seaweedfs/weed/storage/erasure_coding/ecbalancer/place.go

package ecbalancer

import (
	"fmt"
	"math"
	"sort"
	"strings"

	"github.com/seaweedfs/seaweedfs/weed/storage/erasure_coding"
	"github.com/seaweedfs/seaweedfs/weed/storage/super_block"
	storagetypes "github.com/seaweedfs/seaweedfs/weed/storage/types"
)

// Constraints configures a Place call. Ratio resolves a collection's
// (dataShards, parityShards); nil uses the standard scheme. ReplicaPlacement,
// when non-nil, caps shards per rack (DiffRackCount = max shards/rack) and per
// node within a rack (SameRackCount = max shards/node); both digits are direct
// hard caps. The data-center digit (DiffDataCenterCount) is not honored:
// the 1-byte volume ReplicaPlacement can only encode 0-2 there, too small to be a
// meaningful per-DC EC shard cap, so EC relies on the rack/node even spread instead.
//
// DiskTypePolicy controls how DiskType constrains placement (Any / Prefer /
// Require). PreferredTags drives whole-plan tag tiering: Place tries disks
// carrying the earliest tags first and widens to all disks only if a tier cannot
// place every shard.
type Constraints struct {
	DiskType         string
	DiskTypePolicy   DiskTypePolicy
	PreferredTags    []string
	ReplicaPlacement *super_block.ReplicaPlacement
	Ratio            func(collection string) (dataShards, parityShards int)
	// FreeSlotsDeadband (M) is free volume slots. After even-this-volume,
	// Place ranks that evenness class by maxFree, keeps |free−maxFree| ≤
	// M×dataShards, then lower liveShards. 0 is only maxFree then live
	// (old strict freeSlots; liveShards only on an exact free tie).
	FreeSlotsDeadband int
}

// DiskTypePolicy controls how Constraints.DiskType constrains placement.
type DiskTypePolicy int

const (
	DiskTypeAny     DiskTypePolicy = iota // any disk type
	DiskTypePrefer                        // prefer DiskType, spill to other types if needed
	DiskTypeRequire                       // only DiskType (HardDriveType when "")
)

// diskTypeEqual compares disk types after normalization, so "" and "hdd" (both
// HardDriveType) are equal.
func diskTypeEqual(a, b string) bool {
	return storagetypes.ToDiskType(a).String() == storagetypes.ToDiskType(b).String()
}

// diskHasAnyTag reports whether the disk carries any of the given tags.
func diskHasAnyTag(d *disk, tags []string) bool {
	for _, want := range tags {
		for _, have := range d.tags {
			if have == want {
				return true
			}
		}
	}
	return false
}

// Destination is a chosen target for one shard. DataCenter and Rack are kept as
// separate values (matching topology.DiskInfo) rather than a "dc:rack" composite,
// so callers read them directly instead of parsing.
type Destination struct {
	Node       string
	DiskID     uint32
	DataCenter string
	Rack       string // bare rack id within DataCenter
}

// PlaceResult holds the chosen destinations, which constraints had to be relaxed
// (durability-first only), and whether placement spilled outside the preferred
// disk type or tag tiers (for parity with today's logging).
type PlaceResult struct {
	Destinations                map[int]Destination
	Relaxed                     []string
	SpilledToOtherDiskType      bool
	SpilledOutsidePreferredTags bool
}

// PlacementMode selects the strictness/relaxation policy.
type PlacementMode int

const (
	// PlaceStrict: caps, the total-shards-per-rack cap and ReplicaPlacement are
	// hard. Place fails rather than violate them, so the caller can defer (leave
	// the volume as-is and retry).
	PlaceStrict PlacementMode = iota
	// PlaceDurabilityFirst (used by both encode and repair): relax per-type caps ->
	// data/parity anti-affinity -> ReplicaPlacement -> the total-shards-per-rack
	// cap, in that order, until each shard lands, reporting what was relaxed in
	// PlaceResult.Relaxed. The per-disk durability cap (<= parityShards per disk)
	// is never relaxed. Fails only if no eligible disk has room for another shard
	// of the volume. Encode places best-effort this way and rebalancing tightens
	// the spread afterward.
	PlaceDurabilityFirst
)

