storj/satellite/repair/repairer/ec.go
paul cannon 1b8bd6c082 satellite/repair: unify repair logic
The repair checker and repair worker both need to determine which pieces
are healthy, which are retrievable, and which should be replaced, but
they have been doing it in different ways in different code, which has
been the cause of bugs. The same term could have very similar but subtly
different meanings between the two, causing much confusion.

With this change, the piece- and node-classification logic is
consolidated into one place within the satellite/repair package, so that
both subsystems can use it. This ought to make decision-making code more
concise and more readable.

The consolidated classification logic has been expanded to create more
sets, so that the decision-making code does not need to do as much
precalculation. It should now be clearer in comments and code that a
piece can belong to multiple sets arbitrarily (except where the
definition of the sets makes this logically impossible), and what the
precise meaning of each set is. These sets include Missing, Suspended,
Clumped, OutOfPlacement, InExcludedCountry, ForcingRepair,
UnhealthyRetrievable, Unhealthy, Retrievable, and Healthy.

Some other side effects of this change:

* CreatePutRepairOrderLimits no longer needs to special-case excluded
  countries; it can just create as many order limits as requested (by
  way of len(newNodes)).
* The repair checker will now queue a segment for repair when there are
  any pieces out of placement. The code calls this "forcing a repair".
* The checker.ReliabilityCache is now accessed by way of a GetNodes()
  function similar to the one on the overlay. The classification methods
  like MissingPieces(), OutOfPlacementPieces(), and
  PiecesNodesLastNetsInOrder() are removed in favor of the
  classification logic in satellite/repair/classification.go. This
  means the reliability cache no longer needs access to the placement
  rules or excluded countries list.

Change-Id: I105109fb94ee126952f07d747c6e11131164fadb
2023-09-25 09:42:08 -05:00

642 lines
21 KiB
Go

// Copyright (C) 2019 Storj Labs, Inc.
// See LICENSE for copying information.
package repairer
import (
"bytes"
"context"
"errors"
"hash"
"io"
"sort"
"sync"
"sync/atomic"
"time"
"github.com/calebcase/tmpfile"
"github.com/vivint/infectious"
"github.com/zeebo/errs"
"go.uber.org/zap"
"storj.io/common/errs2"
"storj.io/common/pb"
"storj.io/common/rpc"
"storj.io/common/rpc/rpcpool"
"storj.io/common/signing"
"storj.io/common/storj"
"storj.io/common/sync2"
"storj.io/storj/satellite/audit"
"storj.io/storj/satellite/metabase"
"storj.io/storj/satellite/overlay"
"storj.io/uplink/private/eestream"
"storj.io/uplink/private/piecestore"
)
var (
// ErrPieceHashVerifyFailed is the errs class when a piece hash downloaded from storagenode fails to match the original hash.
ErrPieceHashVerifyFailed = errs.Class("piece hashes don't match")
// ErrDialFailed is the errs class when a failure happens during Dial.
ErrDialFailed = errs.Class("dial failure")
)
// ECRepairer allows the repairer to download, verify, and upload pieces from storagenodes.
type ECRepairer struct {
log *zap.Logger
dialer rpc.Dialer
satelliteSignee signing.Signee
dialTimeout time.Duration
downloadTimeout time.Duration
inmemory bool
// used only in tests, where we expect failures and want to wait for them
minFailures int
}
// NewECRepairer creates a new repairer for interfacing with storagenodes.
func NewECRepairer(log *zap.Logger, dialer rpc.Dialer, satelliteSignee signing.Signee, dialTimeout time.Duration, downloadTimeout time.Duration, inmemory bool) *ECRepairer {
return &ECRepairer{
log: log,
dialer: dialer,
satelliteSignee: satelliteSignee,
dialTimeout: dialTimeout,
downloadTimeout: downloadTimeout,
inmemory: inmemory,
}
}
func (ec *ECRepairer) dialPiecestore(ctx context.Context, n storj.NodeURL) (*piecestore.Client, error) {
client, err := piecestore.Dial(rpcpool.WithForceDial(ctx), ec.dialer, n, piecestore.DefaultConfig)
return client, ErrDialFailed.Wrap(err)
}
// TestingSetMinFailures sets the minFailures attribute, which tells the Repair machinery that we _expect_
// there to be failures and that we should wait for them if necessary. This is only used in tests.
func (ec *ECRepairer) TestingSetMinFailures(minFailures int) {
ec.minFailures = minFailures
}
// Get downloads pieces from storagenodes using the provided order limits, and decodes those pieces into a segment.
