1 // Copyright 2009 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
5 // Garbage collector: sweeping
7 // The sweeper consists of two different algorithms:
9 // * The object reclaimer finds and frees unmarked slots in spans. It
10 // can free a whole span if none of the objects are marked, but that
11 // isn't its goal. This can be driven either synchronously by
12 // mcentral.cacheSpan for mcentral spans, or asynchronously by
13 // sweepone, which looks at all the mcentral lists.
15 // * The span reclaimer looks for spans that contain no marked objects
16 // and frees whole spans. This is a separate algorithm because
17 // freeing whole spans is the hardest task for the object reclaimer,
18 // but is critical when allocating new spans. The entry point for
19 // this is mheap_.reclaim and it's driven by a sequential scan of
20 // the page marks bitmap in the heap arenas.
22 // Both algorithms ultimately call mspan.sweep, which sweeps a single
28 "runtime/internal/atomic"
34 // State of background sweep.
35 type sweepdata struct {
44 // centralIndex is the current unswept span class.
45 // It represents an index into the mcentral span
46 // sets. Accessed and updated via its load and
47 // update methods. Not protected by a lock.
49 // Reset at mark termination.
50 // Used by mheap.nextSpanForSweep.
51 centralIndex sweepClass
54 // sweepClass is a spanClass and one bit to represent whether we're currently
55 // sweeping partial or full spans.
56 type sweepClass uint32
59 numSweepClasses = numSpanClasses * 2
60 sweepClassDone sweepClass = sweepClass(^uint32(0))
63 func (s *sweepClass) load() sweepClass {
64 return sweepClass(atomic.Load((*uint32)(s)))
67 func (s *sweepClass) update(sNew sweepClass) {
68 // Only update *s if its current value is less than sNew,
69 // since *s increases monotonically.
71 for sOld < sNew && !atomic.Cas((*uint32)(s), uint32(sOld), uint32(sNew)) {
74 // TODO(mknyszek): This isn't the only place we have
75 // an atomic monotonically increasing counter. It would
76 // be nice to have an "atomic max" which is just implemented
77 // as the above on most architectures. Some architectures
78 // like RISC-V however have native support for an atomic max.
81 func (s *sweepClass) clear() {
82 atomic.Store((*uint32)(s), 0)
85 // split returns the underlying span class as well as
86 // whether we're interested in the full or partial
87 // unswept lists for that class, indicated as a boolean
88 // (true means "full").
89 func (s sweepClass) split() (spc spanClass, full bool) {
90 return spanClass(s >> 1), s&1 == 0
93 // nextSpanForSweep finds and pops the next span for sweeping from the
94 // central sweep buffers. It returns ownership of the span to the caller.
95 // Returns nil if no such span exists.
96 func (h *mheap) nextSpanForSweep() *mspan {
98 for sc := sweep.centralIndex.load(); sc < numSweepClasses; sc++ {
99 spc, full := sc.split()
100 c := &h.central[spc].mcentral
103 s = c.fullUnswept(sg).pop()
105 s = c.partialUnswept(sg).pop()
108 // Write down that we found something so future sweepers
109 // can start from here.
110 sweep.centralIndex.update(sc)
114 // Write down that we found nothing.
115 sweep.centralIndex.update(sweepClassDone)
119 // finishsweep_m ensures that all spans are swept.
121 // The world must be stopped. This ensures there are no sweeps in
125 func finishsweep_m() {
128 // Sweeping must be complete before marking commences, so
129 // sweep any unswept spans. If this is a concurrent GC, there
130 // shouldn't be any spans left to sweep, so this should finish
131 // instantly. If GC was forced before the concurrent sweep
132 // finished, there may be spans to sweep.
133 for sweepone() != ^uintptr(0) {
137 // Reset all the unswept buffers, which should be empty.
138 // Do this in sweep termination as opposed to mark termination
139 // so that we can catch unswept spans and reclaim blocks as
141 sg := mheap_.sweepgen
142 for i := range mheap_.central {
143 c := &mheap_.central[i].mcentral
144 c.partialUnswept(sg).reset()
145 c.fullUnswept(sg).reset()
148 // Sweeping is done, so if the scavenger isn't already awake,
149 // wake it up. There's definitely work for it to do at this
153 nextMarkBitArenaEpoch()
156 func bgsweep(c chan int) {
159 lockInit(&sweep.lock, lockRankSweep)
163 goparkunlock(&sweep.lock, waitReasonGCSweepWait, traceEvGoBlock, 1)
166 for sweepone() != ^uintptr(0) {
170 for freeSomeWbufs(true) {
175 // This can happen if a GC runs between
176 // gosweepone returning ^0 above
177 // and the lock being acquired.