// relaxation controls which placement-quality constraints are enforced on an
// attempt. preferring fresh nodes (repair's "avoid surviving-shard nodes") is not
// listed: pickNodeInRack already selects the node with the fewest shards of the
// volume, so survivors are deprioritized with built-in fallback.
type relaxation struct {
	caps         bool
	antiAffinity bool
	rp           bool
	rackTotal    bool
}

func (r relaxation) relaxedNames() []string {
	var n []string
	if !r.caps {
		n = append(n, "caps")
	}
	if !r.antiAffinity {
		n = append(n, "anti-affinity")
	}
	if !r.rp {
		n = append(n, "replica-placement")
	}
	if !r.rackTotal {
		n = append(n, "rack-total-cap")
	}
	return n
}

var strictAttempts = []relaxation{{caps: true, antiAffinity: true, rp: true, rackTotal: true}}

// The total-shards-per-rack cap is sized from real capacity (see rackTotalCap),
// so it is relaxed last and only as a safety net; with it dropped, the per-disk
// cap is the only constraint left.
var durabilityAttempts = []relaxation{
	{caps: true, antiAffinity: true, rp: true, rackTotal: true},
	{caps: false, antiAffinity: true, rp: true, rackTotal: true},
	{caps: false, antiAffinity: false, rp: true, rackTotal: true},
	{caps: false, antiAffinity: false, rp: false, rackTotal: true},
	{caps: false, antiAffinity: false, rp: false, rackTotal: false},
}

type placedEntry struct {
	node    *Node
	sid     int
	rackKey string
}

// Place assigns destinations for the `need` shard ids of volume (collection,vid),
// reading the volume's already-placed shards from the snapshot (so encode passes
// an empty-for-this-volume snapshot, repair passes one seeded with the surviving
// shards).
//
// Tag tiering (whole-plan retry): it tries the preferred-tag tiers in order, each
// a complete candidate set, and returns the first tier that places every shard;
// only when it falls through to the no-tag tier does it set
// SpilledOutsidePreferredTags. Within a tier, disk-type Prefer spills to other
// types per shard (SpilledToOtherDiskType); Require filters strictly.
func (t *Topology) Place(vid uint32, collection string, need []int, c Constraints, mode PlacementMode) (*PlaceResult, error) {
	if len(need) == 0 {
		return &PlaceResult{Destinations: map[int]Destination{}}, nil
	}

	vk := volKey{collection: collection, vid: vid}
	dataShards, parityShards := erasure_coding.DataShardsCount, erasure_coding.ParityShardsCount
	if c.Ratio != nil {
		if d, p := c.Ratio(collection); d > 0 && p > 0 {
			dataShards, parityShards = d, p
		}
	}

	racks := buildRacks(t.nodes)
	if len(racks) == 0 {
		return nil, fmt.Errorf("no racks available for EC placement")
	}
	rackKeys := sortedKeys(racks)

	// Disk-type eligibility (Require filters; Any/Prefer admit all) and the soft
	// type preference applied in scoring under Prefer.
	typeEligible := func(d *disk) bool {
		if c.DiskTypePolicy == DiskTypeRequire {
			return diskTypeEqual(d.diskType, c.DiskType)
		}
		return true
	}
	var prefer func(*disk) bool
	if c.DiskTypePolicy == DiskTypePrefer {
		prefer = func(d *disk) bool { return diskTypeEqual(d.diskType, c.DiskType) }
	}

