restic/internal/repository/master_index.go
Michael Eischer 6f53ecc1ae adapt workers based on whether an operation is CPU or IO-bound
Use runtime.GOMAXPROCS(0) as worker count for CPU-bound tasks,
repo.Connections() for IO-bound task and a combination if a task can be
both. Streaming packs is treated as IO-bound as adding more worker
cannot provide a speedup.

Typical IO-bound tasks are download / uploading / deleting files.
Decoding / Encoding / Verifying are usually CPU-bound. Several tasks are
a combination of both, e.g. for combined download and decode functions.
In the latter case add both limits together. As the backends have their
own concurrency limits restic still won't download more than
repo.Connections() files in parallel, but the additional workers can
decode already downloaded data in parallel.
2022-07-03 12:19:26 +02:00

470 lines
11 KiB
Go

package repository
import (
"bytes"
"context"
"fmt"
"runtime"
"sync"
"github.com/restic/restic/internal/debug"
"github.com/restic/restic/internal/restic"
"github.com/restic/restic/internal/ui/progress"
"golang.org/x/sync/errgroup"
)
// MasterIndex is a collection of indexes and IDs of chunks that are in the process of being saved.
type MasterIndex struct {
idx []*Index
pendingBlobs restic.BlobSet
idxMutex sync.RWMutex
compress bool
}
// NewMasterIndex creates a new master index.
func NewMasterIndex() *MasterIndex {
// Always add an empty final index, such that MergeFinalIndexes can merge into this.
// Note that removing this index could lead to a race condition in the rare
// sitation that only two indexes exist which are saved and merged concurrently.
idx := []*Index{NewIndex()}
idx[0].Finalize()
return &MasterIndex{idx: idx, pendingBlobs: restic.NewBlobSet()}
}
func (mi *MasterIndex) markCompressed() {
mi.compress = true
}
// Lookup queries all known Indexes for the ID and returns all matches.
func (mi *MasterIndex) Lookup(bh restic.BlobHandle) (pbs []restic.PackedBlob) {
mi.idxMutex.RLock()
defer mi.idxMutex.RUnlock()
for _, idx := range mi.idx {
pbs = idx.Lookup(bh, pbs)
}
return pbs
}
// LookupSize queries all known Indexes for the ID and returns the first match.
func (mi *MasterIndex) LookupSize(bh restic.BlobHandle) (uint, bool) {
mi.idxMutex.RLock()
defer mi.idxMutex.RUnlock()
for _, idx := range mi.idx {
if size, found := idx.LookupSize(bh); found {
return size, found
}
}
return 0, false
}
// AddPending adds a given blob to list of pending Blobs
// Before doing so it checks if this blob is already known.
// Returns true if adding was successful and false if the blob
// was already known
func (mi *MasterIndex) addPending(bh restic.BlobHandle) bool {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
// Check if blob is pending or in index
if mi.pendingBlobs.Has(bh) {
return false
}
for _, idx := range mi.idx {
if idx.Has(bh) {
return false
}
}
// really not known -> insert
mi.pendingBlobs.Insert(bh)
return true
}
// Has queries all known Indexes for the ID and returns the first match.
// Also returns true if the ID is pending.
func (mi *MasterIndex) Has(bh restic.BlobHandle) bool {
mi.idxMutex.RLock()
defer mi.idxMutex.RUnlock()
// also return true if blob is pending
if mi.pendingBlobs.Has(bh) {
return true
}
for _, idx := range mi.idx {
if idx.Has(bh) {
return true
}
}
return false
}
func (mi *MasterIndex) IsMixedPack(packID restic.ID) bool {
mi.idxMutex.RLock()
defer mi.idxMutex.RUnlock()
for _, idx := range mi.idx {
if idx.MixedPacks().Has(packID) {
return true
}
}
return false
}
// IDs returns the IDs of all indexes contained in the index.
func (mi *MasterIndex) IDs() restic.IDSet {
mi.idxMutex.RLock()
defer mi.idxMutex.RUnlock()
ids := restic.NewIDSet()
for _, idx := range mi.idx {
if !idx.Final() {
continue
}
indexIDs, err := idx.IDs()
if err != nil {
debug.Log("not using index, ID() returned error %v", err)
continue
}
for _, id := range indexIDs {
ids.Insert(id)
}
}
return ids
}
// Packs returns all packs that are covered by the index.
