frostfs-node/pkg/services/object_manager/placement/traverser.go

package placement

import (
	"errors"
	"fmt"
	"sync"

	"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/network"
	"git.frostfs.info/TrueCloudLab/frostfs-sdk-go/container"
	cid "git.frostfs.info/TrueCloudLab/frostfs-sdk-go/container/id"
	"git.frostfs.info/TrueCloudLab/frostfs-sdk-go/netmap"
	oid "git.frostfs.info/TrueCloudLab/frostfs-sdk-go/object/id"
)

// Builder is an interface of the
// object placement vector builder.
type Builder interface {
	// BuildPlacement returns the list of placement vectors
	// for object according to the placement policy.
	//
	// Must return all container nodes if object identifier
	// is nil.
	BuildPlacement(cid.ID, *oid.ID, netmap.PlacementPolicy) ([][]netmap.NodeInfo, error)
}

// Option represents placement traverser option.
type Option func(*cfg)

// Traverser represents utility for controlling
// traversal of object placement vectors.
type Traverser struct {
	mtx sync.RWMutex

	vectors [][]netmap.NodeInfo

	rem []int
}

type cfg struct {
	trackCopies bool
	copyNumbers []uint32

	flatSuccess *uint32

	cnr cid.ID

	obj *oid.ID

	policySet bool
	policy    netmap.PlacementPolicy

	builder Builder
}

const invalidOptsMsg = "invalid traverser options"

var errNilBuilder = errors.New("placement builder is nil")

var errNilPolicy = errors.New("placement policy is nil")

var errCopiesNumberLen = errors.New("copies number accepts only one number or array with length " +
	"equal to length of replicas")

func defaultCfg() *cfg {
	return &cfg{
		trackCopies: true,
	}
}

// NewTraverser creates, initializes with options and returns Traverser instance.
func NewTraverser(opts ...Option) (*Traverser, error) {
	cfg := defaultCfg()

	for i := range opts {
		if opts[i] != nil {
			opts[i](cfg)
		}
	}

	cnLen := len(cfg.copyNumbers)
	if cnLen > 0 && cnLen != 1 && cnLen != cfg.policy.NumberOfReplicas() {
		return nil, errCopiesNumberLen
	}

	if cfg.builder == nil {
		return nil, fmt.Errorf("%s: %w", invalidOptsMsg, errNilBuilder)
	} else if !cfg.policySet {
		return nil, fmt.Errorf("%s: %w", invalidOptsMsg, errNilPolicy)
	}

	ns, err := cfg.builder.BuildPlacement(cfg.cnr, cfg.obj, cfg.policy)
	if err != nil {
		return nil, fmt.Errorf("could not build placement: %w", err)
	}

	// backward compatibility for scalar `copies_number`
	if len(cfg.copyNumbers) == 1 && cfg.copyNumbers[0] != 0 {
		cfg.flatSuccess = &cfg.copyNumbers[0]
	}

	var rem []int
	if cfg.flatSuccess != nil {
		ns = flatNodes(ns)
		rem = []int{int(*cfg.flatSuccess)}
	} else {
		rem = defaultCopiesVector(cfg.policy)

		// Bool flag which is set when cfg.copyNumbers contains not only zeros.
		// In this case we should not modify `rem` slice unless track
		// copies are ignored, because [0, ...] means that all copies should be
		// stored before returning OK to the client.
		var considerCopiesNumber bool
		for _, val := range cfg.copyNumbers {
			if val != 0 {
				considerCopiesNumber = true
				break
			}
		}

		for i := range rem {
			if !cfg.trackCopies {
				rem[i] = -1
			} else if considerCopiesNumber && len(cfg.copyNumbers) > i {
				rem[i] = int(cfg.copyNumbers[i])
			}
		}
	}

	return &Traverser{
		rem:     rem,
		vectors: ns,
	}, nil
}

func defaultCopiesVector(policy netmap.PlacementPolicy) []int {
	replNum := policy.NumberOfReplicas()
	copyVector := make([]int, 0, replNum)

	for i := range replNum {
		copyVector = append(copyVector, int(policy.ReplicaDescriptor(i).NumberOfObjects()+policy.ReplicaDescriptor(i).GetECDataCount()+policy.ReplicaDescriptor(i).GetECParityCount()))
	}

	return copyVector
}

func flatNodes(ns [][]netmap.NodeInfo) [][]netmap.NodeInfo {
	sz := 0
	for i := range ns {
		sz += len(ns[i])
	}

	flat := make([]netmap.NodeInfo, 0, sz)
	for i := range ns {
		flat = append(flat, ns[i]...)
	}

	return [][]netmap.NodeInfo{flat}
}

// Node is a descriptor of storage node with information required for intra-container communication.
type Node struct {
	addresses network.AddressGroup

	externalAddresses network.AddressGroup

	key []byte
}

// Addresses returns group of network addresses.
func (x Node) Addresses() network.AddressGroup {
	return x.addresses
}