	// Whole-plan retry over preferred-tag tiers; the first tier that places every
	// shard wins. Reaching the no-tag tier means we spilled outside the tags.
	tiers := tagTiers(c.PreferredTags)
	var lastErr error
	for _, tierTags := range tiers {
		tt := tierTags
		eligible := func(d *disk) bool {
			return typeEligible(d) && (len(tt) == 0 || diskHasAnyTag(d, tt))
		}
		res, err := t.tryPlace(vk, need, dataShards, parityShards, racks, rackKeys, mode, c.ReplicaPlacement, eligible, prefer, volumeDeadbandToShardSlots(c.FreeSlotsDeadband, dataShards))
		if err != nil {
			lastErr = err
			continue
		}
		if len(c.PreferredTags) > 0 && len(tierTags) == 0 {
			res.SpilledOutsidePreferredTags = true
		}
		return res, nil
	}
	return nil, lastErr
}

// tagTiers returns the eligibility tag-sets in increasing breadth, ending with an
// empty set ("any disk"). Empty preferredTags yields a single any-disk tier.
func tagTiers(preferredTags []string) [][]string {
	if len(preferredTags) == 0 {
		return [][]string{nil}
	}
	tiers := make([][]string, 0, len(preferredTags)+1)
	for k := range preferredTags {
		tiers = append(tiers, append([]string(nil), preferredTags[:k+1]...))
	}
	return append(tiers, nil)
}

// tryPlace runs one whole-plan placement attempt restricted to disks satisfying
// `eligible`, with `prefer` (may be nil) ranking soft-preferred disks first. It
// journals reservations and rolls them all back if any shard cannot be placed, so
// a failed tier leaves the snapshot unchanged for the next attempt.
func (t *Topology) tryPlace(vk volKey, need []int, dataShards, parityShards int, racks map[string]*rack, rackKeys []string, mode PlacementMode, rp *super_block.ReplicaPlacement, eligible func(*disk) bool, prefer func(*disk) bool, freeSlotsDeadband int) (*PlaceResult, error) {
	result := &PlaceResult{Destinations: make(map[int]Destination, len(need))}

	// Per-type shard ids per rack (even caps), total shard count per rack
	// (DiffRackCount), and the racks bearing each type (anti-affinity) — all seeded
	// from the volume's existing shards.
	shardsPerRack := map[bool]map[string][]int{true: {}, false: {}}
	rackShardCount := map[string]int{}
	bearing := map[bool]map[string]bool{true: {}, false: {}}
	for _, n := range t.nodes {
		info, ok := n.shards[vk]
		if !ok {
			continue
		}
		for sid := range info.shardBits.All() {
			s := int(sid)
			isData := s < dataShards
			shardsPerRack[isData][n.rack] = append(shardsPerRack[isData][n.rack], s)
			rackShardCount[n.rack]++
			bearing[isData][n.rack] = true
		}
	}

	// Even per-rack caps divide by racks that can actually take another shard of
	// the volume, not all racks (the snapshot keeps every disk type/tag), so a
	// valid tiered cluster — e.g. SSDs in only 2 of 4 racks — is not capped
	// impossibly low.
	rackRoom := make(map[string]int, len(rackKeys))
	rackRoomRP := make(map[string]int, len(rackKeys))
	numEligibleRacks := 0
	for _, rk := range rackKeys {
		rackRoom[rk] = rackShardRoom(racks[rk], vk, eligible, parityShards)
		rackRoomRP[rk] = rackShardRoomUnderRP(racks[rk], vk, eligible, parityShards, rp)
		if rackRoom[rk] > 0 {
			numEligibleRacks++
		}
	}
	if numEligibleRacks < 1 {
		numEligibleRacks = 1
	}

