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 world must be stopped.
    50  //
    51  //go:nowritebarrier
    52  func gcMarkRootPrepare() {
    53  	work.nFlushCacheRoots = 0
    54  
    55  	// Compute how many data and BSS root blocks there are.
    56  	nBlocks := func(bytes uintptr) int {
    57  		return int((bytes + rootBlockBytes - 1) / rootBlockBytes)
    58  	}
    59  
    60  	work.nDataRoots = 0
    61  	work.nBSSRoots = 0
    62  
    63  	// Scan globals.
    64  	for _, datap := range activeModules() {
    65  		nDataRoots := nBlocks(datap.edata - datap.data)
    66  		if nDataRoots > work.nDataRoots {
    67  			work.nDataRoots = nDataRoots
    68  		}
    69  	}
    70  
    71  	for _, datap := range activeModules() {
    72  		nBSSRoots := nBlocks(datap.ebss - datap.bss)
    73  		if nBSSRoots > work.nBSSRoots {
    74  			work.nBSSRoots = nBSSRoots
    75  		}
    76  	}
    77  
    78  	// Scan span roots for finalizer specials.
    79  	//
    80  	// We depend on addfinalizer to mark objects that get
    81  	// finalizers after root marking.
    82  	//
    83  	// We're only interested in scanning the in-use spans,
    84  	// which will all be swept at this point. More spans
    85  	// may be added to this list during concurrent GC, but
    86  	// we only care about spans that were allocated before
    87  	// this mark phase.
    88  	work.nSpanRoots = mheap_.sweepSpans[mheap_.sweepgen/2%2].numBlocks()
    89  
    90  	// Scan stacks.
    91  	//
    92  	// Gs may be created after this point, but it's okay that we
    93  	// ignore them because they begin life without any roots, so
    94  	// there's nothing to scan, and any roots they create during
    95  	// the concurrent phase will be scanned during mark
    96  	// termination.
    97  	work.nStackRoots = int(atomic.Loaduintptr(&allglen))
    98  
    99  	work.markrootNext = 0
   100  	work.markrootJobs = uint32(fixedRootCount + work.nFlushCacheRoots + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
   101  }
   102  
   103  // gcMarkRootCheck checks that all roots have been scanned. It is
   104  // purely for debugging.
   105  func gcMarkRootCheck() {
   106  	if work.markrootNext < work.markrootJobs {
   107  		print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
   108  		throw("left over markroot jobs")
   109  	}
   110  
   111  	lock(&allglock)
   112  	// Check that stacks have been scanned.
   113  	var gp *g
   114  	for i := 0; i < work.nStackRoots; i++ {
   115  		gp = allgs[i]
   116  		if !gp.gcscandone {
   117  			goto fail
   118  		}
   119  	}
   120  	unlock(&allglock)
   121  	return
   122  
   123  fail:
   124  	println("gp", gp, "goid", gp.goid,
   125  		"status", readgstatus(gp),
   126  		"gcscandone", gp.gcscandone)
   127  	unlock(&allglock) // Avoid self-deadlock with traceback.
   128  	throw("scan missed a g")
   129  }
   130  
   131  // ptrmask for an allocation containing a single pointer.
   132  var oneptrmask = [...]uint8{1}
   133  
   134  // markroot scans the i'th root.
   135  //
   136  // Preemption must be disabled (because this uses a gcWork).
   137  //
   138  // nowritebarrier is only advisory here.
   139  //
   140  //go:nowritebarrier
   141  func markroot(gcw *gcWork, i uint32) {
   142  	// TODO(austin): This is a bit ridiculous. Compute and store
   143  	// the bases in gcMarkRootPrepare instead of the counts.
   144  	baseFlushCache := uint32(fixedRootCount)
   145  	baseData := baseFlushCache + uint32(work.nFlushCacheRoots)
   146  	baseBSS := baseData + uint32(work.nDataRoots)
   147  	baseSpans := baseBSS + uint32(work.nBSSRoots)
   148  	baseStacks := baseSpans + uint32(work.nSpanRoots)
   149  	end := baseStacks + uint32(work.nStackRoots)
   150  
   151  	// Note: if you add a case here, please also update heapdump.go:dumproots.
   152  	switch {
   153  	case baseFlushCache <= i && i < baseData:
   154  		flushmcache(int(i - baseFlushCache))
   155  
   156  	case baseData <= i && i < baseBSS:
   157  		for _, datap := range activeModules() {
   158  			markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-baseData))
   159  		}
   160  
   161  	case baseBSS <= i && i < baseSpans:
   162  		for _, datap := range activeModules() {
   163  			markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-baseBSS))
   164  		}
   165  
   166  	case i == fixedRootFinalizers:
   167  		for fb := allfin; fb != nil; fb = fb.alllink {
   168  			cnt := uintptr(atomic.Load(&fb.cnt))
   169  			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
   170  		}
   171  
   172  	case i == fixedRootFreeGStacks:
   173  		// Switch to the system stack so we can call
   174  		// stackfree.
   175  		systemstack(markrootFreeGStacks)
   176  
   177  	case baseSpans <= i && i < baseStacks:
   178  		// mark mspan.specials
   179  		markrootSpans(gcw, int(i-baseSpans))
   180  
   181  	default:
   182  		// the rest is scanning goroutine stacks
   183  		var gp *g
   184  		if baseStacks <= i && i < end {
   185  			gp = allgs[i-baseStacks]
   186  		} else {
   187  			throw("markroot: bad index")
   188  		}
   189  
   190  		// remember when we've first observed the G blocked
   191  		// needed only to output in traceback
   192  		status := readgstatus(gp) // We are not in a scan state
   193  		if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
   194  			gp.waitsince = work.tstart
   195  		}
   196  
   197  		// scanstack must be done on the system stack in case
   198  		// we're trying to scan our own stack.
   199  		systemstack(func() {
   200  			// If this is a self-scan, put the user G in
   201  			// _Gwaiting to prevent self-deadlock. It may
   202  			// already be in _Gwaiting if this is a mark
   203  			// worker or we're in mark termination.
   204  			userG := getg().m.curg
   205  			selfScan := gp == userG && readgstatus(userG) == _Grunning
   206  			if selfScan {
   207  				casgstatus(userG, _Grunning, _Gwaiting)
   208  				userG.waitreason = waitReasonGarbageCollectionScan
   209  			}
   210  
   211  			// TODO: suspendG blocks (and spins) until gp
   212  			// stops, which may take a while for
   213  			// running goroutines. Consider doing this in
   214  			// two phases where the first is non-blocking:
   215  			// we scan the stacks we can and ask running
   216  			// goroutines to scan themselves; and the
   217  			// second blocks.
   218  			stopped := suspendG(gp)
   219  			if stopped.dead {
   220  				gp.gcscandone = true
   221  				return
   222  			}
   223  			if gp.gcscandone {
   224  				throw("g already scanned")
   225  			}
   226  			scanstack(gp, gcw)
   227  			gp.gcscandone = true
   228  			resumeG(stopped)
   229  
   230  			if selfScan {
   231  				casgstatus(userG, _Gwaiting, _Grunning)
   232  			}
   233  		})
   234  	}
   235  }
   236  
   237  // markrootBlock scans the shard'th shard of the block of memory [b0,
   238  // b0+n0), with the given pointer mask.
