summaryrefslogtreecommitdiffstats
path: root/src/runtime/mgcwork.go
blob: b3a068661ebec0caf0bd40465c63945c6b18c9cb (plain)
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
// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package runtime

import (
	"runtime/internal/atomic"
	"runtime/internal/sys"
	"unsafe"
)

const (
	_WorkbufSize = 2048 // in bytes; larger values result in less contention

	// workbufAlloc is the number of bytes to allocate at a time
	// for new workbufs. This must be a multiple of pageSize and
	// should be a multiple of _WorkbufSize.
	//
	// Larger values reduce workbuf allocation overhead. Smaller
	// values reduce heap fragmentation.
	workbufAlloc = 32 << 10
)

func init() {
	if workbufAlloc%pageSize != 0 || workbufAlloc%_WorkbufSize != 0 {
		throw("bad workbufAlloc")
	}
}

// Garbage collector work pool abstraction.
//
// This implements a producer/consumer model for pointers to grey
// objects. A grey object is one that is marked and on a work
// queue. A black object is marked and not on a work queue.
//
// Write barriers, root discovery, stack scanning, and object scanning
// produce pointers to grey objects. Scanning consumes pointers to
// grey objects, thus blackening them, and then scans them,
// potentially producing new pointers to grey objects.

// A gcWork provides the interface to produce and consume work for the
// garbage collector.
//
// A gcWork can be used on the stack as follows:
//
//     (preemption must be disabled)
//     gcw := &getg().m.p.ptr().gcw
//     .. call gcw.put() to produce and gcw.tryGet() to consume ..
//
// It's important that any use of gcWork during the mark phase prevent
// the garbage collector from transitioning to mark termination since
// gcWork may locally hold GC work buffers. This can be done by
// disabling preemption (systemstack or acquirem).
type gcWork struct {
	// wbuf1 and wbuf2 are the primary and secondary work buffers.
	//
	// This can be thought of as a stack of both work buffers'
	// pointers concatenated. When we pop the last pointer, we
	// shift the stack up by one work buffer by bringing in a new
	// full buffer and discarding an empty one. When we fill both
	// buffers, we shift the stack down by one work buffer by
	// bringing in a new empty buffer and discarding a full one.
	// This way we have one buffer's worth of hysteresis, which
	// amortizes the cost of getting or putting a work buffer over
	// at least one buffer of work and reduces contention on the
	// global work lists.
	//
	// wbuf1 is always the buffer we're currently pushing to and
	// popping from and wbuf2 is the buffer that will be discarded
	// next.
	//
	// Invariant: Both wbuf1 and wbuf2 are nil or neither are.
	wbuf1, wbuf2 *workbuf

	// Bytes marked (blackened) on this gcWork. This is aggregated
	// into work.bytesMarked by dispose.
	bytesMarked uint64

	// Scan work performed on this gcWork. This is aggregated into
	// gcController by dispose and may also be flushed by callers.
	scanWork int64

	// flushedWork indicates that a non-empty work buffer was
	// flushed to the global work list since the last gcMarkDone
	// termination check. Specifically, this indicates that this
	// gcWork may have communicated work to another gcWork.
	flushedWork bool
}

// Most of the methods of gcWork are go:nowritebarrierrec because the
// write barrier itself can invoke gcWork methods but the methods are
// not generally re-entrant. Hence, if a gcWork method invoked the
// write barrier while the gcWork was in an inconsistent state, and
// the write barrier in turn invoked a gcWork method, it could
// permanently corrupt the gcWork.

func (w *gcWork) init() {
	w.wbuf1 = getempty()
	wbuf2 := trygetfull()
	if wbuf2 == nil {
		wbuf2 = getempty()
	}
	w.wbuf2 = wbuf2
}

