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, 2013 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
40 -- * Generic functions
42 , computeAllocationDelta
43 -- * First phase functions
45 -- * Second phase functions
50 -- * Display functions
53 -- * Balacing functions
62 -- * IAllocator functions
70 -- * Allocation functions
73 -- * Node group functions
79 import qualified Data.IntSet as IntSet
81 import Data.Maybe (fromJust, isNothing)
82 import Data.Ord (comparing)
83 import Text.Printf (printf)
85 import Ganeti.BasicTypes
86 import qualified Ganeti.HTools.Container as Container
87 import qualified Ganeti.HTools.Instance as Instance
88 import qualified Ganeti.HTools.Node as Node
89 import qualified Ganeti.HTools.Group as Group
90 import Ganeti.HTools.Types
92 import qualified Ganeti.OpCodes as OpCodes
94 import Ganeti.Types (mkNonEmpty)
98 -- | Allocation\/relocation solution.
99 data AllocSolution = AllocSolution
100 { asFailures :: [FailMode] -- ^ Failure counts
101 , asAllocs :: Int -- ^ Good allocation count
102 , asSolution :: Maybe Node.AllocElement -- ^ The actual allocation result
103 , asLog :: [String] -- ^ Informational messages
106 -- | Node evacuation/group change iallocator result type. This result
107 -- type consists of actual opcodes (a restricted subset) that are
108 -- transmitted back to Ganeti.
109 data EvacSolution = EvacSolution
110 { esMoved :: [(Idx, Gdx, [Ndx])] -- ^ Instances moved successfully
111 , esFailed :: [(Idx, String)] -- ^ Instances which were not
113 , esOpCodes :: [[OpCodes.OpCode]] -- ^ List of jobs
116 -- | Allocation results, as used in 'iterateAlloc' and 'tieredAlloc'.
117 type AllocResult = (FailStats, Node.List, Instance.List,
118 [Instance.Instance], [CStats])
120 -- | Type alias for easier handling.
121 type AllocSolutionList = [(Instance.Instance, AllocSolution)]
123 -- | A type denoting the valid allocation mode/pairs.
125 -- For a one-node allocation, this will be a @Left ['Ndx']@, whereas
126 -- for a two-node allocation, this will be a @Right [('Ndx',
127 -- ['Ndx'])]@. In the latter case, the list is basically an
128 -- association list, grouped by primary node and holding the potential
129 -- secondary nodes in the sub-list.
130 type AllocNodes = Either [Ndx] [(Ndx, [Ndx])]
132 -- | The empty solution we start with when computing allocations.
133 emptyAllocSolution :: AllocSolution
134 emptyAllocSolution = AllocSolution { asFailures = [], asAllocs = 0
135 , asSolution = Nothing, asLog = [] }
137 -- | The empty evac solution.
138 emptyEvacSolution :: EvacSolution
139 emptyEvacSolution = EvacSolution { esMoved = []
144 -- | The complete state for the balancing solution.
145 data Table = Table Node.List Instance.List Score [Placement]
148 -- | Cluster statistics data type.
150 { csFmem :: Integer -- ^ Cluster free mem
151 , csFdsk :: Integer -- ^ Cluster free disk
152 , csAmem :: Integer -- ^ Cluster allocatable mem
153 , csAdsk :: Integer -- ^ Cluster allocatable disk
154 , csAcpu :: Integer -- ^ Cluster allocatable cpus
155 , csMmem :: Integer -- ^ Max node allocatable mem
156 , csMdsk :: Integer -- ^ Max node allocatable disk
157 , csMcpu :: Integer -- ^ Max node allocatable cpu
158 , csImem :: Integer -- ^ Instance used mem
159 , csIdsk :: Integer -- ^ Instance used disk
160 , csIcpu :: Integer -- ^ Instance used cpu
161 , csTmem :: Double -- ^ Cluster total mem
162 , csTdsk :: Double -- ^ Cluster total disk
163 , csTcpu :: Double -- ^ Cluster total cpus
164 , csVcpu :: Integer -- ^ Cluster total virtual cpus
165 , csNcpu :: Double -- ^ Equivalent to 'csIcpu' but in terms of
166 -- physical CPUs, i.e. normalised used phys CPUs
167 , csXmem :: Integer -- ^ Unnacounted for mem
168 , csNmem :: Integer -- ^ Node own memory
169 , csScore :: Score -- ^ The cluster score
170 , csNinst :: Int -- ^ The total number of instances
173 -- | A simple type for allocation functions.
174 type AllocMethod = Node.List -- ^ Node list
175 -> Instance.List -- ^ Instance list
176 -> Maybe Int -- ^ Optional allocation limit
177 -> Instance.Instance -- ^ Instance spec for allocation
178 -> AllocNodes -- ^ Which nodes we should allocate on
179 -> [Instance.Instance] -- ^ Allocated instances
180 -> [CStats] -- ^ Running cluster stats
181 -> Result AllocResult -- ^ Allocation result
183 -- | A simple type for the running solution of evacuations.
184 type EvacInnerState =
185 Either String (Node.List, Instance.Instance, Score, Ndx)
187 -- * Utility functions
189 -- | Verifies the N+1 status and return the affected nodes.
190 verifyN1 :: [Node.Node] -> [Node.Node]
191 verifyN1 = filter Node.failN1
193 {-| Computes the pair of bad nodes and instances.
195 The bad node list is computed via a simple 'verifyN1' check, and the
196 bad instance list is the list of primary and secondary instances of
200 computeBadItems :: Node.List -> Instance.List ->
201 ([Node.Node], [Instance.Instance])
202 computeBadItems nl il =
203 let bad_nodes = verifyN1 $ getOnline nl
204 bad_instances = map (`Container.find` il) .
206 concatMap (\ n -> Node.sList n ++ Node.pList n) bad_nodes
208 (bad_nodes, bad_instances)
210 -- | Extracts the node pairs for an instance. This can fail if the
211 -- instance is single-homed. FIXME: this needs to be improved,
212 -- together with the general enhancement for handling non-DRBD moves.
213 instanceNodes :: Node.List -> Instance.Instance ->
214 (Ndx, Ndx, Node.Node, Node.Node)
215 instanceNodes nl inst =
216 let old_pdx = Instance.pNode inst
217 old_sdx = Instance.sNode inst
218 old_p = Container.find old_pdx nl
219 old_s = Container.find old_sdx nl
220 in (old_pdx, old_sdx, old_p, old_s)
222 -- | Zero-initializer for the CStats type.
223 emptyCStats :: CStats
224 emptyCStats = CStats 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
226 -- | Update stats with data from a new node.
227 updateCStats :: CStats -> Node.Node -> CStats
228 updateCStats cs node =
229 let CStats { csFmem = x_fmem, csFdsk = x_fdsk,
230 csAmem = x_amem, csAcpu = x_acpu, csAdsk = x_adsk,
231 csMmem = x_mmem, csMdsk = x_mdsk, csMcpu = x_mcpu,
232 csImem = x_imem, csIdsk = x_idsk, csIcpu = x_icpu,
233 csTmem = x_tmem, csTdsk = x_tdsk, csTcpu = x_tcpu,
234 csVcpu = x_vcpu, csNcpu = x_ncpu,
235 csXmem = x_xmem, csNmem = x_nmem, csNinst = x_ninst
238 inc_amem = Node.fMem node - Node.rMem node
239 inc_amem' = if inc_amem > 0 then inc_amem else 0
240 inc_adsk = Node.availDisk node
241 inc_imem = truncate (Node.tMem node) - Node.nMem node
242 - Node.xMem node - Node.fMem node
243 inc_icpu = Node.uCpu node
244 inc_idsk = truncate (Node.tDsk node) - Node.fDsk node
245 inc_vcpu = Node.hiCpu node
246 inc_acpu = Node.availCpu node
247 inc_ncpu = fromIntegral (Node.uCpu node) /
248 iPolicyVcpuRatio (Node.iPolicy node)
249 in cs { csFmem = x_fmem + fromIntegral (Node.fMem node)
250 , csFdsk = x_fdsk + fromIntegral (Node.fDsk node)
251 , csAmem = x_amem + fromIntegral inc_amem'
252 , csAdsk = x_adsk + fromIntegral inc_adsk
253 , csAcpu = x_acpu + fromIntegral inc_acpu
254 , csMmem = max x_mmem (fromIntegral inc_amem')
255 , csMdsk = max x_mdsk (fromIntegral inc_adsk)
256 , csMcpu = max x_mcpu (fromIntegral inc_acpu)
257 , csImem = x_imem + fromIntegral inc_imem
258 , csIdsk = x_idsk + fromIntegral inc_idsk
259 , csIcpu = x_icpu + fromIntegral inc_icpu
260 , csTmem = x_tmem + Node.tMem node
261 , csTdsk = x_tdsk + Node.tDsk node
262 , csTcpu = x_tcpu + Node.tCpu node
263 , csVcpu = x_vcpu + fromIntegral inc_vcpu
264 , csNcpu = x_ncpu + inc_ncpu
265 , csXmem = x_xmem + fromIntegral (Node.xMem node)
266 , csNmem = x_nmem + fromIntegral (Node.nMem node)
267 , csNinst = x_ninst + length (Node.pList node)
270 -- | Compute the total free disk and memory in the cluster.
