[WIP] Partial bot skeleton, non-functional though

vendor/github.com/cockroachdb/pebble/internal/cache/LICENSE

@@ -0,0 +1,21 @@
+The MIT License
+
+Copyright (c) 2018 Damian Gryski
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

vendor/github.com/cockroachdb/pebble/internal/cache/cgo_disabled.go

@@ -0,0 +1,10 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build !cgo
+// +build !cgo
+
+package cache
+
+const cgoEnabled = false

vendor/github.com/cockroachdb/pebble/internal/cache/cgo_enabled.go

@@ -0,0 +1,10 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build cgo
+// +build cgo
+
+package cache
+
+const cgoEnabled = true

vendor/github.com/cockroachdb/pebble/internal/cache/clockpro.go

@@ -0,0 +1,841 @@
+// Copyright 2018. All rights reserved. Use of this source code is governed by
+// an MIT-style license that can be found in the LICENSE file.
+
+// Package cache implements the CLOCK-Pro caching algorithm.
+//
+// CLOCK-Pro is a patent-free alternative to the Adaptive Replacement Cache,
+// https://en.wikipedia.org/wiki/Adaptive_replacement_cache.
+// It is an approximation of LIRS ( https://en.wikipedia.org/wiki/LIRS_caching_algorithm ),
+// much like the CLOCK page replacement algorithm is an approximation of LRU.
+//
+// This implementation is based on the python code from https://bitbucket.org/SamiLehtinen/pyclockpro .
+//
+// Slides describing the algorithm: http://fr.slideshare.net/huliang64/clockpro
+//
+// The original paper: http://static.usenix.org/event/usenix05/tech/general/full_papers/jiang/jiang_html/html.html
+//
+// It is MIT licensed, like the original.
+package cache // import "github.com/cockroachdb/pebble/internal/cache"
+
+import (
+	"fmt"
+	"os"
+	"runtime"
+	"runtime/debug"
+	"strings"
+	"sync"
+	"sync/atomic"
+
+	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/invariants"
+)
+
+type fileKey struct {
+	// id is the namespace for fileNums.
+	id      uint64
+	fileNum base.FileNum
+}
+
+type key struct {
+	fileKey
+	offset uint64
+}
+
+// file returns the "file key" for the receiver. This is the key used for the
+// shard.files map.
+func (k key) file() key {
+	k.offset = 0
+	return k
+}
+
+func (k key) String() string {
+	return fmt.Sprintf("%d/%d/%d", k.id, k.fileNum, k.offset)
+}
+
+// Handle provides a strong reference to a value in the cache. The reference
+// does not pin the value in the cache, but it does prevent the underlying byte
+// slice from being reused.
+type Handle struct {
+	value *Value
+}
+
+// Get returns the value stored in handle.
+func (h Handle) Get() []byte {
+	if h.value != nil {
+		// NB: We don't increment shard.hits in this code path because we only want
+		// to record a hit when the handle is retrieved from the cache.
+		return h.value.buf
+	}
+	return nil
+}
+
+// Release releases the reference to the cache entry.
+func (h Handle) Release() {
+	if h.value != nil {
+		h.value.release()
+	}
+}
+
+type shard struct {
+	hits   int64
+	misses int64
+
+	mu sync.RWMutex
+
+	reservedSize int64
+	maxSize      int64
+	coldTarget   int64
+	blocks       robinHoodMap // fileNum+offset -> block
+	files        robinHoodMap // fileNum -> list of blocks
+
+	// The blocks and files maps store values in manually managed memory that is
+	// invisible to the Go GC. This is fine for Value and entry objects that are
+	// stored in manually managed memory, but when the "invariants" build tag is
+	// set, all Value and entry objects are Go allocated and the entries map will
+	// contain a reference to every entry.
+	entries map[*entry]struct{}
+
+	handHot  *entry
+	handCold *entry
+	handTest *entry
+
+	sizeHot  int64
+	sizeCold int64
+	sizeTest int64
+
+	// The count fields are used exclusively for asserting expectations.
+	// We've seen infinite looping (cockroachdb/cockroach#70154) that
+	// could be explained by a corrupted sizeCold. Through asserting on
+	// these fields, we hope to gain more insight from any future
+	// reproductions.
+	countHot  int64
+	countCold int64
+	countTest int64
+}
+
+func (c *shard) Get(id uint64, fileNum base.FileNum, offset uint64) Handle {
+	c.mu.RLock()
+	var value *Value
+	if e := c.blocks.Get(key{fileKey{id, fileNum}, offset}); e != nil {
+		value = e.acquireValue()
+		if value != nil {
+			atomic.StoreInt32(&e.referenced, 1)
+		}
+	}
+	c.mu.RUnlock()
+	if value == nil {
+		atomic.AddInt64(&c.misses, 1)
+		return Handle{}
+	}
+	atomic.AddInt64(&c.hits, 1)
+	return Handle{value: value}
+}
+
+func (c *shard) Set(id uint64, fileNum base.FileNum, offset uint64, value *Value) Handle {
+	if n := value.refs(); n != 1 {
+		panic(fmt.Sprintf("pebble: Value has already been added to the cache: refs=%d", n))
+	}
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	k := key{fileKey{id, fileNum}, offset}
+	e := c.blocks.Get(k)
+
+	switch {
+	case e == nil:
+		// no cache entry? add it
+		e = newEntry(c, k, int64(len(value.buf)))
+		e.setValue(value)
+		if c.metaAdd(k, e) {
+			value.ref.trace("add-cold")
+			c.sizeCold += e.size
+			c.countCold++
+		} else {
+			value.ref.trace("skip-cold")
+			e.free()
+			e = nil
+		}
+
+	case e.peekValue() != nil:
+		// cache entry was a hot or cold page
+		e.setValue(value)
+		atomic.StoreInt32(&e.referenced, 1)
+		delta := int64(len(value.buf)) - e.size
+		e.size = int64(len(value.buf))
+		if e.ptype == etHot {
+			value.ref.trace("add-hot")
+			c.sizeHot += delta
+		} else {
+			value.ref.trace("add-cold")
+			c.sizeCold += delta
+		}
+		c.evict()
+
+	default:
+		// cache entry was a test page
+		c.sizeTest -= e.size
+		c.countTest--
+		c.metaDel(e)
+		c.metaCheck(e)
+
+		e.size = int64(len(value.buf))
+		c.coldTarget += e.size
+		if c.coldTarget > c.targetSize() {
+			c.coldTarget = c.targetSize()
+		}
+
+		atomic.StoreInt32(&e.referenced, 0)
+		e.setValue(value)
+		e.ptype = etHot
+		if c.metaAdd(k, e) {
+			value.ref.trace("add-hot")
+			c.sizeHot += e.size
+			c.countHot++
+		} else {
+			value.ref.trace("skip-hot")
+			e.free()
+			e = nil
+		}
+	}
+
+	c.checkConsistency()
+
+	// Values are initialized with a reference count of 1. That reference count
+	// is being transferred to the returned Handle.