182 goparkunlock(&sweep.lock, waitReasonGCSweepWait, traceEvGoBlock, 1)
186 // sweepone sweeps some unswept heap span and returns the number of pages returned
187 // to the heap, or ^uintptr(0) if there was nothing to sweep.
188 func sweepone() uintptr {
190 sweepRatio := mheap_.sweepPagesPerByte // For debugging
192 // increment locks to ensure that the goroutine is not preempted
193 // in the middle of sweep thus leaving the span in an inconsistent state for next GC
195 if atomic.Load(&mheap_.sweepdone) != 0 {
199 atomic.Xadd(&mheap_.sweepers, +1)
201 // Find a span to sweep.
203 sg := mheap_.sweepgen
205 s = mheap_.nextSpanForSweep()
207 atomic.Store(&mheap_.sweepdone, 1)
210 if state := s.state.get(); state != mSpanInUse {
211 // This can happen if direct sweeping already
212 // swept this span, but in that case the sweep
213 // generation should always be up-to-date.
214 if !(s.sweepgen == sg || s.sweepgen == sg+3) {
215 print("runtime: bad span s.state=", state, " s.sweepgen=", s.sweepgen, " sweepgen=", sg, "\n")
216 throw("non in-use span in unswept list")
220 if s.sweepgen == sg-2 && atomic.Cas(&s.sweepgen, sg-2, sg-1) {
225 // Sweep the span we found.
226 npages := ^uintptr(0)
230 // Whole span was freed. Count it toward the
231 // page reclaimer credit since these pages can
232 // now be used for span allocation.
233 atomic.Xadduintptr(&mheap_.reclaimCredit, npages)
235 // Span is still in-use, so this returned no
236 // pages to the heap and the span needs to
237 // move to the swept in-use list.
242 // Decrement the number of active sweepers and if this is the
243 // last one print trace information.
244 if atomic.Xadd(&mheap_.sweepers, -1) == 0 && atomic.Load(&mheap_.sweepdone) != 0 {
245 // Since the sweeper is done, move the scavenge gen forward (signalling
246 // that there's new work to do) and wake the scavenger.
248 // The scavenger is signaled by the last sweeper because once
249 // sweeping is done, we will definitely have useful work for
250 // the scavenger to do, since the scavenger only runs over the
251 // heap once per GC cyle. This update is not done during sweep
252 // termination because in some cases there may be a long delay
253 // between sweep done and sweep termination (e.g. not enough
254 // allocations to trigger a GC) which would be nice to fill in
255 // with scavenging work.
258 mheap_.pages.scavengeStartGen()
261 // Since we might sweep in an allocation path, it's not possible
262 // for us to wake the scavenger directly via wakeScavenger, since
263 // it could allocate. Ask sysmon to do it for us instead.
266 if debug.gcpacertrace > 0 {
267 print("pacer: sweep done at heap size ", memstats.heap_live>>20, "MB; allocated ", (memstats.heap_live-mheap_.sweepHeapLiveBasis)>>20, "MB during sweep; swept ", mheap_.pagesSwept, " pages at ", sweepRatio, " pages/byte\n")
274 // isSweepDone reports whether all spans are swept or currently being swept.
276 // Note that this condition may transition from false to true at any
277 // time as the sweeper runs. It may transition from true to false if a
278 // GC runs; to prevent that the caller must be non-preemptible or must
279 // somehow block GC progress.
280 func isSweepDone() bool {
281 return mheap_.sweepdone != 0
284 // Returns only when span s has been swept.
286 func (s *mspan) ensureSwept() {
287 // Caller must disable preemption.
288 // Otherwise when this function returns the span can become unswept again
289 // (if GC is triggered on another goroutine).
