lotus/blockstore/autobatch.go

263 lines
6.3 KiB
Go
Raw Normal View History

2022-01-11 16:31:59 +00:00
package blockstore
import (
"context"
"sync"
"time"
2022-01-11 16:31:59 +00:00
block "github.com/ipfs/go-block-format"
"github.com/ipfs/go-cid"
ipld "github.com/ipfs/go-ipld-format"
2022-06-14 15:00:51 +00:00
"golang.org/x/xerrors"
2022-01-11 16:31:59 +00:00
)
// autolog is a logger for the autobatching blockstore. It is subscoped from the
// blockstore logger.
var autolog = log.Named("auto")
2022-01-12 17:57:34 +00:00
// contains the same set of blocks twice, once as an ordered list for flushing, and as a map for fast access
type blockBatch struct {
blockList []block.Block
blockMap map[cid.Cid]block.Block
}
2022-01-11 16:31:59 +00:00
type AutobatchBlockstore struct {
// TODO: drop if memory consumption is too high
addedCids map[cid.Cid]struct{}
2022-01-12 20:03:34 +00:00
stateLock sync.Mutex
2022-01-12 17:57:34 +00:00
bufferedBatch blockBatch
flushingBatch blockBatch
flushErr error
2022-01-12 17:57:34 +00:00
flushCh chan struct{}
doFlushLock sync.Mutex
flushRetryDelay time.Duration
doneCh chan struct{}
shutdown context.CancelFunc
backingBs Blockstore
2022-01-11 16:31:59 +00:00
bufferCapacity int
bufferSize int
}
func NewAutobatch(ctx context.Context, backingBs Blockstore, bufferCapacity int) *AutobatchBlockstore {
ctx, cancel := context.WithCancel(ctx)
2022-01-11 16:31:59 +00:00
bs := &AutobatchBlockstore{
2022-01-12 17:57:34 +00:00
addedCids: make(map[cid.Cid]struct{}),
backingBs: backingBs,
bufferCapacity: bufferCapacity,
flushCh: make(chan struct{}, 1),
doneCh: make(chan struct{}),
// could be made configable
2022-01-12 17:57:34 +00:00
flushRetryDelay: time.Millisecond * 100,
shutdown: cancel,
2022-01-11 16:31:59 +00:00
}
2022-01-12 17:57:34 +00:00
bs.bufferedBatch.blockMap = make(map[cid.Cid]block.Block)
go bs.flushWorker(ctx)
2022-01-11 16:31:59 +00:00
return bs
}
func (bs *AutobatchBlockstore) Put(ctx context.Context, blk block.Block) error {
2022-01-12 20:03:34 +00:00
bs.stateLock.Lock()
defer bs.stateLock.Unlock()
2022-01-12 17:57:34 +00:00
2022-01-11 22:17:34 +00:00
_, ok := bs.addedCids[blk.Cid()]
if !ok {
bs.addedCids[blk.Cid()] = struct{}{}
2022-01-12 17:57:34 +00:00
bs.bufferedBatch.blockList = append(bs.bufferedBatch.blockList, blk)
bs.bufferedBatch.blockMap[blk.Cid()] = blk
2022-01-11 22:17:34 +00:00
bs.bufferSize += len(blk.RawData())
if bs.bufferSize >= bs.bufferCapacity {
// signal that a flush is appropriate, may be ignored
select {
case bs.flushCh <- struct{}{}:
default:
// do nothing
}
2022-01-11 22:17:34 +00:00
}
2022-01-11 16:31:59 +00:00
}
2022-01-12 17:57:34 +00:00
2022-01-11 16:31:59 +00:00
return nil
}
func (bs *AutobatchBlockstore) flushWorker(ctx context.Context) {
defer close(bs.doneCh)
for {
select {
case <-bs.flushCh:
// TODO: check if we _should_ actually flush. We could get a spurious wakeup
// here.
putErr := bs.doFlush(ctx, false)
for putErr != nil {
select {
case <-ctx.Done():
return
2022-01-12 17:57:34 +00:00
case <-time.After(bs.flushRetryDelay):
autolog.Errorf("FLUSH ERRORED: %w, retrying after %v", putErr, bs.flushRetryDelay)
putErr = bs.doFlush(ctx, true)
}
}
case <-ctx.Done():
// Do one last flush.
_ = bs.doFlush(ctx, false)
return
}
}
}
2022-01-12 20:03:34 +00:00
// caller must NOT hold stateLock
// set retryOnly to true to only retry a failed flush and not flush anything new.
func (bs *AutobatchBlockstore) doFlush(ctx context.Context, retryOnly bool) error {
2022-01-12 20:03:34 +00:00
bs.doFlushLock.Lock()
defer bs.doFlushLock.Unlock()
// If we failed to flush last time, try flushing again.
if bs.flushErr != nil {
bs.flushErr = bs.backingBs.PutMany(ctx, bs.flushingBatch.blockList)
2022-01-12 17:57:34 +00:00
}
// If we failed, or we're _only_ retrying, bail.
if retryOnly || bs.flushErr != nil {
return bs.flushErr
}
// Then take the current batch...
