Source file src/runtime/mgcmark.go

     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  	"internal/goarch"
    11  	"internal/goexperiment"
    12  	"runtime/internal/atomic"
    13  	"runtime/internal/sys"
    14  	"unsafe"
    15  )
    16  
    17  const (
    18  	fixedRootFinalizers = iota
    19  	fixedRootFreeGStacks
    20  	fixedRootCount
    21  
    22  	// rootBlockBytes is the number of bytes to scan per data or
    23  	// BSS root.
    24  	rootBlockBytes = 256 << 10
    25  
    26  	// maxObletBytes is the maximum bytes of an object to scan at
    27  	// once. Larger objects will be split up into "oblets" of at
    28  	// most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
    29  	// scan preemption at ~100 µs.
    30  	//
    31  	// This must be > _MaxSmallSize so that the object base is the
    32  	// span base.
    33  	maxObletBytes = 128 << 10
    34  
    35  	// drainCheckThreshold specifies how many units of work to do
    36  	// between self-preemption checks in gcDrain. Assuming a scan
    37  	// rate of 1 MB/ms, this is ~100 µs. Lower values have higher
    38  	// overhead in the scan loop (the scheduler check may perform
    39  	// a syscall, so its overhead is nontrivial). Higher values
    40  	// make the system less responsive to incoming work.
    41  	drainCheckThreshold = 100000
    42  
    43  	// pagesPerSpanRoot indicates how many pages to scan from a span root
    44  	// at a time. Used by special root marking.
    45  	//
    46  	// Higher values improve throughput by increasing locality, but
    47  	// increase the minimum latency of a marking operation.
    48  	//
    49  	// Must be a multiple of the pageInUse bitmap element size and
    50  	// must also evenly divide pagesPerArena.
    51  	pagesPerSpanRoot = 512
    52  )
    53  
    54  // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
    55  // some miscellany) and initializes scanning-related state.
    56  //
    57  // The world must be stopped.
    58  func gcMarkRootPrepare() {
    59  	assertWorldStopped()
    60  
    61  	// Compute how many data and BSS root blocks there are.
    62  	nBlocks := func(bytes uintptr) int {
    63  		return int(divRoundUp(bytes, rootBlockBytes))
    64  	}
    65  
    66  	work.nDataRoots = 0
    67  	work.nBSSRoots = 0
    68  
    69  	// Scan globals.
    70  	for _, datap := range activeModules() {
    71  		nDataRoots := nBlocks(datap.edata - datap.data)
    72  		if nDataRoots > work.nDataRoots {
    73  			work.nDataRoots = nDataRoots
    74  		}
    75  	}
    76  
    77  	for _, datap := range activeModules() {
    78  		nBSSRoots := nBlocks(datap.ebss - datap.bss)
    79  		if nBSSRoots > work.nBSSRoots {
    80  			work.nBSSRoots = nBSSRoots
    81  		}
    82  	}
    83  
    84  	// Scan span roots for finalizer specials.
    85  	//
    86  	// We depend on addfinalizer to mark objects that get
    87  	// finalizers after root marking.
    88  	//
    89  	// We're going to scan the whole heap (that was available at the time the
    90  	// mark phase started, i.e. markArenas) for in-use spans which have specials.
    91  	//
    92  	// Break up the work into arenas, and further into chunks.
    93  	//
    94  	// Snapshot allArenas as markArenas. This snapshot is safe because allArenas
    95  	// is append-only.
    96  	mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
    97  	work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
    98  
    99  	// Scan stacks.
   100  	//
   101  	// Gs may be created after this point, but it's okay that we
   102  	// ignore them because they begin life without any roots, so
   103  	// there's nothing to scan, and any roots they create during
   104  	// the concurrent phase will be caught by the write barrier.
   105  	work.stackRoots = allGsSnapshot()
   106  	work.nStackRoots = len(work.stackRoots)
   107  
   108  	work.markrootNext = 0
   109  	work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
   110  
   111  	// Calculate base indexes of each root type
   112  	work.baseData = uint32(fixedRootCount)
   113  	work.baseBSS = work.baseData + uint32(work.nDataRoots)
   114  	work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
   115  	work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
   116  	work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
   117  }
   118  
   119  // gcMarkRootCheck checks that all roots have been scanned. It is
   120  // purely for debugging.
   121  func gcMarkRootCheck() {
   122  	if work.markrootNext < work.markrootJobs {
   123  		print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
   124  		throw("left over markroot jobs")
   125  	}
   126  
   127  	// Check that stacks have been scanned.
   128  	//
   129  	// We only check the first nStackRoots Gs that we should have scanned.
   130  	// Since we don't care about newer Gs (see comment in
   131  	// gcMarkRootPrepare), no locking is required.
   132  	i := 0
   133  	forEachGRace(func(gp *g) {
   134  		if i >= work.nStackRoots {
   135  			return
   136  		}
   137  
   138  		if !gp.gcscandone {
   139  			println("gp", gp, "goid", gp.goid,
   140  				"status", readgstatus(gp),
   141  				"gcscandone", gp.gcscandone)
   142  			throw("scan missed a g")
   143  		}
   144  
   145  		i++
   146  	})
   147  }
   148  
   149  // ptrmask for an allocation containing a single pointer.
   150  var oneptrmask = [...]uint8{1}
   151  
   152  // markroot scans the i'th root.
   153  //
   154  // Preemption must be disabled (because this uses a gcWork).
   155  //
   156  // Returns the amount of GC work credit produced by the operation.
   157  // If flushBgCredit is true, then that credit is also flushed
   158  // to the background credit pool.
   159  //
   160  // nowritebarrier is only advisory here.
   161  //
   162  //go:nowritebarrier
   163  func markroot(gcw *gcWork, i uint32, flushBgCredit bool) int64 {
   164  	// Note: if you add a case here, please also update heapdump.go:dumproots.
   165  	var workDone int64
   166  	var workCounter *atomic.Int64
   167  	switch {
   168  	case work.baseData <= i && i < work.baseBSS:
   169  		workCounter = &gcController.globalsScanWork
   170  		for _, datap := range activeModules() {
   171  			workDone += markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
   172  		}
   173  
   174  	case work.baseBSS <= i && i < work.baseSpans:
   175  		workCounter = &gcController.globalsScanWork
   176  		for _, datap := range activeModules() {
   177  			workDone += markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
   178  		}
   179  
   180  	case i == fixedRootFinalizers:
   181  		for fb := allfin; fb != nil; fb = fb.alllink {
   182  			cnt := uintptr(atomic.Load(&fb.cnt))
   183  			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
   184  		}
   185  
   186  	case i == fixedRootFreeGStacks:
   187  		// Switch to the system stack so we can call
   188  		// stackfree.
   189  		systemstack(markrootFreeGStacks)
   190  
   191  	case work.baseSpans <= i && i < work.baseStacks:
   192  		// mark mspan.specials
   193  		markrootSpans(gcw, int(i-work.baseSpans))
   194  
   195  	default:
   196  		// the rest is scanning goroutine stacks
   197  		workCounter = &gcController.stackScanWork
   198  		if i < work.baseStacks || work.baseEnd <= i {
   199  			printlock()
   200  			print("runtime: markroot index ", i, " not in stack roots range [", work.baseStacks, ", ", work.baseEnd, ")\n")
   201  			throw("markroot: bad index")
   202  		}
   203  		gp := work.stackRoots[i-work.baseStacks]
   204  
   205  		// remember when we've first observed the G blocked
   206  		// needed only to output in traceback
   207  		status := readgstatus(gp) // We are not in a scan state
   208  		if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
   209  			gp.waitsince = work.tstart
   210  		}
   211  
   212  		// scanstack must be done on the system stack in case
   213  		// we're trying to scan our own stack.
