Source file src/runtime/mgcmark.go

Documentation: runtime

     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.
     4  
     5  // Garbage collector: marking and scanning
     6  
     7  package runtime
     8  
     9  import (
    10  	"runtime/internal/atomic"
    11  	"runtime/internal/sys"
    12  	"unsafe"
    13  )
    14  
    15  const (
    16  	fixedRootFinalizers = iota
    17  	fixedRootFreeGStacks
    18  	fixedRootCount
    19  
    20  	// rootBlockBytes is the number of bytes to scan per data or
    21  	// BSS root.
    22  	rootBlockBytes = 256 << 10
    23  
    24  	// rootBlockSpans is the number of spans to scan per span
    25  	// root.
    26  	rootBlockSpans = 8 * 1024 // 64MB worth of spans
    27  
    28  	// maxObletBytes is the maximum bytes of an object to scan at
    29  	// once. Larger objects will be split up into "oblets" of at
    30  	// most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
    31  	// scan preemption at ~100 µs.
    32  	//
    33  	// This must be > _MaxSmallSize so that the object base is the
    34  	// span base.
    35  	maxObletBytes = 128 << 10
    36  
    37  	// drainCheckThreshold specifies how many units of work to do
    38  	// between self-preemption checks in gcDrain. Assuming a scan
    39  	// rate of 1 MB/ms, this is ~100 µs. Lower values have higher
    40  	// overhead in the scan loop (the scheduler check may perform
    41  	// a syscall, so its overhead is nontrivial). Higher values
    42  	// make the system less responsive to incoming work.
    43  	drainCheckThreshold = 100000
    44  )
    45  
    46  // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
    47  // some miscellany) and initializes scanning-related state.
    48  //
    49  // The caller must have call gcCopySpans().
    50  //
    51  // The world must be stopped.
    52  //
    53  //go:nowritebarrier
    54  func gcMarkRootPrepare() {
    55  	if gcphase == _GCmarktermination {
    56  		work.nFlushCacheRoots = int(gomaxprocs)
    57  	} else {
    58  		work.nFlushCacheRoots = 0
    59  	}
    60  
    61  	// Compute how many data and BSS root blocks there are.
    62  	nBlocks := func(bytes uintptr) int {
    63  		return int((bytes + rootBlockBytes - 1) / rootBlockBytes)
    64  	}
    65  
    66  	work.nDataRoots = 0
    67  	work.nBSSRoots = 0
    68  
    69  	// Only scan globals once per cycle; preferably concurrently.
    70  	if !work.markrootDone {
    71  		for _, datap := range activeModules() {
    72  			nDataRoots := nBlocks(datap.edata - datap.data)
    73  			if nDataRoots > work.nDataRoots {
    74  				work.nDataRoots = nDataRoots
    75  			}
    76  		}
    77  
    78  		for _, datap := range activeModules() {
    79  			nBSSRoots := nBlocks(datap.ebss - datap.bss)
    80  			if nBSSRoots > work.nBSSRoots {
    81  				work.nBSSRoots = nBSSRoots
    82  			}
    83  		}
    84  	}
    85  
    86  	if !work.markrootDone {
    87  		// On the first markroot, we need to scan span roots.
    88  		// In concurrent GC, this happens during concurrent
    89  		// mark and we depend on addfinalizer to ensure the
    90  		// above invariants for objects that get finalizers
    91  		// after concurrent mark. In STW GC, this will happen
    92  		// during mark termination.
    93  		//
    94  		// We're only interested in scanning the in-use spans,
    95  		// which will all be swept at this point. More spans
    96  		// may be added to this list during concurrent GC, but
    97  		// we only care about spans that were allocated before
    98  		// this mark phase.
    99  		work.nSpanRoots = mheap_.sweepSpans[mheap_.sweepgen/2%2].numBlocks()
   100  
   101  		// On the first markroot, we need to scan all Gs. Gs
   102  		// may be created after this point, but it's okay that
   103  		// we ignore them because they begin life without any
   104  		// roots, so there's nothing to scan, and any roots
   105  		// they create during the concurrent phase will be
   106  		// scanned during mark termination. During mark
   107  		// termination, allglen isn't changing, so we'll scan
   108  		// all Gs.
   109  		work.nStackRoots = int(atomic.Loaduintptr(&allglen))
   110  	} else {
   111  		// We've already scanned span roots and kept the scan
   112  		// up-to-date during concurrent mark.
   113  		work.nSpanRoots = 0
   114  
   115  		// The hybrid barrier ensures that stacks can't
   116  		// contain pointers to unmarked objects, so on the
   117  		// second markroot, there's no need to scan stacks.
   118  		work.nStackRoots = 0
   119  
   120  		if debug.gcrescanstacks > 0 {
   121  			// Scan stacks anyway for debugging.
   122  			work.nStackRoots = int(atomic.Loaduintptr(&allglen))
   123  		}
   124  	}
   125  
   126  	work.markrootNext = 0
   127  	work.markrootJobs = uint32(fixedRootCount + work.nFlushCacheRoots + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
   128  }
   129  
   130  // gcMarkRootCheck checks that all roots have been scanned. It is
   131  // purely for debugging.
   132  func gcMarkRootCheck() {
   133  	if work.markrootNext < work.markrootJobs {
   134  		print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
   135  		throw("left over markroot jobs")
   136  	}
   137  
   138  	lock(&allglock)
   139  	// Check that stacks have been scanned.
   140  	var gp *g
   141  	if gcphase == _GCmarktermination && debug.gcrescanstacks > 0 {
   142  		for i := 0; i < len(allgs); i++ {
   143  			gp = allgs[i]
   144  			if !(gp.gcscandone && gp.gcscanvalid) && readgstatus(gp) != _Gdead {
   145  				goto fail
   146  			}
   147  		}
   148  	} else {
   149  		for i := 0; i < work.nStackRoots; i++ {
   150  			gp = allgs[i]
   151  			if !gp.gcscandone {
   152  				goto fail
   153  			}
   154  		}
   155  	}
   156  	unlock(&allglock)
   157  	return
   158  
   159  fail:
   160  	println("gp", gp, "goid", gp.goid,
   161  		"status", readgstatus(gp),
   162  		"gcscandone", gp.gcscandone,
   163  		"gcscanvalid", gp.gcscanvalid)
   164  	unlock(&allglock) // Avoid self-deadlock with traceback.
   165  	throw("scan missed a g")
   166  }
   167  
   168  // ptrmask for an allocation containing a single pointer.
   169  var oneptrmask = [...]uint8{1}
   170  
   171  // markroot scans the i'th root.
   172  //
   173  // Preemption must be disabled (because this uses a gcWork).
   174  //
   175  // nowritebarrier is only advisory here.
   176  //
   177  //go:nowritebarrier
   178  func markroot(gcw *gcWork, i uint32) {
   179  	// TODO(austin): This is a bit ridiculous. Compute and store
   180  	// the bases in gcMarkRootPrepare instead of the counts.
