forked from NVIDIA/aistore
-
Notifications
You must be signed in to change notification settings - Fork 0
/
manager.go
439 lines (389 loc) · 12.6 KB
/
manager.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
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
// Package ec provides erasure coding (EC) based data protection for AIStore.
/*
* Copyright (c) 2018-2023, NVIDIA CORPORATION. All rights reserved.
*/
package ec
import (
"errors"
"fmt"
"io"
"sync"
"unsafe"
"github.com/artashesbalabekyan/aistore/3rdparty/glog"
"github.com/artashesbalabekyan/aistore/api/apc"
"github.com/artashesbalabekyan/aistore/cluster"
"github.com/artashesbalabekyan/aistore/cluster/meta"
"github.com/artashesbalabekyan/aistore/cmn"
"github.com/artashesbalabekyan/aistore/cmn/atomic"
"github.com/artashesbalabekyan/aistore/cmn/cos"
"github.com/artashesbalabekyan/aistore/cmn/debug"
"github.com/artashesbalabekyan/aistore/fs"
"github.com/artashesbalabekyan/aistore/transport"
"github.com/artashesbalabekyan/aistore/transport/bundle"
"github.com/artashesbalabekyan/aistore/xact/xreg"
)
type Manager struct {
mu sync.RWMutex
t cluster.Target
smap *meta.Smap
targetCnt atomic.Int32 // atomic, to avoid races between read/write on smap
bmd *meta.BMD // bmd owner
xacts map[string]*BckXacts // bckName -> xctn map, only ais buckets allowed, no naming collisions
bundleEnabled atomic.Bool // to disable and enable on the fly
netReq string // network used to send object request
netResp string // network used to send/receive slices
reqBundle atomic.Pointer
respBundle atomic.Pointer
}
var (
ECM *Manager
errSkipped = errors.New("skipped") // CT is skipped due to EC unsupported for the content type
)
func initManager(t cluster.Target) error {
var (
netReq, netResp = cmn.NetIntraControl, cmn.NetIntraData
sowner = t.Sowner()
smap = sowner.Get()
)
ECM = &Manager{
netReq: netReq,
netResp: netResp,
t: t,
smap: smap,
targetCnt: *atomic.NewInt32(int32(smap.CountActiveTs())),
bmd: t.Bowner().Get(),
xacts: make(map[string]*BckXacts),
}
if ECM.bmd.IsECUsed() {
return ECM.initECBundles()
}
return nil
}
func (mgr *Manager) req() *bundle.Streams {
return (*bundle.Streams)(mgr.reqBundle.Load())
}
func (mgr *Manager) resp() *bundle.Streams {
return (*bundle.Streams)(mgr.respBundle.Load())
}
func (mgr *Manager) initECBundles() error {
if !mgr.bundleEnabled.CAS(false, true) {
return nil
}
if err := transport.HandleObjStream(ReqStreamName, ECM.recvRequest); err != nil {
return fmt.Errorf("failed to register recvRequest: %v", err)
}
if err := transport.HandleObjStream(RespStreamName, ECM.recvResponse); err != nil {
return fmt.Errorf("failed to register respResponse: %v", err)
}
cbReq := func(hdr transport.ObjHdr, reader io.ReadCloser, _ any, err error) {
if err != nil {
glog.Errorf("failed to request %s: %v", hdr.Cname(), err)
}
}
var (
client = transport.NewIntraDataClient()
config = cmn.GCO.Get()
compression = config.EC.Compression
extraReq = transport.Extra{Callback: cbReq, Compression: compression}
)
reqSbArgs := bundle.Args{
Multiplier: config.EC.SbundleMult,
