1// Package dag contains the base common code to define an entity stored
2// in a chain of git objects, supporting actions like Push, Pull and Merge.
3package dag
4
5import (
6 "encoding/json"
7 "fmt"
8 "sort"
9
10 "github.com/pkg/errors"
11
12 "github.com/MichaelMure/git-bug/entity"
13 "github.com/MichaelMure/git-bug/identity"
14 "github.com/MichaelMure/git-bug/repository"
15 "github.com/MichaelMure/git-bug/util/lamport"
16)
17
18const refsPattern = "refs/%s/%s"
19const creationClockPattern = "%s-create"
20const editClockPattern = "%s-edit"
21
22// Definition hold the details defining one specialization of an Entity.
23type Definition struct {
24 // the name of the entity (bug, pull-request, ...)
25 Typename string
26 // the Namespace in git (bugs, prs, ...)
27 Namespace string
28 // a function decoding a JSON message into an Operation
29 OperationUnmarshaler func(author identity.Interface, raw json.RawMessage) (Operation, error)
30 // the expected format version number, that can be used for data migration/upgrade
31 FormatVersion uint
32}
33
34// Entity is a data structure stored in a chain of git objects, supporting actions like Push, Pull and Merge.
35type Entity struct {
36 // A Lamport clock is a logical clock that allow to order event
37 // inside a distributed system.
38 // It must be the first field in this struct due to https://github.com/golang/go/issues/36606
39 createTime lamport.Time
40 editTime lamport.Time
41
42 Definition
43
44 // operations that are already stored in the repository
45 ops []Operation
46 // operations not yet stored in the repository
47 staging []Operation
48
49 lastCommit repository.Hash
50}
51
52// New create an empty Entity
53func New(definition Definition) *Entity {
54 return &Entity{
55 Definition: definition,
56 }
57}
58
59// Read will read and decode a stored local Entity from a repository
60func Read(def Definition, repo repository.ClockedRepo, resolver identity.Resolver, id entity.Id) (*Entity, error) {
61 if err := id.Validate(); err != nil {
62 return nil, errors.Wrap(err, "invalid id")
63 }
64
65 ref := fmt.Sprintf("refs/%s/%s", def.Namespace, id.String())
66
67 return read(def, repo, resolver, ref)
68}
69
70// readRemote will read and decode a stored remote Entity from a repository
71func readRemote(def Definition, repo repository.ClockedRepo, resolver identity.Resolver, remote string, id entity.Id) (*Entity, error) {
72 if err := id.Validate(); err != nil {
73 return nil, errors.Wrap(err, "invalid id")
74 }
75
76 ref := fmt.Sprintf("refs/remotes/%s/%s/%s", def.Namespace, remote, id.String())
77
78 return read(def, repo, resolver, ref)
79}
80
81// read fetch from git and decode an Entity at an arbitrary git reference.
82func read(def Definition, repo repository.ClockedRepo, resolver identity.Resolver, ref string) (*Entity, error) {
83 rootHash, err := repo.ResolveRef(ref)
84 if err != nil {
85 return nil, err
86 }
87
88 // Perform a breadth-first search to get a topological order of the DAG where we discover the
89 // parents commit and go back in time up to the chronological root
90
91 queue := make([]repository.Hash, 0, 32)
92 visited := make(map[repository.Hash]struct{})
93 BFSOrder := make([]repository.Commit, 0, 32)
94
95 queue = append(queue, rootHash)
96 visited[rootHash] = struct{}{}
97
98 for len(queue) > 0 {
99 // pop
100 hash := queue[0]
101 queue = queue[1:]
102
103 commit, err := repo.ReadCommit(hash)
104 if err != nil {
105 return nil, err
106 }
107
108 BFSOrder = append(BFSOrder, commit)
109
110 for _, parent := range commit.Parents {
111 if _, ok := visited[parent]; !ok {
112 queue = append(queue, parent)
113 // mark as visited
114 visited[parent] = struct{}{}
115 }
116 }
117 }
118
119 // Now, we can reverse this topological order and read the commits in an order where
120 // we are sure to have read all the chronological ancestors when we read a commit.
121
122 // Next step is to:
123 // 1) read the operationPacks
124 // 2) make sure that the clocks causality respect the DAG topology.
125
126 oppMap := make(map[repository.Hash]*operationPack)
127 var opsCount int
128
129 for i := len(BFSOrder) - 1; i >= 0; i-- {
130 commit := BFSOrder[i]
131 isFirstCommit := i == len(BFSOrder)-1
132 isMerge := len(commit.Parents) > 1
133
134 // Verify DAG structure: single chronological root, so only the root
135 // can have no parents. Said otherwise, the DAG need to have exactly
136 // one leaf.
