Upgrade versions of vendored libs

vendor/github.com/cockroachdb/pebble/internal/manifest/version_edit.go

@@ -51,7 +51,9 @@ const (
 	tagMaxColumnFamily  = 203
 
 	// Pebble tags.
-	tagNewFile5 = 104 // Range keys.
+	tagNewFile5            = 104 // Range keys.
+	tagCreatedBackingTable = 105
+	tagRemovedBackingTable = 106
 
 	// The custom tags sub-format used by tagNewFile4 and above.
 	customTagTerminate         = 1
@@ -59,6 +61,7 @@ const (
 	customTagCreationTime      = 6
 	customTagPathID            = 65
 	customTagNonSafeIgnoreMask = 1 << 6
+	customTagVirtual           = 66
 )
 
 // DeletedFileEntry holds the state for a file deletion from a level. The file
@@ -73,6 +76,9 @@ type DeletedFileEntry struct {
 type NewFileEntry struct {
 	Level int
 	Meta  *FileMetadata
+	// BackingFileNum is only set during manifest replay, and only for virtual
+	// sstables.
+	BackingFileNum base.DiskFileNum
 }
 
 // VersionEdit holds the state for an edit to a Version along with other
@@ -143,7 +149,9 @@ type VersionEdit struct {
 
 // Decode decodes an edit from the specified reader.
 //
-// TODO(bananabrick): Support decoding of virtual sstable state.
+// Note that the Decode step will not set the FileBacking for virtual sstables
+// and the responsibility is left to the caller. However, the Decode step will
+// populate the NewFileEntry.BackingFileNum in VersionEdit.NewFiles.
 func (v *VersionEdit) Decode(r io.Reader) error {
 	br, ok := r.(byteReader)
 	if !ok {
@@ -196,6 +204,28 @@ func (v *VersionEdit) Decode(r io.Reader) error {
 			}
 			// NB: RocksDB does not use compaction pointers anymore.
 
+		case tagRemovedBackingTable:
+			n, err := d.readUvarint()
+			if err != nil {
+				return err
+			}
+			v.RemovedBackingTables = append(
+				v.RemovedBackingTables, base.FileNum(n).DiskFileNum(),
+			)
+		case tagCreatedBackingTable:
+			dfn, err := d.readUvarint()
+			if err != nil {
+				return err
+			}
+			size, err := d.readUvarint()
+			if err != nil {
+				return err
+			}
+			fileBacking := &FileBacking{
+				DiskFileNum: base.FileNum(dfn).DiskFileNum(),
+				Size:        size,
+			}
+			v.CreatedBackingTables = append(v.CreatedBackingTables, fileBacking)
 		case tagDeletedFile:
 			level, err := d.readLevel()
 			if err != nil {
@@ -301,6 +331,10 @@ func (v *VersionEdit) Decode(r io.Reader) error {
 			}
 			var markedForCompaction bool
 			var creationTime uint64
+			virtualState := struct {
+				virtual        bool
+				backingFileNum uint64
+			}{}
 			if tag == tagNewFile4 || tag == tagNewFile5 {
 				for {
 					customTag, err := d.readUvarint()
@@ -309,7 +343,16 @@ func (v *VersionEdit) Decode(r io.Reader) error {
 					}
 					if customTag == customTagTerminate {
 						break
+					} else if customTag == customTagVirtual {
+						virtualState.virtual = true
+						n, err := d.readUvarint()
+						if err != nil {
+							return err
+						}
+						virtualState.backingFileNum = n
+						continue
 					}
+
 					field, err := d.readBytes()
 					if err != nil {
 						return err
@@ -345,6 +388,7 @@ func (v *VersionEdit) Decode(r io.Reader) error {
 				SmallestSeqNum:      smallestSeqNum,
 				LargestSeqNum:       largestSeqNum,
 				MarkedForCompaction: markedForCompaction,
+				Virtual:             virtualState.virtual,
 			}
 			if tag != tagNewFile5 { // no range keys present
 				m.SmallestPointKey = base.DecodeInternalKey(smallestPointKey)
@@ -376,11 +420,18 @@ func (v *VersionEdit) Decode(r io.Reader) error {
 				}
 			}
 			m.boundsSet = true
-			m.InitPhysicalBacking()
-			v.NewFiles = append(v.NewFiles, NewFileEntry{
+			if !virtualState.virtual {
+				m.InitPhysicalBacking()
+			}
+
+			nfe := NewFileEntry{
 				Level: level,
 				Meta:  m,
-			})
+			}
+			if virtualState.virtual {
+				nfe.BackingFileNum = base.FileNum(virtualState.backingFileNum).DiskFileNum()
+			}
+			v.NewFiles = append(v.NewFiles, nfe)
 