271 totalResources :: Node.List -> CStats
273 let cs = foldl' updateCStats emptyCStats . Container.elems $ nl
274 in cs { csScore = compCV nl }
276 -- | Compute the delta between two cluster state.
278 -- This is used when doing allocations, to understand better the
279 -- available cluster resources. The return value is a triple of the
280 -- current used values, the delta that was still allocated, and what
281 -- was left unallocated.
282 computeAllocationDelta :: CStats -> CStats -> AllocStats
283 computeAllocationDelta cini cfin =
284 let CStats {csImem = i_imem, csIdsk = i_idsk, csIcpu = i_icpu,
285 csNcpu = i_ncpu } = cini
286 CStats {csImem = f_imem, csIdsk = f_idsk, csIcpu = f_icpu,
287 csTmem = t_mem, csTdsk = t_dsk, csVcpu = f_vcpu,
288 csNcpu = f_ncpu, csTcpu = f_tcpu } = cfin
289 rini = AllocInfo { allocInfoVCpus = fromIntegral i_icpu
290 , allocInfoNCpus = i_ncpu
291 , allocInfoMem = fromIntegral i_imem
292 , allocInfoDisk = fromIntegral i_idsk
294 rfin = AllocInfo { allocInfoVCpus = fromIntegral (f_icpu - i_icpu)
295 , allocInfoNCpus = f_ncpu - i_ncpu
296 , allocInfoMem = fromIntegral (f_imem - i_imem)
297 , allocInfoDisk = fromIntegral (f_idsk - i_idsk)
299 runa = AllocInfo { allocInfoVCpus = fromIntegral (f_vcpu - f_icpu)
300 , allocInfoNCpus = f_tcpu - f_ncpu
301 , allocInfoMem = truncate t_mem - fromIntegral f_imem
302 , allocInfoDisk = truncate t_dsk - fromIntegral f_idsk
304 in (rini, rfin, runa)
306 -- | The names and weights of the individual elements in the CV list.
307 detailedCVInfo :: [(Double, String)]
308 detailedCVInfo = [ (1, "free_mem_cv")
309 , (1, "free_disk_cv")
311 , (1, "reserved_mem_cv")
312 , (4, "offline_all_cnt")
313 , (16, "offline_pri_cnt")
314 , (1, "vcpu_ratio_cv")
317 , (1, "disk_load_cv")
319 , (2, "pri_tags_score")
323 -- | Holds the weights used by 'compCVNodes' for each metric.
324 detailedCVWeights :: [Double]
325 detailedCVWeights = map fst detailedCVInfo
327 -- | Compute the mem and disk covariance.
328 compDetailedCV :: [Node.Node] -> [Double]
329 compDetailedCV all_nodes =
330 let (offline, nodes) = partition Node.offline all_nodes
331 mem_l = map Node.pMem nodes
332 dsk_l = map Node.pDsk nodes
333 -- metric: memory covariance
334 mem_cv = stdDev mem_l
335 -- metric: disk covariance
336 dsk_cv = stdDev dsk_l
337 -- metric: count of instances living on N1 failing nodes
338 n1_score = fromIntegral . sum . map (\n -> length (Node.sList n) +
339 length (Node.pList n)) .
340 filter Node.failN1 $ nodes :: Double
341 res_l = map Node.pRem nodes
342 -- metric: reserved memory covariance
343 res_cv = stdDev res_l
344 -- offline instances metrics
345 offline_ipri = sum . map (length . Node.pList) $ offline
346 offline_isec = sum . map (length . Node.sList) $ offline
347 -- metric: count of instances on offline nodes
348 off_score = fromIntegral (offline_ipri + offline_isec)::Double
349 -- metric: count of primary instances on offline nodes (this
350 -- helps with evacuation/failover of primary instances on
351 -- 2-node clusters with one node offline)
352 off_pri_score = fromIntegral offline_ipri::Double
353 cpu_l = map Node.pCpu nodes
354 -- metric: covariance of vcpu/pcpu ratio
355 cpu_cv = stdDev cpu_l
356 -- metrics: covariance of cpu, memory, disk and network load
357 (c_load, m_load, d_load, n_load) =
359 let DynUtil c1 m1 d1 n1 = Node.utilLoad n
360 DynUtil c2 m2 d2 n2 = Node.utilPool n
361 in (c1/c2, m1/m2, d1/d2, n1/n2)) nodes
362 -- metric: conflicting instance count
363 pri_tags_inst = sum $ map Node.conflictingPrimaries nodes
364 pri_tags_score = fromIntegral pri_tags_inst::Double
365 -- metric: spindles %
366 spindles_cv = map (\n -> Node.instSpindles n / Node.hiSpindles n) nodes
367 in [ mem_cv, dsk_cv, n1_score, res_cv, off_score, off_pri_score, cpu_cv
368 , stdDev c_load, stdDev m_load , stdDev d_load, stdDev n_load
369 , pri_tags_score, stdDev spindles_cv ]
371 -- | Compute the /total/ variance.
372 compCVNodes :: [Node.Node] -> Double
373 compCVNodes = sum . zipWith (*) detailedCVWeights . compDetailedCV
375 -- | Wrapper over 'compCVNodes' for callers that have a 'Node.List'.
376 compCV :: Node.List -> Double
377 compCV = compCVNodes . Container.elems
379 -- | Compute online nodes from a 'Node.List'.
380 getOnline :: Node.List -> [Node.Node]
381 getOnline = filter (not . Node.offline) . Container.elems
383 -- * Balancing functions
385 -- | Compute best table. Note that the ordering of the arguments is important.
386 compareTables :: Table -> Table -> Table
387 compareTables a@(Table _ _ a_cv _) b@(Table _ _ b_cv _ ) =
388 if a_cv > b_cv then b else a
390 -- | Applies an instance move to a given node list and instance.