	// The per-type caps spread data and parity independently, so together they
	// can still stack e.g. 2 data + 1 parity on one rack. Cap the TOTAL per rack
	// as well, at the lowest value the racks' free capacity allows. Attempts that
	// enforce ReplicaPlacement size it from the room left under SameRackCount, so
	// a rack of few nodes does not count disk room its nodes may not use.
	totalShards := len(need)
	for _, n := range rackShardCount {
		totalShards += n
	}
	maxTotalPerRack := rackTotalCap(rackKeys, rackShardCount, rackRoom, totalShards)
	maxTotalPerRackRP := rackTotalCap(rackKeys, rackShardCount, rackRoomRP, totalShards)
	rackCap := func(rl relaxation) int {
		if rl.rp {
			return maxTotalPerRackRP
		}
		return maxTotalPerRack
	}

	attempts := strictAttempts
	if mode == PlaceDurabilityFirst {
		attempts = durabilityAttempts
	}

	var journal []placedEntry
	relaxedSeen := map[string]bool{}
	spilledType := false

	placeShard := func(sid int, isData bool) bool {
		typeTotal := dataShards
		if !isData {
			typeTotal = parityShards
		}
		for _, rl := range attempts {
			node, diskID, spilled, ok := chooseShardDest(vk, sid, isData, dataShards, typeTotal, numEligibleRacks, parityShards, rackCap(rl), racks, rackKeys, rp, eligible, prefer, shardsPerRack[isData], rackShardCount, bearing, rl, freeSlotsDeadband)
			if !ok {
				continue
			}
			reserveShard(node, vk, sid, diskID)
			node.freeSlots--
			racks[node.rack].freeSlots--
			shardsPerRack[isData][node.rack] = append(shardsPerRack[isData][node.rack], sid)
			rackShardCount[node.rack]++
			bearing[isData][node.rack] = true
			journal = append(journal, placedEntry{node: node, sid: sid, rackKey: node.rack})
			result.Destinations[sid] = Destination{
				Node:       node.id,
				DiskID:     diskID,
				DataCenter: node.dc,
				Rack:       strings.TrimPrefix(node.rack, node.dc+":"),
			}
			if spilled {
				spilledType = true
			}
			for _, name := range rl.relaxedNames() {
				relaxedSeen[name] = true
			}
			return true
		}
		return false
	}

	// Data shards first, then parity, so parity can avoid data-bearing racks.
	for _, isData := range []bool{true, false} {
		for _, sid := range shardsOfType(need, isData, dataShards) {
			if placeShard(sid, isData) {
				continue
			}
			for _, e := range journal {
				releaseShard(e.node, vk, e.sid)
				e.node.freeSlots++
				racks[e.rackKey].freeSlots++
			}
			return nil, fmt.Errorf("cannot place EC shard %d of volume %d (collection %q) (rack total cap %d, per-disk cap %d)", sid, vk.vid, vk.collection, rackCap(attempts[len(attempts)-1]), parityShards)
		}
	}

	result.SpilledToOtherDiskType = spilledType
	for name := range relaxedSeen {
		result.Relaxed = append(result.Relaxed, name)
	}
	sort.Strings(result.Relaxed)
	return result, nil
}

// chooseShardDest selects a (node, disk) for one shard at the given relaxation
// level: pick a rack (total-shards-per-rack cap + even per-type cap +
// ReplicaPlacement caps + two-pass anti-affinity to the opposite type, then
// freeSlots deadband / liveShards), then the least-loaded eligible node, then
// the best eligible disk. If no node in the chosen rack fits (e.g. every node
// is at SameRackCount), the next-best rack is tried. The third return reports
// whether the disk spilled off the soft-preferred type. ok=false when no
// rack/node/disk fits.
func chooseShardDest(vk volKey, sid int, isData bool, dataShards, typeTotal, numEligibleRacks, maxPerDisk, maxTotalPerRack int, racks map[string]*rack, rackKeys []string, rp *super_block.ReplicaPlacement, eligible func(*disk) bool, prefer func(*disk) bool, shardsPerRackType map[string][]int, rackShardCount map[string]int, bearing map[bool]map[string]bool, rl relaxation, freeSlotsDeadband int) (*Node, uint32, bool, bool) {
	maxPerRack := numEligibleRacks*typeTotal + 1 // effectively unlimited when caps are relaxed
	if rl.caps {
		if maxPerRack = ceilDivide(typeTotal, numEligibleRacks); maxPerRack < 1 {
			maxPerRack = 1
		}
	}

