lotus/chain/store/snapshot.go

787 lines
19 KiB
Go
Raw Normal View History

package store
import (
"bytes"
"context"
"errors"
"fmt"
"io"
"sync"
"time"
blocks "github.com/ipfs/go-block-format"
"github.com/ipfs/go-cid"
format "github.com/ipfs/go-ipld-format"
"github.com/ipld/go-car"
carutil "github.com/ipld/go-car/util"
carv2 "github.com/ipld/go-car/v2"
"github.com/multiformats/go-multicodec"
cbg "github.com/whyrusleeping/cbor-gen"
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
"go.uber.org/atomic"
"golang.org/x/sync/errgroup"
"golang.org/x/xerrors"
"github.com/filecoin-project/go-state-types/abi"
2022-06-14 15:00:51 +00:00
bstore "github.com/filecoin-project/lotus/blockstore"
"github.com/filecoin-project/lotus/build"
"github.com/filecoin-project/lotus/chain/actors/builtin"
"github.com/filecoin-project/lotus/chain/types"
)
2023-01-20 18:27:04 +00:00
const TipsetkeyBackfillRange = 2 * build.Finality
func (cs *ChainStore) UnionStore() bstore.Blockstore {
return bstore.Union(cs.stateBlockstore, cs.chainBlockstore)
}
func (cs *ChainStore) Export(ctx context.Context, ts *types.TipSet, inclRecentRoots abi.ChainEpoch, skipOldMsgs bool, w io.Writer) error {
h := &car.CarHeader{
Roots: ts.Cids(),
Version: 1,
}
if err := car.WriteHeader(h, w); err != nil {
return xerrors.Errorf("failed to write car header: %s", err)
}
unionBs := cs.UnionStore()
return cs.WalkSnapshot(ctx, ts, inclRecentRoots, skipOldMsgs, true, func(c cid.Cid) error {
2021-12-11 21:03:00 +00:00
blk, err := unionBs.Get(ctx, c)
if err != nil {
return xerrors.Errorf("writing object to car, bs.Get: %w", err)
}
if err := carutil.LdWrite(w, c.Bytes(), blk.RawData()); err != nil {
return xerrors.Errorf("failed to write block to car output: %w", err)
}
return nil
})
}
func (cs *ChainStore) Import(ctx context.Context, r io.Reader) (head *types.TipSet, genesis *types.BlockHeader, err error) {
// TODO: writing only to the state blockstore is incorrect.
// At this time, both the state and chain blockstores are backed by the
// universal store. When we physically segregate the stores, we will need
// to route state objects to the state blockstore, and chain objects to
// the chain blockstore.
br, err := carv2.NewBlockReader(r)
if err != nil {
return nil, nil, xerrors.Errorf("loadcar failed: %w", err)
}
s := cs.StateBlockstore()
parallelPuts := 5
putThrottle := make(chan error, parallelPuts)
for i := 0; i < parallelPuts; i++ {
putThrottle <- nil
}
if len(br.Roots) == 0 {
return nil, nil, xerrors.Errorf("no roots in snapshot car file")
}
nextTailCid := br.Roots[0]
var tailBlock types.BlockHeader
tailBlock.Height = abi.ChainEpoch(-1)
var buf []blocks.Block
for {
blk, err := br.Next()
if err != nil {
// we're at the end
if err == io.EOF {
if len(buf) > 0 {
if err := s.PutMany(ctx, buf); err != nil {
return nil, nil, err
}
}
break
}
return nil, nil, err
}
// check for header block, looking for genesis
if blk.Cid() == nextTailCid && tailBlock.Height != 0 {
if err := tailBlock.UnmarshalCBOR(bytes.NewReader(blk.RawData())); err != nil {
return nil, nil, xerrors.Errorf("failed to unmarshal genesis block: %w", err)
}
if len(tailBlock.Parents) > 0 {
nextTailCid = tailBlock.Parents[0]
} else {
// note: even the 0th block has a parent linking to the cbor genesis block
return nil, nil, xerrors.Errorf("current block (epoch %d cid %s) has no parents", tailBlock.Height, tailBlock.Cid())
}
}
// append to batch
buf = append(buf, blk)
if len(buf) > 1000 {
if lastErr := <-putThrottle; lastErr != nil { // consume one error to have the right to add one
return nil, nil, lastErr
}
go func(buf []blocks.Block) {
putThrottle <- s.PutMany(ctx, buf)
}(buf)
buf = nil
}
}
// check errors
for i := 0; i < parallelPuts; i++ {
if lastErr := <-putThrottle; lastErr != nil {
return nil, nil, lastErr
}
}
if tailBlock.Height != 0 {
return nil, nil, xerrors.Errorf("expected genesis block to have height 0 (genesis), got %d: %s", tailBlock.Height, tailBlock.Cid())
}
root, err := cs.LoadTipSet(ctx, types.NewTipSetKey(br.Roots...))