391 applyMove :: Node.List -> Instance.Instance
392 -> IMove -> OpResult (Node.List, Instance.Instance, Ndx, Ndx)
394 applyMove nl inst Failover =
395 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
396 int_p = Node.removePri old_p inst
397 int_s = Node.removeSec old_s inst
398 new_nl = do -- Maybe monad
399 new_p <- Node.addPriEx (Node.offline old_p) int_s inst
400 new_s <- Node.addSec int_p inst old_sdx
401 let new_inst = Instance.setBoth inst old_sdx old_pdx
402 return (Container.addTwo old_pdx new_s old_sdx new_p nl,
403 new_inst, old_sdx, old_pdx)
406 -- Failover to any (fa)
407 applyMove nl inst (FailoverToAny new_pdx) = do
408 let (old_pdx, old_sdx, old_pnode, _) = instanceNodes nl inst
409 new_pnode = Container.find new_pdx nl
410 force_failover = Node.offline old_pnode
411 new_pnode' <- Node.addPriEx force_failover new_pnode inst
412 let old_pnode' = Node.removePri old_pnode inst
413 inst' = Instance.setPri inst new_pdx
414 nl' = Container.addTwo old_pdx old_pnode' new_pdx new_pnode' nl
415 return (nl', inst', new_pdx, old_sdx)
417 -- Replace the primary (f:, r:np, f)
418 applyMove nl inst (ReplacePrimary new_pdx) =
419 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
420 tgt_n = Container.find new_pdx nl
421 int_p = Node.removePri old_p inst
422 int_s = Node.removeSec old_s inst
423 force_p = Node.offline old_p
424 new_nl = do -- Maybe monad
425 -- check that the current secondary can host the instance
426 -- during the migration
427 tmp_s <- Node.addPriEx force_p int_s inst
428 let tmp_s' = Node.removePri tmp_s inst
429 new_p <- Node.addPriEx force_p tgt_n inst
430 new_s <- Node.addSecEx force_p tmp_s' inst new_pdx
431 let new_inst = Instance.setPri inst new_pdx
432 return (Container.add new_pdx new_p $
433 Container.addTwo old_pdx int_p old_sdx new_s nl,
434 new_inst, new_pdx, old_sdx)
437 -- Replace the secondary (r:ns)
438 applyMove nl inst (ReplaceSecondary new_sdx) =
439 let old_pdx = Instance.pNode inst
440 old_sdx = Instance.sNode inst
441 old_s = Container.find old_sdx nl
442 tgt_n = Container.find new_sdx nl
443 int_s = Node.removeSec old_s inst
444 force_s = Node.offline old_s
445 new_inst = Instance.setSec inst new_sdx
446 new_nl = Node.addSecEx force_s tgt_n inst old_pdx >>=
447 \new_s -> return (Container.addTwo new_sdx
448 new_s old_sdx int_s nl,
449 new_inst, old_pdx, new_sdx)
452 -- Replace the secondary and failover (r:np, f)
453 applyMove nl inst (ReplaceAndFailover new_pdx) =
454 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
455 tgt_n = Container.find new_pdx nl
456 int_p = Node.removePri old_p inst
457 int_s = Node.removeSec old_s inst
458 force_s = Node.offline old_s
459 new_nl = do -- Maybe monad
460 new_p <- Node.addPri tgt_n inst
461 new_s <- Node.addSecEx force_s int_p inst new_pdx
462 let new_inst = Instance.setBoth inst new_pdx old_pdx
463 return (Container.add new_pdx new_p $
464 Container.addTwo old_pdx new_s old_sdx int_s nl,
465 new_inst, new_pdx, old_pdx)
468 -- Failver and replace the secondary (f, r:ns)
469 applyMove nl inst (FailoverAndReplace new_sdx) =
470 let (old_pdx, old_sdx, old_p, old_s) = instanceNodes nl inst
471 tgt_n = Container.find new_sdx nl
472 int_p = Node.removePri old_p inst
473 int_s = Node.removeSec old_s inst
474 force_p = Node.offline old_p
475 new_nl = do -- Maybe monad
476 new_p <- Node.addPriEx force_p int_s inst
477 new_s <- Node.addSecEx force_p tgt_n inst old_sdx
478 let new_inst = Instance.setBoth inst old_sdx new_sdx
479 return (Container.add new_sdx new_s $
480 Container.addTwo old_sdx new_p old_pdx int_p nl,
481 new_inst, old_sdx, new_sdx)
484 -- | Tries to allocate an instance on one given node.
485 allocateOnSingle :: Node.List -> Instance.Instance -> Ndx
486 -> OpResult Node.AllocElement
487 allocateOnSingle nl inst new_pdx =
488 let p = Container.find new_pdx nl
489 new_inst = Instance.setBoth inst new_pdx Node.noSecondary
491 Instance.instMatchesPolicy inst (Node.iPolicy p)
492 new_p <- Node.addPri p inst
493 let new_nl = Container.add new_pdx new_p nl
494 new_score = compCV new_nl
495 return (new_nl, new_inst, [new_p], new_score)
497 -- | Tries to allocate an instance on a given pair of nodes.
498 allocateOnPair :: Node.List -> Instance.Instance -> Ndx -> Ndx
499 -> OpResult Node.AllocElement
500 allocateOnPair nl inst new_pdx new_sdx =
501 let tgt_p = Container.find new_pdx nl
502 tgt_s = Container.find new_sdx nl
504 Instance.instMatchesPolicy inst (Node.iPolicy tgt_p)
505 new_p <- Node.addPri tgt_p inst
506 new_s <- Node.addSec tgt_s inst new_pdx
507 let new_inst = Instance.setBoth inst new_pdx new_sdx
508 new_nl = Container.addTwo new_pdx new_p new_sdx new_s nl
509 return (new_nl, new_inst, [new_p, new_s], compCV new_nl)
511 -- | Tries to perform an instance move and returns the best table
512 -- between the original one and the new one.
513 checkSingleStep :: Table -- ^ The original table
514 -> Instance.Instance -- ^ The instance to move
515 -> Table -- ^ The current best table
516 -> IMove -- ^ The move to apply
517 -> Table -- ^ The final best table
518 checkSingleStep ini_tbl target cur_tbl move =
519 let Table ini_nl ini_il _ ini_plc = ini_tbl
520 tmp_resu = applyMove ini_nl target move
523 Ok (upd_nl, new_inst, pri_idx, sec_idx) ->
524 let tgt_idx = Instance.idx target
525 upd_cvar = compCV upd_nl
526 upd_il = Container.add tgt_idx new_inst ini_il
527 upd_plc = (tgt_idx, pri_idx, sec_idx, move, upd_cvar):ini_plc
528 upd_tbl = Table upd_nl upd_il upd_cvar upd_plc
529 in compareTables cur_tbl upd_tbl
531 -- | Given the status of the current secondary as a valid new node and
532 -- the current candidate target node, generate the possible moves for
534 possibleMoves :: MirrorType -- ^ The mirroring type of the instance
535 -> Bool -- ^ Whether the secondary node is a valid new node
536 -> Bool -- ^ Whether we can change the primary node
537 -> Ndx -- ^ Target node candidate
538 -> [IMove] -- ^ List of valid result moves
540 possibleMoves MirrorNone _ _ _ = []
542 possibleMoves MirrorExternal _ False _ = []
544 possibleMoves MirrorExternal _ True tdx =
545 [ FailoverToAny tdx ]
547 possibleMoves MirrorInternal _ False tdx =
548 [ ReplaceSecondary tdx ]
550 possibleMoves MirrorInternal True True tdx =
551 [ ReplaceSecondary tdx
552 , ReplaceAndFailover tdx
554 , FailoverAndReplace tdx
557 possibleMoves MirrorInternal False True tdx =
558 [ ReplaceSecondary tdx
559 , ReplaceAndFailover tdx
562 -- | Compute the best move for a given instance.
563 checkInstanceMove :: [Ndx] -- ^ Allowed target node indices
564 -> Bool -- ^ Whether disk moves are allowed
565 -> Bool -- ^ Whether instance moves are allowed
566 -> Table -- ^ Original table
567 -> Instance.Instance -- ^ Instance to move
568 -> Table -- ^ Best new table for this instance
569 checkInstanceMove nodes_idx disk_moves inst_moves ini_tbl target =
570 let opdx = Instance.pNode target
571 osdx = Instance.sNode target
572 bad_nodes = [opdx, osdx]
573 nodes = filter (`notElem` bad_nodes) nodes_idx
574 mir_type = Instance.mirrorType target
575 use_secondary = elem osdx nodes_idx && inst_moves
576 aft_failover = if mir_type == MirrorInternal && use_secondary
577 -- if drbd and allowed to failover
578 then checkSingleStep ini_tbl target ini_tbl Failover
582 then concatMap (possibleMoves mir_type use_secondary inst_moves)
586 -- iterate over the possible nodes for this instance
587 foldl' (checkSingleStep ini_tbl target) aft_failover all_moves
589 -- | Compute the best next move.
590 checkMove :: [Ndx] -- ^ Allowed target node indices
591 -> Bool -- ^ Whether disk moves are allowed
592 -> Bool -- ^ Whether instance moves are allowed
593 -> Table -- ^ The current solution
594 -> [Instance.Instance] -- ^ List of instances still to move
595 -> Table -- ^ The new solution
596 checkMove nodes_idx disk_moves inst_moves ini_tbl victims =
597 let Table _ _ _ ini_plc = ini_tbl
598 -- we're using rwhnf from the Control.Parallel.Strategies
599 -- package; we don't need to use rnf as that would force too
600 -- much evaluation in single-threaded cases, and in
601 -- multi-threaded case the weak head normal form is enough to
602 -- spark the evaluation
603 tables = parMap rwhnf (checkInstanceMove nodes_idx disk_moves
606 -- iterate over all instances, computing the best move
607 best_tbl = foldl' compareTables ini_tbl tables
608 Table _ _ _ best_plc = best_tbl
609 in if length best_plc == length ini_plc
610 then ini_tbl -- no advancement
613 -- | Check if we are allowed to go deeper in the balancing.
614 doNextBalance :: Table -- ^ The starting table
615 -> Int -- ^ Remaining length
616 -> Score -- ^ Score at which to stop
617 -> Bool -- ^ The resulting table and commands
618 doNextBalance ini_tbl max_rounds min_score =
619 let Table _ _ ini_cv ini_plc = ini_tbl
620 ini_plc_len = length ini_plc
621 in (max_rounds < 0 || ini_plc_len < max_rounds) && ini_cv > min_score
623 -- | Run a balance move.