   214  		systemstack(func() {
   215  			// If this is a self-scan, put the user G in
   216  			// _Gwaiting to prevent self-deadlock. It may
   217  			// already be in _Gwaiting if this is a mark
   218  			// worker or we're in mark termination.
   219  			userG := getg().m.curg
   220  			selfScan := gp == userG && readgstatus(userG) == _Grunning
   221  			if selfScan {
   222  				casgstatus(userG, _Grunning, _Gwaiting)
   223  				userG.waitreason = waitReasonGarbageCollectionScan
   224  			}
   225  
   226  			// TODO: suspendG blocks (and spins) until gp
   227  			// stops, which may take a while for
   228  			// running goroutines. Consider doing this in
   229  			// two phases where the first is non-blocking:
   230  			// we scan the stacks we can and ask running
   231  			// goroutines to scan themselves; and the
   232  			// second blocks.
   233  			stopped := suspendG(gp)
   234  			if stopped.dead {
   235  				gp.gcscandone = true
   236  				return
   237  			}
   238  			if gp.gcscandone {
   239  				throw("g already scanned")
   240  			}
   241  			workDone += scanstack(gp, gcw)
   242  			gp.gcscandone = true
   243  			resumeG(stopped)
   244  
   245  			if selfScan {
   246  				casgstatus(userG, _Gwaiting, _Grunning)
   247  			}
   248  		})
   249  	}
   250  	if goexperiment.PacerRedesign {
   251  		if workCounter != nil && workDone != 0 {
   252  			workCounter.Add(workDone)
   253  			if flushBgCredit {
   254  				gcFlushBgCredit(workDone)
   255  			}
   256  		}
   257  	}
   258  	return workDone
   259  }
   260  
   261  // markrootBlock scans the shard'th shard of the block of memory [b0,
   262  // b0+n0), with the given pointer mask.
   263  //
   264  // Returns the amount of work done.
   265  //
   266  //go:nowritebarrier
   267  func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
   268  	if rootBlockBytes%(8*goarch.PtrSize) != 0 {
   269  		// This is necessary to pick byte offsets in ptrmask0.
   270  		throw("rootBlockBytes must be a multiple of 8*ptrSize")
   271  	}
   272  
   273  	// Note that if b0 is toward the end of the address space,
   274  	// then b0 + rootBlockBytes might wrap around.
   275  	// These tests are written to avoid any possible overflow.
   276  	off := uintptr(shard) * rootBlockBytes
   277  	if off >= n0 {
   278  		return 0
   279  	}
   280  	b := b0 + off
   281  	ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
   282  	n := uintptr(rootBlockBytes)
   283  	if off+n > n0 {
   284  		n = n0 - off
   285  	}
   286  
   287  	// Scan this shard.
   288  	scanblock(b, n, ptrmask, gcw, nil)
   289  	return int64(n)
   290  }
   291  
   292  // markrootFreeGStacks frees stacks of dead Gs.
   293  //
   294  // This does not free stacks of dead Gs cached on Ps, but having a few
   295  // cached stacks around isn't a problem.
   296  func markrootFreeGStacks() {
   297  	// Take list of dead Gs with stacks.
   298  	lock(&sched.gFree.lock)
   299  	list := sched.gFree.stack
   300  	sched.gFree.stack = gList{}
   301  	unlock(&sched.gFree.lock)
   302  	if list.empty() {
   303  		return
   304  	}
   305  
   306  	// Free stacks.
   307  	q := gQueue{list.head, list.head}
   308  	for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
   309  		stackfree(gp.stack)
   310  		gp.stack.lo = 0
   311  		gp.stack.hi = 0
   312  		// Manipulate the queue directly since the Gs are
   313  		// already all linked the right way.
   314  		q.tail.set(gp)
   315  	}
   316  
   317  	// Put Gs back on the free list.
   318  	lock(&sched.gFree.lock)
   319  	sched.gFree.noStack.pushAll(q)
   320  	unlock(&sched.gFree.lock)
   321  }
   322  
   323  // markrootSpans marks roots for one shard of markArenas.
   324  //
   325  //go:nowritebarrier
   326  func markrootSpans(gcw *gcWork, shard int) {
   327  	// Objects with finalizers have two GC-related invariants:
   328  	//
   329  	// 1) Everything reachable from the object must be marked.
   330  	// This ensures that when we pass the object to its finalizer,
   331  	// everything the finalizer can reach will be retained.
   332  	//
   333  	// 2) Finalizer specials (which are not in the garbage
   334  	// collected heap) are roots. In practice, this means the fn
   335  	// field must be scanned.
   336  	sg := mheap_.sweepgen
   337  
   338  	// Find the arena and page index into that arena for this shard.
   339  	ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
   340  	ha := mheap_.arenas[ai.l1()][ai.l2()]
   341  	arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
   342  
   343  	// Construct slice of bitmap which we'll iterate over.
   344  	specialsbits := ha.pageSpecials[arenaPage/8:]
   345  	specialsbits = specialsbits[:pagesPerSpanRoot/8]
   346  	for i := range specialsbits {
   347  		// Find set bits, which correspond to spans with specials.
   348  		specials := atomic.Load8(&specialsbits[i])
   349  		if specials == 0 {
   350  			continue
   351  		}
   352  		for j := uint(0); j < 8; j++ {
   353  			if specials&(1<<j) == 0 {
   354  				continue
   355  			}
   356  			// Find the span for this bit.
   357  			//
   358  			// This value is guaranteed to be non-nil because having
   359  			// specials implies that the span is in-use, and since we're
   360  			// currently marking we can be sure that we don't have to worry
   361  			// about the span being freed and re-used.
   362  			s := ha.spans[arenaPage+uint(i)*8+j]
   363  
   364  			// The state must be mSpanInUse if the specials bit is set, so
   365  			// sanity check that.
   366  			if state := s.state.get(); state != mSpanInUse {
   367  				print("s.state = ", state, "\n")
   368  				throw("non in-use span found with specials bit set")
   369  			}
   370  			// Check that this span was swept (it may be cached or uncached).
   371  			if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
   372  				// sweepgen was updated (+2) during non-checkmark GC pass
   373  				print("sweep ", s.sweepgen, " ", sg, "\n")
   374  				throw("gc: unswept span")
   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  			for sp := s.specials; sp != nil; sp = sp.next {
   381  				if sp.kind != _KindSpecialFinalizer {
   382  					continue
   383  				}
   384  				// don't mark finalized object, but scan it so we
   385  				// retain everything it points to.
   386  				spf := (*specialfinalizer)(unsafe.Pointer(sp))
   387  				// A finalizer can be set for an inner byte of an object, find object beginning.
   388  				p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
   389  
   390  				// Mark everything that can be reached from
   391  				// the object (but *not* the object itself or
   392  				// we'll never collect it).
   393  				scanobject(p, gcw)
   394  
   395  				// The special itself is a root.
   396  				scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
   397  			}
   398  			unlock(&s.speciallock)
   399  		}
   400  	}
   401  }
   402  
   403  // gcAssistAlloc performs GC work to make gp's assist debt positive.
