx) -- ^ Tuple of new
1039 -- primary\/secondary nodes
1040 -> Result (Node.List, Instance.List, [OpCodes.OpCode], Score)
1041 evacDrbdAllInner nl il inst gdx (t_pdx, t_sdx) = do
1042 let primary = Container.find (Instance.pNode inst) nl
1043 idx = Instance.idx inst
1044 -- if the primary is offline, then we first failover
1045 (nl1, inst1, ops1) <-
1046 if Node.offline primary
1048 (nl', inst', _, _) <-
1049 annotateResult "Failing over to the secondary" $
1050 opToResult $ applyMove nl inst Failover
1051 return (nl', inst', [Failover])
1052 else return (nl, inst, [])
1053 let (o1, o2, o3) = (ReplaceSecondary t_pdx,
1055 ReplaceSecondary t_sdx)
1056 -- we now need to execute a replace secondary to the future
1058 (nl2, inst2, _, _) <-
1059 annotateResult "Changing secondary to new primary" $
1061 applyMove nl1 inst1 o1
1063 -- we now execute another failover, the primary stays fixed now
1064 (nl3, inst3, _, _) <- annotateResult "Failing over to new primary" $
1065 opToResult $ applyMove nl2 inst2 o2
1067 -- and finally another replace secondary, to the final secondary
1068 (nl4, inst4, _, _) <-
1069 annotateResult "Changing secondary to final secondary" $
1071 applyMove nl3 inst3 o3
1073 il' = Container.add idx inst4 il
1074 ops = concatMap (iMoveToJob nl4 il' idx) $ reverse ops4
1075 let nodes = Container.elems nl4
1076 -- The fromJust below is ugly (it can fail nastily), but
1077 -- at this point we should have any internal mismatches,
1078 -- and adding a monad here would be quite involved
1079 grpnodes = fromJust (gdx `lookup` Node.computeGroups nodes)
1080 new_cv = compCVNodes grpnodes
1081 return (nl4, il', ops, new_cv)
1083 -- | Computes the nodes in a given group which are available for
1085 availableGroupNodes :: [(Gdx, [Ndx])] -- ^ Group index/node index assoc list
1086 -> IntSet.IntSet -- ^ Nodes that are excluded
1087 -> Gdx -- ^ The group for which we
1089 -> Result [Ndx] -- ^ List of available node indices
1090 availableGroupNodes group_nodes excl_ndx gdx = do
1091 local_nodes <- maybe (Bad $ "Can't find group with index " ++ show gdx)
1092 Ok (lookup gdx group_nodes)
1093 let avail_nodes = filter (not . flip IntSet.member excl_ndx) local_nodes
1096 -- | Updates the evac solution with the results of an instance
1098 updateEvacSolution :: (Node.List, Instance.List, EvacSolution)
1100 -> Result (Node.List, Instance.List, [OpCodes.OpCode])
1101 -> (Node.List, Instance.List, EvacSolution)
1102 updateEvacSolution (nl, il, es) idx (Bad msg) =
1103 (nl, il, es { esFailed = (idx, msg):esFailed es})
1104 updateEvacSolution (_, _, es) idx (Ok (nl, il, opcodes)) =
1105 (nl, il, es { esMoved = new_elem:esMoved es
1106 , esOpCodes = opcodes:esOpCodes es })
1107 where inst = Container.find idx il
1109 instancePriGroup nl inst,
1110 Instance.allNodes inst)
1112 -- | Node-evacuation IAllocator mode main function.