624 tryBalance :: Table -- ^ The starting table
625 -> Bool -- ^ Allow disk moves
626 -> Bool -- ^ Allow instance moves
627 -> Bool -- ^ Only evacuate moves
628 -> Score -- ^ Min gain threshold
629 -> Score -- ^ Min gain
630 -> Maybe Table -- ^ The resulting table and commands
631 tryBalance ini_tbl disk_moves inst_moves evac_mode mg_limit min_gain =
632 let Table ini_nl ini_il ini_cv _ = ini_tbl
633 all_inst = Container.elems ini_il
634 all_nodes = Container.elems ini_nl
635 (offline_nodes, online_nodes) = partition Node.offline all_nodes
636 all_inst' = if evac_mode
637 then let bad_nodes = map Node.idx offline_nodes
638 in filter (any (`elem` bad_nodes) .
639 Instance.allNodes) all_inst
641 reloc_inst = filter (\i -> Instance.movable i &&
642 Instance.autoBalance i) all_inst'
643 node_idx = map Node.idx online_nodes
644 fin_tbl = checkMove node_idx disk_moves inst_moves ini_tbl reloc_inst
645 (Table _ _ fin_cv _) = fin_tbl
647 if fin_cv < ini_cv && (ini_cv > mg_limit || ini_cv - fin_cv >= min_gain)
648 then Just fin_tbl -- this round made success, return the new table
651 -- * Allocation functions
653 -- | Build failure stats out of a list of failures.
654 collapseFailures :: [FailMode] -> FailStats
655 collapseFailures flst =
656 map (\k -> (k, foldl' (\a e -> if e == k then a + 1 else a) 0 flst))
659 -- | Compares two Maybe AllocElement and chooses the best score.
660 bestAllocElement :: Maybe Node.AllocElement
661 -> Maybe Node.AllocElement
662 -> Maybe Node.AllocElement
663 bestAllocElement a Nothing = a
664 bestAllocElement Nothing b = b
665 bestAllocElement a@(Just (_, _, _, ascore)) b@(Just (_, _, _, bscore)) =
666 if ascore < bscore then a else b
668 -- | Update current Allocation solution and failure stats with new
670 concatAllocs :: AllocSolution -> OpResult Node.AllocElement -> AllocSolution
671 concatAllocs as (Bad reason) = as { asFailures = reason : asFailures as }
673 concatAllocs as (Ok ns) =
674 let -- Choose the old or new solution, based on the cluster score
676 osols = asSolution as
677 nsols = bestAllocElement osols (Just ns)
679 -- Note: we force evaluation of nsols here in order to keep the
680 -- memory profile low - we know that we will need nsols for sure
681 -- in the next cycle, so we force evaluation of nsols, since the
682 -- foldl' in the caller will only evaluate the tuple, but not the
683 -- elements of the tuple
684 in nsols `seq` nsuc `seq` as { asAllocs = nsuc, asSolution = nsols }
686 -- | Sums two 'AllocSolution' structures.
687 sumAllocs :: AllocSolution -> AllocSolution -> AllocSolution
688 sumAllocs (AllocSolution aFails aAllocs aSols aLog)
689 (AllocSolution bFails bAllocs bSols bLog) =
690 -- note: we add b first, since usually it will be smaller; when
691 -- fold'ing, a will grow and grow whereas b is the per-group
692 -- result, hence smaller
693 let nFails = bFails ++ aFails
694 nAllocs = aAllocs + bAllocs
695 nSols = bestAllocElement aSols bSols
697 in AllocSolution nFails nAllocs nSols nLog
699 -- | Given a solution, generates a reasonable description for it.
700 describeSolution :: AllocSolution -> String
701 describeSolution as =
702 let fcnt = asFailures as
705 intercalate ", " . map (\(a, b) -> printf "%s: %d" (show a) b) .
706 filter ((> 0) . snd) . collapseFailures $ fcnt
708 Nothing -> "No valid allocation solutions, failure reasons: " ++
709 (if null fcnt then "unknown reasons" else freasons)
710 Just (_, _, nodes, cv) ->
711 printf ("score: %.8f, successes %d, failures %d (%s)" ++
712 " for node(s) %s") cv (asAllocs as) (length fcnt) freasons
713 (intercalate "/" . map Node.name $ nodes)
715 -- | Annotates a solution with the appropriate string.
716 annotateSolution :: AllocSolution -> AllocSolution
717 annotateSolution as = as { asLog = describeSolution as : asLog as }
719 -- | Reverses an evacuation solution.
721 -- Rationale: we always concat the results to the top of the lists, so
722 -- for proper jobset execution, we should reverse all lists.
723 reverseEvacSolution :: EvacSolution -> EvacSolution
724 reverseEvacSolution (EvacSolution f m o) =
725 EvacSolution (reverse f) (reverse m) (reverse o)
727 -- | Generate the valid node allocation singles or pairs for a new instance.
728 genAllocNodes :: Group.List -- ^ Group list
729 -> Node.List -- ^ The node map
730 -> Int -- ^ The number of nodes required
731 -> Bool -- ^ Whether to drop or not
733 -> Result AllocNodes -- ^ The (monadic) result
734 genAllocNodes gl nl count drop_unalloc =
735 let filter_fn = if drop_unalloc
736 then filter (Group.isAllocable .
737 flip Container.find gl . Node.group)
739 all_nodes = filter_fn $ getOnline nl
740 all_pairs = [(Node.idx p,
741 [Node.idx s | s <- all_nodes,
742 Node.idx p /= Node.idx s,
743 Node.group p == Node.group s]) |
746 1 -> Ok (Left (map Node.idx all_nodes))
747 2 -> Ok (Right (filter (not . null . snd) all_pairs))
748 _ -> Bad "Unsupported number of nodes, only one or two supported"
750 -- | Try to allocate an instance on the cluster.
751 tryAlloc :: (Monad m) =>
752 Node.List -- ^ The node list
753 -> Instance.List -- ^ The instance list
754 -> Instance.Instance -- ^ The instance to allocate
755 -> AllocNodes -- ^ The allocation targets
756 -> m AllocSolution -- ^ Possible solution list
757 tryAlloc _ _ _ (Right []) = fail "Not enough online nodes"
758 tryAlloc nl _ inst (Right ok_pairs) =
759 let psols = parMap rwhnf (\(p, ss) ->
761 concatAllocs cstate .
762 allocateOnPair nl inst p)
763 emptyAllocSolution ss) ok_pairs
764 sols = foldl' sumAllocs emptyAllocSolution psols
765 in return $ annotateSolution sols
767 tryAlloc _ _ _ (Left []) = fail "No online nodes"
768 tryAlloc nl _ inst (Left all_nodes) =
769 let sols = foldl' (\cstate ->
770 concatAllocs cstate . allocateOnSingle nl inst
771 ) emptyAllocSolution all_nodes
772 in return $ annotateSolution sols
774 -- | Given a group/result, describe it as a nice (list of) messages.
775 solutionDescription :: Group.List -> (Gdx, Result AllocSolution) -> [String]
776 solutionDescription gl (groupId, result) =
778 Ok solution -> map (printf "Group %s (%s): %s" gname pol) (asLog solution)
779 Bad message -> [printf "Group %s: error %s" gname message]
780 where grp = Container.find groupId gl
781 gname = Group.name grp
782 pol = allocPolicyToRaw (Group.allocPolicy grp)
784 -- | From a list of possibly bad and possibly empty solutions, filter
785 -- only the groups with a valid result. Note that the result will be
786 -- reversed compared to the original list.
787 filterMGResults :: Group.List
788 -> [(Gdx, Result AllocSolution)]
789 -> [(Gdx, AllocSolution)]
790 filterMGResults gl = foldl' fn []
791 where unallocable = not . Group.isAllocable . flip Container.find gl
792 fn accu (gdx, rasol) =
795 Ok sol | isNothing (asSolution sol) -> accu
796 | unallocable gdx -> accu
797 | otherwise -> (gdx, sol):accu
799 -- | Sort multigroup results based on policy and score.
800 sortMGResults :: Group.List
801 -> [(Gdx, AllocSolution)]
802 -> [(Gdx, AllocSolution)]
803 sortMGResults gl sols =
804 let extractScore (_, _, _, x) = x
805 solScore (gdx, sol) = (Group.allocPolicy (Container.find gdx gl),
806 (extractScore . fromJust . asSolution) sol)
807 in sortBy (comparing solScore) sols
809 -- | Finds the best group for an instance on a multi-group cluster.