if err != nil {
return nil, nil, xerrors.Errorf("failed to load root tipset from chainfile: %w", err)
}
ts := root
tssToPersist := make([]*types.TipSet, 0, TipsetkeyBackfillRange)
2023-01-20 18:27:04 +00:00
for i := 0; i < int(TipsetkeyBackfillRange); i++ {
tssToPersist = append(tssToPersist, ts)
parentTsKey := ts.Parents()
ts, err = cs.LoadTipSet(ctx, parentTsKey)
if ts == nil || err != nil {
log.Warnf("Only able to load the last %d tipsets", i)
break
}
}
if err := cs.PersistTipsets(ctx, tssToPersist); err != nil {
return nil, nil, xerrors.Errorf("failed to persist tipsets: %w", err)
}
return root, &tailBlock, nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
type walkSchedTaskType int
const (
finishTask walkSchedTaskType = -1
blockTask walkSchedTaskType = iota
messageTask
receiptTask
stateTask
dagTask
)
func (t walkSchedTaskType) String() string {
switch t {
case finishTask:
return "finish"
case blockTask:
return "block"
case messageTask:
return "message"
case receiptTask:
return "receipt"
case stateTask:
return "state"
case dagTask:
return "dag"
}
panic(fmt.Sprintf("unknown task %d", t))
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
}
type walkTask struct {
c cid.Cid
taskType walkSchedTaskType
topLevelTaskType walkSchedTaskType
blockCid cid.Cid
epoch abi.ChainEpoch
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
}
// an ever growing FIFO
type taskFifo struct {
in chan walkTask
out chan walkTask
fifo []walkTask
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
type taskResult struct {
c cid.Cid
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
b blocks.Block
}
func newTaskFifo(bufferLen int) *taskFifo {
f := taskFifo{
in: make(chan walkTask, bufferLen),
out: make(chan walkTask, bufferLen),
fifo: make([]walkTask, 0),
}
go f.run()
return &f
}
func (f *taskFifo) Close() error {
close(f.in)
return nil
}
func (f *taskFifo) run() {
for {
if len(f.fifo) > 0 {
// we have items in slice
// try to put next out or read something in.
// blocks if nothing works.
next := f.fifo[0]
select {
case f.out <- next:
f.fifo = f.fifo[1:]
case elem, ok := <-f.in:
if !ok {
// drain and close out.
for _, elem := range f.fifo {
f.out <- elem
}
close(f.out)
return
}
f.fifo = append(f.fifo, elem)
}
} else {
// no elements in fifo to put out.
// Try to read in and block.
// When done, try to put out or add to fifo.
select {
case elem, ok := <-f.in:
if !ok {
close(f.out)
return
}
select {
case f.out <- elem:
default:
f.fifo = append(f.fifo, elem)
}
}
}
}
}
type walkSchedulerConfig struct {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
numWorkers int
head *types.TipSet // Tipset to start walking from.
tail *types.TipSet // Tipset to end at.
includeMessages bool
includeReceipts bool
includeState bool
}
type walkScheduler struct {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
ctx context.Context
cancel context.CancelFunc
store bstore.Blockstore
cfg walkSchedulerConfig
writer io.Writer
workerTasks *taskFifo
totalTasks atomic.Int64
results chan taskResult
writeErrorChan chan error
// tracks number of inflight tasks
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
//taskWg sync.WaitGroup
// launches workers and collects errors if any occur
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
workers *errgroup.Group
// set of CIDs already exported
seen sync.Map
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
}
func newWalkScheduler(ctx context.Context, store bstore.Blockstore, cfg walkSchedulerConfig, w io.Writer) (*walkScheduler, error) {
ctx, cancel := context.WithCancel(ctx)
workers, ctx := errgroup.WithContext(ctx)
s := &walkScheduler{
ctx: ctx,
cancel: cancel,
store: store,
cfg: cfg,
writer: w,
results: make(chan taskResult, cfg.numWorkers*64),
workerTasks: newTaskFifo(cfg.numWorkers * 64),
writeErrorChan: make(chan error, 1),
workers: workers,
}
go func() {
defer close(s.writeErrorChan)
for r := range s.results {
// Write
if err := carutil.LdWrite(s.writer, r.c.Bytes(), r.b.RawData()); err != nil {
// abort operations
cancel()
s.writeErrorChan <- err
}
}
}()
// workers
for i := 0; i < cfg.numWorkers; i++ {
f := func(n int) func() error {
return func() error {
return s.workerFunc(n)
}
}(i)
s.workers.Go(f)
}
s.totalTasks.Add(int64(len(cfg.head.Blocks())))
for _, b := range cfg.head.Blocks() {
select {
case <-ctx.Done():
log.Errorw("context done while sending root tasks", ctx.Err())
cancel() // kill workers
return nil, ctx.Err()
case s.workerTasks.in <- walkTask{
c: b.Cid(),
taskType: blockTask,
topLevelTaskType: blockTask,
blockCid: b.Cid(),
epoch: cfg.head.Height(),
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
}:
}
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
return s, nil
}
func (s *walkScheduler) Wait() error {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
err := s.workers.Wait()
// all workers done. One would have reached genesis and notified the
// rest to exit. Yet, there might be some pending tasks in the queue,
// so we need to run a "single worker".