+	return Handle{value: value}
+}
+
+func (c *shard) checkConsistency() {
+	// See the comment above the count{Hot,Cold,Test} fields.
+	switch {
+	case c.sizeHot < 0 || c.sizeCold < 0 || c.sizeTest < 0 || c.countHot < 0 || c.countCold < 0 || c.countTest < 0:
+		panic(fmt.Sprintf("pebble: unexpected negative: %d (%d bytes) hot, %d (%d bytes) cold, %d (%d bytes) test",
+			c.countHot, c.sizeHot, c.countCold, c.sizeCold, c.countTest, c.sizeTest))
+	case c.sizeHot > 0 && c.countHot == 0:
+		panic(fmt.Sprintf("pebble: mismatch %d hot size, %d hot count", c.sizeHot, c.countHot))
+	case c.sizeCold > 0 && c.countCold == 0:
+		panic(fmt.Sprintf("pebble: mismatch %d cold size, %d cold count", c.sizeCold, c.countCold))
+	case c.sizeTest > 0 && c.countTest == 0:
+		panic(fmt.Sprintf("pebble: mismatch %d test size, %d test count", c.sizeTest, c.countTest))
+	}
+}
+
+// Delete deletes the cached value for the specified file and offset.
+func (c *shard) Delete(id uint64, fileNum base.FileNum, offset uint64) {
+	// The common case is there is nothing to delete, so do a quick check with
+	// shared lock.
+	k := key{fileKey{id, fileNum}, offset}
+	c.mu.RLock()
+	exists := c.blocks.Get(k) != nil
+	c.mu.RUnlock()
+	if !exists {
+		return
+	}
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	e := c.blocks.Get(k)
+	if e == nil {
+		return
+	}
+	c.metaEvict(e)
+
+	c.checkConsistency()
+}
+
+// EvictFile evicts all of the cache values for the specified file.
+func (c *shard) EvictFile(id uint64, fileNum base.FileNum) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	fkey := key{fileKey{id, fileNum}, 0}
+	blocks := c.files.Get(fkey)
+	if blocks == nil {
+		return
+	}
+	for b, n := blocks, (*entry)(nil); ; b = n {
+		n = b.fileLink.next
+		c.metaEvict(b)
+		if b == n {
+			break
+		}
+	}
+
+	c.checkConsistency()
+}
+
+func (c *shard) Free() {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	// NB: we use metaDel rather than metaEvict in order to avoid the expensive
+	// metaCheck call when the "invariants" build tag is specified.
+	for c.handHot != nil {
+		e := c.handHot
+		c.metaDel(c.handHot)
+		e.free()
+	}
+
+	c.blocks.free()
+	c.files.free()
+}
+
+func (c *shard) Reserve(n int) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.reservedSize += int64(n)
+
+	// Changing c.reservedSize will either increase or decrease
+	// the targetSize. But we want coldTarget to be in the range
+	// [0, targetSize]. So, if c.targetSize decreases, make sure
+	// that the coldTarget fits within the limits.
+	targetSize := c.targetSize()
+	if c.coldTarget > targetSize {
+		c.coldTarget = targetSize
+	}
+
+	c.evict()
+	c.checkConsistency()
+}
+
+// Size returns the current space used by the cache.
+func (c *shard) Size() int64 {
+	c.mu.RLock()
+	size := c.sizeHot + c.sizeCold
+	c.mu.RUnlock()
+	return size
+}
+
+func (c *shard) targetSize() int64 {
+	target := c.maxSize - c.reservedSize
+	// Always return a positive integer for targetSize. This is so that we don't
+	// end up in an infinite loop in evict(), in cases where reservedSize is
+	// greater than or equal to maxSize.
+	if target < 1 {
+		return 1
+	}
+	return target
+}
+
+// Add the entry to the cache, returning true if the entry was added and false
+// if it would not fit in the cache.
+func (c *shard) metaAdd(key key, e *entry) bool {
+	c.evict()
+	if e.size > c.targetSize() {
+		// The entry is larger than the target cache size.
+		return false
+	}
+
+	c.blocks.Put(key, e)
+	if entriesGoAllocated {
+		// Go allocated entries need to be referenced from Go memory. The entries
+		// map provides that reference.
+		c.entries[e] = struct{}{}
+	}
+
+	if c.handHot == nil {
+		// first element
+		c.handHot = e
+		c.handCold = e
+		c.handTest = e
+	} else {
+		c.handHot.link(e)
+	}
+
+	if c.handCold == c.handHot {
+		c.handCold = c.handCold.prev()
+	}
+
+	fkey := key.file()
+	if fileBlocks := c.files.Get(fkey); fileBlocks == nil {
+		c.files.Put(fkey, e)
+	} else {
+		fileBlocks.linkFile(e)
+	}
+	return true
+}
+
+// Remove the entry from the cache. This removes the entry from the blocks map,
+// the files map, and ensures that hand{Hot,Cold,Test} are not pointing at the
+// entry.
+func (c *shard) metaDel(e *entry) {
+	if value := e.peekValue(); value != nil {
+		value.ref.trace("metaDel")
+	}
+	e.setValue(nil)
+
+	c.blocks.Delete(e.key)
+	if entriesGoAllocated {
+		// Go allocated entries need to be referenced from Go memory. The entries
+		// map provides that reference.
+		delete(c.entries, e)
+	}
+
+	if e == c.handHot {
+		c.handHot = c.handHot.prev()
+	}
+	if e == c.handCold {
+		c.handCold = c.handCold.prev()
+	}
+	if e == c.handTest {
+		c.handTest = c.handTest.prev()
+	}
+
+	if e.unlink() == e {
+		// This was the last entry in the cache.
+		c.handHot = nil
+		c.handCold = nil
+		c.handTest = nil
+	}
+
+	fkey := e.key.file()
+	if next := e.unlinkFile(); e == next {
+		c.files.Delete(fkey)
+	} else {
+		c.files.Put(fkey, next)
+	}
+}
+
+// Check that the specified entry is not referenced by the cache.
+func (c *shard) metaCheck(e *entry) {
+	if invariants.Enabled {
+		if _, ok := c.entries[e]; ok {
+			fmt.Fprintf(os.Stderr, "%p: %s unexpectedly found in entries map\n%s",
+				e, e.key, debug.Stack())
+			os.Exit(1)
+		}
+		if c.blocks.findByValue(e) != nil {
+			fmt.Fprintf(os.Stderr, "%p: %s unexpectedly found in blocks map\n%s\n%s",
+				e, e.key, &c.blocks, debug.Stack())
+			os.Exit(1)
+		}
+		if c.files.findByValue(e) != nil {
+			fmt.Fprintf(os.Stderr, "%p: %s unexpectedly found in files map\n%s\n%s",
+				e, e.key, &c.files, debug.Stack())
+			os.Exit(1)
+		}
+		// NB: c.hand{Hot,Cold,Test} are pointers into a single linked list. We
+		// only have to traverse one of them to check all of them.
+		var countHot, countCold, countTest int64
+		var sizeHot, sizeCold, sizeTest int64
+		for t := c.handHot.next(); t != nil; t = t.next() {
+			// Recompute count{Hot,Cold,Test} and size{Hot,Cold,Test}.
+			switch t.ptype {
+			case etHot:
+				countHot++
+				sizeHot += t.size
+			case etCold:
+				countCold++
+				sizeCold += t.size
+			case etTest:
+				countTest++
+				sizeTest += t.size
+			}