   181  	baseFlushCache := uint32(fixedRootCount)
   182  	baseData := baseFlushCache + uint32(work.nFlushCacheRoots)
   183  	baseBSS := baseData + uint32(work.nDataRoots)
   184  	baseSpans := baseBSS + uint32(work.nBSSRoots)
   185  	baseStacks := baseSpans + uint32(work.nSpanRoots)
   186  	end := baseStacks + uint32(work.nStackRoots)
   187  
   188  	// Note: if you add a case here, please also update heapdump.go:dumproots.
   189  	switch {
   190  	case baseFlushCache <= i && i < baseData:
   191  		flushmcache(int(i - baseFlushCache))
   192  
   193  	case baseData <= i && i < baseBSS:
   194  		for _, datap := range activeModules() {
   195  			markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-baseData))
   196  		}
   197  
   198  	case baseBSS <= i && i < baseSpans:
   199  		for _, datap := range activeModules() {
   200  			markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-baseBSS))
   201  		}
   202  
   203  	case i == fixedRootFinalizers:
   204  		// Only do this once per GC cycle since we don't call
   205  		// queuefinalizer during marking.
   206  		if work.markrootDone {
   207  			break
   208  		}
   209  		for fb := allfin; fb != nil; fb = fb.alllink {
   210  			cnt := uintptr(atomic.Load(&fb.cnt))
   211  			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw)
   212  		}
   213  
   214  	case i == fixedRootFreeGStacks:
   215  		// Only do this once per GC cycle; preferably
   216  		// concurrently.
   217  		if !work.markrootDone {
   218  			// Switch to the system stack so we can call
   219  			// stackfree.
   220  			systemstack(markrootFreeGStacks)
   221  		}
   222  
   223  	case baseSpans <= i && i < baseStacks:
   224  		// mark MSpan.specials
   225  		markrootSpans(gcw, int(i-baseSpans))
   226  
   227  	default:
   228  		// the rest is scanning goroutine stacks
   229  		var gp *g
   230  		if baseStacks <= i && i < end {
   231  			gp = allgs[i-baseStacks]
   232  		} else {
   233  			throw("markroot: bad index")
   234  		}
   235  
   236  		// remember when we've first observed the G blocked
   237  		// needed only to output in traceback
   238  		status := readgstatus(gp) // We are not in a scan state
   239  		if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
   240  			gp.waitsince = work.tstart
   241  		}
   242  
   243  		// scang must be done on the system stack in case
   244  		// we're trying to scan our own stack.
   245  		systemstack(func() {
   246  			// If this is a self-scan, put the user G in
   247  			// _Gwaiting to prevent self-deadlock. It may
   248  			// already be in _Gwaiting if this is a mark
   249  			// worker or we're in mark termination.
   250  			userG := getg().m.curg
   251  			selfScan := gp == userG && readgstatus(userG) == _Grunning
   252  			if selfScan {
   253  				casgstatus(userG, _Grunning, _Gwaiting)
   254  				userG.waitreason = "garbage collection scan"
   255  			}
   256  
   257  			// TODO: scang blocks until gp's stack has
   258  			// been scanned, which may take a while for
   259  			// running goroutines. Consider doing this in
   260  			// two phases where the first is non-blocking:
   261  			// we scan the stacks we can and ask running
   262  			// goroutines to scan themselves; and the
   263  			// second blocks.
   264  			scang(gp, gcw)
   265  
   266  			if selfScan {
   267  				casgstatus(userG, _Gwaiting, _Grunning)
   268  			}
   269  		})
   270  	}
   271  }
   272  
   273  // markrootBlock scans the shard'th shard of the block of memory [b0,
   274  // b0+n0), with the given pointer mask.
   275  //
   276  //go:nowritebarrier
   277  func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
   278  	if rootBlockBytes%(8*sys.PtrSize) != 0 {
   279  		// This is necessary to pick byte offsets in ptrmask0.
   280  		throw("rootBlockBytes must be a multiple of 8*ptrSize")
   281  	}
   282  
   283  	b := b0 + uintptr(shard)*rootBlockBytes
   284  	if b >= b0+n0 {
   285  		return
   286  	}
   287  	ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*sys.PtrSize))))
   288  	n := uintptr(rootBlockBytes)
   289  	if b+n > b0+n0 {
   290  		n = b0 + n0 - b
   291  	}
   292  
   293  	// Scan this shard.
   294  	scanblock(b, n, ptrmask, gcw)
   295  }
   296  
   297  // markrootFreeGStacks frees stacks of dead Gs.
   298  //
   299  // This does not free stacks of dead Gs cached on Ps, but having a few
   300  // cached stacks around isn't a problem.
   301  //
   302  //TODO go:nowritebarrier
   303  func markrootFreeGStacks() {
   304  	// Take list of dead Gs with stacks.
   305  	lock(&sched.gflock)
   306  	list := sched.gfreeStack
   307  	sched.gfreeStack = nil
   308  	unlock(&sched.gflock)
   309  	if list == nil {
   310  		return
   311  	}
   312  
   313  	// Free stacks.
   314  	tail := list
   315  	for gp := list; gp != nil; gp = gp.schedlink.ptr() {
   316  		shrinkstack(gp)
   317  		tail = gp
   318  	}
   319  
   320  	// Put Gs back on the free list.
   321  	lock(&sched.gflock)
   322  	tail.schedlink.set(sched.gfreeNoStack)
   323  	sched.gfreeNoStack = list
   324  	unlock(&sched.gflock)
   325  }
   326  
   327  // markrootSpans marks roots for one shard of work.spans.
   328  //
   329  //go:nowritebarrier
   330  func markrootSpans(gcw *gcWork, shard int) {
   331  	// Objects with finalizers have two GC-related invariants:
   332  	//
   333  	// 1) Everything reachable from the object must be marked.
   334  	// This ensures that when we pass the object to its finalizer,
   335  	// everything the finalizer can reach will be retained.
   336  	//
   337  	// 2) Finalizer specials (which are not in the garbage
   338  	// collected heap) are roots. In practice, this means the fn
   339  	// field must be scanned.
   340  	//
   341  	// TODO(austin): There are several ideas for making this more
   342  	// efficient in issue #11485.
   343  
   344  	if work.markrootDone {
   345  		throw("markrootSpans during second markroot")
   346  	}
   347  
   348  	sg := mheap_.sweepgen
   349  	spans := mheap_.sweepSpans[mheap_.sweepgen/2%2].block(shard)
   350  	// Note that work.spans may not include spans that were
   351  	// allocated between entering the scan phase and now. This is
   352  	// okay because any objects with finalizers in those spans
   353  	// must have been allocated and given finalizers after we
   354  	// entered the scan phase, so addfinalizer will have ensured
   355  	// the above invariants for them.
   356  	for _, s := range spans {
   357  		if s.state != mSpanInUse {
   358  			continue
   359  		}
   360  		if !useCheckmark && s.sweepgen != sg {
   361  			// sweepgen was updated (+2) during non-checkmark GC pass
   362  			print("sweep ", s.sweepgen, " ", sg, "\n")
   363  			throw("gc: unswept span")
   364  		}
   365  
   366  		// Speculatively check if there are any specials
   367  		// without acquiring the span lock. This may race with
   368  		// adding the first special to a span, but in that
   369  		// case addfinalizer will observe that the GC is
   370  		// active (which is globally synchronized) and ensure
   371  		// the above invariants. We may also ensure the
   372  		// invariants, but it's okay to scan an object twice.