   404  // gp must be the calling user goroutine.
   405  //
   406  // This must be called with preemption enabled.
   407  func gcAssistAlloc(gp *g) {
   408  	// Don't assist in non-preemptible contexts. These are
   409  	// generally fragile and won't allow the assist to block.
   410  	if getg() == gp.m.g0 {
   411  		return
   412  	}
   413  	if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
   414  		return
   415  	}
   416  
   417  	traced := false
   418  retry:
   419  	// Compute the amount of scan work we need to do to make the
   420  	// balance positive. When the required amount of work is low,
   421  	// we over-assist to build up credit for future allocations
   422  	// and amortize the cost of assisting.
   423  	assistWorkPerByte := gcController.assistWorkPerByte.Load()
   424  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   425  	debtBytes := -gp.gcAssistBytes
   426  	scanWork := int64(assistWorkPerByte * float64(debtBytes))
   427  	if scanWork < gcOverAssistWork {
   428  		scanWork = gcOverAssistWork
   429  		debtBytes = int64(assistBytesPerWork * float64(scanWork))
   430  	}
   431  
   432  	// Steal as much credit as we can from the background GC's
   433  	// scan credit. This is racy and may drop the background
   434  	// credit below 0 if two mutators steal at the same time. This
   435  	// will just cause steals to fail until credit is accumulated
   436  	// again, so in the long run it doesn't really matter, but we
   437  	// do have to handle the negative credit case.
   438  	bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
   439  	stolen := int64(0)
   440  	if bgScanCredit > 0 {
   441  		if bgScanCredit < scanWork {
   442  			stolen = bgScanCredit
   443  			gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
   444  		} else {
   445  			stolen = scanWork
   446  			gp.gcAssistBytes += debtBytes
   447  		}
   448  		atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
   449  
   450  		scanWork -= stolen
   451  
   452  		if scanWork == 0 {
   453  			// We were able to steal all of the credit we
   454  			// needed.
   455  			if traced {
   456  				traceGCMarkAssistDone()
   457  			}
   458  			return
   459  		}
   460  	}
   461  
   462  	if trace.enabled && !traced {
   463  		traced = true
   464  		traceGCMarkAssistStart()
   465  	}
   466  
   467  	// Perform assist work
   468  	systemstack(func() {
   469  		gcAssistAlloc1(gp, scanWork)
   470  		// The user stack may have moved, so this can't touch
   471  		// anything on it until it returns from systemstack.
   472  	})
   473  
   474  	completed := gp.param != nil
   475  	gp.param = nil
   476  	if completed {
   477  		gcMarkDone()
   478  	}
   479  
   480  	if gp.gcAssistBytes < 0 {
   481  		// We were unable steal enough credit or perform
   482  		// enough work to pay off the assist debt. We need to
   483  		// do one of these before letting the mutator allocate
   484  		// more to prevent over-allocation.
   485  		//
   486  		// If this is because we were preempted, reschedule
   487  		// and try some more.
   488  		if gp.preempt {
   489  			Gosched()
   490  			goto retry
   491  		}
   492  
   493  		// Add this G to an assist queue and park. When the GC
   494  		// has more background credit, it will satisfy queued
   495  		// assists before flushing to the global credit pool.
   496  		//
   497  		// Note that this does *not* get woken up when more
   498  		// work is added to the work list. The theory is that
   499  		// there wasn't enough work to do anyway, so we might
   500  		// as well let background marking take care of the
   501  		// work that is available.
   502  		if !gcParkAssist() {
   503  			goto retry
   504  		}
   505  
   506  		// At this point either background GC has satisfied
   507  		// this G's assist debt, or the GC cycle is over.
   508  	}
   509  	if traced {
   510  		traceGCMarkAssistDone()
   511  	}
   512  }
   513  
   514  // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
   515  // stack. This is a separate function to make it easier to see that
   516  // we're not capturing anything from the user stack, since the user
   517  // stack may move while we're in this function.
   518  //
   519  // gcAssistAlloc1 indicates whether this assist completed the mark
   520  // phase by setting gp.param to non-nil. This can't be communicated on
   521  // the stack since it may move.
   522  //
   523  //go:systemstack
   524  func gcAssistAlloc1(gp *g, scanWork int64) {
   525  	// Clear the flag indicating that this assist completed the
   526  	// mark phase.
   527  	gp.param = nil
   528  
   529  	if atomic.Load(&gcBlackenEnabled) == 0 {
   530  		// The gcBlackenEnabled check in malloc races with the
   531  		// store that clears it but an atomic check in every malloc
   532  		// would be a performance hit.
   533  		// Instead we recheck it here on the non-preemptable system
   534  		// stack to determine if we should perform an assist.
   535  
   536  		// GC is done, so ignore any remaining debt.
   537  		gp.gcAssistBytes = 0
   538  		return
   539  	}
   540  	// Track time spent in this assist. Since we're on the
   541  	// system stack, this is non-preemptible, so we can
   542  	// just measure start and end time.
   543  	startTime := nanotime()
   544  
   545  	decnwait := atomic.Xadd(&work.nwait, -1)
   546  	if decnwait == work.nproc {
   547  		println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
   548  		throw("nwait > work.nprocs")
   549  	}
   550  
   551  	// gcDrainN requires the caller to be preemptible.
   552  	casgstatus(gp, _Grunning, _Gwaiting)
   553  	gp.waitreason = waitReasonGCAssistMarking
   554  
   555  	// drain own cached work first in the hopes that it
   556  	// will be more cache friendly.
   557  	gcw := &getg().m.p.ptr().gcw
   558  	workDone := gcDrainN(gcw, scanWork)
   559  
   560  	casgstatus(gp, _Gwaiting, _Grunning)
   561  
   562  	// Record that we did this much scan work.
   563  	//
   564  	// Back out the number of bytes of assist credit that
   565  	// this scan work counts for. The "1+" is a poor man's
   566  	// round-up, to ensure this adds credit even if
   567  	// assistBytesPerWork is very low.
   568  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   569  	gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
   570  
   571  	// If this is the last worker and we ran out of work,
   572  	// signal a completion point.
   573  	incnwait := atomic.Xadd(&work.nwait, +1)
   574  	if incnwait > work.nproc {
   575  		println("runtime: work.nwait=", incnwait,
   576  			"work.nproc=", work.nproc)
   577  		throw("work.nwait > work.nproc")
   578  	}
   579  
   580  	if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
   581  		// This has reached a background completion point. Set
   582  		// gp.param to a non-nil value to indicate this. It
   583  		// doesn't matter what we set it to (it just has to be
   584  		// a valid pointer).
   585  		gp.param = unsafe.Pointer(gp)
   586  	}
   587  	duration := nanotime() - startTime
   588  	_p_ := gp.m.p.ptr()
   589  	_p_.gcAssistTime += duration
   590  	if _p_.gcAssistTime > gcAssistTimeSlack {
   591  		atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
   592  		_p_.gcAssistTime = 0
   593  	}
   594  }
   595  
   596  // gcWakeAllAssists wakes all currently blocked assists. This is used
   597  // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
   598  // new assists from going to sleep after this point.