if err != nil {
log.Errorw("export workers finished with error", "error", err)
}
for {
if n := s.totalTasks.Load(); n == 0 {
break // finally fully done
}
select {
case task := <-s.workerTasks.out:
s.totalTasks.Add(-1)
if err != nil {
continue // just drain if errors happened.
}
err = s.processTask(task, 0)
}
}
close(s.results)
errWrite := <-s.writeErrorChan
if errWrite != nil {
log.Errorw("error writing to CAR file", "error", err)
return errWrite
}
_ = s.workerTasks.Close()
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
return err
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
func (s *walkScheduler) enqueueIfNew(task walkTask) {
if multicodec.Code(task.c.Prefix().MhType) == multicodec.Identity {
//log.Infow("ignored", "cid", todo.c.String())
return
}
// This lets through RAW, CBOR, and DagCBOR blocks, the only types that we end up writing to
// the exported CAR.
switch multicodec.Code(task.c.Prefix().Codec) {
case multicodec.Cbor, multicodec.DagCbor, multicodec.Raw:
default:
//log.Infow("ignored", "cid", todo.c.String())
return
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
if _, loaded := s.seen.LoadOrStore(task.c, struct{}{}); loaded {
// we already had it on the map
return
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
log.Debugw("enqueue", "type", task.taskType.String(), "cid", task.c.String())
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.totalTasks.Add(1)
s.workerTasks.in <- task
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
func (s *walkScheduler) sendFinish(workerN int) error {
log.Infow("worker finished work", "worker", workerN)
s.totalTasks.Add(1)
s.workerTasks.in <- walkTask{
taskType: finishTask,
}
return nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
func (s *walkScheduler) workerFunc(workerN int) error {
log.Infow("starting worker", "worker", workerN)
for t := range s.workerTasks.out {
s.totalTasks.Add(-1)
select {
case <-s.ctx.Done():
return s.ctx.Err()
default:
// A worker reached genesis, so we wind down and let others do
// the same. Exit.
if t.taskType == finishTask {
return s.sendFinish(workerN)
}
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
err := s.processTask(t, workerN)
if err != nil {
return err
}
// continue
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
return nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
func (s *walkScheduler) processTask(t walkTask, workerN int) error {
if t.taskType == finishTask {
return nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
blk, err := s.store.Get(s.ctx, t.c)
if errors.Is(err, format.ErrNotFound{}) && t.topLevelTaskType == receiptTask {
log.Debugw("ignoring not-found block in Receipts",
"block", t.blockCid,
"epoch", t.epoch,
"cid", t.c)
return nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
if err != nil {
return xerrors.Errorf(
"blockstore.Get(%s). Task: %s. Block: %s (%s). Epoch: %d. Err: %w",
t.c, t.taskType, t.topLevelTaskType, t.blockCid, t.epoch, err)
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.results <- taskResult{
c: t.c,
b: blk,
}
// We exported the ipld block. If it wasn't a CBOR block, there's nothing
// else to do and we can bail out early as it won't have any links
// etc.