1113 tryNodeEvac :: Group.List -- ^ The cluster groups
1114 -> Node.List -- ^ The node list (cluster-wide, not per group)
1115 -> Instance.List -- ^ Instance list (cluster-wide)
1116 -> EvacMode -- ^ The evacuation mode
1117 -> [Idx] -- ^ List of instance (indices) to be evacuated
1118 -> Result (Node.List, Instance.List, EvacSolution)
1119 tryNodeEvac _ ini_nl ini_il mode idxs =
1120 let evac_ndx = nodesToEvacuate ini_il mode idxs
1121 offline = map Node.idx . filter Node.offline $ Container.elems ini_nl
1122 excl_ndx = foldl' (flip IntSet.insert) evac_ndx offline
1123 group_ndx = map (\(gdx, (nl, _)) -> (gdx, map Node.idx
1124 (Container.elems nl))) $
1125 splitCluster ini_nl ini_il
1126 (fin_nl, fin_il, esol) =
1127 foldl' (\state@(nl, il, _) inst ->
1128 let gdx = instancePriGroup nl inst
1129 pdx = Instance.pNode inst in
1130 updateEvacSolution state (Instance.idx inst) $
1131 availableGroupNodes group_ndx
1132 (IntSet.insert pdx excl_ndx) gdx >>=
1133 nodeEvacInstance nl il mode inst gdx
1135 (ini_nl, ini_il, emptyEvacSolution)
1136 (map (`Container.find` ini_il) idxs)
1137 in return (fin_nl, fin_il, reverseEvacSolution esol)
1139 -- | Change-group IAllocator mode main function.
1141 -- This is very similar to 'tryNodeEvac', the only difference is that
1142 -- we don't choose as target group the current instance group, but
1145 -- 1. at the start of the function, we compute which are the target
1146 -- groups; either no groups were passed in, in which case we choose
1147 -- all groups out of which we don't evacuate instance, or there were
1148 -- some groups passed, in which case we use those
1150 -- 2. for each instance, we use 'findBestAllocGroup' to choose the
1151 -- best group to hold the instance, and then we do what
1152 -- 'tryNodeEvac' does, except for this group instead of the current
1155 -- Note that the correct behaviour of this function relies on the
1156 -- function 'nodeEvacInstance' to be able to do correctly both
1157 -- intra-group and inter-group moves when passed the 'ChangeAll' mode.
1158 tryChangeGroup :: Group.List -- ^ The cluster groups
1159 -> Node.List -- ^ The node list (cluster-wide)
1160 -> Instance.List -- ^ Instance list (cluster-wide)
1161 -> [Gdx] -- ^ Target groups; if empty, any
1162 -- groups not being evacuated
1163 -> [Idx] -- ^ List of instance (indices) to be evacuated
1164 -> Result (Node.List, Instance.List, EvacSolution)
1165 tryChangeGroup gl ini_nl ini_il gdxs idxs =
1166 let evac_gdxs = nub $ map (instancePriGroup ini_nl .
1167 flip Container.find ini_il) idxs
1168 target_gdxs = (if null gdxs
1169 then Container.keys gl
1170 else gdxs) \\ evac_gdxs
1171 offline = map Node.idx . filter Node.offline $ Container.elems ini_nl
1172 excl_ndx = foldl' (flip IntSet.insert) IntSet.empty offline
1173 group_ndx = map (\(gdx, (nl, _)) -> (gdx, map Node.idx
1174 (Container.elems nl))) $
1175 splitCluster ini_nl ini_il
1176 (fin_nl, fin_il, esol) =
1177 foldl' (\state@(nl, il, _) inst ->
1179 let ncnt = Instance.requiredNodes $
1180 Instance.diskTemplate inst
1181 (gdx, _, _) <- findBestAllocGroup gl nl il
1182 (Just target_gdxs) inst ncnt
1183 av_nodes <- availableGroupNodes group_ndx
1185 nodeEvacInstance nl il ChangeAll inst gdx av_nodes
1186 in updateEvacSolution state (Instance.idx inst) solution
1188 (ini_nl, ini_il, emptyEvacSolution)
1189 (map (`Container.find` ini_il) idxs)
1190 in return (fin_nl, fin_il, reverseEvacSolution esol)
1192 -- | Standard-sized allocation method.
1194 -- This places instances of the same size on the cluster until we're
1195 -- out of space. The result will be a list of identically-sized
1197 iterateAlloc :: AllocMethod
1198 iterateAlloc nl il limit newinst allocnodes ixes cstats =
1199 let depth = length ixes
1200 newname = printf "new-%d" depth::String
1201 newidx = Container.size il
1202 newi2 = Instance.setIdx (Instance.setName newinst newname) newidx
1203 newlimit = fmap (flip (-) 1) limit
1204 in case tryAlloc nl il newi2 allocnodes of
1206 Ok (AllocSolution { asFailures = errs, asSolution = sols3 }) ->
1207 let newsol = Ok (collapseFailures errs, nl, il, ixes, cstats) in
1210 Just (xnl, xi, _, _) ->
1213 else iterateAlloc xnl (Container.add newidx xi il)
1214 newlimit newinst allocnodes (xi:ixes)
1215 (totalResources xnl:cstats)
1217 -- | Tiered allocation method.