   239  //
   240  //go:nowritebarrier
   241  func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
   242  	if rootBlockBytes%(8*sys.PtrSize) != 0 {
   243  		// This is necessary to pick byte offsets in ptrmask0.
   244  		throw("rootBlockBytes must be a multiple of 8*ptrSize")
   245  	}
   246  
   247  	// Note that if b0 is toward the end of the address space,
   248  	// then b0 + rootBlockBytes might wrap around.
   249  	// These tests are written to avoid any possible overflow.
   250  	off := uintptr(shard) * rootBlockBytes
   251  	if off >= n0 {
   252  		return
   253  	}
   254  	b := b0 + off
   255  	ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*sys.PtrSize))))
   256  	n := uintptr(rootBlockBytes)
   257  	if off+n > n0 {
   258  		n = n0 - off
   259  	}
   260  
   261  	// Scan this shard.
   262  	scanblock(b, n, ptrmask, gcw, nil)
   263  }
   264  
   265  // markrootFreeGStacks frees stacks of dead Gs.
   266  //
   267  // This does not free stacks of dead Gs cached on Ps, but having a few
   268  // cached stacks around isn't a problem.
   269  func markrootFreeGStacks() {
   270  	// Take list of dead Gs with stacks.
   271  	lock(&sched.gFree.lock)
   272  	list := sched.gFree.stack
   273  	sched.gFree.stack = gList{}
   274  	unlock(&sched.gFree.lock)
   275  	if list.empty() {
   276  		return
   277  	}
   278  
   279  	// Free stacks.
   280  	q := gQueue{list.head, list.head}
   281  	for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
   282  		stackfree(gp.stack)
   283  		gp.stack.lo = 0
   284  		gp.stack.hi = 0
   285  		// Manipulate the queue directly since the Gs are
   286  		// already all linked the right way.
   287  		q.tail.set(gp)
   288  	}
   289  
   290  	// Put Gs back on the free list.
   291  	lock(&sched.gFree.lock)
   292  	sched.gFree.noStack.pushAll(q)
   293  	unlock(&sched.gFree.lock)
   294  }
   295  
   296  // markrootSpans marks roots for one shard of work.spans.
   297  //
   298  //go:nowritebarrier
   299  func markrootSpans(gcw *gcWork, shard int) {
   300  	// Objects with finalizers have two GC-related invariants:
   301  	//
   302  	// 1) Everything reachable from the object must be marked.
   303  	// This ensures that when we pass the object to its finalizer,
   304  	// everything the finalizer can reach will be retained.
   305  	//
   306  	// 2) Finalizer specials (which are not in the garbage
   307  	// collected heap) are roots. In practice, this means the fn
   308  	// field must be scanned.
   309  	//
   310  	// TODO(austin): There are several ideas for making this more
   311  	// efficient in issue #11485.
   312  
   313  	sg := mheap_.sweepgen
   314  	spans := mheap_.sweepSpans[mheap_.sweepgen/2%2].block(shard)
   315  	// Note that work.spans may not include spans that were
   316  	// allocated between entering the scan phase and now. We may
   317  	// also race with spans being added into sweepSpans when they're
   318  	// just created, and as a result we may see nil pointers in the
   319  	// spans slice. This is okay because any objects with finalizers
   320  	// in those spans must have been allocated and given finalizers
   321  	// after we entered the scan phase, so addfinalizer will have
   322  	// ensured the above invariants for them.
   323  	for i := 0; i < len(spans); i++ {
   324  		// sweepBuf.block requires that we read pointers from the block atomically.
   325  		// It also requires that we ignore nil pointers.
   326  		s := (*mspan)(atomic.Loadp(unsafe.Pointer(&spans[i])))
   327  
   328  		// This is racing with spans being initialized, so
   329  		// check the state carefully.
   330  		if s == nil || s.state.get() != mSpanInUse {
   331  			continue
   332  		}
   333  		// Check that this span was swept (it may be cached or uncached).
   334  		if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
   335  			// sweepgen was updated (+2) during non-checkmark GC pass
   336  			print("sweep ", s.sweepgen, " ", sg, "\n")
   337  			throw("gc: unswept span")
   338  		}
   339  
   340  		// Speculatively check if there are any specials
   341  		// without acquiring the span lock. This may race with
   342  		// adding the first special to a span, but in that
   343  		// case addfinalizer will observe that the GC is
   344  		// active (which is globally synchronized) and ensure
   345  		// the above invariants. We may also ensure the
   346  		// invariants, but it's okay to scan an object twice.
   347  		if s.specials == nil {
   348  			continue
   349  		}
   350  
   351  		// Lock the specials to prevent a special from being
   352  		// removed from the list while we're traversing it.
   353  		lock(&s.speciallock)
   354  
   355  		for sp := s.specials; sp != nil; sp = sp.next {
   356  			if sp.kind != _KindSpecialFinalizer {
   357  				continue
   358  			}
   359  			// don't mark finalized object, but scan it so we
   360  			// retain everything it points to.
   361  			spf := (*specialfinalizer)(unsafe.Pointer(sp))
   362  			// A finalizer can be set for an inner byte of an object, find object beginning.
   363  			p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
   364  
   365  			// Mark everything that can be reached from
   366  			// the object (but *not* the object itself or
   367  			// we'll never collect it).
   368  			scanobject(p, gcw)
   369  
   370  			// The special itself is a root.
   371  			scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw, nil)
   372  		}
   373  
   374  		unlock(&s.speciallock)
   375  	}
   376  }
   377  
   378  // gcAssistAlloc performs GC work to make gp's assist debt positive.
   379  // gp must be the calling user gorountine.
   380  //
   381  // This must be called with preemption enabled.
   382  func gcAssistAlloc(gp *g) {
   383  	// Don't assist in non-preemptible contexts. These are
   384  	// generally fragile and won't allow the assist to block.
   385  	if getg() == gp.m.g0 {
   386  		return
   387  	}
   388  	if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
   389  		return
   390  	}
   391  
   392  	traced := false
   393  retry:
   394  	// Compute the amount of scan work we need to do to make the
   395  	// balance positive. When the required amount of work is low,
   396  	// we over-assist to build up credit for future allocations
   397  	// and amortize the cost of assisting.
   398  	debtBytes := -gp.gcAssistBytes
   399  	scanWork := int64(gcController.assistWorkPerByte * float64(debtBytes))
   400  	if scanWork < gcOverAssistWork {
   401  		scanWork = gcOverAssistWork
   402  		debtBytes = int64(gcController.assistBytesPerWork * float64(scanWork))
   403  	}
   404  
   405  	// Steal as much credit as we can from the background GC's
   406  	// scan credit. This is racy and may drop the background
   407  	// credit below 0 if two mutators steal at the same time. This
   408  	// will just cause steals to fail until credit is accumulated
   409  	// again, so in the long run it doesn't really matter, but we
   410  	// do have to handle the negative credit case.
   411  	bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
   412  	stolen := int64(0)
   413  	if bgScanCredit > 0 {
   414  		if bgScanCredit < scanWork {
   415  			stolen = bgScanCredit
   416  			gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(stolen))
   417  		} else {
   418  			stolen = scanWork
   419  			gp.gcAssistBytes += debtBytes
   420  		}
   421  		atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
   422  
   423  		scanWork -= stolen
   424  
   425  		if scanWork == 0 {
   426  			// We were able to steal all of the credit we
   427  			// needed.