   373  		if s.specials == nil {
   374  			continue
   375  		}
   376  
   377  		// Lock the specials to prevent a special from being
   378  		// removed from the list while we're traversing it.
   379  		lock(&s.speciallock)
   380  
   381  		for sp := s.specials; sp != nil; sp = sp.next {
   382  			if sp.kind != _KindSpecialFinalizer {
   383  				continue
   384  			}
   385  			// don't mark finalized object, but scan it so we
   386  			// retain everything it points to.
   387  			spf := (*specialfinalizer)(unsafe.Pointer(sp))
   388  			// A finalizer can be set for an inner byte of an object, find object beginning.
   389  			p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
   390  
   391  			// Mark everything that can be reached from
   392  			// the object (but *not* the object itself or
   393  			// we'll never collect it).
   394  			scanobject(p, gcw)
   395  
   396  			// The special itself is a root.
   397  			scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw)
   398  		}
   399  
   400  		unlock(&s.speciallock)
   401  	}
   402  }
   403  
   404  // gcAssistAlloc performs GC work to make gp's assist debt positive.
   405  // gp must be the calling user gorountine.
   406  //
   407  // This must be called with preemption enabled.
   408  func gcAssistAlloc(gp *g) {
   409  	// Don't assist in non-preemptible contexts. These are
   410  	// generally fragile and won't allow the assist to block.
   411  	if getg() == gp.m.g0 {
   412  		return
   413  	}
   414  	if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
   415  		return
   416  	}
   417  
   418  	traced := false
   419  retry:
   420  	// Compute the amount of scan work we need to do to make the
   421  	// balance positive. When the required amount of work is low,
   422  	// we over-assist to build up credit for future allocations
   423  	// and amortize the cost of assisting.
   424  	debtBytes := -gp.gcAssistBytes
   425  	scanWork := int64(gcController.assistWorkPerByte * float64(debtBytes))
   426  	if scanWork < gcOverAssistWork {
   427  		scanWork = gcOverAssistWork
   428  		debtBytes = int64(gcController.assistBytesPerWork * float64(scanWork))
   429  	}
   430  
   431  	// Steal as much credit as we can from the background GC's
   432  	// scan credit. This is racy and may drop the background
   433  	// credit below 0 if two mutators steal at the same time. This
   434  	// will just cause steals to fail until credit is accumulated
   435  	// again, so in the long run it doesn't really matter, but we
   436  	// do have to handle the negative credit case.
   437  	bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
   438  	stolen := int64(0)
   439  	if bgScanCredit > 0 {
   440  		if bgScanCredit < scanWork {
   441  			stolen = bgScanCredit
   442  			gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(stolen))
   443  		} else {
   444  			stolen = scanWork
   445  			gp.gcAssistBytes += debtBytes
   446  		}
   447  		atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
   448  
   449  		scanWork -= stolen
   450  
   451  		if scanWork == 0 {
   452  			// We were able to steal all of the credit we
   453  			// needed.
   454  			if traced {
   455  				traceGCMarkAssistDone()
   456  			}
   457  			return
   458  		}
   459  	}
   460  
   461  	if trace.enabled && !traced {
   462  		traced = true
   463  		traceGCMarkAssistStart()
   464  	}
   465  
   466  	// Perform assist work
   467  	systemstack(func() {
   468  		gcAssistAlloc1(gp, scanWork)
   469  		// The user stack may have moved, so this can't touch
   470  		// anything on it until it returns from systemstack.
   471  	})
   472  
   473  	completed := gp.param != nil
   474  	gp.param = nil
   475  	if completed {
   476  		gcMarkDone()
   477  	}
   478  
   479  	if gp.gcAssistBytes < 0 {
   480  		// We were unable steal enough credit or perform
   481  		// enough work to pay off the assist debt. We need to
   482  		// do one of these before letting the mutator allocate
   483  		// more to prevent over-allocation.
   484  		//
   485  		// If this is because we were preempted, reschedule
   486  		// and try some more.
   487  		if gp.preempt {
   488  			Gosched()
   489  			goto retry
   490  		}
   491  
   492  		// Add this G to an assist queue and park. When the GC
   493  		// has more background credit, it will satisfy queued
   494  		// assists before flushing to the global credit pool.
   495  		//
   496  		// Note that this does *not* get woken up when more
   497  		// work is added to the work list. The theory is that
   498  		// there wasn't enough work to do anyway, so we might
   499  		// as well let background marking take care of the
   500  		// work that is available.
   501  		if !gcParkAssist() {
   502  			goto retry
   503  		}
   504  
   505  		// At this point either background GC has satisfied
   506  		// this G's assist debt, or the GC cycle is over.
   507  	}
   508  	if traced {
   509  		traceGCMarkAssistDone()
   510  	}
   511  }
   512  
   513  // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
   514  // stack. This is a separate function to make it easier to see that
   515  // we're not capturing anything from the user stack, since the user
   516  // stack may move while we're in this function.
   517  //
   518  // gcAssistAlloc1 indicates whether this assist completed the mark
   519  // phase by setting gp.param to non-nil. This can't be communicated on
   520  // the stack since it may move.
   521  //
   522  //go:systemstack
   523  func gcAssistAlloc1(gp *g, scanWork int64) {
   524  	// Clear the flag indicating that this assist completed the
   525  	// mark phase.
   526  	gp.param = nil
   527  
   528  	if atomic.Load(&gcBlackenEnabled) == 0 {
   529  		// The gcBlackenEnabled check in malloc races with the
   530  		// store that clears it but an atomic check in every malloc
   531  		// would be a performance hit.
   532  		// Instead we recheck it here on the non-preemptable system
   533  		// stack to determine if we should preform an assist.
   534  
   535  		// GC is done, so ignore any remaining debt.
   536  		gp.gcAssistBytes = 0
   537  		return
   538  	}
   539  	// Track time spent in this assist. Since we're on the
   540  	// system stack, this is non-preemptible, so we can
   541  	// just measure start and end time.
   542  	startTime := nanotime()
   543  
   544  	decnwait := atomic.Xadd(&work.nwait, -1)
   545  	if decnwait == work.nproc {
   546  		println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
   547  		throw("nwait > work.nprocs")
   548  	}
   549  
   550  	// gcDrainN requires the caller to be preemptible.
   551  	casgstatus(gp, _Grunning, _Gwaiting)
   552  	gp.waitreason = "GC assist marking"
   553  
   554  	// drain own cached work first in the hopes that it
   555  	// will be more cache friendly.
   556  	gcw := &getg().m.p.ptr().gcw
   557  	workDone := gcDrainN(gcw, scanWork)
   558  	// If we are near the end of the mark phase
   559  	// dispose of the gcw.
   560  	if gcBlackenPromptly {
   561  		gcw.dispose()
   562  	}
   563  
   564  	casgstatus(gp, _Gwaiting, _Grunning)
   565  
   566  	// Record that we did this much scan work.