// It attempts to download from the minimum required number based on the redundancy scheme. It will further wait
// for additional error/failure results up to minFailures, for testing purposes. Under normal conditions,
// minFailures will be 0.
//
// After downloading a piece, the ECRepairer will verify the hash and original order limit for that piece.
// If verification fails, another piece will be downloaded until we reach the minimum required or run out of order limits.
// If piece hash verification fails, it will return all failed node IDs.
func (ec *ECRepairer) Get(ctx context.Context, limits []*pb.AddressedOrderLimit, cachedNodesInfo map[storj.NodeID]overlay.NodeReputation, privateKey storj.PiecePrivateKey, es eestream.ErasureScheme, dataSize int64) (_ io.ReadCloser, _ FetchResultReport, err error) {
defer mon.Task()(&ctx)(&err)
if len(limits) != es.TotalCount() {
return nil, FetchResultReport{}, Error.New("number of limits slice (%d) does not match total count (%d) of erasure scheme", len(limits), es.TotalCount())
}
nonNilLimits := nonNilCount(limits)
if nonNilLimits < es.RequiredCount()+ec.minFailures {
return nil, FetchResultReport{}, Error.New("number of non-nil limits (%d) is less than requested result count (%d)", nonNilCount(limits), es.RequiredCount()+ec.minFailures)
}
mon.IntVal("ECRepairer_Get_nonNilLimits").Observe(int64(nonNilLimits))
pieceSize := eestream.CalcPieceSize(dataSize, es)
errorCount := 0
var successfulPieces, inProgress int
unusedLimits := nonNilLimits
pieceReaders := make(map[int]io.ReadCloser)
var pieces FetchResultReport
limiter := sync2.NewLimiter(es.RequiredCount())
cond := sync.NewCond(&sync.Mutex{})
for currentLimitIndex, limit := range limits {
if limit == nil {
continue
}
currentLimitIndex, limit := currentLimitIndex, limit
limiter.Go(ctx, func() {
cond.L.Lock()
defer cond.Signal()
defer cond.L.Unlock()
for {
if successfulPieces >= es.RequiredCount() && errorCount >= ec.minFailures {
// already downloaded required number of pieces
cond.Broadcast()
return
}
if successfulPieces+inProgress+unusedLimits < es.RequiredCount() || errorCount+inProgress+unusedLimits < ec.minFailures {
// not enough available limits left to get required number of pieces
cond.Broadcast()
return
}
if successfulPieces+inProgress >= es.RequiredCount() && errorCount+inProgress >= ec.minFailures {
// we know that inProgress > 0 here, since we didn't return on the
// "successfulPieces >= es.RequiredCount() && errorCount >= ec.minFailures" check earlier.
// There may be enough downloads in progress to meet all of our needs, so we won't
// start any more immediately. Instead, wait until all needs are met (in which case
// cond.Broadcast() will be called) or until one of the inProgress workers exits
// (in which case cond.Signal() will be called, waking up one waiter) so we can
// reevaluate the situation.
