Evgenii Stratonikov
e0309e398c
LRU `Peek`/`Contains` take LRU mutex _inside_ of a `View` transaction.
`View` transaction itself takes `mmapLock` [1], which is lifted after tx
finishes (in `tx.Commit()` -> `tx.close()` -> `tx.db.removeTx`)
When we evict items from LRU cache mutex order is different:
first we take LRU mutex and then execute `Batch` which _does_ take
`mmapLock` in case we need to remap. Thus the deadlock.
[1] 8f4a7e1f92/db.go (L708)
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
321 lines
6.5 KiB
Go
321 lines
6.5 KiB
Go
package writecache
|
|
|
|
import (
|
|
"bytes"
|
|
"errors"
|
|
"time"
|
|
|
|
objectCore "github.com/TrueCloudLab/frostfs-node/pkg/core/object"
|
|
"github.com/TrueCloudLab/frostfs-node/pkg/local_object_storage/blobstor"
|
|
"github.com/TrueCloudLab/frostfs-node/pkg/local_object_storage/blobstor/common"
|
|
meta "github.com/TrueCloudLab/frostfs-node/pkg/local_object_storage/metabase"
|
|
"github.com/TrueCloudLab/frostfs-sdk-go/object"
|
|
oid "github.com/TrueCloudLab/frostfs-sdk-go/object/id"
|
|
"github.com/mr-tron/base58"
|
|
"github.com/nspcc-dev/neo-go/pkg/util/slice"
|
|
"go.etcd.io/bbolt"
|
|
"go.uber.org/zap"
|
|
)
|
|
|
|
const (
|
|
// flushBatchSize is amount of keys which will be read from cache to be flushed
|
|
// to the main storage. It is used to reduce contention between cache put
|
|
// and cache persist.
|
|
flushBatchSize = 512
|
|
// defaultFlushWorkersCount is number of workers for putting objects in main storage.
|
|
defaultFlushWorkersCount = 20
|
|
// defaultFlushInterval is default time interval between successive flushes.
|
|
defaultFlushInterval = time.Second
|
|
)
|
|
|
|
// runFlushLoop starts background workers which periodically flush objects to the blobstor.
|
|
func (c *cache) runFlushLoop() {
|
|
for i := 0; i < c.workersCount; i++ {
|
|
c.wg.Add(1)
|
|
go c.flushWorker(i)
|
|
}
|
|
|
|
c.wg.Add(1)
|
|
go c.flushBigObjects()
|
|
|
|
c.wg.Add(1)
|
|
go func() {
|
|
defer c.wg.Done()
|
|
|
|
tt := time.NewTimer(defaultFlushInterval)
|
|
defer tt.Stop()
|
|
|
|
for {
|
|
select {
|
|
case <-tt.C:
|
|
c.flushDB()
|
|
tt.Reset(defaultFlushInterval)
|
|
case <-c.closeCh:
|
|
return
|
|
}
|
|
}
|
|
}()
|
|
}
|
|
|
|
func (c *cache) flushDB() {
|
|
var lastKey []byte
|
|
var m []objectInfo
|
|
for {
|
|
select {
|
|
case <-c.closeCh:
|
|
return
|
|
default:
|
|
}
|
|
|
|
m = m[:0]
|
|
|
|
c.modeMtx.RLock()
|
|
if c.readOnly() {
|
|
c.modeMtx.RUnlock()
|
|
time.Sleep(time.Second)
|
|
continue
|
|
}
|
|
|
|
// We put objects in batches of fixed size to not interfere with main put cycle a lot.
|
|
_ = c.db.View(func(tx *bbolt.Tx) error {
|
|
b := tx.Bucket(defaultBucket)
|
|
cs := b.Cursor()
|
|
|
|
var k, v []byte
|
|
|
|
if len(lastKey) == 0 {
|
|
k, v = cs.First()
|
|
} else {
|
|
k, v = cs.Seek(lastKey)
|
|
if bytes.Equal(k, lastKey) {
|
|
k, v = cs.Next()
|
|
}
|
|
}
|
|
|
|
for ; k != nil && len(m) < flushBatchSize; k, v = cs.Next() {
|
|
if len(lastKey) == len(k) {
|
|
copy(lastKey, k)
|
|
} else {
|
|
lastKey = slice.Copy(k)
|
|
}
|
|
|
|
m = append(m, objectInfo{
|
|
addr: string(k),
|
|
data: slice.Copy(v),
|
|
})
|
|
}
|
|
return nil
|
|
})
|
|
|
|
var count int
|
|
for i := range m {
|
|
if c.flushed.Contains(m[i].addr) {
|
|
continue
|
|
}
|
|
|
|
obj := object.New()
|
|
if err := obj.Unmarshal(m[i].data); err != nil {
|
|
continue
|
|
}
|
|
|
|
count++
|
|
select {
|
|
case c.flushCh <- obj:
|
|
case <-c.closeCh:
|
|
c.modeMtx.RUnlock()
|
|
return
|
|
}
|
|
}
|
|
|
|
if count == 0 {
|
|
c.modeMtx.RUnlock()
|
|
break
|
|
}
|
|
|
|
c.modeMtx.RUnlock()
|
|
|
|
c.log.Debug("tried to flush items from write-cache",
|
|
zap.Int("count", count),
|
|
zap.String("start", base58.Encode(lastKey)))
|
|
}
|
|
}
|
|
|
|
func (c *cache) flushBigObjects() {
|
|
defer c.wg.Done()
|
|
|
|
tick := time.NewTicker(defaultFlushInterval * 10)
|
|
for {
|
|
select {
|
|
case <-tick.C:
|
|
c.modeMtx.RLock()
|
|
if c.readOnly() {
|
|