   567  	//
   568  	// Back out the number of bytes of assist credit that
   569  	// this scan work counts for. The "1+" is a poor man's
   570  	// round-up, to ensure this adds credit even if
   571  	// assistBytesPerWork is very low.
   572  	gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(workDone))
   573  
   574  	// If this is the last worker and we ran out of work,
   575  	// signal a completion point.
   576  	incnwait := atomic.Xadd(&work.nwait, +1)
   577  	if incnwait > work.nproc {
   578  		println("runtime: work.nwait=", incnwait,
   579  			"work.nproc=", work.nproc,
   580  			"gcBlackenPromptly=", gcBlackenPromptly)
   581  		throw("work.nwait > work.nproc")
   582  	}
   583  
   584  	if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
   585  		// This has reached a background completion point. Set
   586  		// gp.param to a non-nil value to indicate this. It
   587  		// doesn't matter what we set it to (it just has to be
   588  		// a valid pointer).
   589  		gp.param = unsafe.Pointer(gp)
   590  	}
   591  	duration := nanotime() - startTime
   592  	_p_ := gp.m.p.ptr()
   593  	_p_.gcAssistTime += duration
   594  	if _p_.gcAssistTime > gcAssistTimeSlack {
   595  		atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
   596  		_p_.gcAssistTime = 0
   597  	}
   598  }
   599  
   600  // gcWakeAllAssists wakes all currently blocked assists. This is used
   601  // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
   602  // new assists from going to sleep after this point.
   603  func gcWakeAllAssists() {
   604  	lock(&work.assistQueue.lock)
   605  	injectglist(work.assistQueue.head.ptr())
   606  	work.assistQueue.head.set(nil)
   607  	work.assistQueue.tail.set(nil)
   608  	unlock(&work.assistQueue.lock)
   609  }
   610  
   611  // gcParkAssist puts the current goroutine on the assist queue and parks.
   612  //
   613  // gcParkAssist returns whether the assist is now satisfied. If it
   614  // returns false, the caller must retry the assist.
   615  //
   616  //go:nowritebarrier
   617  func gcParkAssist() bool {
   618  	lock(&work.assistQueue.lock)
   619  	// If the GC cycle finished while we were getting the lock,
   620  	// exit the assist. The cycle can't finish while we hold the
   621  	// lock.
   622  	if atomic.Load(&gcBlackenEnabled) == 0 {
   623  		unlock(&work.assistQueue.lock)
   624  		return true
   625  	}
   626  
   627  	gp := getg()
   628  	oldHead, oldTail := work.assistQueue.head, work.assistQueue.tail
   629  	if oldHead == 0 {
   630  		work.assistQueue.head.set(gp)
   631  	} else {
   632  		oldTail.ptr().schedlink.set(gp)
   633  	}
   634  	work.assistQueue.tail.set(gp)
   635  	gp.schedlink.set(nil)
   636  
   637  	// Recheck for background credit now that this G is in
   638  	// the queue, but can still back out. This avoids a
   639  	// race in case background marking has flushed more
   640  	// credit since we checked above.
   641  	if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
   642  		work.assistQueue.head = oldHead
   643  		work.assistQueue.tail = oldTail
   644  		if oldTail != 0 {
   645  			oldTail.ptr().schedlink.set(nil)
   646  		}
   647  		unlock(&work.assistQueue.lock)
   648  		return false
   649  	}
   650  	// Park.
   651  	goparkunlock(&work.assistQueue.lock, "GC assist wait", traceEvGoBlockGC, 2)
   652  	return true
   653  }
   654  
   655  // gcFlushBgCredit flushes scanWork units of background scan work
   656  // credit. This first satisfies blocked assists on the
   657  // work.assistQueue and then flushes any remaining credit to
   658  // gcController.bgScanCredit.
   659  //
   660  // Write barriers are disallowed because this is used by gcDrain after
   661  // it has ensured that all work is drained and this must preserve that
   662  // condition.
   663  //
   664  //go:nowritebarrierrec
   665  func gcFlushBgCredit(scanWork int64) {
   666  	if work.assistQueue.head == 0 {
   667  		// Fast path; there are no blocked assists. There's a
   668  		// small window here where an assist may add itself to
   669  		// the blocked queue and park. If that happens, we'll
   670  		// just get it on the next flush.
   671  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   672  		return
   673  	}
   674  
   675  	scanBytes := int64(float64(scanWork) * gcController.assistBytesPerWork)
   676  
   677  	lock(&work.assistQueue.lock)
   678  	gp := work.assistQueue.head.ptr()
   679  	for gp != nil && scanBytes > 0 {
   680  		// Note that gp.gcAssistBytes is negative because gp
   681  		// is in debt. Think carefully about the signs below.
   682  		if scanBytes+gp.gcAssistBytes >= 0 {
   683  			// Satisfy this entire assist debt.
   684  			scanBytes += gp.gcAssistBytes
   685  			gp.gcAssistBytes = 0
   686  			xgp := gp
   687  			gp = gp.schedlink.ptr()
   688  			// It's important that we *not* put xgp in
   689  			// runnext. Otherwise, it's possible for user
   690  			// code to exploit the GC worker's high
   691  			// scheduler priority to get itself always run
   692  			// before other goroutines and always in the
   693  			// fresh quantum started by GC.
   694  			ready(xgp, 0, false)
   695  		} else {
   696  			// Partially satisfy this assist.
   697  			gp.gcAssistBytes += scanBytes
   698  			scanBytes = 0
   699  			// As a heuristic, we move this assist to the
   700  			// back of the queue so that large assists
   701  			// can't clog up the assist queue and
   702  			// substantially delay small assists.
   703  			xgp := gp
   704  			gp = gp.schedlink.ptr()
   705  			if gp == nil {
   706  				// gp is the only assist in the queue.
   707  				gp = xgp
   708  			} else {
   709  				xgp.schedlink = 0
   710  				work.assistQueue.tail.ptr().schedlink.set(xgp)
   711  				work.assistQueue.tail.set(xgp)
   712  			}
   713  			break
   714  		}
   715  	}
   716  	work.assistQueue.head.set(gp)
   717  	if gp == nil {
   718  		work.assistQueue.tail.set(nil)
   719  	}
   720  
   721  	if scanBytes > 0 {
   722  		// Convert from scan bytes back to work.
   723  		scanWork = int64(float64(scanBytes) * gcController.assistWorkPerByte)
   724  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   725  	}
   726  	unlock(&work.assistQueue.lock)
   727  }
   728  
   729  // scanstack scans gp's stack, greying all pointers found on the stack.
   730  //
   731  // scanstack is marked go:systemstack because it must not be preempted
   732  // while using a workbuf.
