storage: Weave end offset into unsealed reader code

This commit is contained in:
Łukasz Magiera 2023-05-19 14:33:35 +02:00
parent d1d4b35adf
commit 8b2ef40f4e
7 changed files with 75 additions and 50 deletions

View File

@ -0,0 +1,44 @@
package readerutil
import (
"io"
"os"
)
// NewReadSeekerFromReaderAt returns a new io.ReadSeeker from a io.ReaderAt.
// The returned io.ReadSeeker will read from the io.ReaderAt starting at the
// given base offset.
func NewReadSeekerFromReaderAt(readerAt io.ReaderAt, base int64) io.ReadSeeker {
return &readSeekerFromReaderAt{
readerAt: readerAt,
base: base,
pos: 0,
}
}
type readSeekerFromReaderAt struct {
readerAt io.ReaderAt
base int64
pos int64
}
func (rs *readSeekerFromReaderAt) Read(p []byte) (n int, err error) {
n, err = rs.readerAt.ReadAt(p, rs.pos+rs.base)
rs.pos += int64(n)
return n, err
}
func (rs *readSeekerFromReaderAt) Seek(offset int64, whence int) (int64, error) {
switch whence {
case io.SeekStart:
rs.pos = offset
case io.SeekCurrent:
rs.pos += offset
case io.SeekEnd:
return 0, io.ErrUnexpectedEOF
default:
return 0, os.ErrInvalid
}
return rs.pos, nil
}

View File

@ -3,6 +3,7 @@ package paths
import (
"bytes"
"encoding/json"
"io"
"net/http"
"os"
"strconv"
@ -35,7 +36,7 @@ func (d *DefaultPartialFileHandler) HasAllocated(pf *partialfile.PartialFile, of
return pf.HasAllocated(offset, size)
}
func (d *DefaultPartialFileHandler) Reader(pf *partialfile.PartialFile, offset storiface.PaddedByteIndex, size abi.PaddedPieceSize) (*os.File, error) {
func (d *DefaultPartialFileHandler) Reader(pf *partialfile.PartialFile, offset storiface.PaddedByteIndex, size abi.PaddedPieceSize) (io.Reader, error) {
return pf.Reader(offset, size)
}

View File

@ -2,9 +2,8 @@ package paths
import (
"context"
"os"
"github.com/filecoin-project/go-state-types/abi"
"io"
"github.com/filecoin-project/lotus/storage/sealer/fsutil"
"github.com/filecoin-project/lotus/storage/sealer/partialfile"
@ -24,7 +23,7 @@ type PartialFileHandler interface {
HasAllocated(pf *partialfile.PartialFile, offset storiface.UnpaddedByteIndex, size abi.UnpaddedPieceSize) (bool, error)
// Reader returns a file from which we can read the unsealed piece in the partial file.
Reader(pf *partialfile.PartialFile, offset storiface.PaddedByteIndex, size abi.PaddedPieceSize) (*os.File, error)
Reader(pf *partialfile.PartialFile, offset storiface.PaddedByteIndex, size abi.PaddedPieceSize) (io.Reader, error)
// Close closes the partial file
Close(pf *partialfile.PartialFile) error