// put enqueues a pointer for the garbage collector to trace.
// obj must point to the beginning of a heap object or an oblet.
//go:nowritebarrierrec
func (w *gcWork) put(obj uintptr) {
	flushed := false
	wbuf := w.wbuf1
	// Record that this may acquire the wbufSpans or heap lock to
	// allocate a workbuf.
	lockWithRankMayAcquire(&work.wbufSpans.lock, lockRankWbufSpans)
	lockWithRankMayAcquire(&mheap_.lock, lockRankMheap)
	if wbuf == nil {
		w.init()
		wbuf = w.wbuf1
		// wbuf is empty at this point.
	} else if wbuf.nobj == len(wbuf.obj) {
		w.wbuf1, w.wbuf2 = w.wbuf2, w.wbuf1
		wbuf = w.wbuf1
		if wbuf.nobj == len(wbuf.obj) {
			putfull(wbuf)
			w.flushedWork = true
			wbuf = getempty()
			w.wbuf1 = wbuf
			flushed = true
		}
	}

	wbuf.obj[wbuf.nobj] = obj
	wbuf.nobj++

	// If we put a buffer on full, let the GC controller know so
	// it can encourage more workers to run. We delay this until
	// the end of put so that w is in a consistent state, since
	// enlistWorker may itself manipulate w.
	if flushed && gcphase == _GCmark {
		gcController.enlistWorker()
	}
}

// putFast does a put and reports whether it can be done quickly
// otherwise it returns false and the caller needs to call put.
//go:nowritebarrierrec
func (w *gcWork) putFast(obj uintptr) bool {
	wbuf := w.wbuf1
	if wbuf == nil {
		return false
	} else if wbuf.nobj == len(wbuf.obj) {
		return false
	}

	wbuf.obj[wbuf.nobj] = obj
	wbuf.nobj++
	return true
}

// putBatch performs a put on every pointer in obj. See put for
// constraints on these pointers.
//
//go:nowritebarrierrec
func (w *gcWork) putBatch(obj []uintptr) {
	if len(obj) == 0 {
		return
	}

	flushed := false
	wbuf := w.wbuf1
	if wbuf == nil {
		w.init()
		wbuf = w.wbuf1
	}

	for len(obj) > 0 {
		for wbuf.nobj == len(wbuf.obj) {
			putfull(wbuf)
			w.flushedWork = true
			w.wbuf1, w.wbuf2 = w.wbuf2, getempty()
			wbuf = w.wbuf1
			flushed = true
		}
		n := copy(wbuf.obj[wbuf.nobj:], obj)
		wbuf.nobj += n
		obj = obj[n:]
	}

	if flushed && gcphase == _GCmark {
		gcController.enlistWorker()
	}
}

// tryGet dequeues a pointer for the garbage collector to trace.
//
// If there are no pointers remaining in this gcWork or in the global
// queue, tryGet returns 0.  Note that there may still be pointers in
// other gcWork instances or other caches.
//go:nowritebarrierrec
func (w *gcWork) tryGet() uintptr {
	wbuf := w.wbuf1
	if wbuf == nil {
		w.init()
		wbuf = w.wbuf1
		// wbuf is empty at this point.
	}
	if wbuf.nobj == 0 {
		w.wbuf1, w.wbuf2 = w.wbuf2, w.wbuf1
		wbuf = w.wbuf1
		if wbuf.nobj == 0 {
			owbuf := wbuf
			wbuf = trygetfull()
			if wbuf == nil {
				return 0
			}
			putempty(owbuf)
			w.wbuf1 = wbuf
		}
	}

	wbuf.nobj--
	return wbuf.obj[wbuf.nobj]
}

// tryGetFast dequeues a pointer for the garbage collector to trace
// if one is readily available. Otherwise it returns 0 and
// the caller is expected to call tryGet().
//go:nowritebarrierrec
func (w *gcWork) tryGetFast() uintptr {
	wbuf := w.wbuf1
	if wbuf == nil {
		return 0
	}
	if wbuf.nobj == 0 {
		return 0
	}