811 -- Only solutions in @preferred@ and @last_resort@ groups will be
812 -- accepted as valid, and additionally if the allowed groups parameter
813 -- is not null then allocation will only be run for those group
815 findBestAllocGroup :: Group.List -- ^ The group list
816 -> Node.List -- ^ The node list
817 -> Instance.List -- ^ The instance list
818 -> Maybe [Gdx] -- ^ The allowed groups
819 -> Instance.Instance -- ^ The instance to allocate
820 -> Int -- ^ Required number of nodes
821 -> Result (Gdx, AllocSolution, [String])
822 findBestAllocGroup mggl mgnl mgil allowed_gdxs inst cnt =
823 let groups = splitCluster mgnl mgil
824 groups' = maybe groups (\gs -> filter ((`elem` gs) . fst) groups)
826 sols = map (\(gid, (nl, il)) ->
827 (gid, genAllocNodes mggl nl cnt False >>=
828 tryAlloc nl il inst))
829 groups'::[(Gdx, Result AllocSolution)]
830 all_msgs = concatMap (solutionDescription mggl) sols
831 goodSols = filterMGResults mggl sols
832 sortedSols = sortMGResults mggl goodSols
833 in case sortedSols of
834 [] -> Bad $ if null groups'
835 then "no groups for evacuation: allowed groups was" ++
836 show allowed_gdxs ++ ", all groups: " ++
837 show (map fst groups)
838 else intercalate ", " all_msgs
839 (final_group, final_sol):_ -> return (final_group, final_sol, all_msgs)
841 -- | Try to allocate an instance on a multi-group cluster.
842 tryMGAlloc :: Group.List -- ^ The group list
843 -> Node.List -- ^ The node list
844 -> Instance.List -- ^ The instance list
845 -> Instance.Instance -- ^ The instance to allocate
846 -> Int -- ^ Required number of nodes
847 -> Result AllocSolution -- ^ Possible solution list
848 tryMGAlloc mggl mgnl mgil inst cnt = do
849 (best_group, solution, all_msgs) <-
850 findBestAllocGroup mggl mgnl mgil Nothing inst cnt
851 let group_name = Group.name $ Container.find best_group mggl
852 selmsg = "Selected group: " ++ group_name
853 return $ solution { asLog = selmsg:all_msgs }
855 -- | Calculate the new instance list after allocation solution.
856 updateIl :: Instance.List -- ^ The original instance list
857 -> Maybe Node.AllocElement -- ^ The result of the allocation attempt
858 -> Instance.List -- ^ The updated instance list
859 updateIl il Nothing = il
860 updateIl il (Just (_, xi, _, _)) = Container.add (Container.size il) xi il
862 -- | Extract the the new node list from the allocation solution.
863 extractNl :: Node.List -- ^ The original node list
864 -> Maybe Node.AllocElement -- ^ The result of the allocation attempt
865 -> Node.List -- ^ The new node list
866 extractNl nl Nothing = nl
867 extractNl _ (Just (xnl, _, _, _)) = xnl
869 -- | Try to allocate a list of instances on a multi-group cluster.
870 allocList :: Group.List -- ^ The group list
871 -> Node.List -- ^ The node list
872 -> Instance.List -- ^ The instance list
873 -> [(Instance.Instance, Int)] -- ^ The instance to allocate
874 -> AllocSolutionList -- ^ Possible solution list
875 -> Result (Node.List, Instance.List,
876 AllocSolutionList) -- ^ The final solution list
877 allocList _ nl il [] result = Ok (nl, il, result)
878 allocList gl nl il ((xi, xicnt):xies) result = do
879 ares <- tryMGAlloc gl nl il xi xicnt
880 let sol = asSolution ares
881 nl' = extractNl nl sol
882 il' = updateIl il sol
883 allocList gl nl' il' xies ((xi, ares):result)
885 -- | Function which fails if the requested mode is change secondary.
887 -- This is useful since except DRBD, no other disk template can
888 -- execute change secondary; thus, we can just call this function
889 -- instead of always checking for secondary mode. After the call to
890 -- this function, whatever mode we have is just a primary change.
891 failOnSecondaryChange :: (Monad m) => EvacMode -> DiskTemplate -> m ()
892 failOnSecondaryChange ChangeSecondary dt =
893 fail $ "Instances with disk template '" ++ diskTemplateToRaw dt ++
894 "' can't execute change secondary"
895 failOnSecondaryChange _ _ = return ()
897 -- | Run evacuation for a single instance.
899 -- /Note:/ this function should correctly execute both intra-group
900 -- evacuations (in all modes) and inter-group evacuations (in the
901 -- 'ChangeAll' mode). Of course, this requires that the correct list
902 -- of target nodes is passed.
903 nodeEvacInstance :: Node.List -- ^ The node list (cluster-wide)
904 -> Instance.List -- ^ Instance list (cluster-wide)
905 -> EvacMode -- ^ The evacuation mode
906 -> Instance.Instance -- ^ The instance to be evacuated
907 -> Gdx -- ^ The group we're targetting
908 -> [Ndx] -- ^ The list of available nodes
910 -> Result (Node.List, Instance.List, [OpCodes.OpCode])
911 nodeEvacInstance nl il mode inst@(Instance.Instance
912 {Instance.diskTemplate = dt@DTDiskless})
914 failOnSecondaryChange mode dt >>
915 evacOneNodeOnly nl il inst gdx avail_nodes
917 nodeEvacInstance _ _ _ (Instance.Instance
918 {Instance.diskTemplate = DTPlain}) _ _ =
919 fail "Instances of type plain cannot be relocated"
921 nodeEvacInstance _ _ _ (Instance.Instance
922 {Instance.diskTemplate = DTFile}) _ _ =
923 fail "Instances of type file cannot be relocated"
925 nodeEvacInstance nl il mode inst@(Instance.Instance
926 {Instance.diskTemplate = dt@DTSharedFile})
928 failOnSecondaryChange mode dt >>
929 evacOneNodeOnly nl il inst gdx avail_nodes
931 nodeEvacInstance nl il mode inst@(Instance.Instance
932 {Instance.diskTemplate = dt@DTBlock})
934 failOnSecondaryChange mode dt >>
935 evacOneNodeOnly nl il inst gdx avail_nodes
937 nodeEvacInstance nl il mode inst@(Instance.Instance
938 {Instance.diskTemplate = dt@DTRbd})
940 failOnSecondaryChange mode dt >>
941 evacOneNodeOnly nl il inst gdx avail_nodes
943 nodeEvacInstance nl il mode inst@(Instance.Instance
944 {Instance.diskTemplate = dt@DTExt})
946 failOnSecondaryChange mode dt >>
947 evacOneNodeOnly nl il inst gdx avail_nodes
949 nodeEvacInstance nl il ChangePrimary
950 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})
953 (nl', inst', _, _) <- opToResult $ applyMove nl inst Failover
954 let idx = Instance.idx inst
955 il' = Container.add idx inst' il
956 ops = iMoveToJob nl' il' idx Failover
957 return (nl', il', ops)
959 nodeEvacInstance nl il ChangeSecondary
960 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})
962 evacOneNodeOnly nl il inst gdx avail_nodes
964 -- The algorithm for ChangeAll is as follows:
966 -- * generate all (primary, secondary) node pairs for the target groups
967 -- * for each pair, execute the needed moves (r:s, f, r:s) and compute
968 -- the final node list state and group score
969 -- * select the best choice via a foldl that uses the same Either
970 -- String solution as the ChangeSecondary mode
971 nodeEvacInstance nl il ChangeAll
972 inst@(Instance.Instance {Instance.diskTemplate = DTDrbd8})
975 let no_nodes = Left "no nodes available"
976 node_pairs = [(p,s) | p <- avail_nodes, s <- avail_nodes, p /= s]
977 (nl', il', ops, _) <-
978 annotateResult "Can't find any good nodes for relocation" .
981 (\accu nodes -> case evacDrbdAllInner nl il inst gdx nodes of
985 -- we don't need more details (which
986 -- nodes, etc.) as we only selected
987 -- this group if we can allocate on
988 -- it, hence failures will not
989 -- propagate out of this fold loop
990 Left _ -> Left $ "Allocation failed: " ++ msg
991 Ok result@(_, _, _, new_cv) ->
992 let new_accu = Right result in
995 Right (_, _, _, old_cv) ->
999 ) no_nodes node_pairs
1001 return (nl', il', ops)
1003 -- | Generic function for changing one node of an instance.
