-
Notifications
You must be signed in to change notification settings - Fork 1.2k
/
scheduler.go
417 lines (364 loc) · 10 KB
/
scheduler.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
package solver
import (
"context"
"os"
"sync"
"github.com/moby/buildkit/solver/internal/pipe"
"github.com/moby/buildkit/util/bklog"
"github.com/moby/buildkit/util/cond"
"github.com/pkg/errors"
)
var debugScheduler = false // TODO: replace with logs in build trace
func init() {
if os.Getenv("BUILDKIT_SCHEDULER_DEBUG") == "1" {
debugScheduler = true
}
}
func newScheduler(ef edgeFactory) *scheduler {
s := &scheduler{
waitq: map[*edge]struct{}{},
incoming: map[*edge][]*edgePipe{},
outgoing: map[*edge][]*edgePipe{},
stopped: make(chan struct{}),
closed: make(chan struct{}),
ef: ef,
}
s.cond = cond.NewStatefulCond(&s.mu)
go s.loop()
return s
}
type dispatcher struct {
next *dispatcher
e *edge
}
type scheduler struct {
cond *cond.StatefulCond
mu sync.Mutex
muQ sync.Mutex
ef edgeFactory
waitq map[*edge]struct{}
next *dispatcher
last *dispatcher
stopped chan struct{}
stoppedOnce sync.Once
closed chan struct{}
incoming map[*edge][]*edgePipe
outgoing map[*edge][]*edgePipe
}
func (s *scheduler) Stop() {
s.stoppedOnce.Do(func() {
close(s.stopped)
})
<-s.closed
}
func (s *scheduler) loop() {
defer func() {
close(s.closed)
}()
go func() {
<-s.stopped
s.mu.Lock()
s.cond.Signal()
s.mu.Unlock()
}()
s.mu.Lock()
for {
select {
case <-s.stopped:
s.mu.Unlock()
return
default:
}
s.muQ.Lock()
l := s.next
if l != nil {
if l == s.last {
s.last = nil
}
s.next = l.next
delete(s.waitq, l.e)
}
s.muQ.Unlock()
if l == nil {
s.cond.Wait()
continue
}
s.dispatch(l.e)
}
}
// dispatch schedules an edge to be processed
func (s *scheduler) dispatch(e *edge) {
inc := make([]pipe.Sender, len(s.incoming[e]))
for i, p := range s.incoming[e] {
inc[i] = p.Sender
}
out := make([]pipe.Receiver, len(s.outgoing[e]))
for i, p := range s.outgoing[e] {
out[i] = p.Receiver
}
e.hasActiveOutgoing = false
updates := []pipe.Receiver{}
for _, p := range out {
if ok := p.Receive(); ok {
updates = append(updates, p)
}
if !p.Status().Completed {
e.hasActiveOutgoing = true
}
}
pf := &pipeFactory{s: s, e: e}
// unpark the edge
if debugScheduler {
debugSchedulerPreUnpark(e, inc, updates, out)
}
e.unpark(inc, updates, out, pf)
if debugScheduler {
debugSchedulerPostUnpark(e, inc)
}
postUnpark:
// set up new requests that didn't complete/were added by this run
openIncoming := make([]*edgePipe, 0, len(inc))
for _, r := range s.incoming[e] {
if !r.Sender.Status().Completed {
openIncoming = append(openIncoming, r)
}
}
if len(openIncoming) > 0 {
s.incoming[e] = openIncoming
} else {
delete(s.incoming, e)
}
openOutgoing := make([]*edgePipe, 0, len(out))
for _, r := range s.outgoing[e] {
if !r.Receiver.Status().Completed {
openOutgoing = append(openOutgoing, r)
}
}
if len(openOutgoing) > 0 {
s.outgoing[e] = openOutgoing
} else {
delete(s.outgoing, e)
}
// if keys changed there might be possiblity for merge with other edge
if e.keysDidChange {
if k := e.currentIndexKey(); k != nil {
// skip this if not at least 1 key per dep
origEdge := e.index.LoadOrStore(k, e)
if origEdge != nil {
if e.isDep(origEdge) || origEdge.isDep(e) {
bklog.G(context.TODO()).Debugf("skip merge due to dependency")
} else {
bklog.G(context.TODO()).Debugf("merging edge %s to %s\n", e.edge.Vertex.Name(), origEdge.edge.Vertex.Name())
if s.mergeTo(origEdge, e) {
s.ef.setEdge(e.edge, origEdge)
}
}
}
}
e.keysDidChange = false
}
// validation to avoid deadlocks/resource leaks:
// TODO: if these start showing up in error reports they can be changed
// to error the edge instead. They can only appear from algorithm bugs in
// unpark(), not for any external input.