cond.Wait()
continue
}
unusedLimits--
inProgress++
cond.L.Unlock()
info := cachedNodesInfo[limit.GetLimit().StorageNodeId]
address := limit.GetStorageNodeAddress().GetAddress()
var triedLastIPPort bool
if info.LastIPPort != "" && info.LastIPPort != address {
address = info.LastIPPort
triedLastIPPort = true
}
pieceReadCloser, _, _, err := ec.downloadAndVerifyPiece(ctx, limit, address, privateKey, "", pieceSize)
// if piecestore dial with last ip:port failed try again with node address
if triedLastIPPort && ErrDialFailed.Has(err) {
if pieceReadCloser != nil {
_ = pieceReadCloser.Close()
}
pieceReadCloser, _, _, err = ec.downloadAndVerifyPiece(ctx, limit, limit.GetStorageNodeAddress().GetAddress(), privateKey, "", pieceSize)
}
cond.L.Lock()
inProgress--
piece := metabase.Piece{
Number: uint16(currentLimitIndex),
StorageNode: limit.GetLimit().StorageNodeId,
}
if err != nil {
if pieceReadCloser != nil {
_ = pieceReadCloser.Close()
}
// gather nodes where the calculated piece hash doesn't match the uplink signed piece hash
if ErrPieceHashVerifyFailed.Has(err) {
ec.log.Info("audit failed",
zap.Stringer("node ID", limit.GetLimit().StorageNodeId),
zap.Stringer("Piece ID", limit.Limit.PieceId),
zap.String("reason", err.Error()))
pieces.Failed = append(pieces.Failed, PieceFetchResult{Piece: piece, Err: err})
errorCount++
return
}
pieceAudit := audit.PieceAuditFromErr(err)
switch pieceAudit {
case audit.PieceAuditFailure:
ec.log.Debug("Failed to download piece for repair: piece not found (audit failed)",
zap.Stringer("Node ID", limit.GetLimit().StorageNodeId),
zap.Stringer("Piece ID", limit.Limit.PieceId),
zap.Error(err))
pieces.Failed = append(pieces.Failed, PieceFetchResult{Piece: piece, Err: err})
errorCount++
case audit.PieceAuditOffline:
ec.log.Debug("Failed to download piece for repair: dial timeout (offline)",
zap.Stringer("Node ID", limit.GetLimit().StorageNodeId),
zap.Stringer("Piece ID", limit.Limit.PieceId),
zap.Error(err))
pieces.Offline = append(pieces.Offline, PieceFetchResult{Piece: piece, Err: err})
errorCount++
case audit.PieceAuditContained:
ec.log.Info("Failed to download piece for repair: download timeout (contained)",
zap.Stringer("Node ID", limit.GetLimit().StorageNodeId),
zap.Stringer("Piece ID", limit.Limit.PieceId),
zap.Error(err))
pieces.Contained = append(pieces.Contained, PieceFetchResult{Piece: piece, Err: err})
errorCount++
case audit.PieceAuditUnknown:
ec.log.Info("Failed to download piece for repair: unknown transport error (skipped)",
zap.Stringer("Node ID", limit.GetLimit().StorageNodeId),
zap.Stringer("Piece ID", limit.Limit.PieceId),
zap.Error(err))
pieces.Unknown = append(pieces.Unknown, PieceFetchResult{Piece: piece, Err: err})
errorCount++
}
return
}
pieceReaders[currentLimitIndex] = pieceReadCloser
pieces.Successful = append(pieces.Successful, PieceFetchResult{Piece: piece})
successfulPieces++
return
}
})
}
limiter.Wait()
if successfulPieces < es.RequiredCount() {
mon.Meter("download_failed_not_enough_pieces_repair").Mark(1) //mon:locked
return nil, pieces, &irreparableError{
piecesAvailable: int32(successfulPieces),
piecesRequired: int32(es.RequiredCount()),
}
}
if errorCount < ec.minFailures {
return nil, pieces, Error.New("expected %d failures, but only observed %d", ec.minFailures, errorCount)
}
fec, err := infectious.NewFEC(es.RequiredCount(), es.TotalCount())
if err != nil {
return nil, pieces, Error.Wrap(err)
}
esScheme := eestream.NewUnsafeRSScheme(fec, es.ErasureShareSize())
expectedSize := pieceSize * int64(es.RequiredCount())
ctx, cancel := context.WithCancel(ctx)
decodeReader := eestream.DecodeReaders2(ctx, cancel, pieceReaders, esScheme, expectedSize, 0, false)
return decodeReader, pieces, nil
}
// lazyHashWriter is a writer which can get the hash algorithm just before the first write.
type lazyHashWriter struct {
hasher hash.Hash
downloader *piecestore.Download
}
func (l *lazyHashWriter) Write(p []byte) (n int, err error) {
// hash is available only after receiving the first message.
if l.hasher == nil {
h, _ := l.downloader.GetHashAndLimit()
l.hasher = pb.NewHashFromAlgorithm(h.HashAlgorithm)
}
return l.hasher.Write(p)
}
// Sum delegates hash calculation to the real hash algorithm.