	var anti map[string]bool
	if rl.antiAffinity {
		anti = bearing[!isData] // racks already holding the opposite shard type
	}

	if !rl.rp {
		rp = nil
	}
	// A rack is eligible only if it is under the total-shards-per-rack cap and,
	// when set, the per-rack shard cap (DiffRackCount). The total cap does not
	// depend on ReplicaPlacement, so it also holds when rp is nil.
	withinLimit := func(r string) bool {
		if rl.rackTotal && rackShardCount[r] >= maxTotalPerRack {
			return false
		}
		if rp != nil && rp.DiffRackCount > 0 && rackShardCount[r] >= rp.DiffRackCount {
			return false
		}
		return true
	}

	tried := map[string]bool{}
	hasRoom := func(r string) bool {
		return !tried[r] && racks[r].freeSlots > 0 && rackShardRoom(racks[r], vk, eligible, maxPerDisk) > 0
	}
	for {
		destRack, ok := pickRackEligible(rackKeys, shardsPerRackType, maxPerRack, anti, hasRoom, withinLimit, racks, freeSlotsDeadband)
		if !ok {
			return nil, 0, false, false
		}
		if node := pickNodeInRackEligible(racks[destRack], vk, rp, eligible, maxPerDisk, freeSlotsDeadband); node != nil {
			if diskID, ok, spilled := pickBestDiskEligible(node, vk, eligible, prefer, sid, dataShards, maxPerDisk); ok {
				return node, diskID, spilled, true
			}
		}
		tried[destRack] = true
	}
}

// rackTotalCap returns the smallest per-rack total c such that the racks can
// hold totalShards shards of the volume with none above c, given each rack's
// shards already placed (held) and its room for more. On a uniform cluster this
// is ceil(totalShards/racks); a nearly full rack raises it just enough for the
// other racks to absorb its share, so the cap alone never makes a feasible
// placement fail.
//
// It is the most even spread the free capacity allows, not a durability
// guarantee: losing k racks loses up to k*c shards, which the volume survives
// only while k*c <= parityShards. With few racks no placement can achieve that.
func rackTotalCap(rackKeys []string, held, room map[string]int, totalShards int) int {
	for c := 1; c < totalShards; c++ {
		fits := 0
		for _, rk := range rackKeys {
			fits += max(held[rk], min(c, held[rk]+room[rk]))
		}
		if fits >= totalShards {
			return c
		}
	}
	return totalShards
}

// diskShardRoom returns how many more shards of the volume the disk can take:
// its free slots, bounded by the per-disk durability cap (maxPerDisk shards of
// one volume per disk; <= 0 disables it).
func diskShardRoom(n *Node, d *disk, vk volKey, maxPerDisk int) int {
	if d.noReceive {
		return 0
	}
	room := d.freeSlots
	if maxPerDisk > 0 {
		held := 0
		if info := n.shards[vk]; info != nil {
			held = info.diskShardBits[d.diskID].Count()
		}
		room = min(room, maxPerDisk-held)
	}
	return max(room, 0)
}

// nodeShardRoom returns how many more shards of the volume the node's eligible
// disks can take, bounded by the node's free slots.
func nodeShardRoom(n *Node, vk volKey, eligible func(*disk) bool, maxPerDisk int) int {
	if n.noReceive {
		return 0
	}
	room := 0
	for _, d := range n.disks {
		if eligible(d) {
			room += diskShardRoom(n, d, vk, maxPerDisk)
		}
	}
	return max(min(room, n.receiveSlots()), 0)
}

// rackShardRoom returns how many more shards of the volume the rack can take.
func rackShardRoom(r *rack, vk volKey, eligible func(*disk) bool, maxPerDisk int) int {
	room := 0
	for _, n := range r.nodes {
		room += nodeShardRoom(n, vk, eligible, maxPerDisk)
	}
	return room
}

