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
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
// Copyright (C) 2018 Nexedi SA and Contributors.
// Kirill Smelkov <kirr@nexedi.com>
//
// This program is free software: you can Use, Study, Modify and Redistribute
// it under the terms of the GNU General Public License version 3, or (at your
// option) any later version, as published by the Free Software Foundation.
//
// You can also Link and Combine this program with other software covered by
// the terms of any of the Free Software licenses or any of the Open Source
// Initiative approved licenses and Convey the resulting work. Corresponding
// source of such a combination shall include the source code for all other
// software used.
//
// This program is distributed WITHOUT ANY WARRANTY; without even the implied
// warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
//
// See COPYING file for full licensing terms.
// See https://www.nexedi.com/licensing for rationale and options.
// Program wcfs provides filesystem server with file data backed by wendelin.core arrays.
//
// Intro
//
// Each wendelin.core array (ZBigArray) is actually a linear file (ZBigFile)
// and array metadata like dtype, shape and strides associated with it. This
// program exposes as files only ZBigFile data and leaves rest of
// array-specific handling to client. Every ZBigFile is exposed as one separate
// file that represents whole ZBigFile's data.
//
// For a client, the primary way to access a bigfile should be to mmap
// bigfile/<bigfileX>/head/data which represents always latest bigfile data.
// Clients that want to get isolation guarantee should subscribe for
// invalidations and re-mmap invalidated regions to file with pinned bigfile revision for
// the duration of their transaction. See "Invalidation protocol" for details.
//
// In the usual situation when bigfiles are big, and there are O(1)/δt updates,
// there should be no need for any cache besides shared kernel cache of latest
// bigfile data.
//
//
// Filesystem organization
//
// Top-level structure of provided filesystem is as follows:
//
// bigfile/
// <oid(bigfile1)>/
// ...
// <oid(bigfile2)>/
// ...
// ...
//
// where for a bigfileX there is bigfile/<oid(bigfileX)>/ directory, with
// oid(bigfileX) being ZODB object-id of corresponding ZBigFile object formatted with %016x.
//
// Each bigfileX/ has the following structure:
//
// bigfile/<bigfileX>/
// head/ ; latest bigfile revision
// ...
// @<tid1>/ ; bigfile revision as of transaction <tidX>
// ...
// @<tid2>/
// ...
// ...
//
// where head/ represents latest bigfile as stored in upstream ZODB, and
// @<tidX>/ represents bigfile as of transaction <tidX>.
//
// head/ has the following structure:
//
// bigfile/<bigfileX>/head/
// data ; latest bigfile data
// at ; data is bigfile view as of this ZODB transaction
// invalidations ; channel that describes invalidated data regions
//
// where /data represents latest bigfile data as stored in upstream ZODB. As
// there can be some lag receiving updates from the database, /at describes
// precisely ZODB state for which bigfile data is currently exposed. Whenever
// bigfile data is changed in upstream ZODB, information about the changes is
// first propagated to /invalidations, and only after that /data is
// updated. See "Invalidation protocol" for details.
//
// @<tidX>/ has the following structure:
//
// bigfile/<bigfileX>/@<tidX>/
// data ; bigfile data as of transaction <tidX>
//
// where /data represents bigfile data as of transaction <tidX>.
//
// bigfile/<bigfileX>/ should be created by client via mkdir. Unless explicitly
// created bigfile/<bigfileX>/ are not automatically visible in wcfs
// filesystem. Similarly bigfile/<bigfileX>/@<tidX>/ should be too created by
// client.
//
//
// Invalidation protocol
//
// In order to support isolation wcfs implements invalidation protocol that
// must be cooperatively followed by both wcfs and client.
//
// First, before client wants to mmap bigfile, it opens
// bigfile/<bigfileX>/head/invalidations and tells wcfs through it for which
// ZODB state it wants to get bigfile view. The server in turn reports for
// which ZODB state head/data is current, δ describing changed bigfile region
// between those revisions, or "wait" flag if server state is earlier compared
// to what client wants:
//
// C: want <Cat>
// S: have <Sat>, wait ; Sat < Cat
// S: have <Sat>, δR(Cat,Sat) ; Sat ≥ Cat
//
// If server reply was "wait" the client does nothing and waits for next server
// message which must come without "wait" flag set. When client receives have
// message with δR(Cat,Sat) it has the guarantee from wcfs that head/data
// content is for Sat ZODB revision and won't change until client sends ack
// back to the server. The client in turn now can mmap head/data and
// @<Cat>/data to get bigfile view as of Cat:
//
// mmap(bigfile/<bigfileX>/head/data)
// mmap(bigfile/<bigfileX>/@<Cat>/data, δR(Cat,Sat), MAP_FIXED) # mmaped at addresses corresponding to δR(Cat,Sat)
//
// When client completes its initial mmapping it sends ack back to the server:
//
// C: ack
//
// From now on the server will be processing updates to bigfile coming from
// ZODB as follows:
//
//
// The filesystem server itself receives information about changed data
// from ZODB server through regular ZODB invalidation channel (as it is ZODB
// client itself). Then, before actually updating bigfile/<bigfileX>/head/data
// content in changed part, it notifies through bigfile/<bigfileX>/head/invalidations
// to clients that had opened this file (separately to each client) about the changes:
//
// S: have <Sat>, δR(Sat_prev, Sat)
//
// where Sat_prev is ZODB revision last reported to client for this bigfile,
// and waits until they all confirm that changed file part can be updated in
// global OS cache.
