colblk: cache KeyspanDecoder in block metadata

sstable/colblk/data_block.go

@@ -763,9 +763,6 @@ const _ uint = uint(-(dataBlockDecoderSize % 8))
 // Assert that a DataBlockDecoder and a KeySeekerMetadata can fit inside block.Metadata.
 const _ uint = block.MetadataSize - uint(dataBlockDecoderSize) - KeySeekerMetadataSize
 
-// Assert that an IndexBlockDecoder can fit inside block.Metadata.
-const _ uint = block.MetadataSize - uint(unsafe.Sizeof(IndexBlockDecoder{}))
-
 // InitDataBlockMetadata initializes the metadata for a data block.
 func InitDataBlockMetadata(schema KeySchema, md *block.Metadata, data []byte) (err error) {
  if uintptr(unsafe.Pointer(md))%8 != 0 {
@@ -785,6 +782,9 @@ func InitDataBlockMetadata(schema KeySchema, md *block.Metadata, data []byte) (e
  return nil
 }
 
+// Assert that an IndexBlockDecoder can fit inside block.Metadata.
+const _ uint = block.MetadataSize - uint(unsafe.Sizeof(IndexBlockDecoder{}))
+
 // InitIndexBlockMetadata initializes the metadata for an index block.
 func InitIndexBlockMetadata(md *block.Metadata, data []byte) (err error) {
  if uintptr(unsafe.Pointer(md))%8 != 0 {
@@ -795,7 +795,27 @@ func InitIndexBlockMetadata(md *block.Metadata, data []byte) (err error) {
  // layers can add file number and offset information).
  defer func() {
  if r := recover(); r != nil {
- err = base.CorruptionErrorf("error initializing data block metadata: %v", r)
+ err = base.CorruptionErrorf("error initializing index block metadata: %v", r)
+ }
+ }()
+ d.Init(data)
+ return nil
+}
+
+// Assert that a IndexBlockDecoder can fit inside block.Metadata.
+const _ uint = block.MetadataSize - uint(unsafe.Sizeof(KeyspanDecoder{}))
+
+// InitKeyspanBlockMetadata initializes the metadata for a rangedel or range key block.
+func InitKeyspanBlockMetadata(md *block.Metadata, data []byte) (err error) {
+ if uintptr(unsafe.Pointer(md))%8 != 0 {
+ return errors.AssertionFailedf("metadata is not 8-byte aligned")
+ }
+ d := (*KeyspanDecoder)(unsafe.Pointer(md))
+ // Initialization can panic; convert panics to corruption errors (so higher
+ // layers can add file number and offset information).
+ defer func() {
+ if r := recover(); r != nil {
+ err = base.CorruptionErrorf("error initializing keyspan block metadata: %v", r)
  }
  }()
  d.Init(data)

sstable/colblk/keyspan.go

@@ -11,6 +11,7 @@ import (
  "fmt"
  "os"
  "sync"
+ "unsafe"
 
  "github.com/cockroachdb/errors"
  "github.com/cockroachdb/pebble/internal/base"
@@ -327,12 +328,8 @@ func NewKeyspanIter(
  i := keyspanIterPool.Get().(*KeyspanIter)
  i.closeCheck = invariants.CloseChecker{}
  i.handle = h
- // TODO(jackson): We can teach the block cache to stash a *KeyspanDecoder.
- // Then all iters would use the same decoder rather than needing to allocate
- // their own KeyspanDecoder and parse the high-level block structure
- // themselves.
- i.allocDecoder.Init(h.BlockData())
- i.init(cmp, &i.allocDecoder, transforms)
+ d := (*KeyspanDecoder)(unsafe.Pointer(h.BlockMetadata()))
+ i.init(cmp, d, transforms)
  return i
 }
 
@@ -353,8 +350,7 @@ var keyspanIterPool = sync.Pool{
 // keyspan.FragmentIterator interface.
 type KeyspanIter struct {
  keyspanIter
- handle block.BufferHandle
- allocDecoder KeyspanDecoder
+ handle block.BufferHandle
 
  closeCheck invariants.CloseChecker
 }
@@ -372,7 +368,6 @@ func (i *KeyspanIter) Close() {
  }
 
  i.keyspanIter.Close()
- i.allocDecoder = KeyspanDecoder{}
  i.closeCheck.Close()
  keyspanIterPool.Put(i)
 }

sstable/colblk/keyspan_test.go

@@ -13,6 +13,7 @@ import (
  "sync"
  "testing"
  "time"
+ "unsafe"
 
  "github.com/cockroachdb/datadriven"
  "github.com/cockroachdb/pebble/internal/base"
@@ -86,6 +87,8 @@ func TestKeyspanBlockPooling(t *testing.T) {
  defer c.Unref()
  v := block.Alloc(len(b), nil)
  copy(v.BlockData(), b)
+ d := (*KeyspanDecoder)(unsafe.Pointer(v.BlockMetadata()))
+ d.Init(v.BlockData())
  v.MakeHandle(c, cache.ID(1), base.DiskFileNum(1), 0).Release()
 
  getBlockAndIterate := func() {

sstable/reader.go

@@ -431,14 +431,23 @@ func (r *Reader) readRangeDelBlock(
  ctx context.Context, env readBlockEnv, readHandle objstorage.ReadHandle, bh block.Handle,
 ) (block.BufferHandle, error) {
  ctx = objiotracing.WithBlockType(ctx, objiotracing.MetadataBlock)
- return r.readBlockInternal(ctx, env, readHandle, bh, noInitBlockMetadataFn)
+ return r.readBlockInternal(ctx, env, readHandle, bh, r.initKeyspanBlockMetadata)
 }
 
 func (r *Reader) readRangeKeyBlock(
  ctx context.Context, env readBlockEnv, readHandle objstorage.ReadHandle, bh block.Handle,
 ) (block.BufferHandle, error) {
  ctx = objiotracing.WithBlockType(ctx, objiotracing.MetadataBlock)
- return r.readBlockInternal(ctx, env, readHandle, bh, noInitBlockMetadataFn)
+ return r.readBlockInternal(ctx, env, readHandle, bh, r.initKeyspanBlockMetadata)
+}
+
+// initKeyspanBlockMetadata initializes the Metadata for a rangedel or range key
+// block. This will later be used (and reused) when reading from the block.
+func (r *Reader) initKeyspanBlockMetadata(metadata *block.Metadata, data []byte) error {
+ if r.tableFormat.BlockColumnar() {
+ return colblk.InitKeyspanBlockMetadata(metadata, data)
+ }
+ return nil
 }
 
 func (r *Reader) readValueBlock(