// rackShardRoomUnderRP is rackShardRoom with each node further bounded by the
// shards it may still take under ReplicaPlacement's SameRackCount (max shards of
// the volume per node), which pickNodeInRackEligible enforces.
func rackShardRoomUnderRP(r *rack, vk volKey, eligible func(*disk) bool, maxPerDisk int, rp *super_block.ReplicaPlacement) int {
	if rp == nil || rp.SameRackCount <= 0 {
		return rackShardRoom(r, vk, eligible, maxPerDisk)
	}
	room := 0
	for _, n := range r.nodes {
		room += min(nodeShardRoom(n, vk, eligible, maxPerDisk), max(rp.SameRackCount-volumeShardCount(n, vk), 0))
	}
	return room
}

// pickRackEligible is pickTarget for Place racks. Even this-type / cap /
// two-pass anti-affinity stay. Among that evenness class the remainder is
// max rack freeSlots, keep |free−maxFree| ≤ shard-slot deadband, then
// lower liveShards (sum of nodeLiveShardCount). M=0 is only maxFree then
// live (old strict freeSlots). pickTarget itself is unchanged so balancer
// / DC paths stay without M.
func pickRackEligible(candidates []string, shardsPerTarget map[string][]int, maxPerTarget int, antiAffinity map[string]bool, hasFreeSlots, withinLimit func(string) bool, racks map[string]*rack, freeSlotsDeadband int) (string, bool) {
	try := func(skipAnti bool) (string, bool) {
		var keys []string
		var free, live []int
		bestCount := maxPerTarget + 1
		for _, c := range candidates {
			if skipAnti && antiAffinity[c] {
				continue
			}
			if !hasFreeSlots(c) {
				continue
			}
			if len(shardsPerTarget[c]) >= maxPerTarget {
				continue
			}
			if !withinLimit(c) {
				continue
			}
			cnt := len(shardsPerTarget[c])
			if len(keys) == 0 || cnt < bestCount {
				keys, free, live = keys[:0], free[:0], live[:0]
				bestCount = cnt
			} else if cnt > bestCount {
				continue
			}
			f := 0
			if r := racks[c]; r != nil {
				f = r.freeSlots
			}
			keys = append(keys, c)
			free = append(free, f)
			live = append(live, rackLiveShardCount(racks[c]))
		}
		if len(keys) == 0 {
			return "", false
		}
		return keys[pickDeadbandRemainder(free, live, freeSlotsDeadband)], true
	}
	if len(antiAffinity) > 0 {
		if t, ok := try(true); ok {
			return t, true
		}
	}
	return try(false)
}

func rackLiveShardCount(r *rack) int {
	if r == nil {
		return 0
	}
	total := 0
	for _, n := range r.nodes {
		total += nodeLiveShardCount(n)
	}
	return total
}

// pickNodeInRackEligible is pickNodeInRack restricted to nodes with an eligible
// disk that can take another shard of the volume (free slot, under maxPerDisk).
// FromActiveTopology keeps all disk types/tags in the snapshot, so without this a
// node with free volume slots but no eligible disk could be chosen.
//
// Even first (this volume, machine then node). Among that class: max node
// freeSlots, keep |free−maxFree| ≤ shard-slot deadband, then lower
// liveShards. Machine freeSlots (mFree) stays dropped as on origin/master.
// SameRackCount / 2+2 unchanged. M=0 is only maxFree then live. Callers
// convert UI volume M via volumeDeadbandToShardSlots.
func pickNodeInRackEligible(r *rack, vk volKey, rp *super_block.ReplicaPlacement, eligible func(*disk) bool, maxPerDisk, freeSlotsDeadband int) *Node {
	machineShards := countShardsByHost(vk, r.nodes)
	var nodes []*Node
	var bestMCount, bestNCount int
	for _, id := range sortedNodeKeys(r.nodes) {
		node := r.nodes[id]
		if nodeShardRoom(node, vk, eligible, maxPerDisk) <= 0 {
			continue
		}
		count := volumeShardCount(node, vk)
		if rp != nil && rp.SameRackCount > 0 && count >= rp.SameRackCount {
			continue
		}
		mCount := machineShards[node.host]
		if len(nodes) == 0 || mCount < bestMCount || (mCount == bestMCount && count < bestNCount) {
			nodes = nodes[:0]
			bestMCount, bestNCount = mCount, count
		} else if mCount != bestMCount || count != bestNCount {
			continue
		}
		nodes = append(nodes, node)
	}
	if len(nodes) == 0 {
		return nil
	}
	free := make([]int, len(nodes))
	live := make([]int, len(nodes))
	for i, n := range nodes {
		free[i] = n.freeSlots
		live[i] = nodeLiveShardCount(n)
	}
	return nodes[pickDeadbandRemainder(free, live, freeSlotsDeadband)]
}