+			if e == t {
+				fmt.Fprintf(os.Stderr, "%p: %s unexpectedly found in blocks list\n%s",
+					e, e.key, debug.Stack())
+				os.Exit(1)
+			}
+			if t == c.handHot {
+				break
+			}
+		}
+		if countHot != c.countHot || countCold != c.countCold || countTest != c.countTest ||
+			sizeHot != c.sizeHot || sizeCold != c.sizeCold || sizeTest != c.sizeTest {
+			fmt.Fprintf(os.Stderr, `divergence of Hot,Cold,Test statistics
+				cache's statistics: hot %d, %d, cold %d, %d, test %d, %d
+				recalculated statistics: hot %d, %d, cold %d, %d, test %d, %d\n%s`,
+				c.countHot, c.sizeHot, c.countCold, c.sizeCold, c.countTest, c.sizeTest,
+				countHot, sizeHot, countCold, sizeCold, countTest, sizeTest,
+				debug.Stack())
+			os.Exit(1)
+		}
+	}
+}
+
+func (c *shard) metaEvict(e *entry) {
+	switch e.ptype {
+	case etHot:
+		c.sizeHot -= e.size
+		c.countHot--
+	case etCold:
+		c.sizeCold -= e.size
+		c.countCold--
+	case etTest:
+		c.sizeTest -= e.size
+		c.countTest--
+	}
+	c.metaDel(e)
+	c.metaCheck(e)
+	e.free()
+}
+
+func (c *shard) evict() {
+	for c.targetSize() <= c.sizeHot+c.sizeCold && c.handCold != nil {
+		c.runHandCold(c.countCold, c.sizeCold)
+	}
+}
+
+func (c *shard) runHandCold(countColdDebug, sizeColdDebug int64) {
+	// countColdDebug and sizeColdDebug should equal c.countCold and
+	// c.sizeCold. They're parameters only to aid in debugging of
+	// cockroachdb/cockroach#70154. Since they're parameters, their
+	// arguments will appear within stack traces should we encounter
+	// a reproduction.
+	if c.countCold != countColdDebug || c.sizeCold != sizeColdDebug {
+		panic(fmt.Sprintf("runHandCold: cold count and size are %d, %d, arguments are %d and %d",
+			c.countCold, c.sizeCold, countColdDebug, sizeColdDebug))
+	}
+
+	e := c.handCold
+	if e.ptype == etCold {
+		if atomic.LoadInt32(&e.referenced) == 1 {
+			atomic.StoreInt32(&e.referenced, 0)
+			e.ptype = etHot
+			c.sizeCold -= e.size
+			c.countCold--
+			c.sizeHot += e.size
+			c.countHot++
+		} else {
+			e.setValue(nil)
+			e.ptype = etTest
+			c.sizeCold -= e.size
+			c.countCold--
+			c.sizeTest += e.size
+			c.countTest++
+			for c.targetSize() < c.sizeTest && c.handTest != nil {
+				c.runHandTest()
+			}
+		}
+	}
+
+	c.handCold = c.handCold.next()
+
+	for c.targetSize()-c.coldTarget <= c.sizeHot && c.handHot != nil {
+		c.runHandHot()
+	}
+}
+
+func (c *shard) runHandHot() {
+	if c.handHot == c.handTest && c.handTest != nil {
+		c.runHandTest()
+		if c.handHot == nil {
+			return
+		}
+	}
+
+	e := c.handHot
+	if e.ptype == etHot {
+		if atomic.LoadInt32(&e.referenced) == 1 {
+			atomic.StoreInt32(&e.referenced, 0)
+		} else {
+			e.ptype = etCold
+			c.sizeHot -= e.size
+			c.countHot--
+			c.sizeCold += e.size
+			c.countCold++
+		}
+	}
+
+	c.handHot = c.handHot.next()
+}
+
+func (c *shard) runHandTest() {
+	if c.sizeCold > 0 && c.handTest == c.handCold && c.handCold != nil {
+		// sizeCold is > 0, so assert that countCold == 0. See the
+		// comment above count{Hot,Cold,Test}.
+		if c.countCold == 0 {
+			panic(fmt.Sprintf("pebble: mismatch %d cold size, %d cold count", c.sizeCold, c.countCold))
+		}
+
+		c.runHandCold(c.countCold, c.sizeCold)
+		if c.handTest == nil {
+			return
+		}
+	}
+
+	e := c.handTest
+	if e.ptype == etTest {
+		c.sizeTest -= e.size
+		c.countTest--
+		c.coldTarget -= e.size
+		if c.coldTarget < 0 {
+			c.coldTarget = 0
+		}
+		c.metaDel(e)
+		c.metaCheck(e)
+		e.free()
+	}
+
+	c.handTest = c.handTest.next()
+}
+
+// Metrics holds metrics for the cache.
+type Metrics struct {
+	// The number of bytes inuse by the cache.
+	Size int64
+	// The count of objects (blocks or tables) in the cache.
+	Count int64
+	// The number of cache hits.
+	Hits int64
+	// The number of cache misses.
+	Misses int64
+}
+
+// Cache implements Pebble's sharded block cache. The Clock-PRO algorithm is
+// used for page replacement
+// (http://static.usenix.org/event/usenix05/tech/general/full_papers/jiang/jiang_html/html.html). In
+// order to provide better concurrency, 2 x NumCPUs shards are created, with
+// each shard being given 1/n of the target cache size. The Clock-PRO algorithm
+// is run independently on each shard.
+//
+// Blocks are keyed by an (id, fileNum, offset) triple. The ID is a namespace
+// for file numbers and allows a single Cache to be shared between multiple
+// Pebble instances. The fileNum and offset refer to an sstable file number and
+// the offset of the block within the file. Because sstables are immutable and
+// file numbers are never reused, (fileNum,offset) are unique for the lifetime
+// of a Pebble instance.
+//
+// In addition to maintaining a map from (fileNum,offset) to data, each shard
+// maintains a map of the cached blocks for a particular fileNum. This allows
+// efficient eviction of all of the blocks for a file which is used when an
+// sstable is deleted from disk.
+//
+// # Memory Management
+//
+// In order to reduce pressure on the Go GC, manual memory management is
+// performed for the data stored in the cache. Manual memory management is
+// performed by calling into C.{malloc,free} to allocate memory. Cache.Values
+// are reference counted and the memory backing a manual value is freed when
+// the reference count drops to 0.
+//
+// Manual memory management brings the possibility of memory leaks. It is
+// imperative that every Handle returned by Cache.{Get,Set} is eventually
+// released. The "invariants" build tag enables a leak detection facility that
+// places a GC finalizer on cache.Value. When the cache.Value finalizer is run,
+// if the underlying buffer is still present a leak has occurred. The "tracing"
+// build tag enables tracing of cache.Value reference count manipulation and
+// eases finding where a leak has occurred. These two facilities are usually
+// used in combination by specifying `-tags invariants,tracing`. Note that
+// "tracing" produces a significant slowdown, while "invariants" does not.
+type Cache struct {
+	refs    int64
+	maxSize int64
+	idAlloc uint64
+	shards  []shard
+
+	// Traces recorded by Cache.trace. Used for debugging.
+	tr struct {
+		sync.Mutex
+		msgs []string
+	}
