forked from colinmarc/hdfs
-
Notifications
You must be signed in to change notification settings - Fork 0
/
block_reader.go
238 lines (205 loc) · 6.26 KB
/
block_reader.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
package rpc
import (
"code.google.com/p/goprotobuf/proto"
"encoding/binary"
"errors"
"fmt"
hdfs "github.com/colinmarc/hdfs/protocol/hadoop_hdfs"
"hash/crc32"
"io"
"io/ioutil"
"net"
)
// BlockReader implements io.ReadCloser, for reading a block. It abstracts over
// reading from multiple datanodes, in order to be robust to connection
// failures, timeouts, and other shenanigans.
type BlockReader struct {
block *hdfs.LocatedBlockProto
datanodes *datanodeFailover
stream *blockReadStream
conn net.Conn
offset int64
closed bool
}
// NewBlockReader returns a new BlockReader, given the block information and
// security token from the namenode. It will connect (lazily) to one of the
// provided datanode locations based on which datanodes have seen failures.
func NewBlockReader(block *hdfs.LocatedBlockProto, offset int64) *BlockReader {
locs := block.GetLocs()
datanodes := make([]string, len(locs))
for i, loc := range locs {
dn := loc.GetId()
datanodes[i] = fmt.Sprintf("%s:%d", dn.GetIpAddr(), dn.GetXferPort())
}
return &BlockReader{
block: block,
datanodes: newDatanodeFailover(datanodes),
offset: offset,
}
}
// Read implements io.Reader.
//
// In the case that a failure (such as a disconnect) occurs while reading, the
// BlockReader will failover to another datanode and continue reading
// transparently. In the case that all the datanodes fail, the error
// from the most recent attempt will be returned.
//
// Any datanode failures are recorded in a global cache, so subsequent reads,
// even reads for different blocks, will prioritize them lower.
func (br *BlockReader) Read(b []byte) (int, error) {
if br.closed {
return 0, io.ErrClosedPipe
} else if uint64(br.offset) >= br.block.GetB().GetNumBytes() {
br.Close()
return 0, io.EOF
}
// This is the main retry loop.
for br.stream != nil || br.datanodes.numRemaining() > 0 {
// First, we try to connect. If this fails, we can just skip the datanode
// and continue.
if br.stream == nil {
err := br.connectNext()
if err != nil {
br.datanodes.recordFailure(err)
continue
}
}
// Then, try to read. If we fail here after reading some bytes, we return
// a partial read (n < len(b)).
n, err := br.stream.Read(b)
br.offset += int64(n)
if err != nil && err != io.EOF {
br.stream = nil
br.datanodes.recordFailure(err)
if n > 0 {
return n, nil
}
continue
}
return n, err
}
err := br.datanodes.lastError()
if err == nil {
err = errors.New("No available datanodes for block.")
}
return 0, err
}
// Close implements io.Closer.
func (br *BlockReader) Close() error {
br.closed = true
if br.conn != nil {
br.conn.Close()
}
return nil
}
// connectNext pops a datanode from the list based on previous failures, and
// connects to it.
func (br *BlockReader) connectNext() error {
address := br.datanodes.next()
conn, err := net.DialTimeout("tcp", address, connectionTimeout)
if err != nil {
return err
}
err = br.writeBlockReadRequest(conn)
if err != nil {
return err
}
resp, err := br.readBlockReadResponse(conn)
if err != nil {
return err
}
readInfo := resp.GetReadOpChecksumInfo()
checksumInfo := readInfo.GetChecksum()
var checksumTab *crc32.Table
checksumType := checksumInfo.GetType()
switch checksumType {
case hdfs.ChecksumTypeProto_CHECKSUM_CRC32:
checksumTab = crc32.IEEETable
case hdfs.ChecksumTypeProto_CHECKSUM_CRC32C:
checksumTab = crc32.MakeTable(crc32.Castagnoli)
default:
return fmt.Errorf("Unsupported checksum type: %d", checksumType)
}
chunkSize := int(checksumInfo.GetBytesPerChecksum())
stream := newBlockReadStream(conn, chunkSize, checksumTab)
// The read will start aligned to a chunk boundary, so we need to seek forward
// to the requested offset.
amountToDiscard := br.offset - int64(readInfo.GetChunkOffset())
if amountToDiscard > 0 {
_, err := io.CopyN(ioutil.Discard, stream, amountToDiscard)
if err != nil {
if err == io.EOF {
err = io.ErrUnexpectedEOF
}
conn.Close()
return err
}
}
br.stream = stream
br.conn = conn
return nil
}
// A read request to a datanode:
// +-----------------------------------------------------------+
// | Data Transfer Protocol Version, int16 |
// +-----------------------------------------------------------+
// | Op code, 1 byte (READ_BLOCK = 0x51) |
// +-----------------------------------------------------------+
// | varint length + OpReadBlockProto |
// +-----------------------------------------------------------+
func (br *BlockReader) writeBlockReadRequest(w io.Writer) error {
header := []byte{0x00, dataTransferVersion, readBlockOp}
needed := br.block.GetB().GetNumBytes() - uint64(br.offset)
op := newBlockReadOp(br.block, uint64(br.offset), needed)
opBytes, err := makeDelimitedMsg(op)
if err != nil {
return err
}
req := append(header, opBytes...)
_, err = w.Write(req)
if err != nil {
return err
}
return nil
}
// The initial response from the datanode:
// +-----------------------------------------------------------+
// | varint length + BlockOpResponseProto |
// +-----------------------------------------------------------+
func (br *BlockReader) readBlockReadResponse(r io.Reader) (*hdfs.BlockOpResponseProto, error) {
varintBytes := make([]byte, binary.MaxVarintLen32)
_, err := io.ReadFull(r, varintBytes)
if err != nil {
return nil, err
}
respLength, varintLength := binary.Uvarint(varintBytes)
if varintLength < 1 {
return nil, io.ErrUnexpectedEOF
}
// We may have grabbed too many bytes when reading the varint.
respBytes := make([]byte, respLength)
extraLength := copy(respBytes, varintBytes[varintLength:])
_, err = io.ReadFull(r, respBytes[extraLength:])
if err != nil {
return nil, err
}
resp := &hdfs.BlockOpResponseProto{}
err = proto.Unmarshal(respBytes, resp)
if err != nil {
return nil, err
}
return resp, nil
}
func newBlockReadOp(block *hdfs.LocatedBlockProto, offset, length uint64) *hdfs.OpReadBlockProto {
return &hdfs.OpReadBlockProto{
Header: &hdfs.ClientOperationHeaderProto{
BaseHeader: &hdfs.BaseHeaderProto{
Block: block.GetB(),
Token: block.GetBlockToken(),
},
ClientName: proto.String(ClientName),
},
Offset: proto.Uint64(offset),
Len: proto.Uint64(length),
}
}