   599  func gcWakeAllAssists() {
   600  	lock(&work.assistQueue.lock)
   601  	list := work.assistQueue.q.popList()
   602  	injectglist(&list)
   603  	unlock(&work.assistQueue.lock)
   604  }
   605  
   606  // gcParkAssist puts the current goroutine on the assist queue and parks.
   607  //
   608  // gcParkAssist reports whether the assist is now satisfied. If it
   609  // returns false, the caller must retry the assist.
   610  func gcParkAssist() bool {
   611  	lock(&work.assistQueue.lock)
   612  	// If the GC cycle finished while we were getting the lock,
   613  	// exit the assist. The cycle can't finish while we hold the
   614  	// lock.
   615  	if atomic.Load(&gcBlackenEnabled) == 0 {
   616  		unlock(&work.assistQueue.lock)
   617  		return true
   618  	}
   619  
   620  	gp := getg()
   621  	oldList := work.assistQueue.q
   622  	work.assistQueue.q.pushBack(gp)
   623  
   624  	// Recheck for background credit now that this G is in
   625  	// the queue, but can still back out. This avoids a
   626  	// race in case background marking has flushed more
   627  	// credit since we checked above.
   628  	if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
   629  		work.assistQueue.q = oldList
   630  		if oldList.tail != 0 {
   631  			oldList.tail.ptr().schedlink.set(nil)
   632  		}
   633  		unlock(&work.assistQueue.lock)
   634  		return false
   635  	}
   636  	// Park.
   637  	goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
   638  	return true
   639  }
   640  
   641  // gcFlushBgCredit flushes scanWork units of background scan work
   642  // credit. This first satisfies blocked assists on the
   643  // work.assistQueue and then flushes any remaining credit to
   644  // gcController.bgScanCredit.
   645  //
   646  // Write barriers are disallowed because this is used by gcDrain after
   647  // it has ensured that all work is drained and this must preserve that
   648  // condition.
   649  //
   650  //go:nowritebarrierrec
   651  func gcFlushBgCredit(scanWork int64) {
   652  	if work.assistQueue.q.empty() {
   653  		// Fast path; there are no blocked assists. There's a
   654  		// small window here where an assist may add itself to
   655  		// the blocked queue and park. If that happens, we'll
   656  		// just get it on the next flush.
   657  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   658  		return
   659  	}
   660  
   661  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   662  	scanBytes := int64(float64(scanWork) * assistBytesPerWork)
   663  
   664  	lock(&work.assistQueue.lock)
   665  	for !work.assistQueue.q.empty() && scanBytes > 0 {
   666  		gp := work.assistQueue.q.pop()
   667  		// Note that gp.gcAssistBytes is negative because gp
   668  		// is in debt. Think carefully about the signs below.
   669  		if scanBytes+gp.gcAssistBytes >= 0 {
   670  			// Satisfy this entire assist debt.
   671  			scanBytes += gp.gcAssistBytes
   672  			gp.gcAssistBytes = 0
   673  			// It's important that we *not* put gp in
   674  			// runnext. Otherwise, it's possible for user
   675  			// code to exploit the GC worker's high
   676  			// scheduler priority to get itself always run
   677  			// before other goroutines and always in the
   678  			// fresh quantum started by GC.
   679  			ready(gp, 0, false)
   680  		} else {
   681  			// Partially satisfy this assist.
   682  			gp.gcAssistBytes += scanBytes
   683  			scanBytes = 0
   684  			// As a heuristic, we move this assist to the
   685  			// back of the queue so that large assists
   686  			// can't clog up the assist queue and
   687  			// substantially delay small assists.
   688  			work.assistQueue.q.pushBack(gp)
   689  			break
   690  		}
   691  	}
   692  
   693  	if scanBytes > 0 {
   694  		// Convert from scan bytes back to work.
   695  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
   696  		scanWork = int64(float64(scanBytes) * assistWorkPerByte)
   697  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   698  	}
   699  	unlock(&work.assistQueue.lock)
   700  }
   701  
   702  // scanstack scans gp's stack, greying all pointers found on the stack.
   703  //
   704  // For goexperiment.PacerRedesign:
   705  // Returns the amount of scan work performed, but doesn't update
   706  // gcController.stackScanWork or flush any credit. Any background credit produced
   707  // by this function should be flushed by its caller. scanstack itself can't
   708  // safely flush because it may result in trying to wake up a goroutine that
   709  // was just scanned, resulting in a self-deadlock.
   710  //
   711  // scanstack will also shrink the stack if it is safe to do so. If it
   712  // is not, it schedules a stack shrink for the next synchronous safe
   713  // point.
   714  //
   715  // scanstack is marked go:systemstack because it must not be preempted
   716  // while using a workbuf.
   717  //
   718  //go:nowritebarrier
   719  //go:systemstack
   720  func scanstack(gp *g, gcw *gcWork) int64 {
   721  	if readgstatus(gp)&_Gscan == 0 {
   722  		print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
   723  		throw("scanstack - bad status")
   724  	}
   725  
   726  	switch readgstatus(gp) &^ _Gscan {
   727  	default:
   728  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   729  		throw("mark - bad status")
   730  	case _Gdead:
   731  		return 0
   732  	case _Grunning:
   733  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   734  		throw("scanstack: goroutine not stopped")
   735  	case _Grunnable, _Gsyscall, _Gwaiting:
   736  		// ok
   737  	}
   738  
   739  	if gp == getg() {
   740  		throw("can't scan our own stack")
   741  	}
   742  
   743  	// stackSize is the amount of work we'll be reporting.
   744  	//
   745  	// We report the total stack size, more than we scan,
   746  	// because this number needs to line up with gcControllerState's
   747  	// stackScan and scannableStackSize fields.
   748  	//
   749  	// See the documentation on those fields for more information.
   750  	stackSize := gp.stack.hi - gp.stack.lo
   751  
   752  	if isShrinkStackSafe(gp) {
   753  		// Shrink the stack if not much of it is being used.
   754  		shrinkstack(gp)
   755  	} else {
   756  		// Otherwise, shrink the stack at the next sync safe point.
   757  		gp.preemptShrink = true
   758  	}
   759  
   760  	var state stackScanState
   761  	state.stack = gp.stack
   762  
   763  	if stackTraceDebug {
   764  		println("stack trace goroutine", gp.goid)
   765  	}
   766  
   767  	if debugScanConservative && gp.asyncSafePoint {
   768  		print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
   769  	}
   770  
   771  	// Scan the saved context register. This is effectively a live
   772  	// register that gets moved back and forth between the
   773  	// register and sched.ctxt without a write barrier.
   774  	if gp.sched.ctxt != nil {
   775  		scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   776  	}
   777  
   778  	// Scan the stack. Accumulate a list of stack objects.
   779  	scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
   780  		scanframeworker(frame, &state, gcw)
   781  		return true
   782  	}
   783  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
   784  
   785  	// Find additional pointers that point into the stack from the heap.
   786  	// Currently this includes defers and panics. See also function copystack.
   787  
   788  	// Find and trace other pointers in defer records.
   789  	for d := gp._defer; d != nil; d = d.link {
   790  		if d.fn != nil {
   791  			// Scan the func value, which could be a stack allocated closure.
   792  			// See issue 30453.
   793  			scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   794  		}
   795  		if d.link != nil {
   796  			// The link field of a stack-allocated defer record might point
   797  			// to a heap-allocated defer record. Keep that heap record live.
   798  			scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   799  		}
   800  		// Retain defers records themselves.