Extra: &extraReq,
Net: mgr.netReq,
Trname: ReqStreamName,
}
respSbArgs := bundle.Args{
Multiplier: config.EC.SbundleMult,
Trname: RespStreamName,
Net: mgr.netResp,
Extra: &transport.Extra{Compression: compression},
}
sowner := mgr.t.Sowner()
mgr.reqBundle.Store(unsafe.Pointer(bundle.New(sowner, mgr.t.Snode(), client, reqSbArgs)))
mgr.respBundle.Store(unsafe.Pointer(bundle.New(sowner, mgr.t.Snode(), client, respSbArgs)))
mgr.smap = sowner.Get()
mgr.targetCnt.Store(int32(mgr.smap.CountActiveTs()))
sowner.Listeners().Reg(mgr)
return nil
}
func (mgr *Manager) closeECBundles() {
if !mgr.bundleEnabled.CAS(true, false) {
return
}
mgr.t.Sowner().Listeners().Unreg(mgr)
mgr.req().Close(false)
mgr.resp().Close(false)
transport.Unhandle(ReqStreamName)
transport.Unhandle(RespStreamName)
}
func (mgr *Manager) NewGetXact(bck *cmn.Bck) *XactGet {
return NewGetXact(mgr.t, bck, mgr)
}
func (mgr *Manager) NewPutXact(bck *cmn.Bck) *XactPut {
return NewPutXact(mgr.t, bck, mgr)
}
func (mgr *Manager) NewRespondXact(bck *cmn.Bck) *XactRespond {
return NewRespondXact(mgr.t, bck, mgr)
}
func (mgr *Manager) RestoreBckGetXact(bck *meta.Bck) (xget *XactGet) {
xctn, err := _renewXact(bck, apc.ActECGet)
debug.AssertNoErr(err) // TODO: handle, here and elsewhere
xget = xctn.(*XactGet)
mgr.getBckXacts(bck.Name).SetGet(xget)
return
}
func (mgr *Manager) RestoreBckPutXact(bck *meta.Bck) (xput *XactPut) {
xctn, err := _renewXact(bck, apc.ActECPut)
debug.AssertNoErr(err)
xput = xctn.(*XactPut)
mgr.getBckXacts(bck.Name).SetPut(xput)
return
}
func (mgr *Manager) RestoreBckRespXact(bck *meta.Bck) (xrsp *XactRespond) {
xctn, err := _renewXact(bck, apc.ActECRespond)
debug.AssertNoErr(err)
xrsp = xctn.(*XactRespond)
mgr.getBckXacts(bck.Name).SetReq(xrsp)
return
}
func _renewXact(bck *meta.Bck, kind string) (cluster.Xact, error) {
rns := xreg.RenewBucketXact(kind, bck, xreg.Args{})
if rns.Err != nil {
return nil, rns.Err
}
return rns.Entry.Get(), nil
}
func (mgr *Manager) getBckXacts(bckName string) *BckXacts {
mgr.mu.Lock()
defer mgr.mu.Unlock()
return mgr.getBckXactsUnlocked(bckName)
}
func (mgr *Manager) getBckXactsUnlocked(bckName string) *BckXacts {
xacts, ok := mgr.xacts[bckName]
if !ok {
xacts = &BckXacts{}
mgr.xacts[bckName] = xacts
}
return xacts
}
// A function to process command requests from other targets
func (mgr *Manager) recvRequest(hdr transport.ObjHdr, object io.Reader, err error) error {
defer transport.FreeRecv(object)
if err != nil {
glog.Errorf("request failed: %v", err)
return err
}
// check if the header contains a valid request
if len(hdr.Opaque) == 0 {
err := fmt.Errorf("invalid header: [%+v]", hdr)
glog.Error(err)
return err
}
unpacker := cos.NewUnpacker(hdr.Opaque)
iReq := intraReq{}
if err := unpacker.ReadAny(&iReq); err != nil {
glog.Errorf("failed to unmarshal request: %v", err)
return err
}
// command requests should not have a body, but if it has,
// the body must be drained to avoid errors
if hdr.ObjAttrs.Size != 0 {
if _, err := io.ReadAll(object); err != nil {
glog.Errorf("failed to read request body: %v", err)
return err