   428  			if traced {
   429  				traceGCMarkAssistDone()
   430  			}
   431  			return
   432  		}
   433  	}
   434  
   435  	if trace.enabled && !traced {
   436  		traced = true
   437  		traceGCMarkAssistStart()
   438  	}
   439  
   440  	// Perform assist work
   441  	systemstack(func() {
   442  		gcAssistAlloc1(gp, scanWork)
   443  		// The user stack may have moved, so this can't touch
   444  		// anything on it until it returns from systemstack.
   445  	})
   446  
   447  	completed := gp.param != nil
   448  	gp.param = nil
   449  	if completed {
   450  		gcMarkDone()
   451  	}
   452  
   453  	if gp.gcAssistBytes < 0 {
   454  		// We were unable steal enough credit or perform
   455  		// enough work to pay off the assist debt. We need to
   456  		// do one of these before letting the mutator allocate
   457  		// more to prevent over-allocation.
   458  		//
   459  		// If this is because we were preempted, reschedule
   460  		// and try some more.
   461  		if gp.preempt {
   462  			Gosched()
   463  			goto retry
   464  		}
   465  
   466  		// Add this G to an assist queue and park. When the GC
   467  		// has more background credit, it will satisfy queued
   468  		// assists before flushing to the global credit pool.
   469  		//
   470  		// Note that this does *not* get woken up when more
   471  		// work is added to the work list. The theory is that
   472  		// there wasn't enough work to do anyway, so we might
   473  		// as well let background marking take care of the
   474  		// work that is available.
   475  		if !gcParkAssist() {
   476  			goto retry
   477  		}
   478  
   479  		// At this point either background GC has satisfied
   480  		// this G's assist debt, or the GC cycle is over.
   481  	}
   482  	if traced {
   483  		traceGCMarkAssistDone()
   484  	}
   485  }
   486  
   487  // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
   488  // stack. This is a separate function to make it easier to see that
   489  // we're not capturing anything from the user stack, since the user
   490  // stack may move while we're in this function.
   491  //
   492  // gcAssistAlloc1 indicates whether this assist completed the mark
   493  // phase by setting gp.param to non-nil. This can't be communicated on
   494  // the stack since it may move.
   495  //
   496  //go:systemstack
   497  func gcAssistAlloc1(gp *g, scanWork int64) {
   498  	// Clear the flag indicating that this assist completed the
   499  	// mark phase.
   500  	gp.param = nil
   501  
   502  	if atomic.Load(&gcBlackenEnabled) == 0 {
   503  		// The gcBlackenEnabled check in malloc races with the
   504  		// store that clears it but an atomic check in every malloc
   505  		// would be a performance hit.
   506  		// Instead we recheck it here on the non-preemptable system
   507  		// stack to determine if we should perform an assist.
   508  
   509  		// GC is done, so ignore any remaining debt.
   510  		gp.gcAssistBytes = 0
   511  		return
   512  	}
   513  	// Track time spent in this assist. Since we're on the
   514  	// system stack, this is non-preemptible, so we can
   515  	// just measure start and end time.
   516  	startTime := nanotime()
   517  
   518  	decnwait := atomic.Xadd(&work.nwait, -1)
   519  	if decnwait == work.nproc {
   520  		println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
   521  		throw("nwait > work.nprocs")
   522  	}
   523  
   524  	// gcDrainN requires the caller to be preemptible.
   525  	casgstatus(gp, _Grunning, _Gwaiting)
   526  	gp.waitreason = waitReasonGCAssistMarking
   527  
   528  	// drain own cached work first in the hopes that it
   529  	// will be more cache friendly.
   530  	gcw := &getg().m.p.ptr().gcw
   531  	workDone := gcDrainN(gcw, scanWork)
   532  
   533  	casgstatus(gp, _Gwaiting, _Grunning)
   534  
   535  	// Record that we did this much scan work.
   536  	//
   537  	// Back out the number of bytes of assist credit that
   538  	// this scan work counts for. The "1+" is a poor man's
   539  	// round-up, to ensure this adds credit even if
   540  	// assistBytesPerWork is very low.
   541  	gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(workDone))
   542  
   543  	// If this is the last worker and we ran out of work,
   544  	// signal a completion point.
   545  	incnwait := atomic.Xadd(&work.nwait, +1)
   546  	if incnwait > work.nproc {
   547  		println("runtime: work.nwait=", incnwait,
   548  			"work.nproc=", work.nproc)
   549  		throw("work.nwait > work.nproc")
   550  	}
   551  
   552  	if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
   553  		// This has reached a background completion point. Set
   554  		// gp.param to a non-nil value to indicate this. It
   555  		// doesn't matter what we set it to (it just has to be
   556  		// a valid pointer).
   557  		gp.param = unsafe.Pointer(gp)
   558  	}
   559  	duration := nanotime() - startTime
   560  	_p_ := gp.m.p.ptr()
   561  	_p_.gcAssistTime += duration
   562  	if _p_.gcAssistTime > gcAssistTimeSlack {
   563  		atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
   564  		_p_.gcAssistTime = 0
   565  	}
   566  }
   567  
   568  // gcWakeAllAssists wakes all currently blocked assists. This is used
   569  // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
   570  // new assists from going to sleep after this point.
   571  func gcWakeAllAssists() {
   572  	lock(&work.assistQueue.lock)
   573  	list := work.assistQueue.q.popList()
   574  	injectglist(&list)
   575  	unlock(&work.assistQueue.lock)
   576  }
   577  
   578  // gcParkAssist puts the current goroutine on the assist queue and parks.
   579  //
   580  // gcParkAssist reports whether the assist is now satisfied. If it
   581  // returns false, the caller must retry the assist.
   582  //
   583  //go:nowritebarrier
   584  func gcParkAssist() bool {
   585  	lock(&work.assistQueue.lock)
   586  	// If the GC cycle finished while we were getting the lock,
   587  	// exit the assist. The cycle can't finish while we hold the
   588  	// lock.
   589  	if atomic.Load(&gcBlackenEnabled) == 0 {
   590  		unlock(&work.assistQueue.lock)
   591  		return true
   592  	}
   593  
   594  	gp := getg()
   595  	oldList := work.assistQueue.q
   596  	work.assistQueue.q.pushBack(gp)
   597  
   598  	// Recheck for background credit now that this G is in
   599  	// the queue, but can still back out. This avoids a
   600  	// race in case background marking has flushed more
   601  	// credit since we checked above.
   602  	if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
   603  		work.assistQueue.q = oldList
   604  		if oldList.tail != 0 {
   605  			oldList.tail.ptr().schedlink.set(nil)
   606  		}
   607  		unlock(&work.assistQueue.lock)
   608  		return false
   609  	}
   610  	// Park.
   611  	goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
   612  	return true
   613  }
   614  
   615  // gcFlushBgCredit flushes scanWork units of background scan work
   616  // credit. This first satisfies blocked assists on the
   617  // work.assistQueue and then flushes any remaining credit to
   618  // gcController.bgScanCredit.
   619  //
   620  // Write barriers are disallowed because this is used by gcDrain after
   621  // it has ensured that all work is drained and this must preserve that
   622  // condition.