   733  //
   734  //go:nowritebarrier
   735  //go:systemstack
   736  func scanstack(gp *g, gcw *gcWork) {
   737  	if gp.gcscanvalid {
   738  		return
   739  	}
   740  
   741  	if readgstatus(gp)&_Gscan == 0 {
   742  		print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
   743  		throw("scanstack - bad status")
   744  	}
   745  
   746  	switch readgstatus(gp) &^ _Gscan {
   747  	default:
   748  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   749  		throw("mark - bad status")
   750  	case _Gdead:
   751  		return
   752  	case _Grunning:
   753  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   754  		throw("scanstack: goroutine not stopped")
   755  	case _Grunnable, _Gsyscall, _Gwaiting:
   756  		// ok
   757  	}
   758  
   759  	if gp == getg() {
   760  		throw("can't scan our own stack")
   761  	}
   762  	mp := gp.m
   763  	if mp != nil && mp.helpgc != 0 {
   764  		throw("can't scan gchelper stack")
   765  	}
   766  
   767  	// Shrink the stack if not much of it is being used. During
   768  	// concurrent GC, we can do this during concurrent mark.
   769  	if !work.markrootDone {
   770  		shrinkstack(gp)
   771  	}
   772  
   773  	// Scan the saved context register. This is effectively a live
   774  	// register that gets moved back and forth between the
   775  	// register and sched.ctxt without a write barrier.
   776  	if gp.sched.ctxt != nil {
   777  		scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), sys.PtrSize, &oneptrmask[0], gcw)
   778  	}
   779  
   780  	// Scan the stack.
   781  	var cache pcvalueCache
   782  	scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
   783  		scanframeworker(frame, &cache, gcw)
   784  		return true
   785  	}
   786  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
   787  	tracebackdefers(gp, scanframe, nil)
   788  	gp.gcscanvalid = true
   789  }
   790  
   791  // Scan a stack frame: local variables and function arguments/results.
   792  //go:nowritebarrier
   793  func scanframeworker(frame *stkframe, cache *pcvalueCache, gcw *gcWork) {
   794  
   795  	f := frame.fn
   796  	targetpc := frame.continpc
   797  	if targetpc == 0 {
   798  		// Frame is dead.
   799  		return
   800  	}
   801  	if _DebugGC > 1 {
   802  		print("scanframe ", funcname(f), "\n")
   803  	}
   804  	if targetpc != f.entry {
   805  		targetpc--
   806  	}
   807  	pcdata := pcdatavalue(f, _PCDATA_StackMapIndex, targetpc, cache)
   808  	if pcdata == -1 {
   809  		// We do not have a valid pcdata value but there might be a
   810  		// stackmap for this function. It is likely that we are looking
   811  		// at the function prologue, assume so and hope for the best.
   812  		pcdata = 0
   813  	}
   814  
   815  	// Scan local variables if stack frame has been allocated.
   816  	size := frame.varp - frame.sp
   817  	var minsize uintptr
   818  	switch sys.ArchFamily {
   819  	case sys.ARM64:
   820  		minsize = sys.SpAlign
   821  	default:
   822  		minsize = sys.MinFrameSize
   823  	}
   824  	if size > minsize {
   825  		stkmap := (*stackmap)(funcdata(f, _FUNCDATA_LocalsPointerMaps))
   826  		if stkmap == nil || stkmap.n <= 0 {
   827  			print("runtime: frame ", funcname(f), " untyped locals ", hex(frame.varp-size), "+", hex(size), "\n")
   828  			throw("missing stackmap")
   829  		}
   830  
   831  		// Locals bitmap information, scan just the pointers in locals.
   832  		if pcdata < 0 || pcdata >= stkmap.n {
   833  			// don't know where we are
   834  			print("runtime: pcdata is ", pcdata, " and ", stkmap.n, " locals stack map entries for ", funcname(f), " (targetpc=", targetpc, ")\n")
   835  			throw("scanframe: bad symbol table")
   836  		}
   837  		bv := stackmapdata(stkmap, pcdata)
   838  		size = uintptr(bv.n) * sys.PtrSize
   839  		scanblock(frame.varp-size, size, bv.bytedata, gcw)
   840  	}
   841  
   842  	// Scan arguments.
   843  	if frame.arglen > 0 {
   844  		var bv bitvector
   845  		if frame.argmap != nil {
   846  			bv = *frame.argmap
   847  		} else {
   848  			stkmap := (*stackmap)(funcdata(f, _FUNCDATA_ArgsPointerMaps))
   849  			if stkmap == nil || stkmap.n <= 0 {
   850  				print("runtime: frame ", funcname(f), " untyped args ", hex(frame.argp), "+", hex(frame.arglen), "\n")
   851  				throw("missing stackmap")
   852  			}
   853  			if pcdata < 0 || pcdata >= stkmap.n {
   854  				// don't know where we are
   855  				print("runtime: pcdata is ", pcdata, " and ", stkmap.n, " args stack map entries for ", funcname(f), " (targetpc=", targetpc, ")\n")
   856  				throw("scanframe: bad symbol table")
   857  			}
   858  			bv = stackmapdata(stkmap, pcdata)
   859  		}
   860  		scanblock(frame.argp, uintptr(bv.n)*sys.PtrSize, bv.bytedata, gcw)
   861  	}
   862  }
   863  
   864  type gcDrainFlags int
   865  
   866  const (
   867  	gcDrainUntilPreempt gcDrainFlags = 1 << iota
   868  	gcDrainNoBlock
   869  	gcDrainFlushBgCredit
   870  	gcDrainIdle
   871  	gcDrainFractional
   872  
   873  	// gcDrainBlock means neither gcDrainUntilPreempt or
   874  	// gcDrainNoBlock. It is the default, but callers should use
   875  	// the constant for documentation purposes.
   876  	gcDrainBlock gcDrainFlags = 0
   877  )
   878  
   879  // gcDrain scans roots and objects in work buffers, blackening grey
   880  // objects until all roots and work buffers have been drained.
   881  //
   882  // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
   883  // is set. This implies gcDrainNoBlock.
   884  //
   885  // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
   886  // to do. This implies gcDrainNoBlock.
   887  //
   888  // If flags&gcDrainFractional != 0, gcDrain self-preempts when
   889  // pollFractionalWorkerExit() returns true. This implies
   890  // gcDrainNoBlock.
   891  //
   892  // If flags&gcDrainNoBlock != 0, gcDrain returns as soon as it is
   893  // unable to get more work. Otherwise, it will block until all
   894  // blocking calls are blocked in gcDrain.
   895  //
   896  // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
   897  // credit to gcController.bgScanCredit every gcCreditSlack units of
   898  // scan work.
   899  //
   900  //go:nowritebarrier
   901  func gcDrain(gcw *gcWork, flags gcDrainFlags) {
   902  	if !writeBarrier.needed {
   903  		throw("gcDrain phase incorrect")
   904  	}
   905  
   906  	gp := getg().m.curg
   907  	preemptible := flags&gcDrainUntilPreempt != 0
   908  	blocking := flags&(gcDrainUntilPreempt|gcDrainIdle|gcDrainFractional|gcDrainNoBlock) == 0
   909  	flushBgCredit := flags&gcDrainFlushBgCredit != 0
   910  	idle := flags&gcDrainIdle != 0
   911  
   912  	initScanWork := gcw.scanWork
   913  
   914  	// checkWork is the scan work before performing the next
   915  	// self-preempt check.