+}
+
+// New creates a new cache of the specified size. Memory for the cache is
+// allocated on demand, not during initialization. The cache is created with a
+// reference count of 1. Each DB it is associated with adds a reference, so the
+// creator of the cache should usually release their reference after the DB is
+// created.
+//
+//	c := cache.New(...)
+//	defer c.Unref()
+//	d, err := pebble.Open(pebble.Options{Cache: c})
+func New(size int64) *Cache {
+	return newShards(size, 2*runtime.GOMAXPROCS(0))
+}
+
+func newShards(size int64, shards int) *Cache {
+	c := &Cache{
+		refs:    1,
+		maxSize: size,
+		idAlloc: 1,
+		shards:  make([]shard, shards),
+	}
+	c.trace("alloc", c.refs)
+	for i := range c.shards {
+		c.shards[i] = shard{
+			maxSize:    size / int64(len(c.shards)),
+			coldTarget: size / int64(len(c.shards)),
+		}
+		if entriesGoAllocated {
+			c.shards[i].entries = make(map[*entry]struct{})
+		}
+		c.shards[i].blocks.init(16)
+		c.shards[i].files.init(16)
+	}
+
+	// Note: this is a no-op if invariants are disabled or race is enabled.
+	invariants.SetFinalizer(c, func(obj interface{}) {
+		c := obj.(*Cache)
+		if v := atomic.LoadInt64(&c.refs); v != 0 {
+			c.tr.Lock()
+			fmt.Fprintf(os.Stderr,
+				"pebble: cache (%p) has non-zero reference count: %d\n", c, v)
+			if len(c.tr.msgs) > 0 {
+				fmt.Fprintf(os.Stderr, "%s\n", strings.Join(c.tr.msgs, "\n"))
+			}
+			c.tr.Unlock()
+			os.Exit(1)
+		}
+	})
+	return c
+}
+
+func (c *Cache) getShard(id uint64, fileNum base.FileNum, offset uint64) *shard {
+	if id == 0 {
+		panic("pebble: 0 cache ID is invalid")
+	}
+
+	// Inlined version of fnv.New64 + Write.
+	const offset64 = 14695981039346656037
+	const prime64 = 1099511628211
+
+	h := uint64(offset64)
+	for i := 0; i < 8; i++ {
+		h *= prime64
+		h ^= uint64(id & 0xff)
+		id >>= 8
+	}
+	for i := 0; i < 8; i++ {
+		h *= prime64
+		h ^= uint64(fileNum & 0xff)
+		fileNum >>= 8
+	}
+	for i := 0; i < 8; i++ {
+		h *= prime64
+		h ^= uint64(offset & 0xff)
+		offset >>= 8
+	}
+
+	return &c.shards[h%uint64(len(c.shards))]
+}
+
+// Ref adds a reference to the cache. The cache only remains valid as long a
+// reference is maintained to it.
+func (c *Cache) Ref() {
+	v := atomic.AddInt64(&c.refs, 1)
+	if v <= 1 {
+		panic(fmt.Sprintf("pebble: inconsistent reference count: %d", v))
+	}
+	c.trace("ref", v)
+}
+
+// Unref releases a reference on the cache.
+func (c *Cache) Unref() {
+	v := atomic.AddInt64(&c.refs, -1)
+	c.trace("unref", v)
+	switch {
+	case v < 0:
+		panic(fmt.Sprintf("pebble: inconsistent reference count: %d", v))
+	case v == 0:
+		for i := range c.shards {
+			c.shards[i].Free()
+		}
+	}
+}
+
+// Get retrieves the cache value for the specified file and offset, returning
+// nil if no value is present.
+func (c *Cache) Get(id uint64, fileNum base.FileNum, offset uint64) Handle {
+	return c.getShard(id, fileNum, offset).Get(id, fileNum, offset)
+}
+
+// Set sets the cache value for the specified file and offset, overwriting an
+// existing value if present. A Handle is returned which provides faster
+// retrieval of the cached value than Get (lock-free and avoidance of the map
+// lookup). The value must have been allocated by Cache.Alloc.
+func (c *Cache) Set(id uint64, fileNum base.FileNum, offset uint64, value *Value) Handle {
+	return c.getShard(id, fileNum, offset).Set(id, fileNum, offset, value)
+}
+
+// Delete deletes the cached value for the specified file and offset.
+func (c *Cache) Delete(id uint64, fileNum base.FileNum, offset uint64) {
+	c.getShard(id, fileNum, offset).Delete(id, fileNum, offset)
+}
+
+// EvictFile evicts all of the cache values for the specified file.
+func (c *Cache) EvictFile(id uint64, fileNum base.FileNum) {
+	if id == 0 {
+		panic("pebble: 0 cache ID is invalid")
+	}
+	for i := range c.shards {
+		c.shards[i].EvictFile(id, fileNum)
+	}
+}
+
+// MaxSize returns the max size of the cache.
+func (c *Cache) MaxSize() int64 {
+	return c.maxSize
+}
+
+// Size returns the current space used by the cache.
+func (c *Cache) Size() int64 {
+	var size int64
+	for i := range c.shards {
+		size += c.shards[i].Size()
+	}
+	return size
+}
+
+// Alloc allocates a byte slice of the specified size, possibly reusing
+// previously allocated but unused memory. The memory backing the value is
+// manually managed. The caller MUST either add the value to the cache (via
+// Cache.Set), or release the value (via Cache.Free). Failure to do so will
+// result in a memory leak.
+func (c *Cache) Alloc(n int) *Value {
+	return newValue(n)
+}
+
+// Free frees the specified value. The buffer associated with the value will
+// possibly be reused, making it invalid to use the buffer after calling
+// Free. Do not call Free on a value that has been added to the cache.
+func (c *Cache) Free(v *Value) {
+	if n := v.refs(); n > 1 {
+		panic(fmt.Sprintf("pebble: Value has been added to the cache: refs=%d", n))
+	}
+	v.release()
+}
+
+// Reserve N bytes in the cache. This effectively shrinks the size of the cache
+// by N bytes, without actually consuming any memory. The returned closure
+// should be invoked to release the reservation.
+func (c *Cache) Reserve(n int) func() {
+	// Round-up the per-shard reservation. Most reservations should be large, so
+	// this probably doesn't matter in practice.
+	shardN := (n + len(c.shards) - 1) / len(c.shards)
+	for i := range c.shards {
+		c.shards[i].Reserve(shardN)
+	}
+	return func() {
+		if shardN == -1 {
+			panic("pebble: cache reservation already released")
+		}
+		for i := range c.shards {
+			c.shards[i].Reserve(-shardN)
+		}
+		shardN = -1
+	}
+}
+
+// Metrics returns the metrics for the cache.
+func (c *Cache) Metrics() Metrics {
+	var m Metrics
+	for i := range c.shards {
+		s := &c.shards[i]
+		s.mu.RLock()
+		m.Count += int64(s.blocks.Count())
+		m.Size += s.sizeHot + s.sizeCold
+		s.mu.RUnlock()
+		m.Hits += atomic.LoadInt64(&s.hits)
+		m.Misses += atomic.LoadInt64(&s.misses)
+	}
+	return m
+}
+
+// NewID returns a new ID to be used as a namespace for cached file
+// blocks.
+func (c *Cache) NewID() uint64 {
+	return atomic.AddUint64(&c.idAlloc, 1)
+}