291 if _g_.m.locks == 0 && _g_.m.mallocing == 0 && _g_ != _g_.m.g0 {
292 throw("mspan.ensureSwept: m is not locked")
295 sg := mheap_.sweepgen
296 spangen := atomic.Load(&s.sweepgen)
297 if spangen == sg || spangen == sg+3 {
300 // The caller must be sure that the span is a mSpanInUse span.
301 if atomic.Cas(&s.sweepgen, sg-2, sg-1) {
305 // unfortunate condition, and we don't have efficient means to wait
307 spangen := atomic.Load(&s.sweepgen)
308 if spangen == sg || spangen == sg+3 {
315 // Sweep frees or collects finalizers for blocks not marked in the mark phase.
316 // It clears the mark bits in preparation for the next GC round.
317 // Returns true if the span was returned to heap.
318 // If preserve=true, don't return it to heap nor relink in mcentral lists;
319 // caller takes care of it.
320 func (s *mspan) sweep(preserve bool) bool {
321 // It's critical that we enter this function with preemption disabled,
322 // GC must not start while we are in the middle of this function.
324 if _g_.m.locks == 0 && _g_.m.mallocing == 0 && _g_ != _g_.m.g0 {
325 throw("mspan.sweep: m is not locked")
327 sweepgen := mheap_.sweepgen
328 if state := s.state.get(); state != mSpanInUse || s.sweepgen != sweepgen-1 {
329 print("mspan.sweep: state=", state, " sweepgen=", s.sweepgen, " mheap.sweepgen=", sweepgen, "\n")
330 throw("mspan.sweep: bad span state")
334 traceGCSweepSpan(s.npages * _PageSize)
337 atomic.Xadd64(&mheap_.pagesSwept, int64(s.npages))
342 // The allocBits indicate which unmarked objects don't need to be
343 // processed since they were free at the end of the last GC cycle
344 // and were not allocated since then.
345 // If the allocBits index is >= s.freeindex and the bit
346 // is not marked then the object remains unallocated
347 // since the last GC.
348 // This situation is analogous to being on a freelist.
350 // Unlink & free special records for any objects we're about to free.
351 // Two complications here:
352 // 1. An object can have both finalizer and profile special records.
353 // In such case we need to queue finalizer for execution,
354 // mark the object as live and preserve the profile special.
355 // 2. A tiny object can have several finalizers setup for different offsets.
356 // If such object is not marked, we need to queue all finalizers at once.
357 // Both 1 and 2 are possible at the same time.
358 hadSpecials := s.specials != nil
359 specialp := &s.specials
362 // A finalizer can be set for an inner byte of an object, find object beginning.
363 objIndex := uintptr(special.offset) / size
364 p := s.base() + objIndex*size
365 mbits := s.markBitsForIndex(objIndex)
366 if !mbits.isMarked() {
367 // This object is not marked and has at least one special record.
368 // Pass 1: see if it has at least one finalizer.
370 endOffset := p - s.base() + size
371 for tmp := special; tmp != nil && uintptr(tmp.offset) < endOffset; tmp = tmp.next {
372 if tmp.kind == _KindSpecialFinalizer {
373 // Stop freeing of object if it has a finalizer.
374 mbits.setMarkedNonAtomic()
379 // Pass 2: queue all finalizers _or_ handle profile record.
380 for special != nil && uintptr(special.offset) < endOffset {
381 // Find the exact byte for which the special was setup
382 // (as opposed to object beginning).
383 p := s.base() + uintptr(special.offset)
384 if special.kind == _KindSpecialFinalizer || !hasFin {
385 // Splice out special record.
387 special = special.next
389 freespecial(y, unsafe.Pointer(p), size)
391 // This is profile record, but the object has finalizers (so kept alive).
392 // Keep special record.
393 specialp = &special.next
398 // object is still live: keep special record
399 specialp = &special.next
403 if hadSpecials && s.specials == nil {
407 if debug.allocfreetrace != 0 || debug.clobberfree != 0 || raceenabled || msanenabled {
408 // Find all newly freed objects. This doesn't have to
409 // efficient; allocfreetrace has massive overhead.