   801  		// Defer records might not be reachable from the G through regular heap
   802  		// tracing because the defer linked list might weave between the stack and the heap.
   803  		if d.heap {
   804  			scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   805  		}
   806  	}
   807  	if gp._panic != nil {
   808  		// Panics are always stack allocated.
   809  		state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
   810  	}
   811  
   812  	// Find and scan all reachable stack objects.
   813  	//
   814  	// The state's pointer queue prioritizes precise pointers over
   815  	// conservative pointers so that we'll prefer scanning stack
   816  	// objects precisely.
   817  	state.buildIndex()
   818  	for {
   819  		p, conservative := state.getPtr()
   820  		if p == 0 {
   821  			break
   822  		}
   823  		obj := state.findObject(p)
   824  		if obj == nil {
   825  			continue
   826  		}
   827  		r := obj.r
   828  		if r == nil {
   829  			// We've already scanned this object.
   830  			continue
   831  		}
   832  		obj.setRecord(nil) // Don't scan it again.
   833  		if stackTraceDebug {
   834  			printlock()
   835  			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
   836  			if conservative {
   837  				print(" (conservative)")
   838  			}
   839  			println()
   840  			printunlock()
   841  		}
   842  		gcdata := r.gcdata()
   843  		var s *mspan
   844  		if r.useGCProg() {
   845  			// This path is pretty unlikely, an object large enough
   846  			// to have a GC program allocated on the stack.
   847  			// We need some space to unpack the program into a straight
   848  			// bitmask, which we allocate/free here.
   849  			// TODO: it would be nice if there were a way to run a GC
   850  			// program without having to store all its bits. We'd have
   851  			// to change from a Lempel-Ziv style program to something else.
   852  			// Or we can forbid putting objects on stacks if they require
   853  			// a gc program (see issue 27447).
   854  			s = materializeGCProg(r.ptrdata(), gcdata)
   855  			gcdata = (*byte)(unsafe.Pointer(s.startAddr))
   856  		}
   857  
   858  		b := state.stack.lo + uintptr(obj.off)
   859  		if conservative {
   860  			scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
   861  		} else {
   862  			scanblock(b, r.ptrdata(), gcdata, gcw, &state)
   863  		}
   864  
   865  		if s != nil {
   866  			dematerializeGCProg(s)
   867  		}
   868  	}
   869  
   870  	// Deallocate object buffers.
   871  	// (Pointer buffers were all deallocated in the loop above.)
   872  	for state.head != nil {
   873  		x := state.head
   874  		state.head = x.next
   875  		if stackTraceDebug {
   876  			for i := 0; i < x.nobj; i++ {
   877  				obj := &x.obj[i]
   878  				if obj.r == nil { // reachable
   879  					continue
   880  				}
   881  				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
   882  				// Note: not necessarily really dead - only reachable-from-ptr dead.
   883  			}
   884  		}
   885  		x.nobj = 0
   886  		putempty((*workbuf)(unsafe.Pointer(x)))
   887  	}
   888  	if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
   889  		throw("remaining pointer buffers")
   890  	}
   891  	return int64(stackSize)
   892  }
   893  
   894  // Scan a stack frame: local variables and function arguments/results.
   895  //go:nowritebarrier
   896  func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
   897  	if _DebugGC > 1 && frame.continpc != 0 {
   898  		print("scanframe ", funcname(frame.fn), "\n")
   899  	}
   900  
   901  	isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
   902  	isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
   903  	if state.conservative || isAsyncPreempt || isDebugCall {
   904  		if debugScanConservative {
   905  			println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
   906  		}
   907  
   908  		// Conservatively scan the frame. Unlike the precise
   909  		// case, this includes the outgoing argument space
   910  		// since we may have stopped while this function was
   911  		// setting up a call.
   912  		//
   913  		// TODO: We could narrow this down if the compiler
   914  		// produced a single map per function of stack slots
   915  		// and registers that ever contain a pointer.
   916  		if frame.varp != 0 {
   917  			size := frame.varp - frame.sp
   918  			if size > 0 {
   919  				scanConservative(frame.sp, size, nil, gcw, state)
   920  			}
   921  		}
   922  
   923  		// Scan arguments to this frame.
   924  		if frame.arglen != 0 {
   925  			// TODO: We could pass the entry argument map
   926  			// to narrow this down further.
   927  			scanConservative(frame.argp, frame.arglen, nil, gcw, state)
   928  		}
   929  
   930  		if isAsyncPreempt || isDebugCall {
   931  			// This function's frame contained the
   932  			// registers for the asynchronously stopped
   933  			// parent frame. Scan the parent
   934  			// conservatively.
   935  			state.conservative = true
   936  		} else {
   937  			// We only wanted to scan those two frames
   938  			// conservatively. Clear the flag for future
   939  			// frames.
   940  			state.conservative = false
   941  		}
   942  		return
   943  	}
   944  
   945  	locals, args, objs := getStackMap(frame, &state.cache, false)
   946  
   947  	// Scan local variables if stack frame has been allocated.
   948  	if locals.n > 0 {
   949  		size := uintptr(locals.n) * goarch.PtrSize
   950  		scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
   951  	}
   952  
   953  	// Scan arguments.
   954  	if args.n > 0 {
   955  		scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
   956  	}
   957  
   958  	// Add all stack objects to the stack object list.
   959  	if frame.varp != 0 {
   960  		// varp is 0 for defers, where there are no locals.
   961  		// In that case, there can't be a pointer to its args, either.
   962  		// (And all args would be scanned above anyway.)
   963  		for i := range objs {
   964  			obj := &objs[i]
   965  			off := obj.off
   966  			base := frame.varp // locals base pointer
   967  			if off >= 0 {
   968  				base = frame.argp // arguments and return values base pointer
   969  			}
   970  			ptr := base + uintptr(off)
   971  			if ptr < frame.sp {
   972  				// object hasn't been allocated in the frame yet.
   973  				continue
   974  			}
   975  			if stackTraceDebug {
   976  				println("stkobj at", hex(ptr), "of size", obj.size)
   977  			}
   978  			state.addObject(ptr, obj)
   979  		}
   980  	}
   981  }
   982  
   983  type gcDrainFlags int
   984  
   985  const (
   986  	gcDrainUntilPreempt gcDrainFlags = 1 << iota
   987  	gcDrainFlushBgCredit
   988  	gcDrainIdle
   989  	gcDrainFractional
   990  )
   991  
   992  // gcDrain scans roots and objects in work buffers, blackening grey
   993  // objects until it is unable to get more work. It may return before
   994  // GC is done; it's the caller's responsibility to balance work from
   995  // other Ps.
   996  //
   997  // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
   998  // is set.
   999  //
  1000  // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
  1001  // to do.
  1002  //
  1003  // If flags&gcDrainFractional != 0, gcDrain self-preempts when
  1004  // pollFractionalWorkerExit() returns true. This implies
  1005  // gcDrainNoBlock.
  1006  //
  1007  // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
  1008  // credit to gcController.bgScanCredit every gcCreditSlack units of
  1009  // scan work.
  1010  //
  1011  // gcDrain will always return if there is a pending STW.