c.modeMtx.RUnlock()
|
|
break
|
|
}
|
|
|
|
_ = c.flushFSTree(true)
|
|
|
|
c.modeMtx.RUnlock()
|
|
case <-c.closeCh:
|
|
return
|
|
}
|
|
}
|
|
}
|
|
|
|
func (c *cache) reportFlushError(msg string, addr string, err error) {
|
|
if c.reportError != nil {
|
|
c.reportError(msg, err)
|
|
} else {
|
|
c.log.Error(msg,
|
|
zap.String("address", addr),
|
|
zap.Error(err))
|
|
}
|
|
}
|
|
|
|
func (c *cache) flushFSTree(ignoreErrors bool) error {
|
|
var prm common.IteratePrm
|
|
prm.IgnoreErrors = ignoreErrors
|
|
prm.LazyHandler = func(addr oid.Address, f func() ([]byte, error)) error {
|
|
sAddr := addr.EncodeToString()
|
|
|
|
if _, ok := c.store.flushed.Peek(sAddr); ok {
|
|
return nil
|
|
}
|
|
|
|
data, err := f()
|
|
if err != nil {
|
|
c.reportFlushError("can't read a file", sAddr, err)
|
|
if ignoreErrors {
|
|
return nil
|
|
}
|
|
return err
|
|
}
|
|
|
|
var obj object.Object
|
|
err = obj.Unmarshal(data)
|
|
if err != nil {
|
|
c.reportFlushError("can't unmarshal an object", sAddr, err)
|
|
if ignoreErrors {
|
|
return nil
|
|
}
|
|
return err
|
|
}
|
|
|
|
err = c.flushObject(&obj, data)
|
|
if err != nil {
|
|
if ignoreErrors {
|
|
return nil
|
|
}
|
|
return err
|
|
}
|
|
|
|
// mark object as flushed
|
|
c.flushed.Add(sAddr, false)
|
|
|
|
return nil
|
|
}
|
|
|
|
_, err := c.fsTree.Iterate(prm)
|
|
return err
|
|
}
|
|
|
|
// flushWorker writes objects to the main storage.
|
|
func (c *cache) flushWorker(_ int) {
|
|
defer c.wg.Done()
|
|
|
|
var obj *object.Object
|
|
for {
|
|
// Give priority to direct put.
|
|
select {
|
|
case obj = <-c.flushCh:
|
|
case <-c.closeCh:
|
|
return
|
|
}
|
|
|
|
err := c.flushObject(obj, nil)
|
|
if err == nil {
|
|
c.flushed.Add(objectCore.AddressOf(obj).EncodeToString(), true)
|
|
}
|
|
}
|
|
}
|
|
|
|
// flushObject is used to write object directly to the main storage.
|
|
func (c *cache) flushObject(obj *object.Object, data []byte) error {
|
|
addr := objectCore.AddressOf(obj)
|
|
|
|
var prm common.PutPrm
|
|
prm.Object = obj
|
|
prm.RawData = data
|
|
|
|
res, err := c.blobstor.Put(prm)
|
|
if err != nil {
|
|
if !errors.Is(err, common.ErrNoSpace) && !errors.Is(err, common.ErrReadOnly) &&
|
|
!errors.Is(err, blobstor.ErrNoPlaceFound) {
|
|
c.reportFlushError("can't flush an object to blobstor",
|
|
addr.EncodeToString(), err)
|
|
}
|
|
return err
|
|
}
|
|
|
|
var updPrm meta.UpdateStorageIDPrm
|
|
updPrm.SetAddress(addr)
|
|
updPrm.SetStorageID(res.StorageID)
|
|
|
|
_, err = c.metabase.UpdateStorageID(updPrm)
|
|
if err != nil {
|
|
c.reportFlushError("can't update object storage ID",
|
|
addr.EncodeToString(), err)
|
|
}
|
|
return err
|
|
}
|
|
|
|
// Flush flushes all objects from the write-cache to the main storage.
|
|
// Write-cache must be in readonly mode to ensure correctness of an operation and
|
|
// to prevent interference with background flush workers.
|
|
func (c *cache) Flush(ignoreErrors bool) error {
|
|
c.modeMtx.RLock()
|
|
defer c.modeMtx.RUnlock()
|
|
|
|
return c.flush(ignoreErrors)
|
|
}
|
|
|
|
func (c *cache) flush(ignoreErrors bool) error {
|
|
if err := c.flushFSTree(ignoreErrors); err != nil {
|
|
return err
|
|
}
|
|
|
|
return c.db.View(func(tx *bbolt.Tx) error {
|
|
var addr oid.Address
|
|
|
|
b := tx.Bucket(defaultBucket)
|
|
cs := b.Cursor()
|
|
for k, data := cs.Seek(nil); k != nil; k, data = cs.Next() {
|
|
sa := string(k)
|
|
if _, ok := c.flushed.Peek(sa); ok {
|
|
continue
|
|
}
|
|
|
|
if err := addr.DecodeString(sa); err != nil {
|
|
c.reportFlushError("can't decode object address from the DB", sa, err)
|
|
if ignoreErrors {
|
|
continue
|
|
}
|
|
return err
|
|
}
|
|
|
|
var obj object.Object
|
|
if err := obj.Unmarshal(data); err != nil {
|
|
c.reportFlushError("can't unmarshal an object from the DB", sa, err)
|
|
if ignoreErrors {
|
|
continue
|
|
}
|
|
return err
|
|
}
|
|
|
|
if err := c.flushObject(&obj, data); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
return nil
|
|
})
|
|
}
|