410 mbits := s.markBitsForBase()
411 abits := s.allocBitsForIndex(0)
412 for i := uintptr(0); i < s.nelems; i++ {
413 if !mbits.isMarked() && (abits.index < s.freeindex || abits.isMarked()) {
414 x := s.base() + i*s.elemsize
415 if debug.allocfreetrace != 0 {
416 tracefree(unsafe.Pointer(x), size)
418 if debug.clobberfree != 0 {
419 clobberfree(unsafe.Pointer(x), size)
422 racefree(unsafe.Pointer(x), size)
425 msanfree(unsafe.Pointer(x), size)
433 // Check for zombie objects.
434 if s.freeindex < s.nelems {
435 // Everything < freeindex is allocated and hence
436 // cannot be zombies.
438 // Check the first bitmap byte, where we have to be
439 // careful with freeindex.
441 if (*s.gcmarkBits.bytep(obj / 8)&^*s.allocBits.bytep(obj / 8))>>(obj%8) != 0 {
444 // Check remaining bytes.
445 for i := obj/8 + 1; i < divRoundUp(s.nelems, 8); i++ {
446 if *s.gcmarkBits.bytep(i)&^*s.allocBits.bytep(i) != 0 {
452 // Count the number of free objects in this span.
453 nalloc := uint16(s.countAlloc())
454 nfreed := s.allocCount - nalloc
455 if nalloc > s.allocCount {
456 // The zombie check above should have caught this in
458 print("runtime: nelems=", s.nelems, " nalloc=", nalloc, " previous allocCount=", s.allocCount, " nfreed=", nfreed, "\n")
459 throw("sweep increased allocation count")
462 s.allocCount = nalloc
463 s.freeindex = 0 // reset allocation index to start of span.
465 getg().m.p.ptr().traceReclaimed += uintptr(nfreed) * s.elemsize
468 // gcmarkBits becomes the allocBits.
469 // get a fresh cleared gcmarkBits in preparation for next GC
470 s.allocBits = s.gcmarkBits
471 s.gcmarkBits = newMarkBits(s.nelems)
473 // Initialize alloc bits cache.
474 s.refillAllocCache(0)
476 // The span must be in our exclusive ownership until we update sweepgen,
477 // check for potential races.
478 if state := s.state.get(); state != mSpanInUse || s.sweepgen != sweepgen-1 {
479 print("mspan.sweep: state=", state, " sweepgen=", s.sweepgen, " mheap.sweepgen=", sweepgen, "\n")
480 throw("mspan.sweep: bad span state after sweep")
482 if s.sweepgen == sweepgen+1 || s.sweepgen == sweepgen+3 {
483 throw("swept cached span")
486 // We need to set s.sweepgen = h.sweepgen only when all blocks are swept,
487 // because of the potential for a concurrent free/SetFinalizer.
489 // But we need to set it before we make the span available for allocation
490 // (return it to heap or mcentral), because allocation code assumes that a
491 // span is already swept if available for allocation.
493 // Serialization point.
494 // At this point the mark bits are cleared and allocation ready
495 // to go so release the span.
496 atomic.Store(&s.sweepgen, sweepgen)
498 if spc.sizeclass() != 0 {
499 // Handle spans for small objects.
501 // Only mark the span as needing zeroing if we've freed any
502 // objects, because a fresh span that had been allocated into,
503 // wasn't totally filled, but then swept, still has all of its
504 // free slots zeroed.
506 stats := memstats.heapStats.acquire()
507 atomic.Xadduintptr(&stats.smallFreeCount[spc.sizeclass()], uintptr(nfreed))
508 memstats.heapStats.release()
511 // The caller may not have removed this span from whatever
512 // unswept set its on but taken ownership of the span for
513 // sweeping by updating sweepgen. If this span still is in
514 // an unswept set, then the mcentral will pop it off the
515 // set, check its sweepgen, and ignore it.
517 // Free totally free span directly back to the heap.
521 // Return span back to the right mcentral list.
522 if uintptr(nalloc) == s.nelems {
523 mheap_.central[spc].mcentral.fullSwept(sweepgen).push(s)
525 mheap_.central[spc].mcentral.partialSwept(sweepgen).push(s)
528 } else if !preserve {
529 // Handle spans for large objects.
531 // Free large object span to heap.
533 // NOTE(rsc,dvyukov): The original implementation of efence
534 // in CL 22060046 used sysFree instead of sysFault, so that
535 // the operating system would eventually give the memory
536 // back to us again, so that an efence program could run
537 // longer without running out of memory. Unfortunately,
538 // calling sysFree here without any kind of adjustment of the
539 // heap data structures means that when the memory does
540 // come back to us, we have the wrong metadata for it, either in
541 // the mspan structures or in the garbage collection bitmap.