// pickDeadbandRemainder ranks same-evenness candidates: maxFree of the
// class, keep |free−maxFree|≤m, then min live. Ties keep the earlier
// index (scan order). M=0 keeps only maxFree then live. Shared by node
// and rack so the band cannot drift.
func pickDeadbandRemainder(free, live []int, m int) int {
	maxFree := free[0]
	for _, f := range free[1:] {
		if f > maxFree {
			maxFree = f
		}
	}
	best := -1
	for i, f := range free {
		if absIntDiff(f, maxFree) > m {
			continue
		}
		if best < 0 || live[i] < live[best] {
			best = i
		}
	}
	return best
}

func absIntDiff(a, b int) int {
	if a > b {
		return a - b
	}
	return b - a
}

// volumeDeadbandToShardSlots converts UI volume-slot M to the shard-slot
// deadband Place compares against node.freeSlots. OSS 10+4 uses dataShards=10.
func volumeDeadbandToShardSlots(volumes, dataShards int) int {
	if volumes <= 0 || dataShards <= 0 {
		return 0
	}
	if volumes > math.MaxInt/dataShards {
		return math.MaxInt
	}
	return volumes * dataShards
}

// pickBestDiskEligible chooses the best eligible disk on a node, ranking
// soft-preferred disks (prefer != nil && prefer(d)) ahead of others so disk-type
// Prefer uses the preferred type when available but spills otherwise. Returns the
// disk id, whether one was found, and whether the chosen disk spilled off the
// preferred type.
func pickBestDiskEligible(node *Node, vk volKey, eligible func(*disk) bool, prefer func(*disk) bool, shardID, dataShardCount, maxPerDisk int) (uint32, bool, bool) {
	isDataShard := dataShardCount > 0 && shardID < dataShardCount
	info := node.shards[vk]
	var bestDiskID uint32
	bestScore := -1
	bestPreferred := false
	for _, diskID := range sortedDiskKeys(node.disks) {
		d := node.disks[diskID]
		if !eligible(d) || d.freeSlots <= 0 || d.noReceive {
			continue
		}
		existingShards := 0
		hasData := false
		hasParity := false
		if info != nil {
			bits := info.diskShardBits[diskID]
			existingShards = bits.Count()
			if dataShardCount > 0 {
				for sid := range bits.All() {
					if int(sid) < dataShardCount {
						hasData = true
					} else {
						hasParity = true
					}
				}
			}
		}
		// Durability: never put more than maxPerDisk (parityShards) shards of this
		// volume on one disk, or losing that disk would lose more than EC can
		// recover. Hard cap, enforced even under durability-first relaxation.
		if maxPerDisk > 0 && existingShards >= maxPerDisk {
			continue
		}
		score := d.shardCount*10 + existingShards*100
		if dataShardCount > 0 {
			if isDataShard && hasParity {
				score += 1000
			} else if !isDataShard && hasData {
				score += 1000
			}
		}
		preferred := prefer == nil || prefer(d)
		if !preferred {
			score += 100000 // strongly deprioritize spilling to a non-preferred type
		}
		if bestScore == -1 || score < bestScore {
			bestScore = score
			bestDiskID = diskID
			bestPreferred = preferred
		}
	}
	if bestScore == -1 {
		return 0, false, false
	}
	return bestDiskID, true, prefer != nil && !bestPreferred
}