   623  //
   624  //go:nowritebarrierrec
   625  func gcFlushBgCredit(scanWork int64) {
   626  	if work.assistQueue.q.empty() {
   627  		// Fast path; there are no blocked assists. There's a
   628  		// small window here where an assist may add itself to
   629  		// the blocked queue and park. If that happens, we'll
   630  		// just get it on the next flush.
   631  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   632  		return
   633  	}
   634  
   635  	scanBytes := int64(float64(scanWork) * gcController.assistBytesPerWork)
   636  
   637  	lock(&work.assistQueue.lock)
   638  	for !work.assistQueue.q.empty() && scanBytes > 0 {
   639  		gp := work.assistQueue.q.pop()
   640  		// Note that gp.gcAssistBytes is negative because gp
   641  		// is in debt. Think carefully about the signs below.
   642  		if scanBytes+gp.gcAssistBytes >= 0 {
   643  			// Satisfy this entire assist debt.
   644  			scanBytes += gp.gcAssistBytes
   645  			gp.gcAssistBytes = 0
   646  			// It's important that we *not* put gp in
   647  			// runnext. Otherwise, it's possible for user
   648  			// code to exploit the GC worker's high
   649  			// scheduler priority to get itself always run
   650  			// before other goroutines and always in the
   651  			// fresh quantum started by GC.
   652  			ready(gp, 0, false)
   653  		} else {
   654  			// Partially satisfy this assist.
   655  			gp.gcAssistBytes += scanBytes
   656  			scanBytes = 0
   657  			// As a heuristic, we move this assist to the
   658  			// back of the queue so that large assists
   659  			// can't clog up the assist queue and
   660  			// substantially delay small assists.
   661  			work.assistQueue.q.pushBack(gp)
   662  			break
   663  		}
   664  	}
   665  
   666  	if scanBytes > 0 {
   667  		// Convert from scan bytes back to work.
   668  		scanWork = int64(float64(scanBytes) * gcController.assistWorkPerByte)
   669  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   670  	}
   671  	unlock(&work.assistQueue.lock)
   672  }
   673  
   674  // scanstack scans gp's stack, greying all pointers found on the stack.
   675  //
   676  // scanstack will also shrink the stack if it is safe to do so. If it
   677  // is not, it schedules a stack shrink for the next synchronous safe
   678  // point.
   679  //
   680  // scanstack is marked go:systemstack because it must not be preempted
   681  // while using a workbuf.
   682  //
   683  //go:nowritebarrier
   684  //go:systemstack
   685  func scanstack(gp *g, gcw *gcWork) {
   686  	if readgstatus(gp)&_Gscan == 0 {
   687  		print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
   688  		throw("scanstack - bad status")
   689  	}
   690  
   691  	switch readgstatus(gp) &^ _Gscan {
   692  	default:
   693  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   694  		throw("mark - bad status")
   695  	case _Gdead:
   696  		return
   697  	case _Grunning:
   698  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   699  		throw("scanstack: goroutine not stopped")
   700  	case _Grunnable, _Gsyscall, _Gwaiting:
   701  		// ok
   702  	}
   703  
   704  	if gp == getg() {
   705  		throw("can't scan our own stack")
   706  	}
   707  
   708  	if isShrinkStackSafe(gp) {
   709  		// Shrink the stack if not much of it is being used.
   710  		shrinkstack(gp)
   711  	} else {
   712  		// Otherwise, shrink the stack at the next sync safe point.
   713  		gp.preemptShrink = true
   714  	}
   715  
   716  	var state stackScanState
   717  	state.stack = gp.stack
   718  
   719  	if stackTraceDebug {
   720  		println("stack trace goroutine", gp.goid)
   721  	}
   722  
   723  	if debugScanConservative && gp.asyncSafePoint {
   724  		print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
   725  	}
   726  
   727  	// Scan the saved context register. This is effectively a live
   728  	// register that gets moved back and forth between the
   729  	// register and sched.ctxt without a write barrier.
   730  	if gp.sched.ctxt != nil {
   731  		scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   732  	}
   733  
   734  	// Scan the stack. Accumulate a list of stack objects.
   735  	scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
   736  		scanframeworker(frame, &state, gcw)
   737  		return true
   738  	}
   739  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
   740  
   741  	// Find additional pointers that point into the stack from the heap.
   742  	// Currently this includes defers and panics. See also function copystack.
   743  
   744  	// Find and trace all defer arguments.
   745  	tracebackdefers(gp, scanframe, nil)
   746  
   747  	// Find and trace other pointers in defer records.
   748  	for d := gp._defer; d != nil; d = d.link {
   749  		if d.fn != nil {
   750  			// tracebackdefers above does not scan the func value, which could
   751  			// be a stack allocated closure. See issue 30453.
   752  			scanblock(uintptr(unsafe.Pointer(&d.fn)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   753  		}
   754  		if d.link != nil {
   755  			// The link field of a stack-allocated defer record might point
   756  			// to a heap-allocated defer record. Keep that heap record live.
   757  			scanblock(uintptr(unsafe.Pointer(&d.link)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   758  		}
   759  		// Retain defers records themselves.
   760  		// Defer records might not be reachable from the G through regular heap
   761  		// tracing because the defer linked list might weave between the stack and the heap.
   762  		if d.heap {
   763  			scanblock(uintptr(unsafe.Pointer(&d)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   764  		}
   765  	}
   766  	if gp._panic != nil {
   767  		// Panics are always stack allocated.
   768  		state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
   769  	}
   770  
   771  	// Find and scan all reachable stack objects.
   772  	//
   773  	// The state's pointer queue prioritizes precise pointers over
   774  	// conservative pointers so that we'll prefer scanning stack
   775  	// objects precisely.
   776  	state.buildIndex()
   777  	for {
   778  		p, conservative := state.getPtr()
   779  		if p == 0 {
   780  			break
   781  		}
   782  		obj := state.findObject(p)
   783  		if obj == nil {
   784  			continue
   785  		}
   786  		t := obj.typ
   787  		if t == nil {
   788  			// We've already scanned this object.
   789  			continue
   790  		}
   791  		obj.setType(nil) // Don't scan it again.
   792  		if stackTraceDebug {
   793  			printlock()
   794  			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of type", t.string())
   795  			if conservative {
   796  				print(" (conservative)")
   797  			}
   798  			println()
   799  			printunlock()
   800  		}
   801  		gcdata := t.gcdata
   802  		var s *mspan
   803  		if t.kind&kindGCProg != 0 {
   804  			// This path is pretty unlikely, an object large enough
   805  			// to have a GC program allocated on the stack.
   806  			// We need some space to unpack the program into a straight
   807  			// bitmask, which we allocate/free here.
   808  			// TODO: it would be nice if there were a way to run a GC
   809  			// program without having to store all its bits. We'd have
   810  			// to change from a Lempel-Ziv style program to something else.
   811  			// Or we can forbid putting objects on stacks if they require
   812  			// a gc program (see issue 27447).