   916  	checkWork := int64(1<<63 - 1)
   917  	var check func() bool
   918  	if flags&(gcDrainIdle|gcDrainFractional) != 0 {
   919  		checkWork = initScanWork + drainCheckThreshold
   920  		if idle {
   921  			check = pollWork
   922  		} else if flags&gcDrainFractional != 0 {
   923  			check = pollFractionalWorkerExit
   924  		}
   925  	}
   926  
   927  	// Drain root marking jobs.
   928  	if work.markrootNext < work.markrootJobs {
   929  		for !(preemptible && gp.preempt) {
   930  			job := atomic.Xadd(&work.markrootNext, +1) - 1
   931  			if job >= work.markrootJobs {
   932  				break
   933  			}
   934  			markroot(gcw, job)
   935  			if check != nil && check() {
   936  				goto done
   937  			}
   938  		}
   939  	}
   940  
   941  	// Drain heap marking jobs.
   942  	for !(preemptible && gp.preempt) {
   943  		// Try to keep work available on the global queue. We used to
   944  		// check if there were waiting workers, but it's better to
   945  		// just keep work available than to make workers wait. In the
   946  		// worst case, we'll do O(log(_WorkbufSize)) unnecessary
   947  		// balances.
   948  		if work.full == 0 {
   949  			gcw.balance()
   950  		}
   951  
   952  		var b uintptr
   953  		if blocking {
   954  			b = gcw.get()
   955  		} else {
   956  			b = gcw.tryGetFast()
   957  			if b == 0 {
   958  				b = gcw.tryGet()
   959  			}
   960  		}
   961  		if b == 0 {
   962  			// work barrier reached or tryGet failed.
   963  			break
   964  		}
   965  		scanobject(b, gcw)
   966  
   967  		// Flush background scan work credit to the global
   968  		// account if we've accumulated enough locally so
   969  		// mutator assists can draw on it.
   970  		if gcw.scanWork >= gcCreditSlack {
   971  			atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
   972  			if flushBgCredit {
   973  				gcFlushBgCredit(gcw.scanWork - initScanWork)
   974  				initScanWork = 0
   975  			}
   976  			checkWork -= gcw.scanWork
   977  			gcw.scanWork = 0
   978  
   979  			if checkWork <= 0 {
   980  				checkWork += drainCheckThreshold
   981  				if check != nil && check() {
   982  					break
   983  				}
   984  			}
   985  		}
   986  	}
   987  
   988  	// In blocking mode, write barriers are not allowed after this
   989  	// point because we must preserve the condition that the work
   990  	// buffers are empty.
   991  
   992  done:
   993  	// Flush remaining scan work credit.
   994  	if gcw.scanWork > 0 {
   995  		atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
   996  		if flushBgCredit {
   997  			gcFlushBgCredit(gcw.scanWork - initScanWork)
   998  		}
   999  		gcw.scanWork = 0
  1000  	}
  1001  }
  1002  
  1003  // gcDrainN blackens grey objects until it has performed roughly
  1004  // scanWork units of scan work or the G is preempted. This is
  1005  // best-effort, so it may perform less work if it fails to get a work
  1006  // buffer. Otherwise, it will perform at least n units of work, but
  1007  // may perform more because scanning is always done in whole object
  1008  // increments. It returns the amount of scan work performed.
  1009  //
  1010  // The caller goroutine must be in a preemptible state (e.g.,
  1011  // _Gwaiting) to prevent deadlocks during stack scanning. As a
  1012  // consequence, this must be called on the system stack.
  1013  //
  1014  //go:nowritebarrier
  1015  //go:systemstack
  1016  func gcDrainN(gcw *gcWork, scanWork int64) int64 {
  1017  	if !writeBarrier.needed {
  1018  		throw("gcDrainN phase incorrect")
  1019  	}
  1020  
  1021  	// There may already be scan work on the gcw, which we don't
  1022  	// want to claim was done by this call.
  1023  	workFlushed := -gcw.scanWork
  1024  
  1025  	gp := getg().m.curg
  1026  	for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
  1027  		// See gcDrain comment.
  1028  		if work.full == 0 {
  1029  			gcw.balance()
  1030  		}
  1031  
  1032  		// This might be a good place to add prefetch code...
  1033  		// if(wbuf.nobj > 4) {
  1034  		//         PREFETCH(wbuf->obj[wbuf.nobj - 3];
  1035  		//  }
  1036  		//
  1037  		b := gcw.tryGetFast()
  1038  		if b == 0 {
  1039  			b = gcw.tryGet()
  1040  		}
  1041  
  1042  		if b == 0 {
  1043  			// Try to do a root job.
  1044  			//
  1045  			// TODO: Assists should get credit for this
  1046  			// work.
  1047  			if work.markrootNext < work.markrootJobs {
  1048  				job := atomic.Xadd(&work.markrootNext, +1) - 1
  1049  				if job < work.markrootJobs {
  1050  					markroot(gcw, job)
  1051  					continue
  1052  				}
  1053  			}
  1054  			// No heap or root jobs.
  1055  			break
  1056  		}
  1057  		scanobject(b, gcw)
  1058  
  1059  		// Flush background scan work credit.
  1060  		if gcw.scanWork >= gcCreditSlack {
  1061  			atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1062  			workFlushed += gcw.scanWork
  1063  			gcw.scanWork = 0
  1064  		}
  1065  	}
  1066  
  1067  	// Unlike gcDrain, there's no need to flush remaining work
  1068  	// here because this never flushes to bgScanCredit and
  1069  	// gcw.dispose will flush any remaining work to scanWork.
  1070  
  1071  	return workFlushed + gcw.scanWork
  1072  }
  1073  
  1074  // scanblock scans b as scanobject would, but using an explicit
  1075  // pointer bitmap instead of the heap bitmap.
  1076  //
  1077  // This is used to scan non-heap roots, so it does not update
  1078  // gcw.bytesMarked or gcw.scanWork.
  1079  //
  1080  //go:nowritebarrier
  1081  func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork) {
  1082  	// Use local copies of original parameters, so that a stack trace
  1083  	// due to one of the throws below shows the original block
  1084  	// base and extent.
  1085  	b := b0
  1086  	n := n0
  1087  
  1088  	arena_start := mheap_.arena_start
  1089  	arena_used := mheap_.arena_used
  1090  
  1091  	for i := uintptr(0); i < n; {
  1092  		// Find bits for the next word.
  1093  		bits := uint32(*addb(ptrmask, i/(sys.PtrSize*8)))
  1094  		if bits == 0 {
  1095  			i += sys.PtrSize * 8
  1096  			continue
  1097  		}
  1098  		for j := 0; j < 8 && i < n; j++ {
  1099  			if bits&1 != 0 {
  1100  				// Same work as in scanobject; see comments there.