if len(openIncoming) > 0 && len(openOutgoing) == 0 {
e.markFailed(pf, errors.New("buildkit scheduler error: return leaving incoming open. Please report this with BUILDKIT_SCHEDULER_DEBUG=1"))
goto postUnpark
}
if len(openIncoming) == 0 && len(openOutgoing) > 0 {
e.markFailed(pf, errors.New("buildkit scheduler error: return leaving outgoing open. Please report this with BUILDKIT_SCHEDULER_DEBUG=1"))
goto postUnpark
}
}
// signal notifies that an edge needs to be processed again
func (s *scheduler) signal(e *edge) {
s.muQ.Lock()
if _, ok := s.waitq[e]; !ok {
d := &dispatcher{e: e}
if s.last == nil {
s.next = d
} else {
s.last.next = d
}
s.last = d
s.waitq[e] = struct{}{}
s.cond.Signal()
}
s.muQ.Unlock()
}
// build evaluates edge into a result
func (s *scheduler) build(ctx context.Context, edge Edge) (CachedResult, error) {
s.mu.Lock()
e := s.ef.getEdge(edge)
if e == nil {
s.mu.Unlock()
return nil, errors.Errorf("invalid request %v for build", edge)
}
wait := make(chan struct{})
var p *pipe.Pipe
p = s.newPipe(e, nil, pipe.Request{Payload: &edgeRequest{desiredState: edgeStatusComplete}})
p.OnSendCompletion = func() {
p.Receiver.Receive()
if p.Receiver.Status().Completed {
close(wait)
}
}
s.mu.Unlock()
ctx, cancel := context.WithCancel(ctx)
defer cancel()
go func() {
<-ctx.Done()
p.Receiver.Cancel()
}()
<-wait
if err := p.Receiver.Status().Err; err != nil {
return nil, err
}
return p.Receiver.Status().Value.(*edgeState).result.CloneCachedResult(), nil
}
// newPipe creates a new request pipe between two edges
func (s *scheduler) newPipe(target, from *edge, req pipe.Request) *pipe.Pipe {
p := &edgePipe{
Pipe: pipe.New(req),
Target: target,
From: from,
}
s.signal(target)
if from != nil {
p.OnSendCompletion = func() {
p.mu.Lock()
defer p.mu.Unlock()
s.signal(p.From)
}
s.outgoing[from] = append(s.outgoing[from], p)
}
s.incoming[target] = append(s.incoming[target], p)
p.OnReceiveCompletion = func() {
p.mu.Lock()
defer p.mu.Unlock()
s.signal(p.Target)
}
return p.Pipe
}
// newRequestWithFunc creates a new request pipe that invokes a async function
func (s *scheduler) newRequestWithFunc(e *edge, f func(context.Context) (interface{}, error)) pipe.Receiver {
pp, start := pipe.NewWithFunction(f)
p := &edgePipe{
Pipe: pp,
From: e,
}
p.OnSendCompletion = func() {
p.mu.Lock()
defer p.mu.Unlock()
s.signal(p.From)
}
s.outgoing[e] = append(s.outgoing[e], p)
go start()
return p.Receiver
}
// mergeTo merges the state from one edge to another. source edge is discarded.