  1012  //
  1013  //go:nowritebarrier
  1014  func gcDrain(gcw *gcWork, flags gcDrainFlags) {
  1015  	if !writeBarrier.needed {
  1016  		throw("gcDrain phase incorrect")
  1017  	}
  1018  
  1019  	gp := getg().m.curg
  1020  	preemptible := flags&gcDrainUntilPreempt != 0
  1021  	flushBgCredit := flags&gcDrainFlushBgCredit != 0
  1022  	idle := flags&gcDrainIdle != 0
  1023  
  1024  	initScanWork := gcw.heapScanWork
  1025  
  1026  	// checkWork is the scan work before performing the next
  1027  	// self-preempt check.
  1028  	checkWork := int64(1<<63 - 1)
  1029  	var check func() bool
  1030  	if flags&(gcDrainIdle|gcDrainFractional) != 0 {
  1031  		checkWork = initScanWork + drainCheckThreshold
  1032  		if idle {
  1033  			check = pollWork
  1034  		} else if flags&gcDrainFractional != 0 {
  1035  			check = pollFractionalWorkerExit
  1036  		}
  1037  	}
  1038  
  1039  	// Drain root marking jobs.
  1040  	if work.markrootNext < work.markrootJobs {
  1041  		// Stop if we're preemptible or if someone wants to STW.
  1042  		for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
  1043  			job := atomic.Xadd(&work.markrootNext, +1) - 1
  1044  			if job >= work.markrootJobs {
  1045  				break
  1046  			}
  1047  			markroot(gcw, job, flushBgCredit)
  1048  			if check != nil && check() {
  1049  				goto done
  1050  			}
  1051  		}
  1052  	}
  1053  
  1054  	// Drain heap marking jobs.
  1055  	// Stop if we're preemptible or if someone wants to STW.
  1056  	for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
  1057  		// Try to keep work available on the global queue. We used to
  1058  		// check if there were waiting workers, but it's better to
  1059  		// just keep work available than to make workers wait. In the
  1060  		// worst case, we'll do O(log(_WorkbufSize)) unnecessary
  1061  		// balances.
  1062  		if work.full == 0 {
  1063  			gcw.balance()
  1064  		}
  1065  
  1066  		b := gcw.tryGetFast()
  1067  		if b == 0 {
  1068  			b = gcw.tryGet()
  1069  			if b == 0 {
  1070  				// Flush the write barrier
  1071  				// buffer; this may create
  1072  				// more work.
  1073  				wbBufFlush(nil, 0)
  1074  				b = gcw.tryGet()
  1075  			}
  1076  		}
  1077  		if b == 0 {
  1078  			// Unable to get work.
  1079  			break
  1080  		}
  1081  		scanobject(b, gcw)
  1082  
  1083  		// Flush background scan work credit to the global
  1084  		// account if we've accumulated enough locally so
  1085  		// mutator assists can draw on it.
  1086  		if gcw.heapScanWork >= gcCreditSlack {
  1087  			gcController.heapScanWork.Add(gcw.heapScanWork)
  1088  			if flushBgCredit {
  1089  				gcFlushBgCredit(gcw.heapScanWork - initScanWork)
  1090  				initScanWork = 0
  1091  			}
  1092  			checkWork -= gcw.heapScanWork
  1093  			gcw.heapScanWork = 0
  1094  
  1095  			if checkWork <= 0 {
  1096  				checkWork += drainCheckThreshold
  1097  				if check != nil && check() {
  1098  					break
  1099  				}
  1100  			}
  1101  		}
  1102  	}
  1103  
  1104  done:
  1105  	// Flush remaining scan work credit.
  1106  	if gcw.heapScanWork > 0 {
  1107  		gcController.heapScanWork.Add(gcw.heapScanWork)
  1108  		if flushBgCredit {
  1109  			gcFlushBgCredit(gcw.heapScanWork - initScanWork)
  1110  		}
  1111  		gcw.heapScanWork = 0
  1112  	}
  1113  }
  1114  
  1115  // gcDrainN blackens grey objects until it has performed roughly
  1116  // scanWork units of scan work or the G is preempted. This is
  1117  // best-effort, so it may perform less work if it fails to get a work
  1118  // buffer. Otherwise, it will perform at least n units of work, but
  1119  // may perform more because scanning is always done in whole object
  1120  // increments. It returns the amount of scan work performed.
  1121  //
  1122  // The caller goroutine must be in a preemptible state (e.g.,
  1123  // _Gwaiting) to prevent deadlocks during stack scanning. As a
  1124  // consequence, this must be called on the system stack.
  1125  //
  1126  //go:nowritebarrier
  1127  //go:systemstack
  1128  func gcDrainN(gcw *gcWork, scanWork int64) int64 {
  1129  	if !writeBarrier.needed {
  1130  		throw("gcDrainN phase incorrect")
  1131  	}
  1132  
  1133  	// There may already be scan work on the gcw, which we don't
  1134  	// want to claim was done by this call.
  1135  	workFlushed := -gcw.heapScanWork
  1136  
  1137  	gp := getg().m.curg
  1138  	for !gp.preempt && workFlushed+gcw.heapScanWork < scanWork {
  1139  		// See gcDrain comment.
  1140  		if work.full == 0 {
  1141  			gcw.balance()
  1142  		}
  1143  
  1144  		b := gcw.tryGetFast()
  1145  		if b == 0 {
  1146  			b = gcw.tryGet()
  1147  			if b == 0 {
  1148  				// Flush the write barrier buffer;
  1149  				// this may create more work.
  1150  				wbBufFlush(nil, 0)
  1151  				b = gcw.tryGet()
  1152  			}
  1153  		}
  1154  
  1155  		if b == 0 {
  1156  			// Try to do a root job.
  1157  			if work.markrootNext < work.markrootJobs {
  1158  				job := atomic.Xadd(&work.markrootNext, +1) - 1
  1159  				if job < work.markrootJobs {
  1160  					work := markroot(gcw, job, false)
  1161  					if goexperiment.PacerRedesign {
  1162  						workFlushed += work
  1163  					}
  1164  					continue
  1165  				}
  1166  			}
  1167  			// No heap or root jobs.
  1168  			break
  1169  		}
  1170  
  1171  		scanobject(b, gcw)
  1172  
  1173  		// Flush background scan work credit.
  1174  		if gcw.heapScanWork >= gcCreditSlack {
  1175  			gcController.heapScanWork.Add(gcw.heapScanWork)
  1176  			workFlushed += gcw.heapScanWork
  1177  			gcw.heapScanWork = 0
  1178  		}
  1179  	}
  1180  
  1181  	// Unlike gcDrain, there's no need to flush remaining work
  1182  	// here because this never flushes to bgScanCredit and
  1183  	// gcw.dispose will flush any remaining work to scanWork.
  1184  
  1185  	return workFlushed + gcw.heapScanWork
  1186  }
  1187  
  1188  // scanblock scans b as scanobject would, but using an explicit
  1189  // pointer bitmap instead of the heap bitmap.
  1190  //
  1191  // This is used to scan non-heap roots, so it does not update
  1192  // gcw.bytesMarked or gcw.heapScanWork.
  1193  //
  1194  // If stk != nil, possible stack pointers are also reported to stk.putPtr.
  1195  //go:nowritebarrier
  1196  func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
  1197  	// Use local copies of original parameters, so that a stack trace
  1198  	// due to one of the throws below shows the original block
  1199  	// base and extent.
  1200  	b := b0
  1201  	n := n0
  1202  
  1203  	for i := uintptr(0); i < n; {
  1204  		// Find bits for the next word.