137 if !isFirstCommit && len(commit.Parents) == 0 {
138 return nil, fmt.Errorf("multiple leafs in the entity DAG")
139 }
140
141 opp, err := readOperationPack(def, repo, resolver, commit)
142 if err != nil {
143 return nil, err
144 }
145
146 err = opp.Validate()
147 if err != nil {
148 return nil, err
149 }
150
151 // Check that the create lamport clock is set (not checked in Validate() as it's optional)
152 if isFirstCommit && opp.CreateTime <= 0 {
153 return nil, fmt.Errorf("creation lamport time not set")
154 }
155
156 // make sure that the lamport clocks causality match the DAG topology
157 for _, parentHash := range commit.Parents {
158 parentPack, ok := oppMap[parentHash]
159 if !ok {
160 panic("DFS failed")
161 }
162
163 if parentPack.EditTime >= opp.EditTime {
164 return nil, fmt.Errorf("lamport clock ordering doesn't match the DAG")
165 }
166
167 // to avoid an attack where clocks are pushed toward the uint64 rollover, make sure
168 // that the clocks don't jump too far in the future
169 // we ignore merge commits here to allow merging after a loooong time without breaking anything,
170 // as long as there is one valid chain of small hops, it's fine.
171 if !isMerge && opp.EditTime-parentPack.EditTime > 1_000_000 {
172 return nil, fmt.Errorf("lamport clock jumping too far in the future, likely an attack")
173 }
174 }
175
176 oppMap[commit.Hash] = opp
177 opsCount += len(opp.Operations)
178 }
179
180 // The clocks are fine, we witness them
181 for _, opp := range oppMap {
182 err = repo.Witness(fmt.Sprintf(creationClockPattern, def.Namespace), opp.CreateTime)
183 if err != nil {
184 return nil, err
185 }
186 err = repo.Witness(fmt.Sprintf(editClockPattern, def.Namespace), opp.EditTime)
187 if err != nil {
188 return nil, err
189 }
190 }
191
192 // Now that we know that the topological order and clocks are fine, we order the operationPacks
193 // based on the logical clocks, entirely ignoring the DAG topology
194
195 oppSlice := make([]*operationPack, 0, len(oppMap))
196 for _, pack := range oppMap {
197 oppSlice = append(oppSlice, pack)
198 }
199 sort.Slice(oppSlice, func(i, j int) bool {
200 // Primary ordering with the EditTime.
201 if oppSlice[i].EditTime != oppSlice[j].EditTime {
202 return oppSlice[i].EditTime < oppSlice[j].EditTime
203 }
204 // We have equal EditTime, which means we have concurrent edition over different machines and we
205 // can't tell which one came first. So, what now? We still need a total ordering and the most stable possible.
206 // As a secondary ordering, we can order based on a hash of the serialized Operations in the
207 // operationPack. It doesn't carry much meaning but it's unbiased and hard to abuse.
208 // This is a lexicographic ordering on the stringified ID.
209 return oppSlice[i].Id() < oppSlice[j].Id()
210 })
211
212 // Now that we ordered the operationPacks, we have the order of the Operations
213
214 ops := make([]Operation, 0, opsCount)
215 var createTime lamport.Time
216 var editTime lamport.Time
217 for _, pack := range oppSlice {
218 for _, operation := range pack.Operations {
219 ops = append(ops, operation)
220 }
221 if pack.CreateTime > createTime {
222 createTime = pack.CreateTime
223 }
224 if pack.EditTime > editTime {
225 editTime = pack.EditTime
226 }
227 }
228
229 return &Entity{
230 Definition: def,
231 ops: ops,
232 lastCommit: rootHash,
233 createTime: createTime,
234 editTime: editTime,
235 }, nil
236}
237
238type StreamedEntity struct {
239 Entity *Entity
240 Err error
241}
242
243// ReadAll read and parse all local Entity
244func ReadAll(def Definition, repo repository.ClockedRepo, resolver identity.Resolver) <-chan StreamedEntity {
245 out := make(chan StreamedEntity)
246
247 go func() {
248 defer close(out)
249
250 refPrefix := fmt.Sprintf("refs/%s/", def.Namespace)
251
252 refs, err := repo.ListRefs(refPrefix)
253 if err != nil {
254 out <- StreamedEntity{Err: err}
255 return
256 }
257
258 for _, ref := range refs {
259 e, err := read(def, repo, resolver, ref)
260
261 if err != nil {
262 out <- StreamedEntity{Err: err}
263 return
264 }
265
266 out <- StreamedEntity{Entity: e}
267 }
268 }()
269
270 return out
271}
272
273// Id return the Entity identifier
274func (e *Entity) Id() entity.Id {
275 // id is the id of the first operation
276 return e.FirstOp().Id()
277}
278
279// Validate check if the Entity data is valid
280func (e *Entity) Validate() error {
281 // non-empty
282 if len(e.ops) == 0 && len(e.staging) == 0 {
283 return fmt.Errorf("entity has no operations")
284 }
285
286 // check if each operations are valid
287 for _, op := range e.ops {
288 if err := op.Validate(); err != nil {
289 return err
290 }
291 }
292
293 // check if staging is valid if needed
294 for _, op := range e.staging {
295 if err := op.Validate(); err != nil {
296 return err
297 }
298 }
299
300 // Check that there is no colliding operation's ID
301 ids := make(map[entity.Id]struct{})
302 for _, op := range e.Operations() {
303 if _, ok := ids[op.Id()]; ok {
304 return fmt.Errorf("id collision: %s", op.Id())
305 }
306 ids[op.Id()] = struct{}{}
307 }
308
309 return nil
310}
311
312// Operations return the ordered operations
313func (e *Entity) Operations() []Operation {
314 return append(e.ops, e.staging...)