func (s *scheduler) mergeTo(target, src *edge) bool {
if !target.edge.Vertex.Options().IgnoreCache && src.edge.Vertex.Options().IgnoreCache {
return false
}
for _, inc := range s.incoming[src] {
inc.mu.Lock()
inc.Target = target
s.incoming[target] = append(s.incoming[target], inc)
inc.mu.Unlock()
}
for _, out := range s.outgoing[src] {
out.mu.Lock()
out.From = target
s.outgoing[target] = append(s.outgoing[target], out)
out.mu.Unlock()
out.Receiver.Cancel()
}
delete(s.incoming, src)
delete(s.outgoing, src)
s.signal(target)
for i, d := range src.deps {
for _, k := range d.keys {
target.secondaryExporters = append(target.secondaryExporters, expDep{i, CacheKeyWithSelector{CacheKey: k, Selector: src.cacheMap.Deps[i].Selector}})
}
if d.slowCacheKey != nil {
target.secondaryExporters = append(target.secondaryExporters, expDep{i, CacheKeyWithSelector{CacheKey: *d.slowCacheKey}})
}
if d.result != nil {
for _, dk := range d.result.CacheKeys() {
target.secondaryExporters = append(target.secondaryExporters, expDep{i, CacheKeyWithSelector{CacheKey: dk, Selector: src.cacheMap.Deps[i].Selector}})
}
}
}
// TODO(tonistiigi): merge cache providers
return true
}
// edgeFactory allows access to the edges from a shared graph
type edgeFactory interface {
getEdge(Edge) *edge
setEdge(Edge, *edge)
}
type pipeFactory struct {
e *edge
s *scheduler
}
func (pf *pipeFactory) NewInputRequest(ee Edge, req *edgeRequest) pipe.Receiver {
target := pf.s.ef.getEdge(ee)
if target == nil {
return pf.NewFuncRequest(func(_ context.Context) (interface{}, error) {
return nil, errors.Errorf("failed to get edge: inconsistent graph state")
})
}
p := pf.s.newPipe(target, pf.e, pipe.Request{Payload: req})
if debugScheduler {
bklog.G(context.TODO()).Debugf("> newPipe %s %p desiredState=%s", ee.Vertex.Name(), p, req.desiredState)
}
return p.Receiver
}
func (pf *pipeFactory) NewFuncRequest(f func(context.Context) (interface{}, error)) pipe.Receiver {
p := pf.s.newRequestWithFunc(pf.e, f)
if debugScheduler {
bklog.G(context.TODO()).Debugf("> newFunc %p", p)
}
return p
}
func debugSchedulerPreUnpark(e *edge, inc []pipe.Sender, updates, allPipes []pipe.Receiver) {
log := bklog.G(context.TODO())
log.Debugf(">> unpark %s req=%d upt=%d out=%d state=%s %s", e.edge.Vertex.Name(), len(inc), len(updates), len(allPipes), e.state, e.edge.Vertex.Digest())
for i, dep := range e.deps {
des := edgeStatusInitial
if dep.req != nil {
des = dep.req.Request().(*edgeRequest).desiredState
}
log.Debugf(":: dep%d %s state=%s des=%s keys=%d hasslowcache=%v preprocessfunc=%v", i, e.edge.Vertex.Inputs()[i].Vertex.Name(), dep.state, des, len(dep.keys), e.slowCacheFunc(dep) != nil, e.preprocessFunc(dep) != nil)
}
for i, in := range inc {
req := in.Request()
log.Debugf("> incoming-%d: %p dstate=%s canceled=%v", i, in, req.Payload.(*edgeRequest).desiredState, req.Canceled)
}
for i, up := range updates {
if up == e.cacheMapReq {
log.Debugf("> update-%d: %p cacheMapReq complete=%v", i, up, up.Status().Completed)
} else if up == e.execReq {
log.Debugf("> update-%d: %p execReq complete=%v", i, up, up.Status().Completed)
} else {
st, ok := up.Status().Value.(*edgeState)
if ok {
index := -1
if dep, ok := e.depRequests[up]; ok {
index = int(dep.index)
}
log.Debugf("> update-%d: %p input-%d keys=%d state=%s", i, up, index, len(st.keys), st.state)
} else {
log.Debugf("> update-%d: unknown", i)
}
}
}
}
func debugSchedulerPostUnpark(e *edge, inc []pipe.Sender) {
log := bklog.G(context.TODO())
for i, in := range inc {
log.Debugf("< incoming-%d: %p completed=%v", i, in, in.Status().Completed)
}
log.Debugf("<< unpark %s\n", e.edge.Vertex.Name())
}