//
// The client in turn can now re-mmap invalidated regions to bigfile@Cat
//
// # mmapped at addresses corresponding to δR(Sat_prev, Sat)
// mmap(bigfile/<bigfileX>/@<Cat>/data, δR(Sat_prev, Sat), MAP_FIXED)
//
// and must send ack back to the server when it is done:
//
// C: ack
//
// When clients are done with bigfile/<bigfileX>/@<Cat>/data (i.e. Cat
// transaction ends and array is unmapped), the server sees number of opened
// files to bigfile/<bigfileX>/@<Cat>/data drops to zero, and automatically
// destroys bigfile/<bigfileX>/@<Cat>/ directory after reasonable timeout.
//
//
// Protection against slow or faulty clients
//
// If a client, on purpose or due to a bug or being stopped, is slow to
// respond with ack to invalidation notification, it creates a problem because
// head/data updates will be blocked and thus all other clients that try to
// work with current data will get stuck.
//
// To avoid this problem it should be possible for wcfs to stop a client with
// ptrace and change its address space in a style similar to e.g.
// VirtualAllocEx on Windows. Here is hacky example how this could be done on Linux:
//
// https://gist.github.com/rofl0r/1073739/63f0f788a4923e26fcf743dd9a8411d4916f0ac0
//
// This way there should be no possibility for a client to block wcfs
// indefinitely waiting for client's ack.
//
// Similarly for initial mmapings client could first mmap head/data, then open
// head/invalidations and tell the server that it wants Cat revision, with
// the server then remmaping blocks to get to Cat state via ptrace.
//
// However for simplicity the plan is to go first without ptrace and just kill
// a slow client on, say 30 seconds, timeout.
//
//
// Writes
//
// As each bigfile is represented by 1 synthetic file, there can be several
// write schemes:
//
// 1. mmap(MAP_PRIVATE) + writeout by client
//
// In this scheme bigfile data is mmapped in MAP_PRIVATE mode, so that local
// user changes are not automatically propagated back to the file. When there
// is a need to commit, client investigates via some OS mechanism, e.g.
// /proc/self/pagemap or something similar, which pages of this mapping it
// modified. Knowing this it knows which data it dirtied and so can write this
// data back to ZODB itself, without filesystem server providing write support.
//
// 2. mmap(MAP_SHARED, PROT_READ) + write-tracking & writeout by client
//
// In this scheme bigfile data is mmaped in MAP_SHARED mode with read-only pages
// protection. Then whenever write fault occurs, client allocates RAM from
// shmfs, copies faulted page to it, and then mmaps RAM page with RW protection
// in place of original bigfile page. Writeout implementation should be similar
// to "1", only here client already knows the pages it dirtied, and this way
// there is no need to consult /proc/self/pagemap.
//
// The advantage of this scheme over mmap(MAP_PRIVATE) is that in case
// there are several in-process mappings of the same bigfile with overlapping
// in-file ranges, changes in one mapping will be visible in another mapping.
// Contrary: whenever a MAP_PRIVATE mapping is modified, the kernel COWs
// faulted page into a page completely private to this mapping, so that other
// MAP_PRIVATE mappings of this file, including ones created from the same
// process, do not see changes made to the first mapping.
//
// Since wendelin.core needs to provide coherency in between different slices
// of the same array, this is the mode wendelin.core actually uses.
//
// 3. write to wcfs
//
// XXX we later could implement "write-directly" mode where clients would write
// data directly into the file.
package main
// Notes on OS pagecache control:
//
// the cache of snapshotted bigfile can be pre-made hot, if invalidated region
// was already in pagecache of head/data:
//
// - we can retrieve a region from pagecache of head/data with FUSE_NOTIFY_RETRIEVE.