1005 -- This is similar to 'nodeEvacInstance' but will be used in a few of
1006 -- its sub-patterns. It folds the inner function 'evacOneNodeInner'
1007 -- over the list of available nodes, which results in the best choice
1009 evacOneNodeOnly :: Node.List -- ^ The node list (cluster-wide)
1010 -> Instance.List -- ^ Instance list (cluster-wide)
1011 -> Instance.Instance -- ^ The instance to be evacuated
1012 -> Gdx -- ^ The group we're targetting
1013 -> [Ndx] -- ^ The list of available nodes
1015 -> Result (Node.List, Instance.List, [OpCodes.OpCode])
1016 evacOneNodeOnly nl il inst gdx avail_nodes = do
1017 op_fn <- case Instance.mirrorType inst of
1018 MirrorNone -> Bad "Can't relocate/evacuate non-mirrored instances"
1019 MirrorInternal -> Ok ReplaceSecondary
1020 MirrorExternal -> Ok FailoverToAny
1021 (nl', inst', _, ndx) <- annotateResult "Can't find any good node" .
1023 foldl' (evacOneNodeInner nl inst gdx op_fn)
1024 (Left "no nodes available") avail_nodes
1025 let idx = Instance.idx inst
1026 il' = Container.add idx inst' il
1027 ops = iMoveToJob nl' il' idx (op_fn ndx)
1028 return (nl', il', ops)
1030 -- | Inner fold function for changing one node of an instance.
1032 -- Depending on the instance disk template, this will either change
1033 -- the secondary (for DRBD) or the primary node (for shared
1034 -- storage). However, the operation is generic otherwise.
1036 -- The running solution is either a @Left String@, which means we
1037 -- don't have yet a working solution, or a @Right (...)@, which
1038 -- represents a valid solution; it holds the modified node list, the
1039 -- modified instance (after evacuation), the score of that solution,
1040 -- and the new secondary node index.
1041 evacOneNodeInner :: Node.List -- ^ Cluster node list
1042 -> Instance.Instance -- ^ Instance being evacuated
1043 -> Gdx -- ^ The group index of the instance
1044 -> (Ndx -> IMove) -- ^ Operation constructor
1045 -> EvacInnerState -- ^ Current best solution
1046 -> Ndx -- ^ Node we're evaluating as target
1047 -> EvacInnerState -- ^ New best solution
1048 evacOneNodeInner nl inst gdx op_fn accu ndx =
1049 case applyMove nl inst (op_fn ndx) of
1050 Bad fm -> let fail_msg = "Node " ++ Container.nameOf nl ndx ++
1051 " failed: " ++ show fm
1052 in either (const $ Left fail_msg) (const accu) accu
1053 Ok (nl', inst', _, _) ->
1054 let nodes = Container.elems nl'
1055 -- The fromJust below is ugly (it can fail nastily), but
1056 -- at this point we should have any internal mismatches,
1057 -- and adding a monad here would be quite involved
1058 grpnodes = fromJust (gdx `lookup` Node.computeGroups nodes)
1059 new_cv = compCVNodes grpnodes
1060 new_accu = Right (nl', inst', new_cv, ndx)
1063 Right (_, _, old_cv, _) ->
1068 -- | Compute result of changing all nodes of a DRBD instance.
1070 -- Given the target primary and secondary node (which might be in a
1071 -- different group or not), this function will 'execute' all the
1072 -- required steps and assuming all operations succceed, will return
1073 -- the modified node and instance lists, the opcodes needed for this
1074 -- and the new group score.
1075 evacDrbdAllInner :: Node.List -- ^ Cluster node list
1076 -> Instance.List -- ^ Cluster instance list
1077 -> Instance.Instance -- ^ The instance to be moved
1078 -> Gdx -- ^ The target group index
1079 -- (which can differ from the
1080 -- current group of the
1082 -> (Ndx, Ndx) -- ^ Tuple of new
1083 -- primary\/secondary nodes
1084 -> Result (Node.List, Instance.List, [OpCodes.OpCode], Score)
1085 evacDrbdAllInner nl il inst gdx (t_pdx, t_sdx) = do
1086 let primary = Container.find (Instance.pNode inst) nl
1087 idx = Instance.idx inst
1088 -- if the primary is offline, then we first failover
1089 (nl1, inst1, ops1) <-
1090 if Node.offline primary
1092 (nl', inst', _, _) <-
1093 annotateResult "Failing over to the secondary" .
1094 opToResult $ applyMove nl inst Failover
1095 return (nl', inst', [Failover])
1096 else return (nl, inst, [])
1097 let (o1, o2, o3) = (ReplaceSecondary t_pdx,
1099 ReplaceSecondary t_sdx)
1100 -- we now need to execute a replace secondary to the future
1102 (nl2, inst2, _, _) <-
1103 annotateResult "Changing secondary to new primary" .
1105 applyMove nl1 inst1 o1
1107 -- we now execute another failover, the primary stays fixed now
1108 (nl3, inst3, _, _) <- annotateResult "Failing over to new primary" .
1109 opToResult $ applyMove nl2 inst2 o2
1111 -- and finally another replace secondary, to the final secondary
1112 (nl4, inst4, _, _) <-
1113 annotateResult "Changing secondary to final secondary" .
1115 applyMove nl3 inst3 o3
1117 il' = Container.add idx inst4 il
1118 ops = concatMap (iMoveToJob nl4 il' idx) $ reverse ops4
1119 let nodes = Container.elems nl4
1120 -- The fromJust below is ugly (it can fail nastily), but
1121 -- at this point we should have any internal mismatches,
1122 -- and adding a monad here would be quite involved
1123 grpnodes = fromJust (gdx `lookup` Node.computeGroups nodes)
1124 new_cv = compCVNodes grpnodes
1125 return (nl4, il', ops, new_cv)
1127 -- | Computes the nodes in a given group which are available for
1129 availableGroupNodes :: [(Gdx, [Ndx])] -- ^ Group index/node index assoc list
1130 -> IntSet.IntSet -- ^ Nodes that are excluded
1131 -> Gdx -- ^ The group for which we
1133 -> Result [Ndx] -- ^ List of available node indices
1134 availableGroupNodes group_nodes excl_ndx gdx = do
1135 local_nodes <- maybe (Bad $ "Can't find group with index " ++ show gdx)
1136 Ok (lookup gdx group_nodes)
1137 let avail_nodes = filter (not . flip IntSet.member excl_ndx) local_nodes
1140 -- | Updates the evac solution with the results of an instance
1142 updateEvacSolution :: (Node.List, Instance.List, EvacSolution)
1144 -> Result (Node.List, Instance.List, [OpCodes.OpCode])
1145 -> (Node.List, Instance.List, EvacSolution)
1146 updateEvacSolution (nl, il, es) idx (Bad msg) =
1147 (nl, il, es { esFailed = (idx, msg):esFailed es})
1148 updateEvacSolution (_, _, es) idx (Ok (nl, il, opcodes)) =
1149 (nl, il, es { esMoved = new_elem:esMoved es
1150 , esOpCodes = opcodes:esOpCodes es })
1151 where inst = Container.find idx il
1153 instancePriGroup nl inst,
1154 Instance.allNodes inst)
1156 -- | Node-evacuation IAllocator mode main function.
1157 tryNodeEvac :: Group.List -- ^ The cluster groups
1158 -> Node.List -- ^ The node list (cluster-wide, not per group)
1159 -> Instance.List -- ^ Instance list (cluster-wide)
1160 -> EvacMode -- ^ The evacuation mode
1161 -> [Idx] -- ^ List of instance (indices) to be evacuated
1162 -> Result (Node.List, Instance.List, EvacSolution)
1163 tryNodeEvac _ ini_nl ini_il mode idxs =
1164 let evac_ndx = nodesToEvacuate ini_il mode idxs
1165 offline = map Node.idx . filter Node.offline $ Container.elems ini_nl
1166 excl_ndx = foldl' (flip IntSet.insert) evac_ndx offline
1167 group_ndx = map (\(gdx, (nl, _)) -> (gdx, map Node.idx
1168 (Container.elems nl))) $
1169 splitCluster ini_nl ini_il
1170 (fin_nl, fin_il, esol) =
1171 foldl' (\state@(nl, il, _) inst ->
1172 let gdx = instancePriGroup nl inst
1173 pdx = Instance.pNode inst in
1174 updateEvacSolution state (Instance.idx inst) $
1175 availableGroupNodes group_ndx
1176 (IntSet.insert pdx excl_ndx) gdx >>=
1177 nodeEvacInstance nl il mode inst gdx
1179 (ini_nl, ini_il, emptyEvacSolution)
1180 (map (`Container.find` ini_il) idxs)
1181 in return (fin_nl, fin_il, reverseEvacSolution esol)
1183 -- | Change-group IAllocator mode main function.