  1101  				obj := *(*uintptr)(unsafe.Pointer(b + i))
  1102  				if obj != 0 && arena_start <= obj && obj < arena_used {
  1103  					if obj, hbits, span, objIndex := heapBitsForObject(obj, b, i); obj != 0 {
  1104  						greyobject(obj, b, i, hbits, span, gcw, objIndex)
  1105  					}
  1106  				}
  1107  			}
  1108  			bits >>= 1
  1109  			i += sys.PtrSize
  1110  		}
  1111  	}
  1112  }
  1113  
  1114  // scanobject scans the object starting at b, adding pointers to gcw.
  1115  // b must point to the beginning of a heap object or an oblet.
  1116  // scanobject consults the GC bitmap for the pointer mask and the
  1117  // spans for the size of the object.
  1118  //
  1119  //go:nowritebarrier
  1120  func scanobject(b uintptr, gcw *gcWork) {
  1121  	// Note that arena_used may change concurrently during
  1122  	// scanobject and hence scanobject may encounter a pointer to
  1123  	// a newly allocated heap object that is *not* in
  1124  	// [start,used). It will not mark this object; however, we
  1125  	// know that it was just installed by a mutator, which means
  1126  	// that mutator will execute a write barrier and take care of
  1127  	// marking it. This is even more pronounced on relaxed memory
  1128  	// architectures since we access arena_used without barriers
  1129  	// or synchronization, but the same logic applies.
  1130  	arena_start := mheap_.arena_start
  1131  	arena_used := mheap_.arena_used
  1132  
  1133  	// Find the bits for b and the size of the object at b.
  1134  	//
  1135  	// b is either the beginning of an object, in which case this
  1136  	// is the size of the object to scan, or it points to an
  1137  	// oblet, in which case we compute the size to scan below.
  1138  	hbits := heapBitsForAddr(b)
  1139  	s := spanOfUnchecked(b)
  1140  	n := s.elemsize
  1141  	if n == 0 {
  1142  		throw("scanobject n == 0")
  1143  	}
  1144  
  1145  	if n > maxObletBytes {
  1146  		// Large object. Break into oblets for better
  1147  		// parallelism and lower latency.
  1148  		if b == s.base() {
  1149  			// It's possible this is a noscan object (not
  1150  			// from greyobject, but from other code
  1151  			// paths), in which case we must *not* enqueue
  1152  			// oblets since their bitmaps will be
  1153  			// uninitialized.
  1154  			if s.spanclass.noscan() {
  1155  				// Bypass the whole scan.
  1156  				gcw.bytesMarked += uint64(n)
  1157  				return
  1158  			}
  1159  
  1160  			// Enqueue the other oblets to scan later.
  1161  			// Some oblets may be in b's scalar tail, but
  1162  			// these will be marked as "no more pointers",
  1163  			// so we'll drop out immediately when we go to
  1164  			// scan those.
  1165  			for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
  1166  				if !gcw.putFast(oblet) {
  1167  					gcw.put(oblet)
  1168  				}
  1169  			}
  1170  		}
  1171  
  1172  		// Compute the size of the oblet. Since this object
  1173  		// must be a large object, s.base() is the beginning
  1174  		// of the object.
  1175  		n = s.base() + s.elemsize - b
  1176  		if n > maxObletBytes {
  1177  			n = maxObletBytes
  1178  		}
  1179  	}
  1180  
  1181  	var i uintptr
  1182  	for i = 0; i < n; i += sys.PtrSize {
  1183  		// Find bits for this word.
  1184  		if i != 0 {
  1185  			// Avoid needless hbits.next() on last iteration.
  1186  			hbits = hbits.next()
  1187  		}
  1188  		// Load bits once. See CL 22712 and issue 16973 for discussion.
  1189  		bits := hbits.bits()
  1190  		// During checkmarking, 1-word objects store the checkmark
  1191  		// in the type bit for the one word. The only one-word objects
  1192  		// are pointers, or else they'd be merged with other non-pointer
  1193  		// data into larger allocations.
  1194  		if i != 1*sys.PtrSize && bits&bitScan == 0 {
  1195  			break // no more pointers in this object
  1196  		}
  1197  		if bits&bitPointer == 0 {
  1198  			continue // not a pointer
  1199  		}
  1200  
  1201  		// Work here is duplicated in scanblock and above.
  1202  		// If you make changes here, make changes there too.
  1203  		obj := *(*uintptr)(unsafe.Pointer(b + i))
  1204  
  1205  		// At this point we have extracted the next potential pointer.
  1206  		// Check if it points into heap and not back at the current object.
  1207  		if obj != 0 && arena_start <= obj && obj < arena_used && obj-b >= n {
  1208  			// Mark the object.
  1209  			if obj, hbits, span, objIndex := heapBitsForObject(obj, b, i); obj != 0 {
  1210  				greyobject(obj, b, i, hbits, span, gcw, objIndex)
  1211  			}
  1212  		}
  1213  	}
  1214  	gcw.bytesMarked += uint64(n)
  1215  	gcw.scanWork += int64(i)
  1216  }
  1217  
  1218  // Shade the object if it isn't already.
  1219  // The object is not nil and known to be in the heap.
  1220  // Preemption must be disabled.
  1221  //go:nowritebarrier
  1222  func shade(b uintptr) {
  1223  	if obj, hbits, span, objIndex := heapBitsForObject(b, 0, 0); obj != 0 {
  1224  		gcw := &getg().m.p.ptr().gcw
  1225  		greyobject(obj, 0, 0, hbits, span, gcw, objIndex)
  1226  		if gcphase == _GCmarktermination || gcBlackenPromptly {
  1227  			// Ps aren't allowed to cache work during mark
  1228  			// termination.
  1229  			gcw.dispose()
  1230  		}
  1231  	}
  1232  }
  1233  
  1234  // obj is the start of an object with mark mbits.
  1235  // If it isn't already marked, mark it and enqueue into gcw.
  1236  // base and off are for debugging only and could be removed.
  1237  //
  1238  // See also wbBufFlush1, which partially duplicates this logic.
  1239  //
  1240  //go:nowritebarrierrec
  1241  func greyobject(obj, base, off uintptr, hbits heapBits, span *mspan, gcw *gcWork, objIndex uintptr) {
  1242  	// obj should be start of allocation, and so must be at least pointer-aligned.
  1243  	if obj&(sys.PtrSize-1) != 0 {
  1244  		throw("greyobject: obj not pointer-aligned")
  1245  	}
  1246  	mbits := span.markBitsForIndex(objIndex)
  1247  
  1248  	if useCheckmark {
  1249  		if !mbits.isMarked() {
  1250  			printlock()
  1251  			print("runtime:greyobject: checkmarks finds unexpected unmarked object obj=", hex(obj), "\n")
  1252  			print("runtime: found obj at *(", hex(base), "+", hex(off), ")\n")
  1253  
  1254  			// Dump the source (base) object
  1255  			gcDumpObject("base", base, off)
  1256  
  1257  			// Dump the object
  1258  			gcDumpObject("obj", obj, ^uintptr(0))
  1259  
  1260  			getg().m.traceback = 2
  1261  			throw("checkmark found unmarked object")
  1262  		}
  1263  		if hbits.isCheckmarked(span.elemsize) {
  1264  			return
  1265  		}
  1266  		hbits.setCheckmarked(span.elemsize)
  1267  		if !hbits.isCheckmarked(span.elemsize) {
  1268  			throw("setCheckmarked and isCheckmarked disagree")
  1269  		}
  1270  	} else {
  1271  		if debug.gccheckmark > 0 && span.isFree(objIndex) {
  1272  			print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
  1273  			gcDumpObject("base", base, off)
  1274  			gcDumpObject("obj", obj, ^uintptr(0))
  1275  			getg().m.traceback = 2
  1276  			throw("marking free object")
  1277  		}
  1278  
  1279  		// If marked we have nothing to do.