1219 -- This places instances on the cluster, and decreases the spec until
1220 -- we can allocate again. The result will be a list of decreasing
1222 tieredAlloc :: AllocMethod
1223 tieredAlloc nl il limit newinst allocnodes ixes cstats =
1224 case iterateAlloc nl il limit newinst allocnodes ixes cstats of
1226 Ok (errs, nl', il', ixes', cstats') ->
1227 let newsol = Ok (errs, nl', il', ixes', cstats')
1228 ixes_cnt = length ixes'
1229 (stop, newlimit) = case limit of
1230 Nothing -> (False, Nothing)
1231 Just n -> (n <= ixes_cnt,
1232 Just (n - ixes_cnt)) in
1233 if stop then newsol else
1234 case Instance.shrinkByType newinst . fst . last $
1235 sortBy (comparing snd) errs of
1237 Ok newinst' -> tieredAlloc nl' il' newlimit
1238 newinst' allocnodes ixes' cstats'
1240 -- * Formatting functions
1242 -- | Given the original and final nodes, computes the relocation description.
1243 computeMoves :: Instance.Instance -- ^ The instance to be moved
1244 -> String -- ^ The instance name
1245 -> IMove -- ^ The move being performed
1246 -> String -- ^ New primary
1247 -> String -- ^ New secondary
1248 -> (String, [String])
1249 -- ^ Tuple of moves and commands list; moves is containing
1250 -- either @/f/@ for failover or @/r:name/@ for replace
1251 -- secondary, while the command list holds gnt-instance
1252 -- commands (without that prefix), e.g \"@failover instance1@\"
1253 computeMoves i inam mv c d =
1255 Failover -> ("f", [mig])
1256 FailoverToAny _ -> (printf "fa:%s" c, [mig_any])
1257 FailoverAndReplace _ -> (printf "f r:%s" d, [mig, rep d])
1258 ReplaceSecondary _ -> (printf "r:%s" d, [rep d])
1259 ReplaceAndFailover _ -> (printf "r:%s f" c, [rep c, mig])
1260 ReplacePrimary _ -> (printf "f r:%s f" c, [mig, rep c, mig])
1261 where morf = if Instance.isRunning i then "migrate" else "failover"
1262 mig = printf "%s -f %s" morf inam::String
1263 mig_any = printf "%s -f -n %s %s" morf c inam::String
1264 rep n = printf "replace-disks -n %s %s" n inam::String
1266 -- | Converts a placement to string format.
1267 printSolutionLine :: Node.List -- ^ The node list
1268 -> Instance.List -- ^ The instance list
1269 -> Int -- ^ Maximum node name length
1270 -> Int -- ^ Maximum instance name length
1271 -> Placement -- ^ The current placement
1272 -> Int -- ^ The index of the placement in
1274 -> (String, [String])
1275 printSolutionLine nl il nmlen imlen plc pos =
1276 let pmlen = (2*nmlen + 1)
1277 (i, p, s, mv, c) = plc
1278 old_sec = Instance.sNode inst
1279 inst = Container.find i il
1280 inam = Instance.alias inst
1281 npri = Node.alias $ Container.find p nl
1282 nsec = Node.alias $ Container.find s nl
1283 opri = Node.alias $ Container.find (Instance.pNode inst) nl
1284 osec = Node.alias $ Container.find old_sec nl
1285 (moves, cmds) = computeMoves inst inam mv npri nsec
1286 -- FIXME: this should check instead/also the disk template
1287 ostr = if old_sec == Node.noSecondary
1288 then printf "%s" opri::String
1289 else printf "%s:%s" opri osec::String
1290 nstr = if s == Node.noSecondary
1291 then printf "%s" npri::String
1292 else printf "%s:%s" npri nsec::String
1293 in (printf " %3d. %-*s %-*s => %-*s %12.8f a=%s"
1294 pos imlen inam pmlen ostr pmlen nstr c moves,
1297 -- | Return the instance and involved nodes in an instance move.