}
}
bck := meta.CloneBck(&hdr.Bck)
if err = bck.Init(mgr.t.Bowner()); err != nil {
if _, ok := err.(*cmn.ErrRemoteBckNotFound); !ok { // is ais
glog.Errorf("failed to init bucket %s: %v", bck, err)
return err
}
}
mgr.RestoreBckRespXact(bck).DispatchReq(iReq, &hdr, bck)
return nil
}
// A function to process big chunks of data (replica/slice/meta) sent from other targets
func (mgr *Manager) recvResponse(hdr transport.ObjHdr, object io.Reader, err error) error {
defer transport.DrainAndFreeReader(object)
if err != nil {
glog.Errorf("receive failed: %v", err)
return err
}
// check if the request is valid
if len(hdr.Opaque) == 0 {
err := fmt.Errorf("invalid header: [%+v]", hdr)
glog.Error(err)
return err
}
unpacker := cos.NewUnpacker(hdr.Opaque)
iReq := intraReq{}
if err := unpacker.ReadAny(&iReq); err != nil {
glog.Errorf("Failed to unmarshal request: %v", err)
return err
}
bck := meta.CloneBck(&hdr.Bck)
if err = bck.Init(mgr.t.Bowner()); err != nil {
if _, ok := err.(*cmn.ErrRemoteBckNotFound); !ok { // is ais
glog.Error(err)
return err
}
}
switch hdr.Opcode {
case reqPut:
mgr.RestoreBckRespXact(bck).DispatchResp(iReq, &hdr, object)
case respPut:
// Process the request even if the number of targets is insufficient
// (might've started when we had enough)
mgr.RestoreBckGetXact(bck).DispatchResp(iReq, &hdr, bck, object)
default:
debug.Assertf(false, "unknown EC response action %d", hdr.Opcode)
}
return nil
}
// EncodeObject generates slices using Reed-Solom algorithm:
// - lom - object to encode
// - intra - if true, it is internal request and has low priority
// - cb - optional callback that is called after the object is encoded
func (mgr *Manager) EncodeObject(lom *cluster.LOM, cb ...cluster.OnFinishObj) error {
if !lom.Bprops().EC.Enabled {
return ErrorECDisabled
}
if cs := fs.Cap(); cs.Err != nil {
return cs.Err
}
isECCopy := IsECCopy(lom.SizeBytes(), &lom.Bprops().EC)
targetCnt := mgr.targetCnt.Load()
// compromise: encoding a small object requires fewer targets
if required := lom.Bprops().EC.RequiredEncodeTargets(); !isECCopy && int(targetCnt) < required {
return fmt.Errorf("%v: %d targets required to erasure code %s (have %d)",
cmn.ErrNotEnoughTargets, required, lom, targetCnt)
}
spec, _ := fs.CSM.FileSpec(lom.FQN)
if spec != nil && !spec.PermToProcess() {
return errSkipped
}
req := allocateReq(ActSplit, lom.LIF())
req.IsCopy = IsECCopy(lom.SizeBytes(), &lom.Bprops().EC)
if len(cb) != 0 {
req.rebuild = true
req.Callback = cb[0]
}
mgr.RestoreBckPutXact(lom.Bck()).encode(req, lom)
return nil
}
func (mgr *Manager) CleanupObject(lom *cluster.LOM) {
if !lom.Bprops().EC.Enabled {
return
}
debug.Assert(lom.FQN != "" && lom.Mountpath().Path != "")
req := allocateReq(ActDelete, lom.LIF())
mgr.RestoreBckPutXact(lom.Bck()).cleanup(req, lom)
}
func (mgr *Manager) RestoreObject(lom *cluster.LOM) error {
if !lom.Bprops().EC.Enabled {
return ErrorECDisabled
}
if cs := fs.Cap(); cs.Err != nil {
return cs.Err
}
targetCnt := mgr.targetCnt.Load()
// NOTE: Restore replica object is done with GFN, safe to always abort.