// If packBlacklist is given, those packs are only contained in the
// resulting IDSet if they are contained in a non-final (newly written) index.
func (mi *MasterIndex) Packs(packBlacklist restic.IDSet) restic.IDSet {
mi.idxMutex.RLock()
defer mi.idxMutex.RUnlock()
packs := restic.NewIDSet()
for _, idx := range mi.idx {
idxPacks := idx.Packs()
if idx.final {
idxPacks = idxPacks.Sub(packBlacklist)
}
packs.Merge(idxPacks)
}
return packs
}
// Insert adds a new index to the MasterIndex.
func (mi *MasterIndex) Insert(idx *Index) {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
mi.idx = append(mi.idx, idx)
}
// StorePack remembers the id and pack in the index.
func (mi *MasterIndex) StorePack(id restic.ID, blobs []restic.Blob) {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
// delete blobs from pending
for _, blob := range blobs {
mi.pendingBlobs.Delete(restic.BlobHandle{Type: blob.Type, ID: blob.ID})
}
for _, idx := range mi.idx {
if !idx.Final() {
idx.StorePack(id, blobs)
return
}
}
newIdx := NewIndex()
newIdx.StorePack(id, blobs)
mi.idx = append(mi.idx, newIdx)
}
// finalizeNotFinalIndexes finalizes all indexes that
// have not yet been saved and returns that list
func (mi *MasterIndex) finalizeNotFinalIndexes() []*Index {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
var list []*Index
for _, idx := range mi.idx {
if !idx.Final() {
idx.Finalize()
list = append(list, idx)
}
}
debug.Log("return %d indexes", len(list))
return list
}
// finalizeFullIndexes finalizes all indexes that are full and returns that list.
func (mi *MasterIndex) finalizeFullIndexes() []*Index {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
var list []*Index
debug.Log("checking %d indexes", len(mi.idx))
for _, idx := range mi.idx {
if idx.Final() {
continue
}
if IndexFull(idx, mi.compress) {
debug.Log("index %p is full", idx)
idx.Finalize()
list = append(list, idx)
} else {
debug.Log("index %p not full", idx)
}
}
debug.Log("return %d indexes", len(list))
return list
}
// Each returns a channel that yields all blobs known to the index. When the
// context is cancelled, the background goroutine terminates. This blocks any
// modification of the index.
func (mi *MasterIndex) Each(ctx context.Context) <-chan restic.PackedBlob {
mi.idxMutex.RLock()
ch := make(chan restic.PackedBlob)
go func() {
defer mi.idxMutex.RUnlock()
defer close(ch)
for _, idx := range mi.idx {
for pb := range idx.Each(ctx) {
select {
case <-ctx.Done():
return
case ch <- pb:
}
}
}
}()
return ch
}
// MergeFinalIndexes merges all final indexes together.
// After calling, there will be only one big final index in MasterIndex
// containing all final index contents.
// Indexes that are not final are left untouched.
// This merging can only be called after all index files are loaded - as
// removing of superseded index contents is only possible for unmerged indexes.
func (mi *MasterIndex) MergeFinalIndexes() error {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
// The first index is always final and the one to merge into
newIdx := mi.idx[:1]
for i := 1; i < len(mi.idx); i++ {
idx := mi.idx[i]
// clear reference in masterindex as it may become stale
mi.idx[i] = nil
// do not merge indexes that have no id set
ids, _ := idx.IDs()
if !idx.Final() || len(ids) == 0 {
newIdx = append(newIdx, idx)
} else {
err := mi.idx[0].merge(idx)
if err != nil {
return fmt.Errorf("MergeFinalIndexes: %w", err)
}
}
}
mi.idx = newIdx
return nil
}
// Save saves all known indexes to index files, leaving out any
// packs whose ID is contained in packBlacklist from finalized indexes.
// The new index contains the IDs of all known indexes in the "supersedes"
// field. The IDs are also returned in the IDSet obsolete.