1185 -- This is very similar to 'tryNodeEvac', the only difference is that
1186 -- we don't choose as target group the current instance group, but
1189 -- 1. at the start of the function, we compute which are the target
1190 -- groups; either no groups were passed in, in which case we choose
1191 -- all groups out of which we don't evacuate instance, or there were
1192 -- some groups passed, in which case we use those
1194 -- 2. for each instance, we use 'findBestAllocGroup' to choose the
1195 -- best group to hold the instance, and then we do what
1196 -- 'tryNodeEvac' does, except for this group instead of the current
1199 -- Note that the correct behaviour of this function relies on the
1200 -- function 'nodeEvacInstance' to be able to do correctly both
1201 -- intra-group and inter-group moves when passed the 'ChangeAll' mode.
1202 tryChangeGroup :: Group.List -- ^ The cluster groups
1203 -> Node.List -- ^ The node list (cluster-wide)
1204 -> Instance.List -- ^ Instance list (cluster-wide)
1205 -> [Gdx] -- ^ Target groups; if empty, any
1206 -- groups not being evacuated
1207 -> [Idx] -- ^ List of instance (indices) to be evacuated
1208 -> Result (Node.List, Instance.List, EvacSolution)
1209 tryChangeGroup gl ini_nl ini_il gdxs idxs =
1210 let evac_gdxs = nub $ map (instancePriGroup ini_nl .
1211 flip Container.find ini_il) idxs
1212 target_gdxs = (if null gdxs
1213 then Container.keys gl
1214 else gdxs) \\ evac_gdxs
1215 offline = map Node.idx . filter Node.offline $ Container.elems ini_nl
1216 excl_ndx = foldl' (flip IntSet.insert) IntSet.empty offline
1217 group_ndx = map (\(gdx, (nl, _)) -> (gdx, map Node.idx
1218 (Container.elems nl))) $
1219 splitCluster ini_nl ini_il
1220 (fin_nl, fin_il, esol) =
1221 foldl' (\state@(nl, il, _) inst ->
1223 let ncnt = Instance.requiredNodes $
1224 Instance.diskTemplate inst
1225 (gdx, _, _) <- findBestAllocGroup gl nl il
1226 (Just target_gdxs) inst ncnt
1227 av_nodes <- availableGroupNodes group_ndx
1229 nodeEvacInstance nl il ChangeAll inst gdx av_nodes
1230 in updateEvacSolution state (Instance.idx inst) solution
1232 (ini_nl, ini_il, emptyEvacSolution)
1233 (map (`Container.find` ini_il) idxs)
1234 in return (fin_nl, fin_il, reverseEvacSolution esol)
1236 -- | Standard-sized allocation method.
1238 -- This places instances of the same size on the cluster until we're
1239 -- out of space. The result will be a list of identically-sized
1241 iterateAlloc :: AllocMethod
1242 iterateAlloc nl il limit newinst allocnodes ixes cstats =
1243 let depth = length ixes
1244 newname = printf "new-%d" depth::String
1245 newidx = Container.size il
1246 newi2 = Instance.setIdx (Instance.setName newinst newname) newidx
1247 newlimit = fmap (flip (-) 1) limit
1248 in case tryAlloc nl il newi2 allocnodes of
1250 Ok (AllocSolution { asFailures = errs, asSolution = sols3 }) ->
1251 let newsol = Ok (collapseFailures errs, nl, il, ixes, cstats) in
1254 Just (xnl, xi, _, _) ->
1257 else iterateAlloc xnl (Container.add newidx xi il)
1258 newlimit newinst allocnodes (xi:ixes)
1259 (totalResources xnl:cstats)
1261 -- | Tiered allocation method.
1263 -- This places instances on the cluster, and decreases the spec until
1264 -- we can allocate again. The result will be a list of decreasing
1266 tieredAlloc :: AllocMethod
1267 tieredAlloc nl il limit newinst allocnodes ixes cstats =
1268 case iterateAlloc nl il limit newinst allocnodes ixes cstats of
1270 Ok (errs, nl', il', ixes', cstats') ->
1271 let newsol = Ok (errs, nl', il', ixes', cstats')
1272 ixes_cnt = length ixes'
1273 (stop, newlimit) = case limit of
1274 Nothing -> (False, Nothing)
1275 Just n -> (n <= ixes_cnt,
1276 Just (n - ixes_cnt)) in
1277 if stop then newsol else
1278 case Instance.shrinkByType newinst . fst . last $
1279 sortBy (comparing snd) errs of
1281 Ok newinst' -> tieredAlloc nl' il' newlimit
1282 newinst' allocnodes ixes' cstats'
1284 -- * Formatting functions
1286 -- | Given the original and final nodes, computes the relocation description.
1287 computeMoves :: Instance.Instance -- ^ The instance to be moved
1288 -> String -- ^ The instance name
1289 -> IMove -- ^ The move being performed
1290 -> String -- ^ New primary
1291 -> String -- ^ New secondary
1292 -> (String, [String])
1293 -- ^ Tuple of moves and commands list; moves is containing
1294 -- either @/f/@ for failover or @/r:name/@ for replace
1295 -- secondary, while the command list holds gnt-instance
1296 -- commands (without that prefix), e.g \"@failover instance1@\"
1297 computeMoves i inam mv c d =
1299 Failover -> ("f", [mig])
1300 FailoverToAny _ -> (printf "fa:%s" c, [mig_any])
1301 FailoverAndReplace _ -> (printf "f r:%s" d, [mig, rep d])
1302 ReplaceSecondary _ -> (printf "r:%s" d, [rep d])
1303 ReplaceAndFailover _ -> (printf "r:%s f" c, [rep c, mig])
1304 ReplacePrimary _ -> (printf "f r:%s f" c, [mig, rep c, mig])
1305 where morf = if Instance.isRunning i then "migrate" else "failover"
1306 mig = printf "%s -f %s" morf inam::String
1307 mig_any = printf "%s -f -n %s %s" morf c inam::String
1308 rep n = printf "replace-disks -n %s %s" n inam::String
1310 -- | Converts a placement to string format.
1311 printSolutionLine :: Node.List -- ^ The node list
1312 -> Instance.List -- ^ The instance list
1313 -> Int -- ^ Maximum node name length
1314 -> Int -- ^ Maximum instance name length
1315 -> Placement -- ^ The current placement
1316 -> Int -- ^ The index of the placement in
1318 -> (String, [String])
1319 printSolutionLine nl il nmlen imlen plc pos =
1320 let pmlen = (2*nmlen + 1)
1321 (i, p, s, mv, c) = plc
1322 old_sec = Instance.sNode inst
1323 inst = Container.find i il
1324 inam = Instance.alias inst
1325 npri = Node.alias $ Container.find p nl
1326 nsec = Node.alias $ Container.find s nl
1327 opri = Node.alias $ Container.find (Instance.pNode inst) nl
1328 osec = Node.alias $ Container.find old_sec nl
1329 (moves, cmds) = computeMoves inst inam mv npri nsec
1330 -- FIXME: this should check instead/also the disk template
1331 ostr = if old_sec == Node.noSecondary
1332 then printf "%s" opri::String
1333 else printf "%s:%s" opri osec::String
1334 nstr = if s == Node.noSecondary
1335 then printf "%s" npri::String
1336 else printf "%s:%s" npri nsec::String
1337 in (printf " %3d. %-*s %-*s => %-*s %12.8f a=%s"
1338 pos imlen inam pmlen ostr pmlen nstr c moves,
1341 -- | Return the instance and involved nodes in an instance move.
1343 -- Note that the output list length can vary, and is not required nor
1344 -- guaranteed to be of any specific length.
1345 involvedNodes :: Instance.List -- ^ Instance list, used for retrieving
1346 -- the instance from its index; note
1347 -- that this /must/ be the original
1348 -- instance list, so that we can
1349 -- retrieve the old nodes
1350 -> Placement -- ^ The placement we're investigating,
1351 -- containing the new nodes and
1353 -> [Ndx] -- ^ Resulting list of node indices
1354 involvedNodes il plc =
1355 let (i, np, ns, _, _) = plc
1356 inst = Container.find i il
1357 in nub $ [np, ns] ++ Instance.allNodes inst
1359 -- | Inner function for splitJobs, that either appends the next job to
1360 -- the current jobset, or starts a new jobset.