if multicodec.Code(t.c.Prefix().Codec) != multicodec.DagCbor ||
multicodec.Code(t.c.Prefix().MhType) == multicodec.Identity {
return nil
}
rawData := blk.RawData()
// extract relevant dags to walk from the block
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
if t.taskType == blockTask {
var b types.BlockHeader
if err := b.UnmarshalCBOR(bytes.NewBuffer(rawData)); err != nil {
return xerrors.Errorf("unmarshalling block header (cid=%s): %w", blk, err)
}
if b.Height%1_000 == 0 {
log.Infow("block export", "height", b.Height)
}
if b.Height == 0 {
log.Info("exporting genesis block")
for i := range b.Parents {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: b.Parents[i],
taskType: dagTask,
topLevelTaskType: blockTask,
blockCid: b.Parents[i],
epoch: 0,
})
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: b.ParentStateRoot,
taskType: stateTask,
topLevelTaskType: stateTask,
blockCid: t.c,
epoch: 0,
})
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
return s.sendFinish(workerN)
}
// enqueue block parents
for i := range b.Parents {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: b.Parents[i],
taskType: blockTask,
topLevelTaskType: blockTask,
blockCid: b.Parents[i],
epoch: b.Height,
})
}
if s.cfg.tail.Height() >= b.Height {
log.Debugw("tail reached: only blocks will be exported from now until genesis", "cid", t.c.String())
return nil
}
if s.cfg.includeMessages {
// enqueue block messages
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: b.Messages,
taskType: messageTask,
topLevelTaskType: messageTask,
blockCid: t.c,
epoch: b.Height,
})
}
if s.cfg.includeReceipts {
// enqueue block receipts
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: b.ParentMessageReceipts,
taskType: receiptTask,
topLevelTaskType: receiptTask,
blockCid: t.c,
epoch: b.Height,
})
}
if s.cfg.includeState {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: b.ParentStateRoot,
taskType: stateTask,
topLevelTaskType: stateTask,
blockCid: t.c,
epoch: b.Height,
})
}
return nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
// Not a chain-block: we scan for CIDs in the raw block-data
err = cbg.ScanForLinks(bytes.NewReader(rawData), func(c cid.Cid) {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
s.enqueueIfNew(walkTask{
c: c,
taskType: dagTask,
topLevelTaskType: t.topLevelTaskType,
blockCid: t.blockCid,
epoch: t.epoch,
})
})
if err != nil {
return xerrors.Errorf(
"ScanForLinks(%s). Task: %s. Block: %s (%s). Epoch: %d. Err: %w",
t.c, t.taskType, t.topLevelTaskType, t.blockCid, t.epoch, err)
}
return nil
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
func (cs *ChainStore) ExportRange(
ctx context.Context,
w io.Writer,
head, tail *types.TipSet,
messages, receipts, stateroots bool,
workers int) error {
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
h := &car.CarHeader{
Roots: head.Cids(),
Version: 1,
}
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
if err := car.WriteHeader(h, w); err != nil {
return xerrors.Errorf("failed to write car header: %s", err)
}
start := time.Now()
log.Infow("walking snapshot range",
"head", head.Key(),
"tail", tail.Key(),
"messages", messages,
"receipts", receipts,
"stateroots",
stateroots,
"workers", workers)
cfg := walkSchedulerConfig{
numWorkers: workers,
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
head: head,
tail: tail,
includeMessages: messages,
includeState: stateroots,
includeReceipts: receipts,
}
pw, err := newWalkScheduler(ctx, cs.UnionStore(), cfg, w)
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
if err != nil {
return err
}
// wait until all workers are done.
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
err = pw.Wait()
if err != nil {
log.Errorw("walker scheduler", "error", err)
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
return err
}
log.Infow("walking snapshot range complete", "duration", time.Since(start), "success", err == nil)
Chain ranged export: rework and address current shortcomings This commit moderately refactors the ranged export code. It addresses several problems: * Code does not finish cleanly and things hang on ctrl-c * Same block is read multiple times in a row (artificially increasing cached blockstore metrics to 50%) * It is unclear whether there are additional races (a single worker quits when reaching height 0) * CARs produced have duplicated blocks (~400k for an 80M-blocks CAR or so). Some blocks appear up to 5 times. * Using pointers for tasks where it is not necessary. The changes: * Use a FIFO instead of stack: simpler implementation as its own type. This has not proven to be much more memory-friendly, but it has not made things worse either. * We avoid a probably not small amount of allocations by not using unnecessary pointers. * Fix duplicated blocks by atomically checking+adding to CID set. * Context-termination now works correctly. Worker lifetime is correctly tracked and all channels are closed, avoiding any memory leaks and deadlocks. * We ensure all work is finished before finishing, something that might have been broken in some edge cases previously. In practice, we would not have seen this except perhaps in very early snapshots close to genesis. Initial testing shows the code is currently about 5% faster. Resulting snapshots do not have duplicates so they are a bit smaller. We have manually verified that no CID is lost versus previous results, with both old and recent snapshots.