 		case tagPrevLogNumber:
 			n, err := d.readUvarint()
@@ -399,8 +450,7 @@ func (v *VersionEdit) Decode(r io.Reader) error {
 	return nil
 }
 
-// String implements fmt.Stringer for a VersionEdit.
-func (v *VersionEdit) String() string {
+func (v *VersionEdit) string(verbose bool, fmtKey base.FormatKey) string {
 	var buf bytes.Buffer
 	if v.ComparerName != "" {
 		fmt.Fprintf(&buf, "  comparer:     %s", v.ComparerName)
@@ -431,7 +481,12 @@ func (v *VersionEdit) String() string {
 		fmt.Fprintf(&buf, "  deleted:       L%d %s\n", df.Level, df.FileNum)
 	}
 	for _, nf := range v.NewFiles {
-		fmt.Fprintf(&buf, "  added:         L%d %s", nf.Level, nf.Meta.String())
+		fmt.Fprintf(&buf, "  added:         L%d", nf.Level)
+		if verbose {
+			fmt.Fprintf(&buf, " %s", nf.Meta.DebugString(fmtKey, true /* verbose */))
+		} else {
+			fmt.Fprintf(&buf, " %s", nf.Meta.String())
+		}
 		if nf.Meta.CreationTime != 0 {
 			fmt.Fprintf(&buf, " (%s)",
 				time.Unix(nf.Meta.CreationTime, 0).UTC().Format(time.RFC3339))
@@ -441,9 +496,17 @@ func (v *VersionEdit) String() string {
 	return buf.String()
 }
 
+// DebugString is a more verbose version of String(). Use this in tests.
+func (v *VersionEdit) DebugString(fmtKey base.FormatKey) string {
+	return v.string(true /* verbose */, fmtKey)
+}
+
+// String implements fmt.Stringer for a VersionEdit.
+func (v *VersionEdit) String() string {
+	return v.string(false /* verbose */, base.DefaultFormatter)
+}
+
 // Encode encodes an edit to the specified writer.
-//
-// TODO(bananabrick): Support encoding of virtual sstable state.
 func (v *VersionEdit) Encode(w io.Writer) error {
 	e := versionEditEncoder{new(bytes.Buffer)}
 
@@ -463,6 +526,15 @@ func (v *VersionEdit) Encode(w io.Writer) error {
 		e.writeUvarint(tagNextFileNumber)
 		e.writeUvarint(uint64(v.NextFileNum))
 	}
+	for _, dfn := range v.RemovedBackingTables {
+		e.writeUvarint(tagRemovedBackingTable)
+		e.writeUvarint(uint64(dfn.FileNum()))
+	}
+	for _, fileBacking := range v.CreatedBackingTables {
+		e.writeUvarint(tagCreatedBackingTable)
+		e.writeUvarint(uint64(fileBacking.DiskFileNum.FileNum()))
+		e.writeUvarint(fileBacking.Size)
+	}
 	// RocksDB requires LastSeqNum to be encoded for the first MANIFEST entry,
 	// even though its value is zero. We detect this by encoding LastSeqNum when
 	// ComparerName is set.
@@ -476,7 +548,7 @@ func (v *VersionEdit) Encode(w io.Writer) error {
 		e.writeUvarint(uint64(x.FileNum))
 	}
 	for _, x := range v.NewFiles {
-		customFields := x.Meta.MarkedForCompaction || x.Meta.CreationTime != 0
+		customFields := x.Meta.MarkedForCompaction || x.Meta.CreationTime != 0 || x.Meta.Virtual
 		var tag uint64
 		switch {
 		case x.Meta.HasRangeKeys:
@@ -529,6 +601,10 @@ func (v *VersionEdit) Encode(w io.Writer) error {
 				e.writeUvarint(customTagNeedsCompaction)
 				e.writeBytes([]byte{1})
 			}
+			if x.Meta.Virtual {
+				e.writeUvarint(customTagVirtual)
+				e.writeUvarint(uint64(x.Meta.FileBacking.DiskFileNum.FileNum()))
+			}
 			e.writeUvarint(customTagTerminate)
 		}
 	}
@@ -536,6 +612,7 @@ func (v *VersionEdit) Encode(w io.Writer) error {
 	return err
 }
 