vendor/github.com/cockroachdb/pebble/internal/cache/clockpro_normal.go

@@ -0,0 +1,10 @@
+// Copyright 2022 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build !tracing
+// +build !tracing
+
+package cache
+
+func (c *Cache) trace(_ string, _ int64) {}

vendor/github.com/cockroachdb/pebble/internal/cache/clockpro_tracing.go

@@ -0,0 +1,20 @@
+// Copyright 2022 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build tracing
+// +build tracing
+
+package cache
+
+import (
+	"fmt"
+	"runtime/debug"
+)
+
+func (c *Cache) trace(msg string, refs int64) {
+	s := fmt.Sprintf("%s: refs=%d\n%s", msg, refs, debug.Stack())
+	c.tr.Lock()
+	c.tr.msgs = append(c.tr.msgs, s)
+	c.tr.Unlock()
+}

vendor/github.com/cockroachdb/pebble/internal/cache/entry.go

@@ -0,0 +1,155 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+package cache
+
+type entryType int8
+
+const (
+	etTest entryType = iota
+	etCold
+	etHot
+)
+
+func (p entryType) String() string {
+	switch p {
+	case etTest:
+		return "test"
+	case etCold:
+		return "cold"
+	case etHot:
+		return "hot"
+	}
+	return "unknown"
+}
+
+// entry holds the metadata for a cache entry. The memory for an entry is
+// allocated from manually managed memory.
+//
+// Using manual memory management for entries is technically a volation of the
+// Cgo pointer rules:
+//
+//	https://golang.org/cmd/cgo/#hdr-Passing_pointers
+//
+// Specifically, Go pointers should not be stored in C allocated memory. The
+// reason for this rule is that the Go GC will not look at C allocated memory
+// to find pointers to Go objects. If the only reference to a Go object is
+// stored in C allocated memory, the object will be reclaimed. The shard field
+// of the entry struct points to a Go allocated object, thus the
+// violation. What makes this "safe" is that the Cache guarantees that there
+// are other pointers to the shard which will keep it alive.
+type entry struct {
+	key key
+	// The value associated with the entry. The entry holds a reference on the
+	// value which is maintained by entry.setValue().
+	val       *Value
+	blockLink struct {
+		next *entry
+		prev *entry
+	}
+	fileLink struct {
+		next *entry
+		prev *entry
+	}
+	size  int64
+	ptype entryType
+	// referenced is atomically set to indicate that this entry has been accessed
+	// since the last time one of the clock hands swept it.
+	referenced int32
+	shard      *shard
+	// Reference count for the entry. The entry is freed when the reference count
+	// drops to zero.
+	ref refcnt
+}
+
+func newEntry(s *shard, key key, size int64) *entry {
+	e := entryAllocNew()
+	*e = entry{
+		key:   key,
+		size:  size,
+		ptype: etCold,
+		shard: s,
+	}
+	e.blockLink.next = e
+	e.blockLink.prev = e
+	e.fileLink.next = e
+	e.fileLink.prev = e
+	e.ref.init(1)
+	return e
+}
+
+func (e *entry) free() {
+	e.setValue(nil)
+	*e = entry{}
+	entryAllocFree(e)
+}
+
+func (e *entry) next() *entry {
+	if e == nil {
+		return nil
+	}
+	return e.blockLink.next
+}
+
+func (e *entry) prev() *entry {
+	if e == nil {
+		return nil
+	}
+	return e.blockLink.prev
+}
+
+func (e *entry) link(s *entry) {
+	s.blockLink.prev = e.blockLink.prev
+	s.blockLink.prev.blockLink.next = s
+	s.blockLink.next = e
+	s.blockLink.next.blockLink.prev = s
+}
+
+func (e *entry) unlink() *entry {
+	next := e.blockLink.next
+	e.blockLink.prev.blockLink.next = e.blockLink.next
+	e.blockLink.next.blockLink.prev = e.blockLink.prev
+	e.blockLink.prev = e
+	e.blockLink.next = e
+	return next
+}
+
+func (e *entry) linkFile(s *entry) {
+	s.fileLink.prev = e.fileLink.prev
+	s.fileLink.prev.fileLink.next = s
+	s.fileLink.next = e
+	s.fileLink.next.fileLink.prev = s
+}
+
+func (e *entry) unlinkFile() *entry {
+	next := e.fileLink.next
+	e.fileLink.prev.fileLink.next = e.fileLink.next
+	e.fileLink.next.fileLink.prev = e.fileLink.prev
+	e.fileLink.prev = e
+	e.fileLink.next = e
+	return next
+}
+
+func (e *entry) setValue(v *Value) {
+	if v != nil {
+		v.acquire()
+	}
+	old := e.val
+	e.val = v
+	if old != nil {
+		old.release()
+	}
+}
+
+func (e *entry) peekValue() *Value {
+	return e.val
+}
+
+func (e *entry) acquireValue() *Value {
+	v := e.val
+	if v != nil {
+		v.acquire()
+	}
+	return v
+}

vendor/github.com/cockroachdb/pebble/internal/cache/entry_invariants.go

@@ -0,0 +1,38 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+//
+//go:build (invariants && !race) || (tracing && !race)
+// +build invariants,!race tracing,!race
+
+package cache
+
+import (
+	"fmt"
+	"os"
+
+	"github.com/cockroachdb/pebble/internal/invariants"
+)
+
+// When the "invariants" or "tracing" build tags are enabled, we need to
+// allocate entries using the Go allocator so entry.val properly maintains a
+// reference to the Value.
+const entriesGoAllocated = true
+
+func entryAllocNew() *entry {
+	e := &entry{}
+	// Note: this is a no-op if invariants and tracing are disabled or race is
+	// enabled.
+	invariants.SetFinalizer(e, func(obj interface{}) {
+		e := obj.(*entry)
+		if v := e.ref.refs(); v != 0 {
+			fmt.Fprintf(os.Stderr, "%p: cache entry has non-zero reference count: %d\n%s",
+				e, v, e.ref.traces())
+			os.Exit(1)
+		}
+	})
+	return e
+}
+
+func entryAllocFree(e *entry) {
+}