// - we can store that retrieved data into pagecache region of @<tidX>/ with FUSE_NOTIFY_STORE.
// - we can invalidate a region from pagecache of head/data with FUSE_NOTIFY_INVAL_INODE.
//
// we have to disable FUSE_AUTO_INVAL_DATA to tell the kernel we are fully
// responsible for invalidating pagecache. If we don't, the kernel will be
// clearing whole cache of head/data on e.g. its mtime change.
//
// XXX FUSE_AUTO_INVAL_DATA does not fully prevent kernel from automatically
// invalidating pagecache - e.g. it will invalidate whole cache on file size changes:
//
// https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/fuse/inode.c?id=e0bc833d10#n233
//
// we can currently workaround it with using writeback mode (see !is_wb in the
// link above), but better we have proper FUSE flag for filesystem server to
// tell the kernel it is fully responsible for invalidating pagecache.
// TODO implementation organization.
// - 1 ZODB connection per 1 bigfile (each bigfile can be at its different @at,
// because invalidations for different bigfiles can be processed with different
// timings depending on clients). No harm here as different bigfiles use
// completely different ZODB BTree and data objects.
import (
"context"
"flag"
"log"
"os"
"sync"
"syscall"
"lab.nexedi.com/kirr/neo/go/zodb"
"lab.nexedi.com/kirr/neo/go/zodb/btree"
_ "lab.nexedi.com/kirr/neo/go/zodb/wks"
"github.com/hanwen/go-fuse/fuse"
"github.com/hanwen/go-fuse/fuse/nodefs"
//"github.com/pkg/errors"
//pickle "github.com/kisielk/og-rek" // XXX should be temp here?
)
// BigFileRoot represents "/bigfile"
type BigFileRoot struct {
nodefs.Node
zstor zodb.IStorage
mu sync.Mutex
tab map[zodb.Oid]*BigFileX
}
// BigFileX represents "/bigfile/<bigfileX>"
// XXX -> BigFileDir ?
type BigFileX struct {
nodefs.Node
oid zodb.Oid
root *BigFileRoot
}
// BigFileHead represents "/bigfile/<bigfileX>/head"
type BigFileHead struct {
nodefs.Node
x *BigFileX
data *BigFile
//at *BigFileAt
//inv *BigFileInvalidations
}
// BigFile represents "/bigfile/<bigfileX>/head/data"
// XXX also @<tidX>/data ?
// XXX -> BigFileData ?
type BigFile struct {
nodefs.Node
parent *BigFileHead // XXX name
topoid zodb.Oid // oid of ZBigFile
blksize int64 // ZBigFile.blksize XXX if it is changed - invalidate all? allowed to change?
head zodb.Tid // current view of ZODB
lastChange zodb.Tid // last change to whole bigfile as of .head view
}
// ---- ctors ---- XXX to down?
func NewBigFileRoot(zstor zodb.IStorage) *BigFileRoot {
return &BigFileRoot{
Node: nodefs.NewDefaultNode(),
zstor: zstor,
tab: make(map[zodb.Oid]*BigFileX),
}
}
func NewBigFileX(oid zodb.Oid, root *BigFileRoot) *BigFileX {
bx := &BigFileX{
Node: nodefs.NewDefaultNode(),
oid: oid,
root: root,
}
return bx
}
func NewBigFileHead(x *BigFileX) *BigFileHead {
f := &BigFileHead{Node: nodefs.NewDefaultNode(), x: x}
f.data = NewBigFile(f)
// XXX + .at
return f
}
func NewBigFile(head *BigFileHead) *BigFile {
return &BigFile{Node: nodefs.NewDefaultNode(), parent: head}
}
// Mkdir receives client request to create /bigfile/<bigfileX>.
func (br *BigFileRoot) Mkdir(name string, mode uint32, fctx *fuse.Context) (*nodefs.Inode, fuse.Status) {
oid, err := zodb.ParseOid(name)
if err != nil {
log.Printf("/bigfile: mkdir %q: not-oid", name)
return nil, fuse.EINVAL
}
// XXX ok to ignore mode?
br.mu.Lock()
if _, already := br.tab[oid]; already {
return nil, fuse.Status(syscall.EEXIST)
}
br.mu.Unlock()
ctx := asctx(fctx)
_ = ctx
return nil, fuse.ENOSYS // XXX temp
// xbf, err := zconn.Get(ctx, oid)
// check err
// check if xbf.(*ZBigFile)
/* XXX kill
buf, _, err := br.zstor.Load(ctx, zodb.Xid{Oid: oid, At: zodb.TidMax}) // FIXME At, use serial
if err != nil {
switch errors.Cause(err).(type) {
case *zodb.NoObjectError:
return nil, fuse.EINVAL
case *zodb.NoDataError:
return nil, fuse.EINVAL // XXX ok?