// clearShardAccounting removes one shard copy of a volume from the snapshot's
// per-domain accounting (the volume's shard bits) WITHOUT crediting disk capacity.
// It clears only the given physical disk's bit, then recomputes the node-level
// union from the remaining disk bits, so a kept copy of the same shard on another
// disk of the same node still counts toward caps / ReplicaPlacement / anti-affinity.
//
// Repair uses this to drop the duplicate/mismatched copies it plans to delete
// before placing missing shards, so those copies do not inflate placement
// accounting. Capacity is deliberately NOT credited: the deletes run only after
// the rebuilt shards are distributed, so the slots are not free at plan time. This
// is distinct from releaseShard, which credits freeSlots and clears the union.
func clearShardAccounting(node *Node, vk volKey, shardID int, diskID uint32) {
	info, ok := node.shards[vk]
	if !ok {
		return
	}
	sid := erasure_coding.ShardId(shardID)
	if bits, ok := info.diskShardBits[diskID]; ok {
		info.diskShardBits[diskID] = bits.Clear(sid)
	}
	var union erasure_coding.ShardBits
	for _, b := range info.diskShardBits {
		union |= b
	}
	info.shardBits = union
}

// ClearShardAccounting drops one shard copy of a volume from placement accounting
// without crediting capacity (see clearShardAccounting). Repair calls it for each
// copy it plans to delete before placing missing shards, so those copies do not
// inflate caps/RP/anti-affinity. No-op for an unknown node.
func (t *Topology) ClearShardAccounting(nodeID, collection string, vid uint32, shardID int, diskID uint32) {
	n, ok := t.nodes[nodeID]
	if !ok {
		return
	}
	clearShardAccounting(n, volKey{collection: collection, vid: vid}, shardID, diskID)
}

// ReleaseVolumeShards removes every shard of a volume from the snapshot and
// credits the freed disk capacity. A greenfield encode calls this so any stale
// EC shards left by a prior failed attempt (which the encode task deletes before
// distributing the new shards) neither occupy capacity nor skew anti-affinity /
// per-disk caps during planning. Unlike repair's ClearShardAccounting, it credits
// freeSlots because the deletes run before the new writes.
//
// Other volumes' disk.shardCount stay put: zeroing the snapshot here is what
// made within-rack Place see a green field and 50/50 the current volume only.
func (t *Topology) ReleaseVolumeShards(collection string, vid uint32) {
	vk := volKey{collection: collection, vid: vid}
	for _, n := range t.nodes {
		info, ok := n.shards[vk]
		if !ok {
			continue
		}
		// freed is the total disk-slots the volume occupies on this node (a shard may
		// sit on more than one disk). releaseShard credits each disk's freeSlots;
		// credit the node's freeSlots by the same total, since rack capacity is summed
		// from node freeSlots (buildRacks) and node freeSlots gates node eligibility.
		freed := 0
		for _, bits := range info.diskShardBits {
			freed += bits.Count()
		}
		sids := make([]int, 0, info.shardBits.Count())
		for sid := range info.shardBits.All() {
			sids = append(sids, int(sid))
		}
		for _, sid := range sids {
			releaseShard(n, vk, sid)
		}
		n.freeSlots += freed
		delete(n.shards, vk)
	}
}

// shardsOfType returns the sorted subset of need that are data shards (id <
// dataShards) when isData, else the parity subset.
func shardsOfType(need []int, isData bool, dataShards int) []int {
	var out []int
	for _, s := range need {
		if (s < dataShards) == isData {
			out = append(out, s)
		}
	}
	sort.Ints(out)
	return out
}