   813  			s = materializeGCProg(t.ptrdata, gcdata)
   814  			gcdata = (*byte)(unsafe.Pointer(s.startAddr))
   815  		}
   816  
   817  		b := state.stack.lo + uintptr(obj.off)
   818  		if conservative {
   819  			scanConservative(b, t.ptrdata, gcdata, gcw, &state)
   820  		} else {
   821  			scanblock(b, t.ptrdata, gcdata, gcw, &state)
   822  		}
   823  
   824  		if s != nil {
   825  			dematerializeGCProg(s)
   826  		}
   827  	}
   828  
   829  	// Deallocate object buffers.
   830  	// (Pointer buffers were all deallocated in the loop above.)
   831  	for state.head != nil {
   832  		x := state.head
   833  		state.head = x.next
   834  		if stackTraceDebug {
   835  			for _, obj := range x.obj[:x.nobj] {
   836  				if obj.typ == nil { // reachable
   837  					continue
   838  				}
   839  				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of type", obj.typ.string())
   840  				// Note: not necessarily really dead - only reachable-from-ptr dead.
   841  			}
   842  		}
   843  		x.nobj = 0
   844  		putempty((*workbuf)(unsafe.Pointer(x)))
   845  	}
   846  	if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
   847  		throw("remaining pointer buffers")
   848  	}
   849  }
   850  
   851  // Scan a stack frame: local variables and function arguments/results.
   852  //go:nowritebarrier
   853  func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
   854  	if _DebugGC > 1 && frame.continpc != 0 {
   855  		print("scanframe ", funcname(frame.fn), "\n")
   856  	}
   857  
   858  	isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
   859  	if state.conservative || isAsyncPreempt {
   860  		if debugScanConservative {
   861  			println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
   862  		}
   863  
   864  		// Conservatively scan the frame. Unlike the precise
   865  		// case, this includes the outgoing argument space
   866  		// since we may have stopped while this function was
   867  		// setting up a call.
   868  		//
   869  		// TODO: We could narrow this down if the compiler
   870  		// produced a single map per function of stack slots
   871  		// and registers that ever contain a pointer.
   872  		if frame.varp != 0 {
   873  			size := frame.varp - frame.sp
   874  			if size > 0 {
   875  				scanConservative(frame.sp, size, nil, gcw, state)
   876  			}
   877  		}
   878  
   879  		// Scan arguments to this frame.
   880  		if frame.arglen != 0 {
   881  			// TODO: We could pass the entry argument map
   882  			// to narrow this down further.
   883  			scanConservative(frame.argp, frame.arglen, nil, gcw, state)
   884  		}
   885  
   886  		if isAsyncPreempt {
   887  			// This function's frame contained the
   888  			// registers for the asynchronously stopped
   889  			// parent frame. Scan the parent
   890  			// conservatively.
   891  			state.conservative = true
   892  		} else {
   893  			// We only wanted to scan those two frames
   894  			// conservatively. Clear the flag for future
   895  			// frames.
   896  			state.conservative = false
   897  		}
   898  		return
   899  	}
   900  
   901  	locals, args, objs := getStackMap(frame, &state.cache, false)
   902  
   903  	// Scan local variables if stack frame has been allocated.
   904  	if locals.n > 0 {
   905  		size := uintptr(locals.n) * sys.PtrSize
   906  		scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
   907  	}
   908  
   909  	// Scan arguments.
   910  	if args.n > 0 {
   911  		scanblock(frame.argp, uintptr(args.n)*sys.PtrSize, args.bytedata, gcw, state)
   912  	}
   913  
   914  	// Add all stack objects to the stack object list.
   915  	if frame.varp != 0 {
   916  		// varp is 0 for defers, where there are no locals.
   917  		// In that case, there can't be a pointer to its args, either.
   918  		// (And all args would be scanned above anyway.)
   919  		for _, obj := range objs {
   920  			off := obj.off
   921  			base := frame.varp // locals base pointer
   922  			if off >= 0 {
   923  				base = frame.argp // arguments and return values base pointer
   924  			}
   925  			ptr := base + uintptr(off)
   926  			if ptr < frame.sp {
   927  				// object hasn't been allocated in the frame yet.
   928  				continue
   929  			}
   930  			if stackTraceDebug {
   931  				println("stkobj at", hex(ptr), "of type", obj.typ.string())
   932  			}
   933  			state.addObject(ptr, obj.typ)
   934  		}
   935  	}
   936  }
   937  
   938  type gcDrainFlags int
   939  
   940  const (
   941  	gcDrainUntilPreempt gcDrainFlags = 1 << iota
   942  	gcDrainFlushBgCredit
   943  	gcDrainIdle
   944  	gcDrainFractional
   945  )
   946  
   947  // gcDrain scans roots and objects in work buffers, blackening grey
   948  // objects until it is unable to get more work. It may return before
   949  // GC is done; it's the caller's responsibility to balance work from
   950  // other Ps.
   951  //
   952  // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
   953  // is set.
   954  //
   955  // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
   956  // to do.
   957  //
   958  // If flags&gcDrainFractional != 0, gcDrain self-preempts when
   959  // pollFractionalWorkerExit() returns true. This implies
   960  // gcDrainNoBlock.
   961  //
   962  // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
   963  // credit to gcController.bgScanCredit every gcCreditSlack units of
   964  // scan work.
   965  //
   966  //go:nowritebarrier
   967  func gcDrain(gcw *gcWork, flags gcDrainFlags) {
   968  	if !writeBarrier.needed {
   969  		throw("gcDrain phase incorrect")
   970  	}
   971  
   972  	gp := getg().m.curg
   973  	preemptible := flags&gcDrainUntilPreempt != 0
   974  	flushBgCredit := flags&gcDrainFlushBgCredit != 0
   975  	idle := flags&gcDrainIdle != 0
   976  
   977  	initScanWork := gcw.scanWork
   978  
   979  	// checkWork is the scan work before performing the next
   980  	// self-preempt check.
   981  	checkWork := int64(1<<63 - 1)
   982  	var check func() bool
   983  	if flags&(gcDrainIdle|gcDrainFractional) != 0 {
   984  		checkWork = initScanWork + drainCheckThreshold
   985  		if idle {
   986  			check = pollWork
   987  		} else if flags&gcDrainFractional != 0 {
   988  			check = pollFractionalWorkerExit
   989  		}
   990  	}
   991  
   992  	// Drain root marking jobs.
   993  	if work.markrootNext < work.markrootJobs {
   994  		for !(preemptible && gp.preempt) {
   995  			job := atomic.Xadd(&work.markrootNext, +1) - 1
   996  			if job >= work.markrootJobs {
   997  				break
   998  			}
   999  			markroot(gcw, job)
  1000  			if check != nil && check() {
  1001  				goto done
  1002  			}
  1003  		}
  1004  	}
  1005  
  1006  	// Drain heap marking jobs.
  1007  	for !(preemptible && gp.preempt) {
  1008  		// Try to keep work available on the global queue. We used to
  1009  		// check if there were waiting workers, but it's better to
  1010  		// just keep work available than to make workers wait. In the
  1011  		// worst case, we'll do O(log(_WorkbufSize)) unnecessary
  1012  		// balances.
  1013  		if work.full == 0 {
  1014  			gcw.balance()
  1015  		}
  1016  
  1017  		b := gcw.tryGetFast()
  1018  		if b == 0 {
  1019  			b = gcw.tryGet()
  1020  			if b == 0 {
  1021  				// Flush the write barrier
  1022  				// buffer; this may create
  1023  				// more work.