View File

@ -559,7 +559,7 @@ func (r *Remote) CheckIsUnsealed(ctx context.Context, s storiface.SectorRef, off
// 1. no worker(local worker included) has an unsealed file for the given sector OR
// 2. no worker(local worker included) has the unsealed piece in their unsealed sector file.
// Will return a nil reader and a nil error in such a case.
func (r *Remote) Reader(ctx context.Context, s storiface.SectorRef, offset, size abi.PaddedPieceSize) (func(startOffsetAligned storiface.PaddedByteIndex) (io.ReadCloser, error), error) {
func (r *Remote) Reader(ctx context.Context, s storiface.SectorRef, offset, size abi.PaddedPieceSize) (func(startOffsetAligned, endOffsetAligned storiface.PaddedByteIndex) (io.ReadCloser, error), error) {
ft := storiface.FTUnsealed
// check if we have the unsealed sector file locally
@ -598,20 +598,8 @@ func (r *Remote) Reader(ctx context.Context, s storiface.SectorRef, offset, size
if has {
log.Infof("returning piece reader for local unsealed piece sector=%+v, (offset=%d, size=%d)", s.ID, offset, size)
return func(startOffsetAligned storiface.PaddedByteIndex) (io.ReadCloser, error) {
// don't reuse between readers unless closed
f := pf
pf = nil
if f == nil {
f, err = r.pfHandler.OpenPartialFile(abi.PaddedPieceSize(ssize), path)
if err != nil {
return nil, xerrors.Errorf("opening partial file: %w", err)
}
log.Debugf("local partial file (re)opened %s (+%d,%d)", path, offset, size)
}
r, err := r.pfHandler.Reader(f, storiface.PaddedByteIndex(offset)+startOffsetAligned, size-abi.PaddedPieceSize(startOffsetAligned))
return func(startOffsetAligned, endOffsetAligned storiface.PaddedByteIndex) (io.ReadCloser, error) {
r, err := r.pfHandler.Reader(pf, storiface.PaddedByteIndex(offset)+startOffsetAligned, abi.PaddedPieceSize(endOffsetAligned-startOffsetAligned))
if err != nil {
return nil, err
}
@ -622,22 +610,7 @@ func (r *Remote) Reader(ctx context.Context, s storiface.SectorRef, offset, size
}{
Reader: r,
Closer: funcCloser(func() error {
// if we already have a reader cached, close this one
if pf != nil {
if f == nil {
return nil
}
if pf == f {
pf = nil
}
tmp := f
f = nil
return tmp.Close()
}
// otherwise stash it away for reuse
pf = f
// todo keep some refcount, close pf (or push to some lru) when it hits 0
return nil
}),
}, nil
@ -685,10 +658,10 @@ func (r *Remote) Reader(ctx context.Context, s storiface.SectorRef, offset, size
continue
}
return func(startOffsetAligned storiface.PaddedByteIndex) (io.ReadCloser, error) {
return func(startOffsetAligned, endOffsetAligned storiface.PaddedByteIndex) (io.ReadCloser, error) {
// readRemote fetches a reader that we can use to read the unsealed piece from the remote worker.
// It uses a ranged HTTP query to ensure we ONLY read the unsealed piece and not the entire unsealed file.
rd, err := r.readRemote(ctx, url, offset+abi.PaddedPieceSize(startOffsetAligned), size)
rd, err := r.readRemote(ctx, url, offset+abi.PaddedPieceSize(startOffsetAligned), offset+abi.PaddedPieceSize(endOffsetAligned))
if err != nil {
log.Warnw("reading from remote", "url", url, "error", err)
return nil, err

View File

@ -2,6 +2,7 @@ package partialfile
import (
"encoding/binary"
"github.com/filecoin-project/lotus/lib/readerutil"
"io"
"os"
"syscall"
@ -249,7 +250,10 @@ func (pf *PartialFile) Free(offset storiface.PaddedByteIndex, size abi.PaddedPie
return nil
}
func (pf *PartialFile) Reader(offset storiface.PaddedByteIndex, size abi.PaddedPieceSize) (*os.File, error) {
// Reader forks off a new reader from the underlying file, and returns a reader
// starting at the given offset and reading the given size. Safe for concurrent
// use.
func (pf *PartialFile) Reader(offset storiface.PaddedByteIndex, size abi.PaddedPieceSize) (io.Reader, error) {
if _, err := pf.file.Seek(int64(offset), io.SeekStart); err != nil {
return nil, xerrors.Errorf("seek piece start: %w", err)
}
@ -275,7 +279,7 @@ func (pf *PartialFile) Reader(offset storiface.PaddedByteIndex, size abi.PaddedP
}
}
return pf.file, nil
return io.LimitReader(readerutil.NewReadSeekerFromReaderAt(pf.file, int64(offset)), int64(size)), nil
}
func (pf *PartialFile) Allocated() (rlepluslazy.RunIterator, error) {