+// versionEditDecoder should be used to decode version edits.
 type versionEditDecoder struct {
 	byteReader
 }
@@ -634,7 +711,9 @@ type BulkVersionEdit struct {
 	Added   [NumLevels]map[base.FileNum]*FileMetadata
 	Deleted [NumLevels]map[base.FileNum]*FileMetadata
 
-	AddedFileBacking   []*FileBacking
+	// AddedFileBacking is a map to support lookup so that we can populate the
+	// FileBacking of virtual sstables during manifest replay.
+	AddedFileBacking   map[base.DiskFileNum]*FileBacking
 	RemovedFileBacking []base.DiskFileNum
 
 	// AddedByFileNum maps file number to file metadata for all added files
@@ -697,6 +776,22 @@ func (b *BulkVersionEdit) Accumulate(ve *VersionEdit) error {
 		}
 	}
 
+	// Generate state for Added backing files. Note that these must be generated
+	// before we loop through the NewFiles, because we need to populate the
+	// FileBackings which might be used by the NewFiles loop.
+	if b.AddedFileBacking == nil {
+		b.AddedFileBacking = make(map[base.DiskFileNum]*FileBacking)
+	}
+	for _, fb := range ve.CreatedBackingTables {
+		if _, ok := b.AddedFileBacking[fb.DiskFileNum]; ok {
+			// There is already a FileBacking associated with fb.DiskFileNum.
+			// This should never happen. There must always be only one FileBacking
+			// associated with a backing sstable.
+			panic(fmt.Sprintf("pebble: duplicate file backing %s", fb.DiskFileNum.String()))
+		}
+		b.AddedFileBacking[fb.DiskFileNum] = fb
+	}
+
 	for _, nf := range ve.NewFiles {
 		// A new file should not have been deleted in this or a preceding
 		// VersionEdit at the same level (though files can move across levels).
@@ -705,6 +800,17 @@ func (b *BulkVersionEdit) Accumulate(ve *VersionEdit) error {
 				return base.CorruptionErrorf("pebble: file deleted L%d.%s before it was inserted", nf.Level, nf.Meta.FileNum)
 			}
 		}
+		if nf.Meta.Virtual && nf.Meta.FileBacking == nil {
+			// FileBacking for a virtual sstable must only be nil if we're performing
+			// manifest replay.
+			nf.Meta.FileBacking = b.AddedFileBacking[nf.BackingFileNum]
+			if nf.Meta.FileBacking == nil {
+				return errors.Errorf("FileBacking for virtual sstable must not be nil")
+			}
+		} else if nf.Meta.FileBacking == nil {
+			return errors.Errorf("Added file L%d.%s's has no FileBacking", nf.Level, nf.Meta.FileNum)
+		}
+
 		if b.Added[nf.Level] == nil {
 			b.Added[nf.Level] = make(map[base.FileNum]*FileMetadata)
 		}
@@ -717,9 +823,6 @@ func (b *BulkVersionEdit) Accumulate(ve *VersionEdit) error {
 		}
 	}
 
-	// Generate state for the backing files.
-	b.AddedFileBacking = append(b.AddedFileBacking, ve.CreatedBackingTables...)
-
 	// Since a file can be removed from backing files in exactly one version
 	// edit it is safe to just append without any de-duplication.
 	b.RemovedFileBacking = append(b.RemovedFileBacking, ve.RemovedBackingTables...)
@@ -749,6 +852,9 @@ func AccumulateIncompleteAndApplySingleVE(
 	flushSplitBytes int64,
 	readCompactionRate int64,
 	backingStateMap map[base.DiskFileNum]*FileBacking,
+	addBackingFunc func(*FileBacking),
+	removeBackingFunc func(base.DiskFileNum),
+	orderingInvariants OrderingInvariants,
 ) (_ *Version, zombies map[base.DiskFileNum]uint64, _ error) {
 	if len(ve.RemovedBackingTables) != 0 {
 		panic("pebble: invalid incomplete version edit")
@@ -760,14 +866,14 @@ func AccumulateIncompleteAndApplySingleVE(
 	}
 	zombies = make(map[base.DiskFileNum]uint64)
 	v, err := b.Apply(
-		curr, cmp, formatKey, flushSplitBytes, readCompactionRate, zombies,
+		curr, cmp, formatKey, flushSplitBytes, readCompactionRate, zombies, orderingInvariants,
 	)
 	if err != nil {
 		return nil, nil, err
 	}
 