vendor/github.com/cockroachdb/pebble/internal/cache/entry_normal.go

@@ -0,0 +1,103 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+//
+//go:build (!invariants && !tracing) || race
+// +build !invariants,!tracing race
+
+package cache
+
+import (
+	"runtime"
+	"sync"
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/invariants"
+	"github.com/cockroachdb/pebble/internal/manual"
+)
+
+const (
+	entrySize            = int(unsafe.Sizeof(entry{}))
+	entryAllocCacheLimit = 128
+	// Avoid using runtime.SetFinalizer in race builds as finalizers tickle a bug
+	// in the Go race detector in go1.15 and earlier versions. This requires that
+	// entries are Go allocated rather than manually allocated.
+	//
+	// If cgo is disabled we need to allocate the entries using the Go allocator
+	// and is violates the Go GC rules to put Go pointers (such as the entry
+	// pointer fields) into untyped memory (i.e. a []byte).
+	entriesGoAllocated = invariants.RaceEnabled || !cgoEnabled
+)
+
+var entryAllocPool = sync.Pool{
+	New: func() interface{} {
+		return newEntryAllocCache()
+	},
+}
+
+func entryAllocNew() *entry {
+	a := entryAllocPool.Get().(*entryAllocCache)
+	e := a.alloc()
+	entryAllocPool.Put(a)
+	return e
+}
+
+func entryAllocFree(e *entry) {
+	a := entryAllocPool.Get().(*entryAllocCache)
+	a.free(e)
+	entryAllocPool.Put(a)
+}
+
+type entryAllocCache struct {
+	entries []*entry
+}
+
+func newEntryAllocCache() *entryAllocCache {
+	c := &entryAllocCache{}
+	if !entriesGoAllocated {
+		// Note the use of a "real" finalizer here (as opposed to a build tag-gated
+		// no-op finalizer). Without the finalizer, objects released from the pool
+		// and subsequently GC'd by the Go runtime would fail to have their manually
+		// allocated memory freed, which results in a memory leak.
+		// lint:ignore SetFinalizer
+		runtime.SetFinalizer(c, freeEntryAllocCache)
+	}
+	return c
+}
+
+func freeEntryAllocCache(obj interface{}) {
+	c := obj.(*entryAllocCache)
+	for i, e := range c.entries {
+		c.dealloc(e)
+		c.entries[i] = nil
+	}
+}
+
+func (c *entryAllocCache) alloc() *entry {
+	n := len(c.entries)
+	if n == 0 {
+		if entriesGoAllocated {
+			return &entry{}
+		}
+		b := manual.New(entrySize)
+		return (*entry)(unsafe.Pointer(&b[0]))
+	}
+	e := c.entries[n-1]
+	c.entries = c.entries[:n-1]
+	return e
+}
+
+func (c *entryAllocCache) dealloc(e *entry) {
+	if !entriesGoAllocated {
+		buf := (*[manual.MaxArrayLen]byte)(unsafe.Pointer(e))[:entrySize:entrySize]
+		manual.Free(buf)
+	}
+}
+
+func (c *entryAllocCache) free(e *entry) {
+	if len(c.entries) == entryAllocCacheLimit {
+		c.dealloc(e)
+		return
+	}
+	c.entries = append(c.entries, e)
+}

vendor/github.com/cockroachdb/pebble/internal/cache/refcnt_normal.go

@@ -0,0 +1,59 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build !tracing
+// +build !tracing
+
+package cache
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"github.com/cockroachdb/redact"
+)
+
+// refcnt provides an atomic reference count. This version is used when the
+// "tracing" build tag is not enabled. See refcnt_tracing.go for the "tracing"
+// enabled version.
+type refcnt struct {
+	val atomic.Int32
+}
+
+// initialize the reference count to the specified value.
+func (v *refcnt) init(val int32) {
+	v.val.Store(val)
+}
+
+func (v *refcnt) refs() int32 {
+	return v.val.Load()
+}
+
+func (v *refcnt) acquire() {
+	switch v := v.val.Add(1); {
+	case v <= 1:
+		panic(redact.Safe(fmt.Sprintf("pebble: inconsistent reference count: %d", v)))
+	}
+}
+
+func (v *refcnt) release() bool {
+	switch v := v.val.Add(-1); {
+	case v < 0:
+		panic(redact.Safe(fmt.Sprintf("pebble: inconsistent reference count: %d", v)))
+	case v == 0:
+		return true
+	default:
+		return false
+	}
+}
+
+func (v *refcnt) trace(msg string) {
+}
+
+func (v *refcnt) traces() string {
+	return ""
+}
+
+// Silence unused warning.
+var _ = (*refcnt)(nil).traces

vendor/github.com/cockroachdb/pebble/internal/cache/refcnt_tracing.go

@@ -0,0 +1,66 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build tracing
+// +build tracing
+
+package cache
+
+import (
+	"fmt"
+	"runtime/debug"
+	"strings"
+	"sync"
+	"sync/atomic"
+)
+
+// refcnt provides an atomic reference count, along with a tracing facility for
+// debugging logic errors in manipulating the reference count. This version is
+// used when the "tracing" build tag is enabled.
+type refcnt struct {
+	val atomic.Int32
+	sync.Mutex
+	msgs []string
+}
+
+func (v *refcnt) init(val int32) {
+	v.val.Store(val)
+	v.trace("init")
+}
+
+func (v *refcnt) refs() int32 {
+	return v.val.Load()
+}
+
+func (v *refcnt) acquire() {
+	switch n := v.val.Add(1); {
+	case n <= 1:
+		panic(fmt.Sprintf("pebble: inconsistent reference count: %d", n))
+	}
+	v.trace("acquire")
+}
+
+func (v *refcnt) release() bool {
+	n := v.val.Add(-1)
+	switch {
+	case n < 0:
+		panic(fmt.Sprintf("pebble: inconsistent reference count: %d", n))
+	}
+	v.trace("release")
+	return n == 0
+}
+
+func (v *refcnt) trace(msg string) {
+	s := fmt.Sprintf("%s: refs=%d\n%s", msg, v.refs(), debug.Stack())
+	v.Lock()
+	v.msgs = append(v.msgs, s)
+	v.Unlock()
+}
+
+func (v *refcnt) traces() string {
+	v.Lock()
+	s := strings.Join(v.msgs, "\n")
+	v.Unlock()
+	return s
+}