// ExternalAddresses returns group of network addresses.
func (x Node) ExternalAddresses() network.AddressGroup {
	return x.externalAddresses
}

// PublicKey returns public key in a binary format. Should not be mutated.
func (x Node) PublicKey() []byte {
	return x.key
}

// NewNode creates new Node.
func NewNode(addresses network.AddressGroup, externalAddresses network.AddressGroup, key []byte) Node {
	return Node{
		addresses:         addresses,
		externalAddresses: externalAddresses,
		key:               key,
	}
}

// Next returns next unprocessed address of the object placement.
//
// Returns nil if no nodes left or traversal operation succeeded.
func (t *Traverser) Next() []Node {
	t.mtx.Lock()
	defer t.mtx.Unlock()

	t.skipEmptyVectors()

	if len(t.vectors) == 0 {
		return nil
	} else if len(t.vectors[0]) < t.rem[0] {
		return nil
	}

	count := t.rem[0]
	if count < 0 {
		count = len(t.vectors[0])
	}

	nodes := make([]Node, count)

	for i := range count {
		err := nodes[i].addresses.FromIterator(network.NodeEndpointsIterator(t.vectors[0][i]))
		if err != nil {
			return nil
		}

		ext := t.vectors[0][i].ExternalAddresses()
		if len(ext) > 0 {
			// Ignore the error if this field is incorrectly formed.
			_ = nodes[i].externalAddresses.FromStringSlice(ext)
		}

		nodes[i].key = t.vectors[0][i].PublicKey()
	}

	t.vectors[0] = t.vectors[0][count:]

	return nodes
}

func (t *Traverser) skipEmptyVectors() {
	for i := 0; i < len(t.vectors); i++ { // don't use range, slice changes in body
		if len(t.vectors[i]) == 0 && t.rem[i] <= 0 || t.rem[0] == 0 {
			t.vectors = append(t.vectors[:i], t.vectors[i+1:]...)
			t.rem = append(t.rem[:i], t.rem[i+1:]...)
			i--
		} else {
			break
		}
	}
}

// SubmitSuccess writes single succeeded node operation.
func (t *Traverser) SubmitSuccess() {
	t.mtx.Lock()
	if len(t.rem) > 0 {
		t.rem[0]--
	}
	t.mtx.Unlock()
}

// Success returns true if traversal operation succeeded.
func (t *Traverser) Success() bool {
	t.mtx.RLock()
	defer t.mtx.RUnlock()

	for i := range t.rem {
		if t.rem[i] > 0 {
			return false
		}
	}

	return true
}

// UseBuilder is a placement builder setting option.
//
// Overlaps UseNetworkMap option.
func UseBuilder(b Builder) Option {
	return func(c *cfg) {
		c.builder = b
	}
}

// ForContainer is a traversal container setting option.
func ForContainer(cnr container.Container) Option {
	return func(c *cfg) {
		c.policy = cnr.PlacementPolicy()
		c.policySet = true
		container.CalculateID(&c.cnr, cnr)
	}
}

// ForObject is a processing object setting option.
func ForObject(id oid.ID) Option {
	return func(c *cfg) {
		c.obj = &id
	}
}

// SuccessAfter is a flat success number setting option.
//
// Option has no effect if the number is not positive.
func SuccessAfter(v uint32) Option {
	return func(c *cfg) {
		if v > 0 {
			c.flatSuccess = &v
		}
	}
}

// ResetSuccessAfter resets flat success number setting option.
func ResetSuccessAfter() Option {
	return func(c *cfg) {
		c.flatSuccess = nil
	}
}

// WithoutSuccessTracking disables success tracking in traversal.
func WithoutSuccessTracking() Option {
	return func(c *cfg) {
		c.trackCopies = false
	}
}

func WithCopyNumbers(v []uint32) Option {
	return func(c *cfg) {
		c.copyNumbers = v
	}
}