 	for _, s := range b.AddedFileBacking {
-		backingStateMap[s.DiskFileNum] = s
+		addBackingFunc(s)
 	}
 
 	for fileNum := range zombies {
@@ -778,10 +884,9 @@ func AccumulateIncompleteAndApplySingleVE(
 			ve.RemovedBackingTables = append(
 				ve.RemovedBackingTables, fileNum,
 			)
-			delete(backingStateMap, fileNum)
+			removeBackingFunc(fileNum)
 		}
 	}
-
 	return v, zombies, nil
 }
 
@@ -802,6 +907,7 @@ func (b *BulkVersionEdit) Apply(
 	flushSplitBytes int64,
 	readCompactionRate int64,
 	zombies map[base.DiskFileNum]uint64,
+	orderingInvariants OrderingInvariants,
 ) (*Version, error) {
 	addZombie := func(state *FileBacking) {
 		if zombies != nil {
@@ -870,7 +976,7 @@ func (b *BulkVersionEdit) Apply(
 		// level.
 
 		for _, f := range deletedFilesMap {
-			if obsolete := v.Levels[level].tree.Delete(f); obsolete {
+			if obsolete := v.Levels[level].remove(f); obsolete {
 				// Deleting a file from the B-Tree may decrement its
 				// reference count. However, because we cloned the
 				// previous level's B-Tree, this should never result in a
@@ -879,7 +985,7 @@ func (b *BulkVersionEdit) Apply(
 				return nil, err
 			}
 			if f.HasRangeKeys {
-				if obsolete := v.RangeKeyLevels[level].tree.Delete(f); obsolete {
+				if obsolete := v.RangeKeyLevels[level].remove(f); obsolete {
 					// Deleting a file from the B-Tree may decrement its
 					// reference count. However, because we cloned the
 					// previous level's B-Tree, this should never result in a
@@ -932,7 +1038,7 @@ func (b *BulkVersionEdit) Apply(
 			f.AllowedSeeks.Store(allowedSeeks)
 			f.InitAllowedSeeks = allowedSeeks
 
-			err := lm.tree.Insert(f)
+			err := lm.insert(f)
 			// We're adding this file to the new version, so increment the
 			// latest refs count.
 			f.LatestRef()
@@ -940,7 +1046,7 @@ func (b *BulkVersionEdit) Apply(
 				return nil, errors.Wrap(err, "pebble")
 			}
 			if f.HasRangeKeys {
-				err = lmRange.tree.Insert(f)
+				err = lmRange.insert(f)
 				if err != nil {
 					return nil, errors.Wrap(err, "pebble")
 				}
@@ -985,7 +1091,7 @@ func (b *BulkVersionEdit) Apply(
 			} else if err := v.InitL0Sublevels(cmp, formatKey, flushSplitBytes); err != nil {
 				return nil, errors.Wrap(err, "pebble: internal error")
 			}
-			if err := CheckOrdering(cmp, formatKey, Level(0), v.Levels[level].Iter()); err != nil {
+			if err := CheckOrdering(cmp, formatKey, Level(0), v.Levels[level].Iter(), orderingInvariants); err != nil {
 				return nil, errors.Wrap(err, "pebble: internal error")
 			}
 			continue
@@ -1006,7 +1112,7 @@ func (b *BulkVersionEdit) Apply(
 					end.Prev()
 				}
 			})
-			if err := CheckOrdering(cmp, formatKey, Level(level), check.Iter()); err != nil {
+			if err := CheckOrdering(cmp, formatKey, Level(level), check.Iter(), orderingInvariants); err != nil {
 				return nil, errors.Wrap(err, "pebble: internal error")
 			}
 		}