  1024  				wbBufFlush(nil, 0)
  1025  				b = gcw.tryGet()
  1026  			}
  1027  		}
  1028  		if b == 0 {
  1029  			// Unable to get work.
  1030  			break
  1031  		}
  1032  		scanobject(b, gcw)
  1033  
  1034  		// Flush background scan work credit to the global
  1035  		// account if we've accumulated enough locally so
  1036  		// mutator assists can draw on it.
  1037  		if gcw.scanWork >= gcCreditSlack {
  1038  			atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1039  			if flushBgCredit {
  1040  				gcFlushBgCredit(gcw.scanWork - initScanWork)
  1041  				initScanWork = 0
  1042  			}
  1043  			checkWork -= gcw.scanWork
  1044  			gcw.scanWork = 0
  1045  
  1046  			if checkWork <= 0 {
  1047  				checkWork += drainCheckThreshold
  1048  				if check != nil && check() {
  1049  					break
  1050  				}
  1051  			}
  1052  		}
  1053  	}
  1054  
  1055  done:
  1056  	// Flush remaining scan work credit.
  1057  	if gcw.scanWork > 0 {
  1058  		atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1059  		if flushBgCredit {
  1060  			gcFlushBgCredit(gcw.scanWork - initScanWork)
  1061  		}
  1062  		gcw.scanWork = 0
  1063  	}
  1064  }
  1065  
  1066  // gcDrainN blackens grey objects until it has performed roughly
  1067  // scanWork units of scan work or the G is preempted. This is
  1068  // best-effort, so it may perform less work if it fails to get a work
  1069  // buffer. Otherwise, it will perform at least n units of work, but
  1070  // may perform more because scanning is always done in whole object
  1071  // increments. It returns the amount of scan work performed.
  1072  //
  1073  // The caller goroutine must be in a preemptible state (e.g.,
  1074  // _Gwaiting) to prevent deadlocks during stack scanning. As a
  1075  // consequence, this must be called on the system stack.
  1076  //
  1077  //go:nowritebarrier
  1078  //go:systemstack
  1079  func gcDrainN(gcw *gcWork, scanWork int64) int64 {
  1080  	if !writeBarrier.needed {
  1081  		throw("gcDrainN phase incorrect")
  1082  	}
  1083  
  1084  	// There may already be scan work on the gcw, which we don't
  1085  	// want to claim was done by this call.
  1086  	workFlushed := -gcw.scanWork
  1087  
  1088  	gp := getg().m.curg
  1089  	for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
  1090  		// See gcDrain comment.
  1091  		if work.full == 0 {
  1092  			gcw.balance()
  1093  		}
  1094  
  1095  		// This might be a good place to add prefetch code...
  1096  		// if(wbuf.nobj > 4) {
  1097  		//         PREFETCH(wbuf->obj[wbuf.nobj - 3];
  1098  		//  }
  1099  		//
  1100  		b := gcw.tryGetFast()
  1101  		if b == 0 {
  1102  			b = gcw.tryGet()
  1103  			if b == 0 {
  1104  				// Flush the write barrier buffer;
  1105  				// this may create more work.
  1106  				wbBufFlush(nil, 0)
  1107  				b = gcw.tryGet()
  1108  			}
  1109  		}
  1110  
  1111  		if b == 0 {
  1112  			// Try to do a root job.
  1113  			//
  1114  			// TODO: Assists should get credit for this
  1115  			// work.
  1116  			if work.markrootNext < work.markrootJobs {
  1117  				job := atomic.Xadd(&work.markrootNext, +1) - 1
  1118  				if job < work.markrootJobs {
  1119  					markroot(gcw, job)
  1120  					continue
  1121  				}
  1122  			}
  1123  			// No heap or root jobs.
  1124  			break
  1125  		}
  1126  		scanobject(b, gcw)
  1127  
  1128  		// Flush background scan work credit.
  1129  		if gcw.scanWork >= gcCreditSlack {
  1130  			atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1131  			workFlushed += gcw.scanWork
  1132  			gcw.scanWork = 0
  1133  		}
  1134  	}
  1135  
  1136  	// Unlike gcDrain, there's no need to flush remaining work
  1137  	// here because this never flushes to bgScanCredit and
  1138  	// gcw.dispose will flush any remaining work to scanWork.
  1139  
  1140  	return workFlushed + gcw.scanWork
  1141  }
  1142  
  1143  // scanblock scans b as scanobject would, but using an explicit
  1144  // pointer bitmap instead of the heap bitmap.
  1145  //
  1146  // This is used to scan non-heap roots, so it does not update
  1147  // gcw.bytesMarked or gcw.scanWork.
  1148  //
  1149  // If stk != nil, possible stack pointers are also reported to stk.putPtr.
  1150  //go:nowritebarrier
  1151  func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
  1152  	// Use local copies of original parameters, so that a stack trace
  1153  	// due to one of the throws below shows the original block
  1154  	// base and extent.
  1155  	b := b0
  1156  	n := n0
  1157  
  1158  	for i := uintptr(0); i < n; {
  1159  		// Find bits for the next word.
  1160  		bits := uint32(*addb(ptrmask, i/(sys.PtrSize*8)))
  1161  		if bits == 0 {
  1162  			i += sys.PtrSize * 8
  1163  			continue
  1164  		}
  1165  		for j := 0; j < 8 && i < n; j++ {
  1166  			if bits&1 != 0 {
  1167  				// Same work as in scanobject; see comments there.
  1168  				p := *(*uintptr)(unsafe.Pointer(b + i))
  1169  				if p != 0 {
  1170  					if obj, span, objIndex := findObject(p, b, i); obj != 0 {
  1171  						greyobject(obj, b, i, span, gcw, objIndex)
  1172  					} else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
  1173  						stk.putPtr(p, false)
  1174  					}
  1175  				}
  1176  			}
  1177  			bits >>= 1
  1178  			i += sys.PtrSize
  1179  		}
  1180  	}
  1181  }
  1182  
  1183  // scanobject scans the object starting at b, adding pointers to gcw.
  1184  // b must point to the beginning of a heap object or an oblet.
  1185  // scanobject consults the GC bitmap for the pointer mask and the
  1186  // spans for the size of the object.
  1187  //
  1188  //go:nowritebarrier
  1189  func scanobject(b uintptr, gcw *gcWork) {
  1190  	// Find the bits for b and the size of the object at b.
  1191  	//
  1192  	// b is either the beginning of an object, in which case this
  1193  	// is the size of the object to scan, or it points to an
  1194  	// oblet, in which case we compute the size to scan below.
  1195  	hbits := heapBitsForAddr(b)
  1196  	s := spanOfUnchecked(b)
  1197  	n := s.elemsize
  1198  	if n == 0 {
  1199  		throw("scanobject n == 0")
  1200  	}
  1201  
  1202  	if n > maxObletBytes {
  1203  		// Large object. Break into oblets for better
  1204  		// parallelism and lower latency.
  1205  		if b == s.base() {
  1206  			// It's possible this is a noscan object (not
  1207  			// from greyobject, but from other code
  1208  			// paths), in which case we must *not* enqueue
  1209  			// oblets since their bitmaps will be
  1210  			// uninitialized.
  1211  			if s.spanclass.noscan() {
  1212  				// Bypass the whole scan.