vendor/github.com/cockroachdb/pebble/internal/cache/robin_hood.go

@@ -0,0 +1,320 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+package cache
+
+import (
+	"fmt"
+	"math/bits"
+	"os"
+	"runtime/debug"
+	"strings"
+	"time"
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/invariants"
+	"github.com/cockroachdb/pebble/internal/manual"
+)
+
+var hashSeed = uint64(time.Now().UnixNano())
+
+// Fibonacci hash: https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
+func robinHoodHash(k key, shift uint32) uint32 {
+	const m = 11400714819323198485
+	h := hashSeed
+	h ^= k.id * m
+	h ^= uint64(k.fileNum) * m
+	h ^= k.offset * m
+	return uint32(h >> shift)
+}
+
+type robinHoodEntry struct {
+	key key
+	// Note that value may point to a Go allocated object (if the "invariants"
+	// build tag was specified), even though the memory for the entry itself is
+	// manually managed. This is technically a volation of the Cgo pointer rules:
+	//
+	//   https://golang.org/cmd/cgo/#hdr-Passing_pointers
+	//
+	// Specifically, Go pointers should not be stored in C allocated memory. The
+	// reason for this rule is that the Go GC will not look at C allocated memory
+	// to find pointers to Go objects. If the only reference to a Go object is
+	// stored in C allocated memory, the object will be reclaimed. What makes
+	// this "safe" is that the Cache guarantees that there are other pointers to
+	// the entry and shard which will keep them alive. In particular, every Go
+	// allocated entry in the cache is referenced by the shard.entries map. And
+	// every shard is referenced by the Cache.shards map.
+	value *entry
+	// The distance the entry is from its desired position.
+	dist uint32
+}
+
+type robinHoodEntries struct {
+	ptr unsafe.Pointer
+	len uint32
+}
+
+func newRobinHoodEntries(n uint32) robinHoodEntries {
+	size := uintptr(n) * unsafe.Sizeof(robinHoodEntry{})
+	return robinHoodEntries{
+		ptr: unsafe.Pointer(&(manual.New(int(size)))[0]),
+		len: n,
+	}
+}
+
+func (e robinHoodEntries) at(i uint32) *robinHoodEntry {
+	return (*robinHoodEntry)(unsafe.Pointer(uintptr(e.ptr) +
+		uintptr(i)*unsafe.Sizeof(robinHoodEntry{})))
+}
+
+func (e robinHoodEntries) free() {
+	size := uintptr(e.len) * unsafe.Sizeof(robinHoodEntry{})
+	buf := (*[manual.MaxArrayLen]byte)(e.ptr)[:size:size]
+	manual.Free(buf)
+}
+
+// robinHoodMap is an implementation of Robin Hood hashing. Robin Hood hashing
+// is an open-address hash table using linear probing. The twist is that the
+// linear probe distance is reduced by moving existing entries when inserting
+// and deleting. This is accomplished by keeping track of how far an entry is
+// from its "desired" slot (hash of key modulo number of slots). During
+// insertion, if the new entry being inserted is farther from its desired slot
+// than the target entry, we swap the target and new entry. This effectively
+// steals from the "rich" target entry and gives to the "poor" new entry (thus
+// the origin of the name).
+//
+// An extension over the base Robin Hood hashing idea comes from
+// https://probablydance.com/2017/02/26/i-wrote-the-fastest-hashtable/. A cap
+// is placed on the max distance an entry can be from its desired slot. When
+// this threshold is reached during insertion, the size of the table is doubled
+// and insertion is restarted. Additionally, the entries slice is given "max
+// dist" extra entries on the end. The very last entry in the entries slice is
+// never used and acts as a sentinel which terminates loops. The previous
+// maxDist-1 entries act as the extra entries. For example, if the size of the
+// table is 2, maxDist is computed as 4 and the actual size of the entry slice
+// is 6.
+//
+//	+---+---+---+---+---+---+
+//	| 0 | 1 | 2 | 3 | 4 | 5 |
+//	+---+---+---+---+---+---+
+//	        ^
+//	       size
+//
+// In this scenario, the target entry for a key will always be in the range
+// [0,1]. Valid entries may reside in the range [0,4] due to the linear probing
+// of up to maxDist entries. The entry at index 5 will never contain a value,
+// and instead acts as a sentinel (its distance is always 0). The max distance
+// threshold is set to log2(num-entries). This ensures that retrieval is O(log
+// N), though note that N is the number of total entries, not the count of
+// valid entries.
+//
+// Deletion is implemented via the backward shift delete mechanism instead of
+// tombstones. This preserves the performance of the table in the presence of
+// deletions. See
+// http://codecapsule.com/2013/11/17/robin-hood-hashing-backward-shift-deletion
+// for details.
+type robinHoodMap struct {
+	entries robinHoodEntries
+	size    uint32
+	shift   uint32
+	count   uint32
+	maxDist uint32
+}
+
+func maxDistForSize(size uint32) uint32 {
+	desired := uint32(bits.Len32(size))
+	if desired < 4 {
+		desired = 4
+	}
+	return desired
+}
+
+func newRobinHoodMap(initialCapacity int) *robinHoodMap {
+	m := &robinHoodMap{}
+	m.init(initialCapacity)
+
+	// Note: this is a no-op if invariants are disabled or race is enabled.
+	invariants.SetFinalizer(m, func(obj interface{}) {
+		m := obj.(*robinHoodMap)
+		if m.entries.ptr != nil {
+			fmt.Fprintf(os.Stderr, "%p: robin-hood map not freed\n", m)
+			os.Exit(1)
+		}
+	})
+	return m
+}
+
+func (m *robinHoodMap) init(initialCapacity int) {
+	if initialCapacity < 1 {
+		initialCapacity = 1
+	}
+	targetSize := 1 << (uint(bits.Len(uint(2*initialCapacity-1))) - 1)
+	m.rehash(uint32(targetSize))
+}
+
+func (m *robinHoodMap) free() {
+	if m.entries.ptr != nil {
+		m.entries.free()
+		m.entries.ptr = nil
+	}
+}
+
+func (m *robinHoodMap) rehash(size uint32) {
+	oldEntries := m.entries
+
+	m.size = size
+	m.shift = uint32(64 - bits.Len32(m.size-1))
+	m.maxDist = maxDistForSize(size)
+	m.entries = newRobinHoodEntries(size + m.maxDist)
+	m.count = 0
+
+	for i := uint32(0); i < oldEntries.len; i++ {
+		e := oldEntries.at(i)
+		if e.value != nil {
+			m.Put(e.key, e.value)
+		}
+	}
+
+	if oldEntries.ptr != nil {
+		oldEntries.free()
+	}
+}
+
+// Find an entry containing the specified value. This is intended to be used
+// from debug and test code.
+func (m *robinHoodMap) findByValue(v *entry) *robinHoodEntry {
+	for i := uint32(0); i < m.entries.len; i++ {
+		e := m.entries.at(i)
+		if e.value == v {
+			return e
+		}
+	}
+	return nil
+}
+
+func (m *robinHoodMap) Count() int {
+	return int(m.count)
+}
+
+func (m *robinHoodMap) Put(k key, v *entry) {
+	maybeExists := true
+	n := robinHoodEntry{key: k, value: v, dist: 0}
+	for i := robinHoodHash(k, m.shift); ; i++ {
+		e := m.entries.at(i)
+		if maybeExists && k == e.key {
+			// Entry already exists: overwrite.
+			e.value = n.value
+			m.checkEntry(i)
+			return
+		}
+
+		if e.value == nil {
+			// Found an empty entry: insert here.
+			*e = n
+			m.count++
+			m.checkEntry(i)
+			return
+		}
+
+		if e.dist < n.dist {
+			// Swap the new entry with the current entry because the current is
+			// rich. We then continue to loop, looking for a new location for the
+			// current entry. Note that this is also the not-found condition for
+			// retrieval, which means that "k" is not present in the map. See Get().
+			n, *e = *e, n
+			m.checkEntry(i)
+			maybeExists = false
+		}
+
+		// The new entry gradually moves away from its ideal position.
+		n.dist++
+
+		// If we've reached the max distance threshold, grow the table and restart
+		// the insertion.
+		if n.dist == m.maxDist {
+			m.rehash(2 * m.size)
+			i = robinHoodHash(n.key, m.shift) - 1
+			n.dist = 0
+			maybeExists = false
+		}
+	}
+}
+
+func (m *robinHoodMap) Get(k key) *entry {
+	var dist uint32
+	for i := robinHoodHash(k, m.shift); ; i++ {
+		e := m.entries.at(i)
+		if k == e.key {
+			// Found.
+			return e.value
+		}
+		if e.dist < dist {
+			// Not found.
+			return nil
+		}
+		dist++
+	}
+}
+
+func (m *robinHoodMap) Delete(k key) {
+	var dist uint32
+	for i := robinHoodHash(k, m.shift); ; i++ {
+		e := m.entries.at(i)
+		if k == e.key {
+			m.checkEntry(i)
+			// We found the entry to delete. Shift the following entries backwards
+			// until the next empty value or entry with a zero distance. Note that
+			// empty values are guaranteed to have "dist == 0".
+			m.count--
+			for j := i + 1; ; j++ {
+				t := m.entries.at(j)
+				if t.dist == 0 {
+					*e = robinHoodEntry{}
+					return
+				}
+				e.key = t.key
+				e.value = t.value
+				e.dist = t.dist - 1
+				e = t
+				m.checkEntry(j)
+			}
+		}
+		if dist > e.dist {
+			// Not found.
+			return
+		}
+		dist++
+	}
+}
+
+func (m *robinHoodMap) checkEntry(i uint32) {
+	if invariants.Enabled {
+		e := m.entries.at(i)
+		if e.value != nil {
+			pos := robinHoodHash(e.key, m.shift)
+			if (uint32(i) - pos) != e.dist {
+				fmt.Fprintf(os.Stderr, "%d: invalid dist=%d, expected %d: %s\n%s",
+					i, e.dist, uint32(i)-pos, e.key, debug.Stack())
+				os.Exit(1)
+			}
+			if e.dist > m.maxDist {
+				fmt.Fprintf(os.Stderr, "%d: invalid dist=%d > maxDist=%d: %s\n%s",
+					i, e.dist, m.maxDist, e.key, debug.Stack())
+				os.Exit(1)
+			}
+		}
+	}
+}
+
+func (m *robinHoodMap) String() string {
+	var buf strings.Builder
+	fmt.Fprintf(&buf, "count: %d\n", m.count)
+	for i := uint32(0); i < m.entries.len; i++ {
+		e := m.entries.at(i)
+		if e.value != nil {
+			fmt.Fprintf(&buf, "%d: [%s,%p,%d]\n", i, e.key, e.value, e.dist)
+		}
+	}
+	return buf.String()
+}