	wbuf.nobj--
	return wbuf.obj[wbuf.nobj]
}

// dispose returns any cached pointers to the global queue.
// The buffers are being put on the full queue so that the
// write barriers will not simply reacquire them before the
// GC can inspect them. This helps reduce the mutator's
// ability to hide pointers during the concurrent mark phase.
//
//go:nowritebarrierrec
func (w *gcWork) dispose() {
	if wbuf := w.wbuf1; wbuf != nil {
		if wbuf.nobj == 0 {
			putempty(wbuf)
		} else {
			putfull(wbuf)
			w.flushedWork = true
		}
		w.wbuf1 = nil

		wbuf = w.wbuf2
		if wbuf.nobj == 0 {
			putempty(wbuf)
		} else {
			putfull(wbuf)
			w.flushedWork = true
		}
		w.wbuf2 = nil
	}
	if w.bytesMarked != 0 {
		// dispose happens relatively infrequently. If this
		// atomic becomes a problem, we should first try to
		// dispose less and if necessary aggregate in a per-P
		// counter.
		atomic.Xadd64(&work.bytesMarked, int64(w.bytesMarked))
		w.bytesMarked = 0
	}
	if w.scanWork != 0 {
		atomic.Xaddint64(&gcController.scanWork, w.scanWork)
		w.scanWork = 0
	}
}

// balance moves some work that's cached in this gcWork back on the
// global queue.
//go:nowritebarrierrec
func (w *gcWork) balance() {
	if w.wbuf1 == nil {
		return
	}
	if wbuf := w.wbuf2; wbuf.nobj != 0 {
		putfull(wbuf)
		w.flushedWork = true
		w.wbuf2 = getempty()
	} else if wbuf := w.wbuf1; wbuf.nobj > 4 {
		w.wbuf1 = handoff(wbuf)
		w.flushedWork = true // handoff did putfull
	} else {
		return
	}
	// We flushed a buffer to the full list, so wake a worker.
	if gcphase == _GCmark {
		gcController.enlistWorker()
	}
}

// empty reports whether w has no mark work available.
//go:nowritebarrierrec
func (w *gcWork) empty() bool {
	return w.wbuf1 == nil || (w.wbuf1.nobj == 0 && w.wbuf2.nobj == 0)
}

// Internally, the GC work pool is kept in arrays in work buffers.
// The gcWork interface caches a work buffer until full (or empty) to
// avoid contending on the global work buffer lists.

type workbufhdr struct {
	node lfnode // must be first
	nobj int
}

//go:notinheap
type workbuf struct {
	workbufhdr
	// account for the above fields
	obj [(_WorkbufSize - unsafe.Sizeof(workbufhdr{})) / sys.PtrSize]uintptr
}

// workbuf factory routines. These funcs are used to manage the
// workbufs.
// If the GC asks for some work these are the only routines that
// make wbufs available to the GC.

func (b *workbuf) checknonempty() {
	if b.nobj == 0 {
		throw("workbuf is empty")
	}
}

func (b *workbuf) checkempty() {
	if b.nobj != 0 {
		throw("workbuf is not empty")
	}
}