View File

@ -71,7 +71,7 @@ func (p *pieceProvider) IsUnsealed(ctx context.Context, sector storiface.SectorR
// It will NOT try to schedule an Unseal of a sealed sector file for the read.
//
// Returns a nil reader if the piece does NOT exist in any unsealed file or there is no unsealed file for the given sector on any of the workers.
func (p *pieceProvider) tryReadUnsealedPiece(ctx context.Context, pc cid.Cid, sector storiface.SectorRef, pieceOffset storiface.UnpaddedByteIndex, size abi.UnpaddedPieceSize) (mount.Reader, error) {
func (p *pieceProvider) tryReadUnsealedPiece(ctx context.Context, pc cid.Cid, sector storiface.SectorRef, pieceOffset storiface.UnpaddedByteIndex, pieceSize abi.UnpaddedPieceSize) (mount.Reader, error) {
// acquire a lock purely for reading unsealed sectors
ctx, cancel := context.WithCancel(ctx)
if err := p.index.StorageLock(ctx, sector.ID, storiface.FTUnsealed, storiface.FTNone); err != nil {
@ -82,7 +82,7 @@ func (p *pieceProvider) tryReadUnsealedPiece(ctx context.Context, pc cid.Cid, se
// Reader returns a reader getter for an unsealed piece at the given offset in the given sector.
// The returned reader will be nil if none of the workers has an unsealed sector file containing
// the unsealed piece.
rg, err := p.storage.Reader(ctx, sector, abi.PaddedPieceSize(pieceOffset.Padded()), size.Padded())
rg, err := p.storage.Reader(ctx, sector, abi.PaddedPieceSize(pieceOffset.Padded()), pieceSize.Padded())
if err != nil {
cancel()
log.Debugf("did not get storage reader;sector=%+v, err:%s", sector.ID, err)
@ -93,19 +93,23 @@ func (p *pieceProvider) tryReadUnsealedPiece(ctx context.Context, pc cid.Cid, se
return nil, nil
}
buf := make([]byte, fr32.BufSize(size.Padded()))
buf := make([]byte, fr32.BufSize(pieceSize.Padded()))
pr, err := (&pieceReader{
ctx: ctx,
getReader: func(ctx context.Context, startOffset uint64) (io.ReadCloser, error) {
getReader: func(ctx context.Context, startOffset, readSize uint64) (io.ReadCloser, error) {
startOffsetAligned := storiface.UnpaddedByteIndex(startOffset / 127 * 127) // floor to multiple of 127
startOffsetDiff := int(startOffset - uint64(startOffsetAligned))
r, err := rg(startOffsetAligned.Padded())
endOffset := startOffset + readSize
endOffsetAligned := storiface.UnpaddedByteIndex((endOffset + 126) / 127 * 127) // ceil to multiple of 127
r, err := rg(startOffsetAligned.Padded(), endOffsetAligned.Padded())
if err != nil {
return nil, xerrors.Errorf("getting reader at +%d: %w", startOffsetAligned, err)
}
upr, err := fr32.NewUnpadReaderBuf(r, size.Padded(), buf)
upr, err := fr32.NewUnpadReaderBuf(r, pieceSize.Padded(), buf)
if err != nil {
r.Close() // nolint
return nil, xerrors.Errorf("creating unpadded reader: %w", err)
@ -113,7 +117,7 @@ func (p *pieceProvider) tryReadUnsealedPiece(ctx context.Context, pc cid.Cid, se
bir := bufio.NewReaderSize(upr, 127)
if startOffset > uint64(startOffsetAligned) {
if _, err := bir.Discard(int(startOffset - uint64(startOffsetAligned))); err != nil {
if _, err := bir.Discard(startOffsetDiff); err != nil {
r.Close() // nolint
return nil, xerrors.Errorf("discarding bytes for startOffset: %w", err)
}
@ -129,7 +133,7 @@ func (p *pieceProvider) tryReadUnsealedPiece(ctx context.Context, pc cid.Cid, se
}),
}, nil
},
len: size,
len: pieceSize,
onClose: cancel,
pieceCid: pc,
}).init()

View File

@ -21,7 +21,7 @@ import (
var MaxPieceReaderBurnBytes int64 = 1 << 20 // 1M
var ReadBuf = 128 * (127 * 8) // unpadded(128k)
type pieceGetter func(ctx context.Context, offset uint64) (io.ReadCloser, error)
type pieceGetter func(ctx context.Context, offset, size uint64) (io.ReadCloser, error)
type pieceReader struct {
ctx context.Context
@ -43,7 +43,7 @@ func (p *pieceReader) init() (_ *pieceReader, err error) {
stats.Record(p.ctx, metrics.DagStorePRInitCount.M(1))
p.rAt = 0
p.r, err = p.getReader(p.ctx, uint64(p.rAt))
p.r, err = p.getReader(p.ctx, uint64(p.rAt), uint64(p.len))
if err != nil {
return nil, err
}
@ -160,7 +160,7 @@ func (p *pieceReader) readAtUnlocked(b []byte, off int64) (n int, err error) {
}
p.rAt = off
p.r, err = p.getReader(p.ctx, uint64(p.rAt))
p.r, err = p.getReader(p.ctx, uint64(p.rAt), uint64(p.len))
p.br = bufio.NewReaderSize(p.r, ReadBuf)
if err != nil {
return 0, xerrors.Errorf("getting backing reader: %w", err)