default:
log.Printf("/bigfile: mkdir %q: %s", name, err)
return nil, fuse.EIO
}
}
pybf, err := zodb.PyData(buf.Data).Decode()
if err != nil {
log.Printf("/bigfile: mkdir %q: %s", name, err)
return nil, fuse.EIO
}
*/
/* XXX reenable (.zpy not yet here)
pybf, err := br.zpy.Load(ctx, zodb.Xid{Oid: oid, At: zodb.TidMax}) // FIXME At, use serial
if err != nil {
switch errors.Cause(err).(type) {
case *zodb.NoObjectError:
// XXX log?
return nil, fuse.EINVAL
case *zodb.NoDataError:
// XXX log?
return nil, fuse.EINVAL // XXX ok?
default:
log.Printf("/bigfile: mkdir %q: %s", name, err)
return nil, fuse.EIO
}
}
// XXX -> pyclass.FullName() != "wendelin.bigfile.file_zodb" + ".ZBigFile"
pybfClass := pickle.Class{Module: "wendelin.bigfile.file_zodb", Name: ".ZBigFile"}
if pybf.PyClass() != pybfClass {
// XXX log?
return nil, fuse.EINVAL
}
// XXX other checks?
br.mu.Lock()
defer br.mu.Unlock()
// XXX recheck - maybe it was already created while we were not holding br.mu
bx := NewBigFileX(oid, br)
br.tab[oid] = bx
bh := NewBigFileHead(bx)
mkdir(br, name, bx) // XXX takes treeLock - ok under br.mu ?
mkdir(bx, "head", bh)
mkfile(bh, "data", bh.data)
// XXX mkfile(bh, "at", bh.at)
// XXX mkfile(bh, "invalidations", bh.inv)
return bx.Inode(), fuse.OK
*/
}
// XXX do we need to support rmdir? (probably no)
// module: "wendelin.bigfile.file_zodb"
//
// ZBigFile
// .blksize xint
// .blktab LOBtree{} blk -> ZBlk*(blkdata)
//
// ZBlk0 (aliased as ZBlk)
// str with trailing '\0' removed.
//
// ZBlk1
// .chunktab IOBtree{} offset -> ZData(chunk)
//
// ZData
// str (chunk)
// Read implements reading from /bigfile/<bigfileX>/head/data.
// XXX and from /bigfile/<bigfileX>/@<tidX>/data.
/*
func (bf *BigFile) Read(_ nodefs.File, dest []byte, off int64, _ fuse.Context) (fuse.ReadResult, fuse.Status) {
.at
.topoid
// XXX
}
*/
// zodbCacheControl implements LiveCacheControl to tune ZODB to never evict
// LOBTree/LOBucket from live cache. We want to keep LOBTree/LOBucket always alive
// becuse it is essentially the index where to find ZBigFile data.
//
// For the data itself - we put it to kernel cache and always deactivate from
// ZODB right after that.
//
// TODO set it to Connection.CacheControl
type zodbCacheControl struct {}
func (cc *zodbCacheControl) WantEvict(obj zodb.IPersistent) bool {
switch obj.(type) {
default:
return true
case *btree.LOBTree:
case *btree.LOBucket:
}
return false
}
// XXX option to prevent starting if wcfs was already started ?
func main() {
log.SetPrefix("wcfs: ")
debug := flag.Bool("d", false, "debug")
autoexit := flag.Bool("autoexit", false, "automatically stop service when there is no client activity")
flag.Parse()
if len(flag.Args()) != 2 {
log.Fatalf("Usage: %s [OPTIONS] zurl mntpt", os.Args[0])
}
zurl := flag.Args()[0]
mntpt := flag.Args()[1]
// open zodb storage
ctx := context.Background() // XXX + timeout?
zstor, err := zodb.OpenStorage(ctx, zurl, &zodb.OpenOptions{ReadOnly: true})
if err != nil {
log.Fatal(err)
}
defer zstor.Close()
// mount root
opts := nodefs.NewOptions()
opts.Debug = *debug
root := nodefs.NewDefaultNode()
server, _, err := nodefs.MountRoot(mntpt, root, opts)
if err != nil {
log.Fatal(err)
}
// add entries to /
mkfile(root, ".wcfs", NewStaticFile([]byte(zurl)))
mkdir(root, "bigfile", NewBigFileRoot(zstor))
// TODO handle autoexit
_ = autoexit
// serve client requests
server.Serve() // XXX Serve returns no error
}