542 // Using sysFault here means that the program will run out of
543 // memory fairly quickly in efence mode, but at least it won't
544 // have mysterious crashes due to confused memory reuse.
545 // It should be possible to switch back to sysFree if we also
546 // implement and then call some kind of mheap.deleteSpan.
547 if debug.efence > 0 {
548 s.limit = 0 // prevent mlookup from finding this span
549 sysFault(unsafe.Pointer(s.base()), size)
553 stats := memstats.heapStats.acquire()
554 atomic.Xadduintptr(&stats.largeFreeCount, 1)
555 atomic.Xadduintptr(&stats.largeFree, size)
556 memstats.heapStats.release()
560 // Add a large span directly onto the full+swept list.
561 mheap_.central[spc].mcentral.fullSwept(sweepgen).push(s)
566 // reportZombies reports any marked but free objects in s and throws.
568 // This generally means one of the following:
570 // 1. User code converted a pointer to a uintptr and then back
571 // unsafely, and a GC ran while the uintptr was the only reference to
574 // 2. User code (or a compiler bug) constructed a bad pointer that
575 // points to a free slot, often a past-the-end pointer.
577 // 3. The GC two cycles ago missed a pointer and freed a live object,
578 // but it was still live in the last cycle, so this GC cycle found a
579 // pointer to that object and marked it.
580 func (s *mspan) reportZombies() {
582 print("runtime: marked free object in span ", s, ", elemsize=", s.elemsize, " freeindex=", s.freeindex, " (bad use of unsafe.Pointer? try -d=checkptr)\n")
583 mbits := s.markBitsForBase()
584 abits := s.allocBitsForIndex(0)
585 for i := uintptr(0); i < s.nelems; i++ {
586 addr := s.base() + i*s.elemsize
588 alloc := i < s.freeindex || abits.isMarked()
594 if mbits.isMarked() {
599 zombie := mbits.isMarked() && !alloc
609 hexdumpWords(addr, addr+length, nil)
614 throw("found pointer to free object")
617 // deductSweepCredit deducts sweep credit for allocating a span of
618 // size spanBytes. This must be performed *before* the span is
619 // allocated to ensure the system has enough credit. If necessary, it
620 // performs sweeping to prevent going in to debt. If the caller will
621 // also sweep pages (e.g., for a large allocation), it can pass a
622 // non-zero callerSweepPages to leave that many pages unswept.
624 // deductSweepCredit makes a worst-case assumption that all spanBytes
625 // bytes of the ultimately allocated span will be available for object
628 // deductSweepCredit is the core of the "proportional sweep" system.
629 // It uses statistics gathered by the garbage collector to perform
630 // enough sweeping so that all pages are swept during the concurrent
631 // sweep phase between GC cycles.
633 // mheap_ must NOT be locked.
634 func deductSweepCredit(spanBytes uintptr, callerSweepPages uintptr) {
635 if mheap_.sweepPagesPerByte == 0 {
636 // Proportional sweep is done or disabled.
645 sweptBasis := atomic.Load64(&mheap_.pagesSweptBasis)
647 // Fix debt if necessary.
648 newHeapLive := uintptr(atomic.Load64(&memstats.heap_live)-mheap_.sweepHeapLiveBasis) + spanBytes
649 pagesTarget := int64(mheap_.sweepPagesPerByte*float64(newHeapLive)) - int64(callerSweepPages)
650 for pagesTarget > int64(atomic.Load64(&mheap_.pagesSwept)-sweptBasis) {
651 if sweepone() == ^uintptr(0) {
652 mheap_.sweepPagesPerByte = 0
655 if atomic.Load64(&mheap_.pagesSweptBasis) != sweptBasis {
656 // Sweep pacing changed. Recompute debt.
666 // clobberfree sets the memory content at x to bad content, for debugging
668 func clobberfree(x unsafe.Pointer, size uintptr) {
669 // size (span.elemsize) is always a multiple of 4.
670 for i := uintptr(0); i < size; i += 4 {
671 *(*uint32)(add(x, i)) = 0xdeadbeef