  1280  		if mbits.isMarked() {
  1281  			return
  1282  		}
  1283  		// mbits.setMarked() // Avoid extra call overhead with manual inlining.
  1284  		atomic.Or8(mbits.bytep, mbits.mask)
  1285  		// If this is a noscan object, fast-track it to black
  1286  		// instead of greying it.
  1287  		if span.spanclass.noscan() {
  1288  			gcw.bytesMarked += uint64(span.elemsize)
  1289  			return
  1290  		}
  1291  	}
  1292  
  1293  	// Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
  1294  	// seems like a nice optimization that can be added back in.
  1295  	// There needs to be time between the PREFETCH and the use.
  1296  	// Previously we put the obj in an 8 element buffer that is drained at a rate
  1297  	// to give the PREFETCH time to do its work.
  1298  	// Use of PREFETCHNTA might be more appropriate than PREFETCH
  1299  	if !gcw.putFast(obj) {
  1300  		gcw.put(obj)
  1301  	}
  1302  }
  1303  
  1304  // gcDumpObject dumps the contents of obj for debugging and marks the
  1305  // field at byte offset off in obj.
  1306  func gcDumpObject(label string, obj, off uintptr) {
  1307  	if obj < mheap_.arena_start || obj >= mheap_.arena_used {
  1308  		print(label, "=", hex(obj), " is not in the Go heap\n")
  1309  		return
  1310  	}
  1311  	k := obj >> _PageShift
  1312  	x := k
  1313  	x -= mheap_.arena_start >> _PageShift
  1314  	s := mheap_.spans[x]
  1315  	print(label, "=", hex(obj), " k=", hex(k))
  1316  	if s == nil {
  1317  		print(" s=nil\n")
  1318  		return
  1319  	}
  1320  	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
  1321  	if 0 <= s.state && int(s.state) < len(mSpanStateNames) {
  1322  		print(mSpanStateNames[s.state], "\n")
  1323  	} else {
  1324  		print("unknown(", s.state, ")\n")
  1325  	}
  1326  
  1327  	skipped := false
  1328  	size := s.elemsize
  1329  	if s.state == _MSpanManual && size == 0 {
  1330  		// We're printing something from a stack frame. We
  1331  		// don't know how big it is, so just show up to an
  1332  		// including off.
  1333  		size = off + sys.PtrSize
  1334  	}
  1335  	for i := uintptr(0); i < size; i += sys.PtrSize {
  1336  		// For big objects, just print the beginning (because
  1337  		// that usually hints at the object's type) and the
  1338  		// fields around off.
  1339  		if !(i < 128*sys.PtrSize || off-16*sys.PtrSize < i && i < off+16*sys.PtrSize) {
  1340  			skipped = true
  1341  			continue
  1342  		}
  1343  		if skipped {
  1344  			print(" ...\n")
  1345  			skipped = false
  1346  		}
  1347  		print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
  1348  		if i == off {
  1349  			print(" <==")
  1350  		}
  1351  		print("\n")
  1352  	}
  1353  	if skipped {
  1354  		print(" ...\n")
  1355  	}
  1356  }
  1357  
  1358  // gcmarknewobject marks a newly allocated object black. obj must
  1359  // not contain any non-nil pointers.
  1360  //
  1361  // This is nosplit so it can manipulate a gcWork without preemption.
  1362  //
  1363  //go:nowritebarrier
  1364  //go:nosplit
  1365  func gcmarknewobject(obj, size, scanSize uintptr) {
  1366  	if useCheckmark && !gcBlackenPromptly { // The world should be stopped so this should not happen.
  1367  		throw("gcmarknewobject called while doing checkmark")
  1368  	}
  1369  	markBitsForAddr(obj).setMarked()
  1370  	gcw := &getg().m.p.ptr().gcw
  1371  	gcw.bytesMarked += uint64(size)
  1372  	gcw.scanWork += int64(scanSize)
  1373  	if gcBlackenPromptly {
  1374  		// There shouldn't be anything in the work queue, but
  1375  		// we still need to flush stats.
  1376  		gcw.dispose()
  1377  	}
  1378  }
  1379  
  1380  // gcMarkTinyAllocs greys all active tiny alloc blocks.
  1381  //
  1382  // The world must be stopped.
  1383  func gcMarkTinyAllocs() {
  1384  	for _, p := range allp {
  1385  		c := p.mcache
  1386  		if c == nil || c.tiny == 0 {
  1387  			continue
  1388  		}
  1389  		_, hbits, span, objIndex := heapBitsForObject(c.tiny, 0, 0)
  1390  		gcw := &p.gcw
  1391  		greyobject(c.tiny, 0, 0, hbits, span, gcw, objIndex)
  1392  		if gcBlackenPromptly {
  1393  			gcw.dispose()
  1394  		}
  1395  	}
  1396  }
  1397  
  1398  // Checkmarking
  1399  
  1400  // To help debug the concurrent GC we remark with the world
  1401  // stopped ensuring that any object encountered has their normal
  1402  // mark bit set. To do this we use an orthogonal bit
  1403  // pattern to indicate the object is marked. The following pattern
  1404  // uses the upper two bits in the object's boundary nibble.
  1405  // 01: scalar  not marked
  1406  // 10: pointer not marked
  1407  // 11: pointer     marked
  1408  // 00: scalar      marked
  1409  // Xoring with 01 will flip the pattern from marked to unmarked and vica versa.
  1410  // The higher bit is 1 for pointers and 0 for scalars, whether the object
  1411  // is marked or not.
  1412  // The first nibble no longer holds the typeDead pattern indicating that the
  1413  // there are no more pointers in the object. This information is held
  1414  // in the second nibble.
  1415  
  1416  // If useCheckmark is true, marking of an object uses the
  1417  // checkmark bits (encoding above) instead of the standard
  1418  // mark bits.
  1419  var useCheckmark = false
  1420  
  1421  //go:nowritebarrier
  1422  func initCheckmarks() {
  1423  	useCheckmark = true
  1424  	for _, s := range mheap_.allspans {
  1425  		if s.state == _MSpanInUse {
  1426  			heapBitsForSpan(s.base()).initCheckmarkSpan(s.layout())
  1427  		}
  1428  	}
  1429  }
  1430  
  1431  func clearCheckmarks() {
  1432  	useCheckmark = false
  1433  	for _, s := range mheap_.allspans {
  1434  		if s.state == _MSpanInUse {
  1435  			heapBitsForSpan(s.base()).clearCheckmarkSpan(s.layout())
  1436  		}
  1437  	}
  1438  }
  1439  

View as plain text