vendor/github.com/cockroachdb/pebble/internal/cache/value.go

@@ -0,0 +1,46 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+package cache
+
+// Value holds a reference counted immutable value.
+type Value struct {
+	buf []byte
+	// Reference count for the value. The value is freed when the reference count
+	// drops to zero.
+	ref refcnt
+}
+
+// Buf returns the buffer associated with the value. The contents of the buffer
+// should not be changed once the value has been added to the cache. Instead, a
+// new Value should be created and added to the cache to replace the existing
+// value.
+func (v *Value) Buf() []byte {
+	if v == nil {
+		return nil
+	}
+	return v.buf
+}
+
+// Truncate the buffer to the specified length. The buffer length should not be
+// changed once the value has been added to the cache as there may be
+// concurrent readers of the Value. Instead, a new Value should be created and
+// added to the cache to replace the existing value.
+func (v *Value) Truncate(n int) {
+	v.buf = v.buf[:n]
+}
+
+func (v *Value) refs() int32 {
+	return v.ref.refs()
+}
+
+func (v *Value) acquire() {
+	v.ref.acquire()
+}
+
+func (v *Value) release() {
+	if v.ref.release() {
+		v.free()
+	}
+}

vendor/github.com/cockroachdb/pebble/internal/cache/value_invariants.go

@@ -0,0 +1,55 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build (invariants && !race) || (tracing && !race)
+// +build invariants,!race tracing,!race
+
+package cache
+
+import (
+	"fmt"
+	"os"
+
+	"github.com/cockroachdb/pebble/internal/invariants"
+	"github.com/cockroachdb/pebble/internal/manual"
+)
+
+// newValue creates a Value with a manually managed buffer of size n.
+//
+// This definition of newValue is used when either the "invariants" or
+// "tracing" build tags are specified. It hooks up a finalizer to the returned
+// Value that checks for memory leaks when the GC determines the Value is no
+// longer reachable.
+func newValue(n int) *Value {
+	if n == 0 {
+		return nil
+	}
+	b := manual.New(n)
+	v := &Value{buf: b}
+	v.ref.init(1)
+	// Note: this is a no-op if invariants and tracing are disabled or race is
+	// enabled.
+	invariants.SetFinalizer(v, func(obj interface{}) {
+		v := obj.(*Value)
+		if v.buf != nil {
+			fmt.Fprintf(os.Stderr, "%p: cache value was not freed: refs=%d\n%s",
+				v, v.refs(), v.ref.traces())
+			os.Exit(1)
+		}
+	})
+	return v
+}
+
+func (v *Value) free() {
+	// When "invariants" are enabled set the value contents to 0xff in order to
+	// cache use-after-free bugs.
+	for i := range v.buf {
+		v.buf[i] = 0xff
+	}
+	manual.Free(v.buf)
+	// Setting Value.buf to nil is needed for correctness of the leak checking
+	// that is performed when the "invariants" or "tracing" build tags are
+	// enabled.
+	v.buf = nil
+}

vendor/github.com/cockroachdb/pebble/internal/cache/value_normal.go

@@ -0,0 +1,57 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+//go:build (!invariants && !tracing) || race
+// +build !invariants,!tracing race
+
+package cache
+
+import (
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/manual"
+)
+
+const valueSize = int(unsafe.Sizeof(Value{}))
+
+func newValue(n int) *Value {
+	if n == 0 {
+		return nil
+	}
+
+	if !cgoEnabled {
+		// If Cgo is disabled then all memory is allocated from the Go heap and we
+		// can't play the trick below to combine the Value and buffer allocation.
+		v := &Value{buf: make([]byte, n)}
+		v.ref.init(1)
+		return v
+	}
+
+	// When we're not performing leak detection, the lifetime of the returned
+	// Value is exactly the lifetime of the backing buffer and we can manually
+	// allocate both.
+	//
+	// TODO(peter): It may be better to separate the allocation of the value and
+	// the buffer in order to reduce internal fragmentation in malloc. If the
+	// buffer is right at a power of 2, adding valueSize might push the
+	// allocation over into the next larger size.
+	b := manual.New(valueSize + n)
+	v := (*Value)(unsafe.Pointer(&b[0]))
+	v.buf = b[valueSize:]
+	v.ref.init(1)
+	return v
+}
+
+func (v *Value) free() {
+	if !cgoEnabled {
+		return
+	}
+
+	// When we're not performing leak detection, the Value and buffer were
+	// allocated contiguously.
+	n := valueSize + cap(v.buf)
+	buf := (*[manual.MaxArrayLen]byte)(unsafe.Pointer(v))[:n:n]
+	v.buf = nil
+	manual.Free(buf)
+}