func (l *lazyHashWriter) Sum(b []byte) []byte {
if l.hasher == nil {
return []byte{}
}
return l.hasher.Sum(b)
}
var _ io.Writer = &lazyHashWriter{}
// downloadAndVerifyPiece downloads a piece from a storagenode,
// expects the original order limit to have the correct piece public key,
// and expects the hash of the data to match the signed hash provided by the storagenode.
func (ec *ECRepairer) downloadAndVerifyPiece(ctx context.Context, limit *pb.AddressedOrderLimit, address string, privateKey storj.PiecePrivateKey, tmpDir string, pieceSize int64) (pieceReadCloser io.ReadCloser, hash *pb.PieceHash, originalLimit *pb.OrderLimit, err error) {
defer mon.Task()(&ctx)(&err)
// contact node
dialCtx, dialCancel := context.WithTimeout(ctx, ec.dialTimeout)
defer dialCancel()
ps, err := ec.dialPiecestore(dialCtx, storj.NodeURL{
ID: limit.GetLimit().StorageNodeId,
Address: address,
})
if err != nil {
return nil, nil, nil, err
}
defer func() { err = errs.Combine(err, ps.Close()) }()
downloadCtx, cancel := context.WithTimeout(ctx, ec.downloadTimeout)
defer cancel()
downloader, err := ps.Download(downloadCtx, limit.GetLimit(), privateKey, 0, pieceSize)
if err != nil {
return nil, nil, nil, err
}
defer func() { err = errs.Combine(err, downloader.Close()) }()
hashWriter := &lazyHashWriter{
downloader: downloader,
}
downloadReader := io.TeeReader(downloader, hashWriter)
var downloadedPieceSize int64
if ec.inmemory {
pieceBytes, err := io.ReadAll(downloadReader)
if err != nil {
return nil, nil, nil, err
}
downloadedPieceSize = int64(len(pieceBytes))
pieceReadCloser = io.NopCloser(bytes.NewReader(pieceBytes))
} else {
tempfile, err := tmpfile.New(tmpDir, "satellite-repair-*")
if err != nil {
return nil, nil, nil, err
}
// no defer tempfile.Close() here; caller is responsible for closing
// the file, even if an error results (the caller might want the data
// even if there is a verification error).
downloadedPieceSize, err = io.Copy(tempfile, downloadReader)
if err != nil {
return tempfile, nil, nil, err
}
// seek to beginning of file so the repair job starts at the beginning of the piece
_, err = tempfile.Seek(0, io.SeekStart)
if err != nil {
return tempfile, nil, nil, err
}
pieceReadCloser = tempfile
}
mon.Meter("repair_bytes_downloaded").Mark64(downloadedPieceSize) //mon:locked
if downloadedPieceSize != pieceSize {
return pieceReadCloser, nil, nil, Error.New("didn't download the correct amount of data, want %d, got %d", pieceSize, downloadedPieceSize)
}
// get signed piece hash and original order limit
hash, originalLimit = downloader.GetHashAndLimit()
if hash == nil {
return pieceReadCloser, hash, originalLimit, Error.New("hash was not sent from storagenode")
}
if originalLimit == nil {
return pieceReadCloser, hash, originalLimit, Error.New("original order limit was not sent from storagenode")
}
// verify order limit from storage node is signed by the satellite
if err := verifyOrderLimitSignature(ctx, ec.satelliteSignee, originalLimit); err != nil {
return pieceReadCloser, hash, originalLimit, err
}
// verify the hashes from storage node
calculatedHash := hashWriter.Sum(nil)
if err := verifyPieceHash(ctx, originalLimit, hash, calculatedHash); err != nil {
return pieceReadCloser, hash, originalLimit, ErrPieceHashVerifyFailed.Wrap(err)
}
return pieceReadCloser, hash, originalLimit, nil
}
func verifyPieceHash(ctx context.Context, limit *pb.OrderLimit, hash *pb.PieceHash, expectedHash []byte) (err error) {
defer mon.Task()(&ctx)(&err)
if limit == nil || hash == nil || len(expectedHash) == 0 {
return Error.New("invalid arguments")
}
if limit.PieceId != hash.PieceId {
return Error.New("piece id changed")
}
if !bytes.Equal(hash.Hash, expectedHash) {
return Error.New("hash from storage node, %x, does not match calculated hash, %x", hash.Hash, expectedHash)
}
if err := signing.VerifyUplinkPieceHashSignature(ctx, limit.UplinkPublicKey, hash); err != nil {
return Error.New("invalid piece hash signature")
}
return nil
}
func verifyOrderLimitSignature(ctx context.Context, satellite signing.Signee, limit *pb.OrderLimit) (err error) {
if err := signing.VerifyOrderLimitSignature(ctx, satellite, limit); err != nil {
return Error.New("invalid order limit signature: %v", err)
}
return nil
}
// Repair takes a provided segment, encodes it with the provided redundancy strategy,
// and uploads the pieces in need of repair to new nodes provided by order limits.