1361 mergeJobs :: ([JobSet], [Ndx]) -> MoveJob -> ([JobSet], [Ndx])
1362 mergeJobs ([], _) n@(ndx, _, _, _) = ([[n]], ndx)
1363 mergeJobs (cjs@(j:js), nbuf) n@(ndx, _, _, _)
1364 | null (ndx `intersect` nbuf) = ((n:j):js, ndx ++ nbuf)
1365 | otherwise = ([n]:cjs, ndx)
1367 -- | Break a list of moves into independent groups. Note that this
1368 -- will reverse the order of jobs.
1369 splitJobs :: [MoveJob] -> [JobSet]
1370 splitJobs = fst . foldl mergeJobs ([], [])
1372 -- | Given a list of commands, prefix them with @gnt-instance@ and
1373 -- also beautify the display a little.
1374 formatJob :: Int -> Int -> (Int, MoveJob) -> [String]
1375 formatJob jsn jsl (sn, (_, _, _, cmds)) =
1377 printf " echo job %d/%d" jsn sn:
1379 map (" gnt-instance " ++) cmds
1381 then ["", printf "echo jobset %d, %d jobs" jsn jsl] ++ out
1384 -- | Given a list of commands, prefix them with @gnt-instance@ and
1385 -- also beautify the display a little.
1386 formatCmds :: [JobSet] -> String
1389 concatMap (\(jsn, js) -> concatMap (formatJob jsn (length js))
1393 -- | Print the node list.
1394 printNodes :: Node.List -> [String] -> String
1396 let fields = case fs of
1397 [] -> Node.defaultFields
1398 "+":rest -> Node.defaultFields ++ rest
1400 snl = sortBy (comparing Node.idx) (Container.elems nl)
1401 (header, isnum) = unzip $ map Node.showHeader fields
1402 in printTable "" header (map (Node.list fields) snl) isnum
1404 -- | Print the instance list.
1405 printInsts :: Node.List -> Instance.List -> String
1407 let sil = sortBy (comparing Instance.idx) (Container.elems il)
1408 helper inst = [ if Instance.isRunning inst then "R" else " "
1409 , Instance.name inst
1410 , Container.nameOf nl (Instance.pNode inst)
1411 , let sdx = Instance.sNode inst
1412 in if sdx == Node.noSecondary
1414 else Container.nameOf nl sdx
1415 , if Instance.autoBalance inst then "Y" else "N"
1416 , printf "%3d" $ Instance.vcpus inst
1417 , printf "%5d" $ Instance.mem inst
1418 , printf "%5d" $ Instance.dsk inst `div` 1024
1424 where DynUtil lC lM lD lN = Instance.util inst
1425 header = [ "F", "Name", "Pri_node", "Sec_node", "Auto_bal"
1426 , "vcpu", "mem" , "dsk", "lCpu", "lMem", "lDsk", "lNet" ]
1427 isnum = False:False:False:False:False:repeat True
1428 in printTable "" header (map helper sil) isnum
1430 -- | Shows statistics for a given node list.
1431 printStats :: String -> Node.List -> String
1433 let dcvs = compDetailedCV $ Container.elems nl
1434 (weights, names) = unzip detailedCVInfo
1435 hd = zip3 (weights ++ repeat 1) (names ++ repeat "unknown") dcvs
1436 header = [ "Field", "Value", "Weight" ]
1437 formatted = map (\(w, h, val) ->
1442 in printTable lp header formatted $ False:repeat True
1444 -- | Convert a placement into a list of OpCodes (basically a job).
1445 iMoveToJob :: Node.List -- ^ The node list; only used for node
1446 -- names, so any version is good
1447 -- (before or after the operation)
1448 -> Instance.List -- ^ The instance list; also used for
1450 -> Idx -- ^ The index of the instance being
1452 -> IMove -- ^ The actual move to be described
1453 -> [OpCodes.OpCode] -- ^ The list of opcodes equivalent to
1455 iMoveToJob nl il idx move =
1456 let inst = Container.find idx il
1457 iname = Instance.name inst
1458 lookNode n = case mkNonEmpty (Container.nameOf nl n) of
1459 -- FIXME: convert htools codebase to non-empty strings
1460 Bad msg -> error $ "Empty node name for idx " ++
1461 show n ++ ": " ++ msg ++ "??"
1463 opF = OpCodes.OpInstanceMigrate
1464 { OpCodes.opInstanceName = iname
1465 , OpCodes.opMigrationMode = Nothing -- default
1466 , OpCodes.opOldLiveMode = Nothing -- default as well
1467 , OpCodes.opTargetNode = Nothing -- this is drbd
1468 , OpCodes.opAllowRuntimeChanges = False
1469 , OpCodes.opIgnoreIpolicy = False
1470 , OpCodes.opMigrationCleanup = False
1471 , OpCodes.opIallocator = Nothing
1472 , OpCodes.opAllowFailover = True }
1473 opFA n = opF { OpCodes.opTargetNode = lookNode n } -- not drbd
1474 opR n = OpCodes.OpInstanceReplaceDisks
1475 { OpCodes.opInstanceName = iname
1476 , OpCodes.opEarlyRelease = False
1477 , OpCodes.opIgnoreIpolicy = False
1478 , OpCodes.opReplaceDisksMode = OpCodes.ReplaceNewSecondary
1479 , OpCodes.opReplaceDisksList = []
1480 , OpCodes.opRemoteNode = lookNode n
1481 , OpCodes.opIallocator = Nothing
1485 FailoverToAny np -> [ opFA np ]
1486 ReplacePrimary np -> [ opF, opR np, opF ]
1487 ReplaceSecondary ns -> [ opR ns ]
1488 ReplaceAndFailover np -> [ opR np, opF ]
1489 FailoverAndReplace ns -> [ opF, opR ns ]
1491 -- * Node group functions
1493 -- | Computes the group of an instance.
1494 instanceGroup :: Node.List -> Instance.Instance -> Result Gdx
1495 instanceGroup nl i =
1496 let sidx = Instance.sNode i
1497 pnode = Container.find (Instance.pNode i) nl
1498 snode = if sidx == Node.noSecondary
1500 else Container.find sidx nl
1501 pgroup = Node.group pnode
1502 sgroup = Node.group snode
1503 in if pgroup /= sgroup
1504 then fail ("Instance placed accross two node groups, primary " ++
1505 show pgroup ++ ", secondary " ++ show sgroup)
1508 -- | Computes the group of an instance per the primary node.
1509 instancePriGroup :: Node.List -> Instance.Instance -> Gdx
1510 instancePriGroup nl i =
1511 let pnode = Container.find (Instance.pNode i) nl
1514 -- | Compute the list of badly allocated instances (split across node
1516 findSplitInstances :: Node.List -> Instance.List -> [Instance.Instance]
1517 findSplitInstances nl =
1518 filter (not . isOk . instanceGroup nl) . Container.elems
1520 -- | Splits a cluster into the component node groups.
1521 splitCluster :: Node.List -> Instance.List ->
1522 [(Gdx, (Node.List, Instance.List))]
1523 splitCluster nl il =
1524 let ngroups = Node.computeGroups (Container.elems nl)
1525 in map (\(guuid, nodes) ->
1526 let nidxs = map Node.idx nodes
1527 nodes' = zip nidxs nodes
1528 instances = Container.filter ((`elem` nidxs) . Instance.pNode) il
1529 in (guuid, (Container.fromList nodes', instances))) ngroups
1531 -- | Compute the list of nodes that are to be evacuated, given a list
1532 -- of instances and an evacuation mode.
1533 nodesToEvacuate :: Instance.List -- ^ The cluster-wide instance list
1534 -> EvacMode -- ^ The evacuation mode we're using
1535 -> [Idx] -- ^ List of instance indices being evacuated
1536 -> IntSet.IntSet -- ^ Set of node indices
1537 nodesToEvacuate il mode =
1538 IntSet.delete Node.noSecondary .
1540 let i = Container.find idx il
1541 pdx = Instance.pNode i
1542 sdx = Instance.sNode i
1543 dt = Instance.diskTemplate i
1544 withSecondary = case dt of
1545 DTDrbd8 -> IntSet.insert sdx ns
1548 ChangePrimary -> IntSet.insert pdx ns
1549 ChangeSecondary -> withSecondary
1550 ChangeAll -> IntSet.insert pdx withSecondary