  1205  		bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
  1206  		if bits == 0 {
  1207  			i += goarch.PtrSize * 8
  1208  			continue
  1209  		}
  1210  		for j := 0; j < 8 && i < n; j++ {
  1211  			if bits&1 != 0 {
  1212  				// Same work as in scanobject; see comments there.
  1213  				p := *(*uintptr)(unsafe.Pointer(b + i))
  1214  				if p != 0 {
  1215  					if obj, span, objIndex := findObject(p, b, i); obj != 0 {
  1216  						greyobject(obj, b, i, span, gcw, objIndex)
  1217  					} else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
  1218  						stk.putPtr(p, false)
  1219  					}
  1220  				}
  1221  			}
  1222  			bits >>= 1
  1223  			i += goarch.PtrSize
  1224  		}
  1225  	}
  1226  }
  1227  
  1228  // scanobject scans the object starting at b, adding pointers to gcw.
  1229  // b must point to the beginning of a heap object or an oblet.
  1230  // scanobject consults the GC bitmap for the pointer mask and the
  1231  // spans for the size of the object.
  1232  //
  1233  //go:nowritebarrier
  1234  func scanobject(b uintptr, gcw *gcWork) {
  1235  	// Prefetch object before we scan it.
  1236  	//
  1237  	// This will overlap fetching the beginning of the object with initial
  1238  	// setup before we start scanning the object.
  1239  	sys.Prefetch(b)
  1240  
  1241  	// Find the bits for b and the size of the object at b.
  1242  	//
  1243  	// b is either the beginning of an object, in which case this
  1244  	// is the size of the object to scan, or it points to an
  1245  	// oblet, in which case we compute the size to scan below.
  1246  	hbits := heapBitsForAddr(b)
  1247  	s := spanOfUnchecked(b)
  1248  	n := s.elemsize
  1249  	if n == 0 {
  1250  		throw("scanobject n == 0")
  1251  	}
  1252  
  1253  	if n > maxObletBytes {
  1254  		// Large object. Break into oblets for better
  1255  		// parallelism and lower latency.
  1256  		if b == s.base() {
  1257  			// It's possible this is a noscan object (not
  1258  			// from greyobject, but from other code
  1259  			// paths), in which case we must *not* enqueue
  1260  			// oblets since their bitmaps will be
  1261  			// uninitialized.
  1262  			if s.spanclass.noscan() {
  1263  				// Bypass the whole scan.
  1264  				gcw.bytesMarked += uint64(n)
  1265  				return
  1266  			}
  1267  
  1268  			// Enqueue the other oblets to scan later.
  1269  			// Some oblets may be in b's scalar tail, but
  1270  			// these will be marked as "no more pointers",
  1271  			// so we'll drop out immediately when we go to
  1272  			// scan those.
  1273  			for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
  1274  				if !gcw.putFast(oblet) {
  1275  					gcw.put(oblet)
  1276  				}
  1277  			}
  1278  		}
  1279  
  1280  		// Compute the size of the oblet. Since this object
  1281  		// must be a large object, s.base() is the beginning
  1282  		// of the object.
  1283  		n = s.base() + s.elemsize - b
  1284  		if n > maxObletBytes {
  1285  			n = maxObletBytes
  1286  		}
  1287  	}
  1288  
  1289  	var i uintptr
  1290  	for i = 0; i < n; i, hbits = i+goarch.PtrSize, hbits.next() {
  1291  		// Load bits once. See CL 22712 and issue 16973 for discussion.
  1292  		bits := hbits.bits()
  1293  		if bits&bitScan == 0 {
  1294  			break // no more pointers in this object
  1295  		}
  1296  		if bits&bitPointer == 0 {
  1297  			continue // not a pointer
  1298  		}
  1299  
  1300  		// Work here is duplicated in scanblock and above.
  1301  		// If you make changes here, make changes there too.
  1302  		obj := *(*uintptr)(unsafe.Pointer(b + i))
  1303  
  1304  		// At this point we have extracted the next potential pointer.
  1305  		// Quickly filter out nil and pointers back to the current object.
  1306  		if obj != 0 && obj-b >= n {
  1307  			// Test if obj points into the Go heap and, if so,
  1308  			// mark the object.
  1309  			//
  1310  			// Note that it's possible for findObject to
  1311  			// fail if obj points to a just-allocated heap
  1312  			// object because of a race with growing the
  1313  			// heap. In this case, we know the object was
  1314  			// just allocated and hence will be marked by
  1315  			// allocation itself.
  1316  			if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
  1317  				greyobject(obj, b, i, span, gcw, objIndex)
  1318  			}
  1319  		}
  1320  	}
  1321  	gcw.bytesMarked += uint64(n)
  1322  	gcw.heapScanWork += int64(i)
  1323  }
  1324  
  1325  // scanConservative scans block [b, b+n) conservatively, treating any
  1326  // pointer-like value in the block as a pointer.
  1327  //
  1328  // If ptrmask != nil, only words that are marked in ptrmask are
  1329  // considered as potential pointers.
  1330  //
  1331  // If state != nil, it's assumed that [b, b+n) is a block in the stack
  1332  // and may contain pointers to stack objects.
  1333  func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
  1334  	if debugScanConservative {
  1335  		printlock()
  1336  		print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
  1337  		hexdumpWords(b, b+n, func(p uintptr) byte {
  1338  			if ptrmask != nil {
  1339  				word := (p - b) / goarch.PtrSize
  1340  				bits := *addb(ptrmask, word/8)
  1341  				if (bits>>(word%8))&1 == 0 {
  1342  					return '$'
  1343  				}
  1344  			}
  1345  
  1346  			val := *(*uintptr)(unsafe.Pointer(p))
  1347  			if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1348  				return '@'
  1349  			}
  1350  
  1351  			span := spanOfHeap(val)
  1352  			if span == nil {
  1353  				return ' '
  1354  			}
  1355  			idx := span.objIndex(val)
  1356  			if span.isFree(idx) {
  1357  				return ' '
  1358  			}
  1359  			return '*'
  1360  		})
  1361  		printunlock()
  1362  	}
  1363  
  1364  	for i := uintptr(0); i < n; i += goarch.PtrSize {
  1365  		if ptrmask != nil {
  1366  			word := i / goarch.PtrSize
  1367  			bits := *addb(ptrmask, word/8)
  1368  			if bits == 0 {
  1369  				// Skip 8 words (the loop increment will do the 8th)
  1370  				//
  1371  				// This must be the first time we've
  1372  				// seen this word of ptrmask, so i
  1373  				// must be 8-word-aligned, but check
  1374  				// our reasoning just in case.
  1375  				if i%(goarch.PtrSize*8) != 0 {
  1376  					throw("misaligned mask")
  1377  				}
  1378  				i += goarch.PtrSize*8 - goarch.PtrSize
  1379  				continue
  1380  			}
  1381  			if (bits>>(word%8))&1 == 0 {
  1382  				continue
  1383  			}
  1384  		}
  1385  
  1386  		val := *(*uintptr)(unsafe.Pointer(b + i))
  1387  
  1388  		// Check if val points into the stack.
  1389  		if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1390  			// val may point to a stack object. This
  1391  			// object may be dead from last cycle and
  1392  			// hence may contain pointers to unallocated
  1393  			// objects, but unlike heap objects we can't
  1394  			// tell if it's already dead. Hence, if all
  1395  			// pointers to this object are from
  1396  			// conservative scanning, we have to scan it
  1397  			// defensively, too.