315}
316
317// FirstOp lookup for the very first operation of the Entity
318func (e *Entity) FirstOp() Operation {
319 for _, op := range e.ops {
320 return op
321 }
322 for _, op := range e.staging {
323 return op
324 }
325 return nil
326}
327
328// LastOp lookup for the very last operation of the Entity
329func (e *Entity) LastOp() Operation {
330 if len(e.staging) > 0 {
331 return e.staging[len(e.staging)-1]
332 }
333 if len(e.ops) > 0 {
334 return e.ops[len(e.ops)-1]
335 }
336 return nil
337}
338
339// Append add a new Operation to the Entity
340func (e *Entity) Append(op Operation) {
341 e.staging = append(e.staging, op)
342}
343
344// NeedCommit indicate if the in-memory state changed and need to be commit in the repository
345func (e *Entity) NeedCommit() bool {
346 return len(e.staging) > 0
347}
348
349// CommitAsNeeded execute a Commit only if necessary. This function is useful to avoid getting an error if the Entity
350// is already in sync with the repository.
351func (e *Entity) CommitAsNeeded(repo repository.ClockedRepo) error {
352 if e.NeedCommit() {
353 return e.Commit(repo)
354 }
355 return nil
356}
357
358// Commit write the appended operations in the repository
359func (e *Entity) Commit(repo repository.ClockedRepo) error {
360 if !e.NeedCommit() {
361 return fmt.Errorf("can't commit an entity with no pending operation")
362 }
363
364 err := e.Validate()
365 if err != nil {
366 return errors.Wrapf(err, "can't commit a %s with invalid data", e.Definition.Typename)
367 }
368
369 for len(e.staging) > 0 {
370 var author identity.Interface
371 var toCommit []Operation
372
373 // Split into chunks with the same author
374 for len(e.staging) > 0 {
375 op := e.staging[0]
376 if author != nil && op.Author().Id() != author.Id() {
377 break
378 }
379 author = e.staging[0].Author()
380 toCommit = append(toCommit, op)
381 e.staging = e.staging[1:]
382 }
383
384 e.editTime, err = repo.Increment(fmt.Sprintf(editClockPattern, e.Namespace))
385 if err != nil {
386 return err
387 }
388
389 opp := &operationPack{
390 Author: author,
391 Operations: toCommit,
392 EditTime: e.editTime,
393 }
394
395 if e.lastCommit == "" {
396 e.createTime, err = repo.Increment(fmt.Sprintf(creationClockPattern, e.Namespace))
397 if err != nil {
398 return err
399 }
400 opp.CreateTime = e.createTime
401 }
402
403 var parentCommit []repository.Hash
404 if e.lastCommit != "" {
405 parentCommit = []repository.Hash{e.lastCommit}
406 }
407
408 commitHash, err := opp.Write(e.Definition, repo, parentCommit...)
409 if err != nil {
410 return err
411 }
412
413 e.lastCommit = commitHash
414 e.ops = append(e.ops, toCommit...)
415 }
416
417 // not strictly necessary but make equality testing easier in tests
418 e.staging = nil
419
420 // Create or update the Git reference for this entity
421 // When pushing later, the remote will ensure that this ref update
422 // is fast-forward, that is no data has been overwritten.
423 ref := fmt.Sprintf(refsPattern, e.Namespace, e.Id().String())
424 return repo.UpdateRef(ref, e.lastCommit)
425}
426
427// CreateLamportTime return the Lamport time of creation
428func (e *Entity) CreateLamportTime() lamport.Time {
429 return e.createTime
430}
431
432// EditLamportTime return the Lamport time of the last edition
433func (e *Entity) EditLamportTime() lamport.Time {
434 return e.editTime
435}