  1213  				gcw.bytesMarked += uint64(n)
  1214  				return
  1215  			}
  1216  
  1217  			// Enqueue the other oblets to scan later.
  1218  			// Some oblets may be in b's scalar tail, but
  1219  			// these will be marked as "no more pointers",
  1220  			// so we'll drop out immediately when we go to
  1221  			// scan those.
  1222  			for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
  1223  				if !gcw.putFast(oblet) {
  1224  					gcw.put(oblet)
  1225  				}
  1226  			}
  1227  		}
  1228  
  1229  		// Compute the size of the oblet. Since this object
  1230  		// must be a large object, s.base() is the beginning
  1231  		// of the object.
  1232  		n = s.base() + s.elemsize - b
  1233  		if n > maxObletBytes {
  1234  			n = maxObletBytes
  1235  		}
  1236  	}
  1237  
  1238  	var i uintptr
  1239  	for i = 0; i < n; i += sys.PtrSize {
  1240  		// Find bits for this word.
  1241  		if i != 0 {
  1242  			// Avoid needless hbits.next() on last iteration.
  1243  			hbits = hbits.next()
  1244  		}
  1245  		// Load bits once. See CL 22712 and issue 16973 for discussion.
  1246  		bits := hbits.bits()
  1247  		// During checkmarking, 1-word objects store the checkmark
  1248  		// in the type bit for the one word. The only one-word objects
  1249  		// are pointers, or else they'd be merged with other non-pointer
  1250  		// data into larger allocations.
  1251  		if i != 1*sys.PtrSize && bits&bitScan == 0 {
  1252  			break // no more pointers in this object
  1253  		}
  1254  		if bits&bitPointer == 0 {
  1255  			continue // not a pointer
  1256  		}
  1257  
  1258  		// Work here is duplicated in scanblock and above.
  1259  		// If you make changes here, make changes there too.
  1260  		obj := *(*uintptr)(unsafe.Pointer(b + i))
  1261  
  1262  		// At this point we have extracted the next potential pointer.
  1263  		// Quickly filter out nil and pointers back to the current object.
  1264  		if obj != 0 && obj-b >= n {
  1265  			// Test if obj points into the Go heap and, if so,
  1266  			// mark the object.
  1267  			//
  1268  			// Note that it's possible for findObject to
  1269  			// fail if obj points to a just-allocated heap
  1270  			// object because of a race with growing the
  1271  			// heap. In this case, we know the object was
  1272  			// just allocated and hence will be marked by
  1273  			// allocation itself.
  1274  			if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
  1275  				greyobject(obj, b, i, span, gcw, objIndex)
  1276  			}
  1277  		}
  1278  	}
  1279  	gcw.bytesMarked += uint64(n)
  1280  	gcw.scanWork += int64(i)
  1281  }
  1282  
  1283  // scanConservative scans block [b, b+n) conservatively, treating any
  1284  // pointer-like value in the block as a pointer.
  1285  //
  1286  // If ptrmask != nil, only words that are marked in ptrmask are
  1287  // considered as potential pointers.
  1288  //
  1289  // If state != nil, it's assumed that [b, b+n) is a block in the stack
  1290  // and may contain pointers to stack objects.
  1291  func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
  1292  	if debugScanConservative {
  1293  		printlock()
  1294  		print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
  1295  		hexdumpWords(b, b+n, func(p uintptr) byte {
  1296  			if ptrmask != nil {
  1297  				word := (p - b) / sys.PtrSize
  1298  				bits := *addb(ptrmask, word/8)
  1299  				if (bits>>(word%8))&1 == 0 {
  1300  					return '$'
  1301  				}
  1302  			}
  1303  
  1304  			val := *(*uintptr)(unsafe.Pointer(p))
  1305  			if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1306  				return '@'
  1307  			}
  1308  
  1309  			span := spanOfHeap(val)
  1310  			if span == nil {
  1311  				return ' '
  1312  			}
  1313  			idx := span.objIndex(val)
  1314  			if span.isFree(idx) {
  1315  				return ' '
  1316  			}
  1317  			return '*'
  1318  		})
  1319  		printunlock()
  1320  	}
  1321  
  1322  	for i := uintptr(0); i < n; i += sys.PtrSize {
  1323  		if ptrmask != nil {
  1324  			word := i / sys.PtrSize
  1325  			bits := *addb(ptrmask, word/8)
  1326  			if bits == 0 {
  1327  				// Skip 8 words (the loop increment will do the 8th)
  1328  				//
  1329  				// This must be the first time we've
  1330  				// seen this word of ptrmask, so i
  1331  				// must be 8-word-aligned, but check
  1332  				// our reasoning just in case.
  1333  				if i%(sys.PtrSize*8) != 0 {
  1334  					throw("misaligned mask")
  1335  				}
  1336  				i += sys.PtrSize*8 - sys.PtrSize
  1337  				continue
  1338  			}
  1339  			if (bits>>(word%8))&1 == 0 {
  1340  				continue
  1341  			}
  1342  		}
  1343  
  1344  		val := *(*uintptr)(unsafe.Pointer(b + i))
  1345  
  1346  		// Check if val points into the stack.
  1347  		if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1348  			// val may point to a stack object. This
  1349  			// object may be dead from last cycle and
  1350  			// hence may contain pointers to unallocated
  1351  			// objects, but unlike heap objects we can't
  1352  			// tell if it's already dead. Hence, if all
  1353  			// pointers to this object are from
  1354  			// conservative scanning, we have to scan it
  1355  			// defensively, too.
  1356  			state.putPtr(val, true)
  1357  			continue
  1358  		}
  1359  
  1360  		// Check if val points to a heap span.
  1361  		span := spanOfHeap(val)
  1362  		if span == nil {
  1363  			continue
  1364  		}
  1365  
  1366  		// Check if val points to an allocated object.
  1367  		idx := span.objIndex(val)
  1368  		if span.isFree(idx) {
  1369  			continue
  1370  		}
  1371  
  1372  		// val points to an allocated object. Mark it.
  1373  		obj := span.base() + idx*span.elemsize
  1374  		greyobject(obj, b, i, span, gcw, idx)
  1375  	}
  1376  }
  1377  
  1378  // Shade the object if it isn't already.
  1379  // The object is not nil and known to be in the heap.
  1380  // Preemption must be disabled.
  1381  //go:nowritebarrier
  1382  func shade(b uintptr) {
  1383  	if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
  1384  		gcw := &getg().m.p.ptr().gcw
  1385  		greyobject(obj, 0, 0, span, gcw, objIndex)
  1386  	}
  1387  }
  1388  
  1389  // obj is the start of an object with mark mbits.
  1390  // If it isn't already marked, mark it and enqueue into gcw.
  1391  // base and off are for debugging only and could be removed.
  1392  //
  1393  // See also wbBufFlush1, which partially duplicates this logic.
  1394  //
  1395  //go:nowritebarrierrec
  1396  func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
  1397  	// obj should be start of allocation, and so must be at least pointer-aligned.