  1398  			state.putPtr(val, true)
  1399  			continue
  1400  		}
  1401  
  1402  		// Check if val points to a heap span.
  1403  		span := spanOfHeap(val)
  1404  		if span == nil {
  1405  			continue
  1406  		}
  1407  
  1408  		// Check if val points to an allocated object.
  1409  		idx := span.objIndex(val)
  1410  		if span.isFree(idx) {
  1411  			continue
  1412  		}
  1413  
  1414  		// val points to an allocated object. Mark it.
  1415  		obj := span.base() + idx*span.elemsize
  1416  		greyobject(obj, b, i, span, gcw, idx)
  1417  	}
  1418  }
  1419  
  1420  // Shade the object if it isn't already.
  1421  // The object is not nil and known to be in the heap.
  1422  // Preemption must be disabled.
  1423  //go:nowritebarrier
  1424  func shade(b uintptr) {
  1425  	if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
  1426  		gcw := &getg().m.p.ptr().gcw
  1427  		greyobject(obj, 0, 0, span, gcw, objIndex)
  1428  	}
  1429  }
  1430  
  1431  // obj is the start of an object with mark mbits.
  1432  // If it isn't already marked, mark it and enqueue into gcw.
  1433  // base and off are for debugging only and could be removed.
  1434  //
  1435  // See also wbBufFlush1, which partially duplicates this logic.
  1436  //
  1437  //go:nowritebarrierrec
  1438  func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
  1439  	// obj should be start of allocation, and so must be at least pointer-aligned.
  1440  	if obj&(goarch.PtrSize-1) != 0 {
  1441  		throw("greyobject: obj not pointer-aligned")
  1442  	}
  1443  	mbits := span.markBitsForIndex(objIndex)
  1444  
  1445  	if useCheckmark {
  1446  		if setCheckmark(obj, base, off, mbits) {
  1447  			// Already marked.
  1448  			return
  1449  		}
  1450  	} else {
  1451  		if debug.gccheckmark > 0 && span.isFree(objIndex) {
  1452  			print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
  1453  			gcDumpObject("base", base, off)
  1454  			gcDumpObject("obj", obj, ^uintptr(0))
  1455  			getg().m.traceback = 2
  1456  			throw("marking free object")
  1457  		}
  1458  
  1459  		// If marked we have nothing to do.
  1460  		if mbits.isMarked() {
  1461  			return
  1462  		}
  1463  		mbits.setMarked()
  1464  
  1465  		// Mark span.
  1466  		arena, pageIdx, pageMask := pageIndexOf(span.base())
  1467  		if arena.pageMarks[pageIdx]&pageMask == 0 {
  1468  			atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1469  		}
  1470  
  1471  		// If this is a noscan object, fast-track it to black
  1472  		// instead of greying it.
  1473  		if span.spanclass.noscan() {
  1474  			gcw.bytesMarked += uint64(span.elemsize)
  1475  			return
  1476  		}
  1477  	}
  1478  
  1479  	// We're adding obj to P's local workbuf, so it's likely
  1480  	// this object will be processed soon by the same P.
  1481  	// Even if the workbuf gets flushed, there will likely still be
  1482  	// some benefit on platforms with inclusive shared caches.
  1483  	sys.Prefetch(obj)
  1484  	// Queue the obj for scanning.
  1485  	if !gcw.putFast(obj) {
  1486  		gcw.put(obj)
  1487  	}
  1488  }
  1489  
  1490  // gcDumpObject dumps the contents of obj for debugging and marks the
  1491  // field at byte offset off in obj.
  1492  func gcDumpObject(label string, obj, off uintptr) {
  1493  	s := spanOf(obj)
  1494  	print(label, "=", hex(obj))
  1495  	if s == nil {
  1496  		print(" s=nil\n")
  1497  		return
  1498  	}
  1499  	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
  1500  	if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
  1501  		print(mSpanStateNames[state], "\n")
  1502  	} else {
  1503  		print("unknown(", state, ")\n")
  1504  	}
  1505  
  1506  	skipped := false
  1507  	size := s.elemsize
  1508  	if s.state.get() == mSpanManual && size == 0 {
  1509  		// We're printing something from a stack frame. We
  1510  		// don't know how big it is, so just show up to an
  1511  		// including off.
  1512  		size = off + goarch.PtrSize
  1513  	}
  1514  	for i := uintptr(0); i < size; i += goarch.PtrSize {
  1515  		// For big objects, just print the beginning (because
  1516  		// that usually hints at the object's type) and the
  1517  		// fields around off.
  1518  		if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
  1519  			skipped = true
  1520  			continue
  1521  		}
  1522  		if skipped {
  1523  			print(" ...\n")
  1524  			skipped = false
  1525  		}
  1526  		print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
  1527  		if i == off {
  1528  			print(" <==")
  1529  		}
  1530  		print("\n")
  1531  	}
  1532  	if skipped {
  1533  		print(" ...\n")
  1534  	}
  1535  }
  1536  
  1537  // gcmarknewobject marks a newly allocated object black. obj must
  1538  // not contain any non-nil pointers.
  1539  //
  1540  // This is nosplit so it can manipulate a gcWork without preemption.
  1541  //
  1542  //go:nowritebarrier
  1543  //go:nosplit
  1544  func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
  1545  	if useCheckmark { // The world should be stopped so this should not happen.
  1546  		throw("gcmarknewobject called while doing checkmark")
  1547  	}
  1548  
  1549  	// Mark object.
  1550  	objIndex := span.objIndex(obj)
  1551  	span.markBitsForIndex(objIndex).setMarked()
  1552  
  1553  	// Mark span.
  1554  	arena, pageIdx, pageMask := pageIndexOf(span.base())
  1555  	if arena.pageMarks[pageIdx]&pageMask == 0 {
  1556  		atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1557  	}
  1558  
  1559  	gcw := &getg().m.p.ptr().gcw
  1560  	gcw.bytesMarked += uint64(size)
  1561  	if !goexperiment.PacerRedesign {
  1562  		// The old pacer counts newly allocated memory toward
  1563  		// heapScanWork because heapScan is continuously updated
  1564  		// throughout the GC cycle with newly allocated memory. However,
  1565  		// these objects are never actually scanned, so we need
  1566  		// to account for them in heapScanWork here, "faking" their work.
  1567  		// Otherwise the pacer will think it's always behind, potentially
  1568  		// by a large margin.
  1569  		//
  1570  		// The new pacer doesn't care about this because it ceases to updated
  1571  		// heapScan once a GC cycle starts, effectively snapshotting it.
  1572  		gcw.heapScanWork += int64(scanSize)
  1573  	}
  1574  }
  1575  
  1576  // gcMarkTinyAllocs greys all active tiny alloc blocks.
  1577  //
  1578  // The world must be stopped.
  1579  func gcMarkTinyAllocs() {
  1580  	assertWorldStopped()
  1581  
  1582  	for _, p := range allp {
  1583  		c := p.mcache
  1584  		if c == nil || c.tiny == 0 {
  1585  			continue
  1586  		}
  1587  		_, span, objIndex := findObject(c.tiny, 0, 0)
  1588  		gcw := &p.gcw
  1589  		greyobject(c.tiny, 0, 0, span, gcw, objIndex)
  1590  	}
  1591  }
  1592  

View as plain text