2023-02-02 16:51:52 +00:00
return nil
}
func (cs *ChainStore) WalkSnapshot(ctx context.Context, ts *types.TipSet, inclRecentRoots abi.ChainEpoch, skipOldMsgs, skipMsgReceipts bool, cb func(cid.Cid) error) error {
if ts == nil {
ts = cs.GetHeaviestTipSet()
}
seen := cid.NewSet()
walked := cid.NewSet()
blocksToWalk := ts.Cids()
currentMinHeight := ts.Height()
walkChain := func(blk cid.Cid) error {
if !seen.Visit(blk) {
return nil
}
if err := cb(blk); err != nil {
return err
}
2021-12-11 21:03:00 +00:00
data, err := cs.chainBlockstore.Get(ctx, blk)
if err != nil {
return xerrors.Errorf("getting block: %w", err)
}
var b types.BlockHeader
if err := b.UnmarshalCBOR(bytes.NewBuffer(data.RawData())); err != nil {
return xerrors.Errorf("unmarshaling block header (cid=%s): %w", blk, err)
}
if currentMinHeight > b.Height {
currentMinHeight = b.Height
if currentMinHeight%builtin.EpochsInDay == 0 {
log.Infow("export", "height", currentMinHeight)
}
}
var cids []cid.Cid
if !skipOldMsgs || b.Height > ts.Height()-inclRecentRoots {
if walked.Visit(b.Messages) {
2021-12-11 21:03:00 +00:00
mcids, err := recurseLinks(ctx, cs.chainBlockstore, walked, b.Messages, []cid.Cid{b.Messages})
if err != nil {
return xerrors.Errorf("recursing messages failed: %w", err)
}
cids = mcids
}
}
if b.Height > 0 {
2023-10-24 17:02:28 +00:00
blocksToWalk = append(blocksToWalk, b.Parents...)
} else {
// include the genesis block
cids = append(cids, b.Parents...)
}
out := cids
if b.Height == 0 || b.Height > ts.Height()-inclRecentRoots {
if walked.Visit(b.ParentStateRoot) {
2021-12-11 21:03:00 +00:00
cids, err := recurseLinks(ctx, cs.stateBlockstore, walked, b.ParentStateRoot, []cid.Cid{b.ParentStateRoot})
if err != nil {
return xerrors.Errorf("recursing genesis state failed: %w", err)
}
out = append(out, cids...)
}
if !skipMsgReceipts && walked.Visit(b.ParentMessageReceipts) {
out = append(out, b.ParentMessageReceipts)
}
}
for _, c := range out {
if seen.Visit(c) {
prefix := c.Prefix()
// Don't include identity CIDs.
if multicodec.Code(prefix.MhType) == multicodec.Identity {
continue
}
// We only include raw, cbor, and dagcbor, for now.
switch multicodec.Code(prefix.Codec) {
case multicodec.Cbor, multicodec.DagCbor, multicodec.Raw:
default:
continue
}
if err := cb(c); err != nil {
return err
}
}
}
return nil
}
log.Infow("export started")
exportStart := build.Clock.Now()
for len(blocksToWalk) > 0 {
next := blocksToWalk[0]
blocksToWalk = blocksToWalk[1:]
if err := walkChain(next); err != nil {
return xerrors.Errorf("walk chain failed: %w", err)
}
}
log.Infow("export finished", "duration", build.Clock.Now().Sub(exportStart).Seconds())
return nil
}
2021-12-11 21:03:00 +00:00
func recurseLinks(ctx context.Context, bs bstore.Blockstore, walked *cid.Set, root cid.Cid, in []cid.Cid) ([]cid.Cid, error) {
if multicodec.Code(root.Prefix().Codec) != multicodec.DagCbor {
return in, nil
}
2021-12-11 21:03:00 +00:00
data, err := bs.Get(ctx, root)
if err != nil {
return nil, xerrors.Errorf("recurse links get (%s) failed: %w", root, err)
}
var rerr error
err = cbg.ScanForLinks(bytes.NewReader(data.RawData()), func(c cid.Cid) {
if rerr != nil {
// No error return on ScanForLinks :(
return
}
// traversed this already...
if !walked.Visit(c) {
return
}
in = append(in, c)
var err error
2021-12-11 21:03:00 +00:00
in, err = recurseLinks(ctx, bs, walked, c, in)
if err != nil {
rerr = err
}
})
if err != nil {
return nil, xerrors.Errorf("scanning for links failed: %w", err)
}
return in, rerr
}