// After calling this function, you should remove the obsolete index files.
func (mi *MasterIndex) Save(ctx context.Context, repo restic.SaverUnpacked, packBlacklist restic.IDSet, extraObsolete restic.IDs, p *progress.Counter) (obsolete restic.IDSet, err error) {
p.SetMax(uint64(len(mi.Packs(packBlacklist))))
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
debug.Log("start rebuilding index of %d indexes, pack blacklist: %v", len(mi.idx), packBlacklist)
newIndex := NewIndex()
obsolete = restic.NewIDSet()
// track spawned goroutines using wg, create a new context which is
// cancelled as soon as an error occurs.
wg, ctx := errgroup.WithContext(ctx)
ch := make(chan *Index)
wg.Go(func() error {
defer close(ch)
for i, idx := range mi.idx {
if idx.Final() {
ids, err := idx.IDs()
if err != nil {
debug.Log("index %d does not have an ID: %v", err)
return err
}
debug.Log("adding index ids %v to supersedes field", ids)
err = newIndex.AddToSupersedes(ids...)
if err != nil {
return err
}
obsolete.Merge(restic.NewIDSet(ids...))
} else {
debug.Log("index %d isn't final, don't add to supersedes field", i)
}
debug.Log("adding index %d", i)
for pbs := range idx.EachByPack(ctx, packBlacklist) {
newIndex.StorePack(pbs.PackID, pbs.Blobs)
p.Add(1)
if IndexFull(newIndex, mi.compress) {
select {
case ch <- newIndex:
case <-ctx.Done():
return ctx.Err()
}
newIndex = NewIndex()
}
}
}
err = newIndex.AddToSupersedes(extraObsolete...)
if err != nil {
return err
}
obsolete.Merge(restic.NewIDSet(extraObsolete...))
select {
case ch <- newIndex:
case <-ctx.Done():
}
return nil
})
// a worker receives an index from ch, and saves the index
worker := func() error {
for idx := range ch {
idx.Finalize()
if _, err := SaveIndex(ctx, repo, idx); err != nil {
return err
}
}
return nil
}
// encoding an index can take quite some time such that this can be both CPU- or IO-bound
workerCount := int(repo.Connections()) + runtime.GOMAXPROCS(0)
// run workers on ch
for i := 0; i < workerCount; i++ {
wg.Go(worker)
}
err = wg.Wait()
return obsolete, err
}
// SaveIndex saves an index in the repository.
func SaveIndex(ctx context.Context, repo restic.SaverUnpacked, index *Index) (restic.ID, error) {
buf := bytes.NewBuffer(nil)
err := index.Encode(buf)
if err != nil {
return restic.ID{}, err
}
id, err := repo.SaveUnpacked(ctx, restic.IndexFile, buf.Bytes())
ierr := index.SetID(id)
if ierr != nil {
// logic bug
panic(ierr)
}
return id, err
}
// saveIndex saves all indexes in the backend.
func (mi *MasterIndex) saveIndex(ctx context.Context, r restic.SaverUnpacked, indexes ...*Index) error {
for i, idx := range indexes {
debug.Log("Saving index %d", i)
sid, err := SaveIndex(ctx, r, idx)
if err != nil {
return err
}
debug.Log("Saved index %d as %v", i, sid)
}
return mi.MergeFinalIndexes()
}
// SaveIndex saves all new indexes in the backend.
func (mi *MasterIndex) SaveIndex(ctx context.Context, r restic.SaverUnpacked) error {
return mi.saveIndex(ctx, r, mi.finalizeNotFinalIndexes()...)
}
// SaveFullIndex saves all full indexes in the backend.
func (mi *MasterIndex) SaveFullIndex(ctx context.Context, r restic.SaverUnpacked) error {
return mi.saveIndex(ctx, r, mi.finalizeFullIndexes()...)
}
// ListPacks returns the blobs of the specified pack files grouped by pack file.
func (mi *MasterIndex) ListPacks(ctx context.Context, packs restic.IDSet) <-chan restic.PackBlobs {
out := make(chan restic.PackBlobs)
go func() {
defer close(out)
// only resort a part of the index to keep the memory overhead bounded
for i := byte(0); i < 16; i++ {
if ctx.Err() != nil {
return
}
packBlob := make(map[restic.ID][]restic.Blob)
for pack := range packs {
if pack[0]&0xf == i {
packBlob[pack] = nil
}
}
if len(packBlob) == 0 {
continue
}
for pb := range mi.Each(ctx) {
if packs.Has(pb.PackID) && pb.PackID[0]&0xf == i {
packBlob[pb.PackID] = append(packBlob[pb.PackID], pb.Blob)
}
}
// pass on packs
for packID, pbs := range packBlob {
select {
case out <- restic.PackBlobs{PackID: packID, Blobs: pbs}:
case <-ctx.Done():
return
}
}
}
}()
return out
}