1299 -- Note that the output list length can vary, and is not required nor
1300 -- guaranteed to be of any specific length.
1301 involvedNodes :: Instance.List -- ^ Instance list, used for retrieving
1302 -- the instance from its index; note
1303 -- that this /must/ be the original
1304 -- instance list, so that we can
1305 -- retrieve the old nodes
1306 -> Placement -- ^ The placement we're investigating,
1307 -- containing the new nodes and
1309 -> [Ndx] -- ^ Resulting list of node indices
1310 involvedNodes il plc =
1311 let (i, np, ns, _, _) = plc
1312 inst = Container.find i il
1313 in nub $ [np, ns] ++ Instance.allNodes inst
1315 -- | Inner function for splitJobs, that either appends the next job to
1316 -- the current jobset, or starts a new jobset.
1317 mergeJobs :: ([JobSet], [Ndx]) -> MoveJob -> ([JobSet], [Ndx])
1318 mergeJobs ([], _) n@(ndx, _, _, _) = ([[n]], ndx)
1319 mergeJobs (cjs@(j:js), nbuf) n@(ndx, _, _, _)
1320 | null (ndx `intersect` nbuf) = ((n:j):js, ndx ++ nbuf)
1321 | otherwise = ([n]:cjs, ndx)
1323 -- | Break a list of moves into independent groups. Note that this
1324 -- will reverse the order of jobs.
1325 splitJobs :: [MoveJob] -> [JobSet]
1326 splitJobs = fst . foldl mergeJobs ([], [])
1328 -- | Given a list of commands, prefix them with @gnt-instance@ and
1329 -- also beautify the display a little.
1330 formatJob :: Int -> Int -> (Int, MoveJob) -> [String]
1331 formatJob jsn jsl (sn, (_, _, _, cmds)) =
1333 printf " echo job %d/%d" jsn sn:
1335 map (" gnt-instance " ++) cmds
1337 then ["", printf "echo jobset %d, %d jobs" jsn jsl] ++ out
1340 -- | Given a list of commands, prefix them with @gnt-instance@ and
1341 -- also beautify the display a little.
1342 formatCmds :: [JobSet] -> String
1345 concatMap (\(jsn, js) -> concatMap (formatJob jsn (length js))
1349 -- | Print the node list.
1350 printNodes :: Node.List -> [String] -> String
1352 let fields = case fs of
1353 [] -> Node.defaultFields
1354 "+":rest -> Node.defaultFields ++ rest
1356 snl = sortBy (comparing Node.idx) (Container.elems nl)
1357 (header, isnum) = unzip $ map Node.showHeader fields
1358 in printTable "" header (map (Node.list fields) snl) isnum
1360 -- | Print the instance list.
1361 printInsts :: Node.List -> Instance.List -> String
1363 let sil = sortBy (comparing Instance.idx) (Container.elems il)
1364 helper inst = [ if Instance.isRunning inst then "R" else " "
1365 , Instance.name inst
1366 , Container.nameOf nl (Instance.pNode inst)
1367 , let sdx = Instance.sNode inst
1368 in if sdx == Node.noSecondary
1370 else Container.nameOf nl sdx
1371 , if Instance.autoBalance inst then "Y" else "N"
1372 , printf "%3d" $ Instance.vcpus inst
1373 , printf "%5d" $ Instance.mem inst
1374 , printf "%5d" $ Instance.dsk inst `div` 1024
1380 where DynUtil lC lM lD lN = Instance.util inst
1381 header = [ "F", "Name", "Pri_node", "Sec_node", "Auto_bal"
1382 , "vcpu", "mem" , "dsk", "lCpu", "lMem", "lDsk", "lNet" ]
1383 isnum = False:False:False:False:False:repeat True
1384 in printTable "" header (map helper sil) isnum
1386 -- | Shows statistics for a given node list.