if required := lom.Bprops().EC.RequiredRestoreTargets(); int(targetCnt) < required {
return fmt.Errorf("%v: %d targets required to EC-restore %s (have %d)",
cmn.ErrNotEnoughTargets, required, lom, targetCnt)
}
debug.Assert(lom.Mountpath() != nil && lom.Mountpath().Path != "")
req := allocateReq(ActRestore, lom.LIF())
errCh := make(chan error) // unbuffered
req.ErrCh = errCh
mgr.RestoreBckGetXact(lom.Bck()).decode(req, lom)
// wait for EC completes restoring the object
return <-errCh
}
// disableBck starts to reject new EC requests, rejects pending ones
func (mgr *Manager) disableBck(bck *meta.Bck) {
mgr.RestoreBckGetXact(bck).ClearRequests()
mgr.RestoreBckPutXact(bck).ClearRequests()
}
// enableBck aborts xctn disable and starts to accept new EC requests
// enableBck uses the same channel as disableBck, so order of executing them is the same as
// order which they arrived to a target in
func (mgr *Manager) enableBck(bck *meta.Bck) {
mgr.RestoreBckGetXact(bck).EnableRequests()
mgr.RestoreBckPutXact(bck).EnableRequests()
}
func (mgr *Manager) BucketsMDChanged() error {
mgr.mu.Lock()
newBckMD := mgr.t.Bowner().Get()
oldBckMD := mgr.bmd
if newBckMD.Version <= mgr.bmd.Version {
mgr.mu.Unlock()
return nil
}
mgr.bmd = newBckMD
mgr.mu.Unlock()
if newBckMD.IsECUsed() && !oldBckMD.IsECUsed() {
if err := mgr.initECBundles(); err != nil {
return err
}
} else if !newBckMD.IsECUsed() && oldBckMD.IsECUsed() {
mgr.closeECBundles()
}
provider := apc.AIS
newBckMD.Range(&provider, nil, func(nbck *meta.Bck) bool {
oprops, ok := oldBckMD.Get(nbck)
if !ok {
if nbck.Props.EC.Enabled {
mgr.enableBck(nbck)
}
return false
}
if !oprops.EC.Enabled && nbck.Props.EC.Enabled {
mgr.enableBck(nbck)
} else if oprops.EC.Enabled && !nbck.Props.EC.Enabled {
mgr.disableBck(nbck)
}
return false
})
return nil
}
func (mgr *Manager) ListenSmapChanged() {
smap := mgr.t.Sowner().Get()
if smap.Version <= mgr.smap.Version {
return
}
mgr.smap = mgr.t.Sowner().Get()
targetCnt := mgr.smap.CountActiveTs()
mgr.targetCnt.Store(int32(targetCnt))
mgr.mu.Lock()
// Manager is initialized before being registered for smap changes
// bckMD will be present at this point
// stopping relevant EC xactions which can't be satisfied with current number of targets
// respond xaction is never stopped as it should respond regardless of the other targets
provider := apc.AIS
mgr.bmd.Range(&provider, nil, func(bck *meta.Bck) bool {
bckName, bckProps := bck.Name, bck.Props
bckXacts := mgr.getBckXactsUnlocked(bckName)
if !bckProps.EC.Enabled {
return false
}
if required := bckProps.EC.RequiredEncodeTargets(); targetCnt < required {
glog.Warningf("not enough targets for EC encoding for bucket %s; actual: %v, expected: %v",
bckName, targetCnt, required)
bckXacts.AbortPut()
}
// NOTE: this doesn't guarantee that present targets are sufficient to restore an object
// if one target was killed, and a new one joined, this condition will be satisfied even though
// slices of the object are not present on the new target
if required := bckProps.EC.RequiredRestoreTargets(); targetCnt < required {
glog.Warningf("not enough targets for EC restoring for bucket %s; actual: %v, expected: %v",
bckName, targetCnt, required)
bckXacts.AbortGet()
}
return false
})
mgr.mu.Unlock()
}
// implementing cluster.Slistener interface
func (*Manager) String() string { return "ecmanager" }