  1398  	if obj&(sys.PtrSize-1) != 0 {
  1399  		throw("greyobject: obj not pointer-aligned")
  1400  	}
  1401  	mbits := span.markBitsForIndex(objIndex)
  1402  
  1403  	if useCheckmark {
  1404  		if !mbits.isMarked() {
  1405  			printlock()
  1406  			print("runtime:greyobject: checkmarks finds unexpected unmarked object obj=", hex(obj), "\n")
  1407  			print("runtime: found obj at *(", hex(base), "+", hex(off), ")\n")
  1408  
  1409  			// Dump the source (base) object
  1410  			gcDumpObject("base", base, off)
  1411  
  1412  			// Dump the object
  1413  			gcDumpObject("obj", obj, ^uintptr(0))
  1414  
  1415  			getg().m.traceback = 2
  1416  			throw("checkmark found unmarked object")
  1417  		}
  1418  		hbits := heapBitsForAddr(obj)
  1419  		if hbits.isCheckmarked(span.elemsize) {
  1420  			return
  1421  		}
  1422  		hbits.setCheckmarked(span.elemsize)
  1423  		if !hbits.isCheckmarked(span.elemsize) {
  1424  			throw("setCheckmarked and isCheckmarked disagree")
  1425  		}
  1426  	} else {
  1427  		if debug.gccheckmark > 0 && span.isFree(objIndex) {
  1428  			print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
  1429  			gcDumpObject("base", base, off)
  1430  			gcDumpObject("obj", obj, ^uintptr(0))
  1431  			getg().m.traceback = 2
  1432  			throw("marking free object")
  1433  		}
  1434  
  1435  		// If marked we have nothing to do.
  1436  		if mbits.isMarked() {
  1437  			return
  1438  		}
  1439  		mbits.setMarked()
  1440  
  1441  		// Mark span.
  1442  		arena, pageIdx, pageMask := pageIndexOf(span.base())
  1443  		if arena.pageMarks[pageIdx]&pageMask == 0 {
  1444  			atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1445  		}
  1446  
  1447  		// If this is a noscan object, fast-track it to black
  1448  		// instead of greying it.
  1449  		if span.spanclass.noscan() {
  1450  			gcw.bytesMarked += uint64(span.elemsize)
  1451  			return
  1452  		}
  1453  	}
  1454  
  1455  	// Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
  1456  	// seems like a nice optimization that can be added back in.
  1457  	// There needs to be time between the PREFETCH and the use.
  1458  	// Previously we put the obj in an 8 element buffer that is drained at a rate
  1459  	// to give the PREFETCH time to do its work.
  1460  	// Use of PREFETCHNTA might be more appropriate than PREFETCH
  1461  	if !gcw.putFast(obj) {
  1462  		gcw.put(obj)
  1463  	}
  1464  }
  1465  
  1466  // gcDumpObject dumps the contents of obj for debugging and marks the
  1467  // field at byte offset off in obj.
  1468  func gcDumpObject(label string, obj, off uintptr) {
  1469  	s := spanOf(obj)
  1470  	print(label, "=", hex(obj))
  1471  	if s == nil {
  1472  		print(" s=nil\n")
  1473  		return
  1474  	}
  1475  	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
  1476  	if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
  1477  		print(mSpanStateNames[state], "\n")
  1478  	} else {
  1479  		print("unknown(", state, ")\n")
  1480  	}
  1481  
  1482  	skipped := false
  1483  	size := s.elemsize
  1484  	if s.state.get() == mSpanManual && size == 0 {
  1485  		// We're printing something from a stack frame. We
  1486  		// don't know how big it is, so just show up to an
  1487  		// including off.
  1488  		size = off + sys.PtrSize
  1489  	}
  1490  	for i := uintptr(0); i < size; i += sys.PtrSize {
  1491  		// For big objects, just print the beginning (because
  1492  		// that usually hints at the object's type) and the
  1493  		// fields around off.
  1494  		if !(i < 128*sys.PtrSize || off-16*sys.PtrSize < i && i < off+16*sys.PtrSize) {
  1495  			skipped = true
  1496  			continue
  1497  		}
  1498  		if skipped {
  1499  			print(" ...\n")
  1500  			skipped = false
  1501  		}
  1502  		print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
  1503  		if i == off {
  1504  			print(" <==")
  1505  		}
  1506  		print("\n")
  1507  	}
  1508  	if skipped {
  1509  		print(" ...\n")
  1510  	}
  1511  }
  1512  
  1513  // gcmarknewobject marks a newly allocated object black. obj must
  1514  // not contain any non-nil pointers.
  1515  //
  1516  // This is nosplit so it can manipulate a gcWork without preemption.
  1517  //
  1518  //go:nowritebarrier
  1519  //go:nosplit
  1520  func gcmarknewobject(obj, size, scanSize uintptr) {
  1521  	if useCheckmark { // The world should be stopped so this should not happen.
  1522  		throw("gcmarknewobject called while doing checkmark")
  1523  	}
  1524  	markBitsForAddr(obj).setMarked()
  1525  	gcw := &getg().m.p.ptr().gcw
  1526  	gcw.bytesMarked += uint64(size)
  1527  	gcw.scanWork += int64(scanSize)
  1528  }
  1529  
  1530  // gcMarkTinyAllocs greys all active tiny alloc blocks.
  1531  //
  1532  // The world must be stopped.
  1533  func gcMarkTinyAllocs() {
  1534  	for _, p := range allp {
  1535  		c := p.mcache
  1536  		if c == nil || c.tiny == 0 {
  1537  			continue
  1538  		}
  1539  		_, span, objIndex := findObject(c.tiny, 0, 0)
  1540  		gcw := &p.gcw
  1541  		greyobject(c.tiny, 0, 0, span, gcw, objIndex)
  1542  	}
  1543  }
  1544  
  1545  // Checkmarking
  1546  
  1547  // To help debug the concurrent GC we remark with the world
  1548  // stopped ensuring that any object encountered has their normal
  1549  // mark bit set. To do this we use an orthogonal bit
  1550  // pattern to indicate the object is marked. The following pattern
  1551  // uses the upper two bits in the object's boundary nibble.
  1552  // 01: scalar  not marked
  1553  // 10: pointer not marked
  1554  // 11: pointer     marked
  1555  // 00: scalar      marked
  1556  // Xoring with 01 will flip the pattern from marked to unmarked and vica versa.
  1557  // The higher bit is 1 for pointers and 0 for scalars, whether the object
  1558  // is marked or not.
  1559  // The first nibble no longer holds the typeDead pattern indicating that the
  1560  // there are no more pointers in the object. This information is held
  1561  // in the second nibble.
  1562  
  1563  // If useCheckmark is true, marking of an object uses the
  1564  // checkmark bits (encoding above) instead of the standard
  1565  // mark bits.
  1566  var useCheckmark = false
  1567  
  1568  //go:nowritebarrier
  1569  func initCheckmarks() {
  1570  	useCheckmark = true
  1571  	for _, s := range mheap_.allspans {
  1572  		if s.state.get() == mSpanInUse {
  1573  			heapBitsForAddr(s.base()).initCheckmarkSpan(s.layout())
  1574  		}
  1575  	}
  1576  }
  1577  
  1578  func clearCheckmarks() {
  1579  	useCheckmark = false
  1580  	for _, s := range mheap_.allspans {
  1581  		if s.state.get() == mSpanInUse {
  1582  			heapBitsForAddr(s.base()).clearCheckmarkSpan(s.layout())
  1583  		}
  1584  	}
  1585  }
  1586  

View as plain text