func (ec *ECRepairer) Repair(ctx context.Context, limits []*pb.AddressedOrderLimit, privateKey storj.PiecePrivateKey, rs eestream.RedundancyStrategy, data io.Reader, timeout time.Duration, successfulNeeded int) (successfulNodes []*pb.Node, successfulHashes []*pb.PieceHash, err error) {
defer mon.Task()(&ctx)(&err)
pieceCount := len(limits)
if pieceCount != rs.TotalCount() {
return nil, nil, Error.New("size of limits slice (%d) does not match total count (%d) of erasure scheme", pieceCount, rs.TotalCount())
}
if !unique(limits) {
return nil, nil, Error.New("duplicated nodes are not allowed")
}
readers, err := eestream.EncodeReader2(ctx, io.NopCloser(data), rs)
if err != nil {
return nil, nil, err
}
// info contains data about a single piece transfer
type info struct {
i int
err error
hash *pb.PieceHash
}
// this channel is used to synchronize concurrently uploaded pieces with the overall repair
infos := make(chan info, pieceCount)
psCtx, cancel := context.WithCancel(ctx)
defer cancel()
for i, addressedLimit := range limits {
go func(i int, addressedLimit *pb.AddressedOrderLimit) {
hash, err := ec.putPiece(psCtx, ctx, addressedLimit, privateKey, readers[i])
infos <- info{i: i, err: err, hash: hash}
}(i, addressedLimit)
}
ec.log.Debug("Starting a timer for repair so that the number of pieces will be closer to the success threshold",
zap.Duration("Timer", timeout),
zap.Int("Node Count", nonNilCount(limits)),
zap.Int("Optimal Threshold", rs.OptimalThreshold()),
)
var successfulCount, failureCount, cancellationCount int32
timer := time.AfterFunc(timeout, func() {
if !errors.Is(ctx.Err(), context.Canceled) {
ec.log.Debug("Timer expired. Canceling the long tail...",
zap.Int32("Successfully repaired", atomic.LoadInt32(&successfulCount)),
)
cancel()
}
})
successfulNodes = make([]*pb.Node, pieceCount)
successfulHashes = make([]*pb.PieceHash, pieceCount)
for range limits {
info := <-infos
if limits[info.i] == nil {
continue
}
if info.err != nil {
if !errs2.IsCanceled(info.err) {
failureCount++
ec.log.Warn("Repair to a storage node failed",
zap.Stringer("Node ID", limits[info.i].GetLimit().StorageNodeId),
zap.Error(info.err),
)
} else {
cancellationCount++
ec.log.Debug("Repair to storage node cancelled",
zap.Stringer("Node ID", limits[info.i].GetLimit().StorageNodeId),
zap.Error(info.err),
)
}
continue
}
successfulNodes[info.i] = &pb.Node{
Id: limits[info.i].GetLimit().StorageNodeId,
Address: limits[info.i].GetStorageNodeAddress(),
}
successfulHashes[info.i] = info.hash
successfulCount++
if successfulCount >= int32(successfulNeeded) {
// if this is logged more than once for a given repair operation, it is because
// an upload succeeded right after we called cancel(), before that upload could
// actually be canceled. So, successfulCount should increase by one with each
// repeated logging.