// getempty pops an empty work buffer off the work.empty list,
// allocating new buffers if none are available.
//go:nowritebarrier
func getempty() *workbuf {
	var b *workbuf
	if work.empty != 0 {
		b = (*workbuf)(work.empty.pop())
		if b != nil {
			b.checkempty()
		}
	}
	// Record that this may acquire the wbufSpans or heap lock to
	// allocate a workbuf.
	lockWithRankMayAcquire(&work.wbufSpans.lock, lockRankWbufSpans)
	lockWithRankMayAcquire(&mheap_.lock, lockRankMheap)
	if b == nil {
		// Allocate more workbufs.
		var s *mspan
		if work.wbufSpans.free.first != nil {
			lock(&work.wbufSpans.lock)
			s = work.wbufSpans.free.first
			if s != nil {
				work.wbufSpans.free.remove(s)
				work.wbufSpans.busy.insert(s)
			}
			unlock(&work.wbufSpans.lock)
		}
		if s == nil {
			systemstack(func() {
				s = mheap_.allocManual(workbufAlloc/pageSize, spanAllocWorkBuf)
			})
			if s == nil {
				throw("out of memory")
			}
			// Record the new span in the busy list.
			lock(&work.wbufSpans.lock)
			work.wbufSpans.busy.insert(s)
			unlock(&work.wbufSpans.lock)
		}
		// Slice up the span into new workbufs. Return one and
		// put the rest on the empty list.
		for i := uintptr(0); i+_WorkbufSize <= workbufAlloc; i += _WorkbufSize {
			newb := (*workbuf)(unsafe.Pointer(s.base() + i))
			newb.nobj = 0
			lfnodeValidate(&newb.node)
			if i == 0 {
				b = newb
			} else {
				putempty(newb)
			}
		}
	}
	return b
}

// putempty puts a workbuf onto the work.empty list.
// Upon entry this go routine owns b. The lfstack.push relinquishes ownership.
//go:nowritebarrier
func putempty(b *workbuf) {
	b.checkempty()
	work.empty.push(&b.node)
}

// putfull puts the workbuf on the work.full list for the GC.
// putfull accepts partially full buffers so the GC can avoid competing
// with the mutators for ownership of partially full buffers.
//go:nowritebarrier
func putfull(b *workbuf) {
	b.checknonempty()
	work.full.push(&b.node)
}

// trygetfull tries to get a full or partially empty workbuffer.
// If one is not immediately available return nil
//go:nowritebarrier
func trygetfull() *workbuf {
	b := (*workbuf)(work.full.pop())
	if b != nil {
		b.checknonempty()
		return b
	}
	return b
}

//go:nowritebarrier
func handoff(b *workbuf) *workbuf {
	// Make new buffer with half of b's pointers.
	b1 := getempty()
	n := b.nobj / 2
	b.nobj -= n
	b1.nobj = n
	memmove(unsafe.Pointer(&b1.obj[0]), unsafe.Pointer(&b.obj[b.nobj]), uintptr(n)*unsafe.Sizeof(b1.obj[0]))

	// Put b on full list - let first half of b get stolen.
	putfull(b)
	return b1
}

// prepareFreeWorkbufs moves busy workbuf spans to free list so they
// can be freed to the heap. This must only be called when all
// workbufs are on the empty list.
func prepareFreeWorkbufs() {
	lock(&work.wbufSpans.lock)
	if work.full != 0 {
		throw("cannot free workbufs when work.full != 0")
	}
	// Since all workbufs are on the empty list, we don't care
	// which ones are in which spans. We can wipe the entire empty
	// list and move all workbuf spans to the free list.
	work.empty = 0
	work.wbufSpans.free.takeAll(&work.wbufSpans.busy)
	unlock(&work.wbufSpans.lock)
}

// freeSomeWbufs frees some workbufs back to the heap and returns
// true if it should be called again to free more.
func freeSomeWbufs(preemptible bool) bool {
	const batchSize = 64 // ~1–2 µs per span.
	lock(&work.wbufSpans.lock)
	if gcphase != _GCoff || work.wbufSpans.free.isEmpty() {
		unlock(&work.wbufSpans.lock)
		return false
	}
	systemstack(func() {
		gp := getg().m.curg
		for i := 0; i < batchSize && !(preemptible && gp.preempt); i++ {
			span := work.wbufSpans.free.first
			if span == nil {
				break
			}
			work.wbufSpans.free.remove(span)
			mheap_.freeManual(span, spanAllocWorkBuf)
		}
	})
	more := !work.wbufSpans.free.isEmpty()
	unlock(&work.wbufSpans.lock)
	return more
}