1387 printStats :: String -> Node.List -> String
1389 let dcvs = compDetailedCV $ Container.elems nl
1390 (weights, names) = unzip detailedCVInfo
1391 hd = zip3 (weights ++ repeat 1) (names ++ repeat "unknown") dcvs
1392 header = [ "Field", "Value", "Weight" ]
1393 formatted = map (\(w, h, val) ->
1398 in printTable lp header formatted $ False:repeat True
1400 -- | Convert a placement into a list of OpCodes (basically a job).
1401 iMoveToJob :: Node.List -- ^ The node list; only used for node
1402 -- names, so any version is good
1403 -- (before or after the operation)
1404 -> Instance.List -- ^ The instance list; also used for
1406 -> Idx -- ^ The index of the instance being
1408 -> IMove -- ^ The actual move to be described
1409 -> [OpCodes.OpCode] -- ^ The list of opcodes equivalent to
1411 iMoveToJob nl il idx move =
1412 let inst = Container.find idx il
1413 iname = Instance.name inst
1414 lookNode = Just . Container.nameOf nl
1415 opF = OpCodes.OpInstanceMigrate iname True False True Nothing
1416 opFA n = OpCodes.OpInstanceMigrate iname True False True (lookNode n)
1417 opR n = OpCodes.OpInstanceReplaceDisks iname (lookNode n)
1418 OpCodes.ReplaceNewSecondary [] Nothing
1421 FailoverToAny np -> [ opFA np ]
1422 ReplacePrimary np -> [ opF, opR np, opF ]
1423 ReplaceSecondary ns -> [ opR ns ]
1424 ReplaceAndFailover np -> [ opR np, opF ]
1425 FailoverAndReplace ns -> [ opF, opR ns ]
1427 -- * Node group functions
1429 -- | Computes the group of an instance.
1430 instanceGroup :: Node.List -> Instance.Instance -> Result Gdx
1431 instanceGroup nl i =
1432 let sidx = Instance.sNode i
1433 pnode = Container.find (Instance.pNode i) nl
1434 snode = if sidx == Node.noSecondary
1436 else Container.find sidx nl
1437 pgroup = Node.group pnode
1438 sgroup = Node.group snode
1439 in if pgroup /= sgroup
1440 then fail ("Instance placed accross two node groups, primary " ++
1441 show pgroup ++ ", secondary " ++ show sgroup)
1444 -- | Computes the group of an instance per the primary node.
1445 instancePriGroup :: Node.List -> Instance.Instance -> Gdx
1446 instancePriGroup nl i =
1447 let pnode = Container.find (Instance.pNode i) nl
1450 -- | Compute the list of badly allocated instances (split across node
1452 findSplitInstances :: Node.List -> Instance.List -> [Instance.Instance]
1453 findSplitInstances nl =
1454 filter (not . isOk . instanceGroup nl) . Container.elems
1456 -- | Splits a cluster into the component node groups.
1457 splitCluster :: Node.List -> Instance.List ->
1458 [(Gdx, (Node.List, Instance.List))]
1459 splitCluster nl il =
1460 let ngroups = Node.computeGroups (Container.elems nl)
1461 in map (\(guuid, nodes) ->
1462 let nidxs = map Node.idx nodes
1463 nodes' = zip nidxs nodes
1464 instances = Container.filter ((`elem` nidxs) . Instance.pNode) il
1465 in (guuid, (Container.fromList nodes', instances))) ngroups
1467 -- | Compute the list of nodes that are to be evacuated, given a list
1468 -- of instances and an evacuation mode.
1469 nodesToEvacuate :: Instance.List -- ^ The cluster-wide instance list
1470 -> EvacMode -- ^ The evacuation mode we're using
1471 -> [Idx] -- ^ List of instance indices being evacuated
1472 -> IntSet.IntSet -- ^ Set of node indices
1473 nodesToEvacuate il mode =
1474 IntSet.delete Node.noSecondary .
1476 let i = Container.find idx il
1477 pdx = Instance.pNode i
1478 sdx = Instance.sNode i
1479 dt = Instance.diskTemplate i
1480 withSecondary = case dt of
1481 DTDrbd8 -> IntSet.insert sdx ns
1484 ChangePrimary -> IntSet.insert pdx ns
1485 ChangeSecondary -> withSecondary
1486 ChangeAll -> IntSet.insert pdx withSecondary