ec.log.Debug("Number of successful uploads met. Canceling the long tail...",
zap.Int32("Successfully repaired", atomic.LoadInt32(&successfulCount)),
)
cancel()
}
}
// Ensure timer is stopped
_ = timer.Stop()
// TODO: clean up the partially uploaded segment's pieces
defer func() {
select {
case <-ctx.Done():
err = Error.New("repair cancelled")
default:
}
}()
if successfulCount == 0 {
return nil, nil, Error.New("repair to all nodes failed")
}
ec.log.Debug("Successfully repaired",
zap.Int32("Success Count", atomic.LoadInt32(&successfulCount)),
)
mon.IntVal("repair_segment_pieces_total").Observe(int64(pieceCount)) //mon:locked
mon.IntVal("repair_segment_pieces_successful").Observe(int64(successfulCount)) //mon:locked
mon.IntVal("repair_segment_pieces_failed").Observe(int64(failureCount)) //mon:locked
mon.IntVal("repair_segment_pieces_canceled").Observe(int64(cancellationCount)) //mon:locked
return successfulNodes, successfulHashes, nil
}
func (ec *ECRepairer) putPiece(ctx, parent context.Context, limit *pb.AddressedOrderLimit, privateKey storj.PiecePrivateKey, data io.ReadCloser) (hash *pb.PieceHash, err error) {
defer mon.Task()(&ctx)(&err)
nodeName := "nil"
if limit != nil {
nodeName = limit.GetLimit().StorageNodeId.String()[0:8]
}
defer mon.Task()(&ctx, "node: "+nodeName)(&err)
defer func() { err = errs.Combine(err, data.Close()) }()
if limit == nil {
_, _ = io.Copy(io.Discard, data)
return nil, nil
}
storageNodeID := limit.GetLimit().StorageNodeId
pieceID := limit.GetLimit().PieceId
dialCtx, dialCancel := context.WithTimeout(ctx, ec.dialTimeout)
defer dialCancel()
ps, err := ec.dialPiecestore(dialCtx, storj.NodeURL{
ID: storageNodeID,
Address: limit.GetStorageNodeAddress().Address,
})
if err != nil {
ec.log.Debug("Failed dialing for putting piece to node",
zap.Stringer("Piece ID", pieceID),
zap.Stringer("Node ID", storageNodeID),
zap.Error(err),
)
return nil, err
}
defer func() { err = errs.Combine(err, ps.Close()) }()
hash, err = ps.UploadReader(ctx, limit.GetLimit(), privateKey, data)
if err != nil {
if errors.Is(ctx.Err(), context.Canceled) {
// Canceled context means the piece upload was interrupted by user or due
// to slow connection. No error logging for this case.
if errors.Is(parent.Err(), context.Canceled) {
ec.log.Debug("Upload to node canceled by user",
zap.Stringer("Node ID", storageNodeID),
zap.Stringer("Piece ID", pieceID))
} else {
ec.log.Debug("Node cut from upload due to slow connection",
zap.Stringer("Node ID", storageNodeID),
zap.Stringer("Piece ID", pieceID))
}
// make sure context.Canceled is the primary error in the error chain
// for later errors.Is/errs2.IsCanceled checking
err = errs.Combine(context.Canceled, err)
} else {
nodeAddress := "nil"
if limit.GetStorageNodeAddress() != nil {
nodeAddress = limit.GetStorageNodeAddress().GetAddress()
}
ec.log.Debug("Failed uploading piece to node",
zap.Stringer("Piece ID", pieceID),
zap.Stringer("Node ID", storageNodeID),
zap.String("Node Address", nodeAddress),
zap.Error(err),
)
}
}
return hash, err
}
func nonNilCount(limits []*pb.AddressedOrderLimit) int {
total := 0
for _, limit := range limits {
if limit != nil {
total++
}
}
return total
}
func unique(limits []*pb.AddressedOrderLimit) bool {
if len(limits) < 2 {
return true
}
ids := make(storj.NodeIDList, len(limits))
for i, addressedLimit := range limits {
if addressedLimit != nil {
ids[i] = addressedLimit.GetLimit().StorageNodeId
}
}
// sort the ids and check for identical neighbors
sort.Sort(ids)
// sort.Slice(ids, func(i, k int) bool { return ids[i].Less(ids[k]) })
for i := 1; i < len(ids); i++ {
if ids[i] != (storj.NodeID{}) && ids[i] == ids[i-1] {
return false
}
}
return true
}