2022-01-12 20:03:34 +00:00
bs.stateLock.Lock()
// We do NOT clear addedCids here, because its purpose is to expedite Puts
bs.flushingBatch = bs.bufferedBatch
bs.bufferedBatch.blockList = make([]block.Block, 0, len(bs.flushingBatch.blockList))
bs.bufferedBatch.blockMap = make(map[cid.Cid]block.Block, len(bs.flushingBatch.blockMap))
2022-01-12 20:03:34 +00:00
bs.stateLock.Unlock()
// And try to flush it.
bs.flushErr = bs.backingBs.PutMany(ctx, bs.flushingBatch.blockList)
// If we succeeded, reset the batch. Otherwise, we'll try again next time.
if bs.flushErr == nil {
bs.stateLock.Lock()
bs.flushingBatch = blockBatch{}
bs.stateLock.Unlock()
}
2022-01-12 17:57:34 +00:00
return bs.flushErr
}
2022-01-12 20:03:34 +00:00
// caller must NOT hold stateLock
2022-01-12 17:57:34 +00:00
func (bs *AutobatchBlockstore) Flush(ctx context.Context) error {
return bs.doFlush(ctx, false)
}
func (bs *AutobatchBlockstore) Shutdown(ctx context.Context) error {
// TODO: Prevent puts after we call this to avoid losing data.
bs.shutdown()
select {
case <-bs.doneCh:
case <-ctx.Done():
return ctx.Err()
2022-01-12 20:03:34 +00:00
}
bs.doFlushLock.Lock()
defer bs.doFlushLock.Unlock()
2022-01-12 17:57:34 +00:00
return bs.flushErr
2022-01-11 22:44:45 +00:00
}
func (bs *AutobatchBlockstore) Get(ctx context.Context, c cid.Cid) (block.Block, error) {
// may seem backward to check the backingBs first, but that is the likeliest case
2022-01-11 22:44:45 +00:00
blk, err := bs.backingBs.Get(ctx, c)
if err == nil {
return blk, nil
}
if !ipld.IsNotFound(err) {
2022-01-11 22:44:45 +00:00
return blk, err
}
2022-01-12 22:22:54 +00:00
bs.stateLock.Lock()
defer bs.stateLock.Unlock()
2022-01-12 17:57:34 +00:00
v, ok := bs.flushingBatch.blockMap[c]
if ok {
return v, nil
}
2022-01-12 20:03:34 +00:00
v, ok = bs.bufferedBatch.blockMap[c]
if ok {
return v, nil
}
2022-01-12 21:04:47 +00:00
return bs.Get(ctx, c)
2022-01-11 22:44:45 +00:00
}
func (bs *AutobatchBlockstore) DeleteBlock(context.Context, cid.Cid) error {
2022-01-12 17:57:34 +00:00
// if we wanted to support this, we would have to:
// - flush
// - delete from the backingBs (if present)
// - remove from addedCids (if present)
// - if present in addedCids, also walk the ordered lists and remove if present
2022-01-11 22:44:45 +00:00
return xerrors.New("deletion is unsupported")
}
func (bs *AutobatchBlockstore) DeleteMany(ctx context.Context, cids []cid.Cid) error {
// see note in DeleteBlock()
return xerrors.New("deletion is unsupported")
}
func (bs *AutobatchBlockstore) Has(ctx context.Context, c cid.Cid) (bool, error) {
_, err := bs.Get(ctx, c)
if err == nil {
return true, nil
}
if ipld.IsNotFound(err) {
2022-01-11 22:44:45 +00:00
return false, nil
}
return false, err
}
func (bs *AutobatchBlockstore) GetSize(ctx context.Context, c cid.Cid) (int, error) {
blk, err := bs.Get(ctx, c)
if err != nil {
return 0, err
}
return len(blk.RawData()), nil
}
func (bs *AutobatchBlockstore) PutMany(ctx context.Context, blks []block.Block) error {
for _, blk := range blks {
if err := bs.Put(ctx, blk); err != nil {
return err
}
}
return nil
}
func (bs *AutobatchBlockstore) AllKeysChan(ctx context.Context) (<-chan cid.Cid, error) {
2022-01-12 17:57:34 +00:00
if err := bs.Flush(ctx); err != nil {
return nil, err
}
2022-01-12 17:57:34 +00:00
return bs.backingBs.AllKeysChan(ctx)
2022-01-11 22:44:45 +00:00
}
func (bs *AutobatchBlockstore) HashOnRead(enabled bool) {
bs.backingBs.HashOnRead(enabled)
}
func (bs *AutobatchBlockstore) View(ctx context.Context, cid cid.Cid, callback func([]byte) error) error {
2022-01-12 20:03:34 +00:00
blk, err := bs.Get(ctx, cid)
if err != nil {
2022-01-12 17:57:34 +00:00
return err
}
2022-01-12 20:03:34 +00:00
return callback(blk.RawData())
2022-01-11 16:31:59 +00:00
}