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/abi"
    11  	"internal/goarch"
    12  	"internal/runtime/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  		nBSSRoots := nBlocks(datap.ebss - datap.bss)
    77  		if nBSSRoots > work.nBSSRoots {
    78  			work.nBSSRoots = nBSSRoots
    79  		}
    80  	}
    81  
    82  	// Scan span roots for finalizer specials.
    83  	//
    84  	// We depend on addfinalizer to mark objects that get
    85  	// finalizers after root marking.
    86  	//
    87  	// We're going to scan the whole heap (that was available at the time the
    88  	// mark phase started, i.e. markArenas) for in-use spans which have specials.
    89  	//
    90  	// Break up the work into arenas, and further into chunks.
    91  	//
    92  	// Snapshot allArenas as markArenas. This snapshot is safe because allArenas
    93  	// is append-only.
    94  	mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
    95  	work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
    96  
    97  	// Scan stacks.
    98  	//
    99  	// Gs may be created after this point, but it's okay that we
   100  	// ignore them because they begin life without any roots, so
   101  	// there's nothing to scan, and any roots they create during
   102  	// the concurrent phase will be caught by the write barrier.
   103  	work.stackRoots = allGsSnapshot()
   104  	work.nStackRoots = len(work.stackRoots)
   105  
   106  	work.markrootNext = 0
   107  	work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
   108  
   109  	// Calculate base indexes of each root type
   110  	work.baseData = uint32(fixedRootCount)
   111  	work.baseBSS = work.baseData + uint32(work.nDataRoots)
   112  	work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
   113  	work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
   114  	work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
   115  }
   116  
   117  // gcMarkRootCheck checks that all roots have been scanned. It is
   118  // purely for debugging.
   119  func gcMarkRootCheck() {
   120  	if work.markrootNext < work.markrootJobs {
   121  		print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
   122  		throw("left over markroot jobs")
   123  	}
   124  
   125  	// Check that stacks have been scanned.
   126  	//
   127  	// We only check the first nStackRoots Gs that we should have scanned.
   128  	// Since we don't care about newer Gs (see comment in
   129  	// gcMarkRootPrepare), no locking is required.
   130  	i := 0
   131  	forEachGRace(func(gp *g) {
   132  		if i >= work.nStackRoots {
   133  			return
   134  		}
   135  
   136  		if !gp.gcscandone {
   137  			println("gp", gp, "goid", gp.goid,
   138  				"status", readgstatus(gp),
   139  				"gcscandone", gp.gcscandone)
   140  			throw("scan missed a g")
   141  		}
   142  
   143  		i++
   144  	})
   145  }
   146  
   147  // ptrmask for an allocation containing a single pointer.
   148  var oneptrmask = [...]uint8{1}
   149  
   150  // markroot scans the i'th root.
   151  //
   152  // Preemption must be disabled (because this uses a gcWork).
   153  //
   154  // Returns the amount of GC work credit produced by the operation.
   155  // If flushBgCredit is true, then that credit is also flushed
   156  // to the background credit pool.
   157  //
   158  // nowritebarrier is only advisory here.
   159  //
   160  //go:nowritebarrier
   161  func markroot(gcw *gcWork, i uint32, flushBgCredit bool) int64 {
   162  	// Note: if you add a case here, please also update heapdump.go:dumproots.
   163  	var workDone int64
   164  	var workCounter *atomic.Int64
   165  	switch {
   166  	case work.baseData <= i && i < work.baseBSS:
   167  		workCounter = &gcController.globalsScanWork
   168  		for _, datap := range activeModules() {
   169  			workDone += markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
   170  		}
   171  
   172  	case work.baseBSS <= i && i < work.baseSpans:
   173  		workCounter = &gcController.globalsScanWork
   174  		for _, datap := range activeModules() {
   175  			workDone += markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
   176  		}
   177  
   178  	case i == fixedRootFinalizers:
   179  		for fb := allfin; fb != nil; fb = fb.alllink {
   180  			cnt := uintptr(atomic.Load(&fb.cnt))
   181  			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
   182  		}
   183  
   184  	case i == fixedRootFreeGStacks:
   185  		// Switch to the system stack so we can call
   186  		// stackfree.
   187  		systemstack(markrootFreeGStacks)
   188  
   189  	case work.baseSpans <= i && i < work.baseStacks:
   190  		// mark mspan.specials
   191  		markrootSpans(gcw, int(i-work.baseSpans))
   192  
   193  	default:
   194  		// the rest is scanning goroutine stacks
   195  		workCounter = &gcController.stackScanWork
   196  		if i < work.baseStacks || work.baseEnd <= i {
   197  			printlock()
   198  			print("runtime: markroot index ", i, " not in stack roots range [", work.baseStacks, ", ", work.baseEnd, ")\n")
   199  			throw("markroot: bad index")
   200  		}
   201  		gp := work.stackRoots[i-work.baseStacks]
   202  
   203  		// remember when we've first observed the G blocked
   204  		// needed only to output in traceback
   205  		status := readgstatus(gp) // We are not in a scan state
   206  		if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
   207  			gp.waitsince = work.tstart
   208  		}
   209  
   210  		// scanstack must be done on the system stack in case
   211  		// we're trying to scan our own stack.
   212  		systemstack(func() {
   213  			// If this is a self-scan, put the user G in
   214  			// _Gwaiting to prevent self-deadlock. It may
   215  			// already be in _Gwaiting if this is a mark
   216  			// worker or we're in mark termination.
   217  			userG := getg().m.curg
   218  			selfScan := gp == userG && readgstatus(userG) == _Grunning
   219  			if selfScan {
   220  				casGToWaitingForGC(userG, _Grunning, waitReasonGarbageCollectionScan)
   221  			}
   222  
   223  			// TODO: suspendG blocks (and spins) until gp
   224  			// stops, which may take a while for
   225  			// running goroutines. Consider doing this in
   226  			// two phases where the first is non-blocking:
   227  			// we scan the stacks we can and ask running
   228  			// goroutines to scan themselves; and the
   229  			// second blocks.
   230  			stopped := suspendG(gp)
   231  			if stopped.dead {
   232  				gp.gcscandone = true
   233  				return
   234  			}
   235  			if gp.gcscandone {
   236  				throw("g already scanned")
   237  			}
   238  			workDone += scanstack(gp, gcw)
   239  			gp.gcscandone = true
   240  			resumeG(stopped)
   241  
   242  			if selfScan {
   243  				casgstatus(userG, _Gwaiting, _Grunning)
   244  			}
   245  		})
   246  	}
   247  	if workCounter != nil && workDone != 0 {
   248  		workCounter.Add(workDone)
   249  		if flushBgCredit {
   250  			gcFlushBgCredit(workDone)
   251  		}
   252  	}
   253  	return workDone
   254  }
   255  
   256  // markrootBlock scans the shard'th shard of the block of memory [b0,
   257  // b0+n0), with the given pointer mask.
   258  //
   259  // Returns the amount of work done.
   260  //
   261  //go:nowritebarrier
   262  func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
   263  	if rootBlockBytes%(8*goarch.PtrSize) != 0 {
   264  		// This is necessary to pick byte offsets in ptrmask0.
   265  		throw("rootBlockBytes must be a multiple of 8*ptrSize")
   266  	}
   267  
   268  	// Note that if b0 is toward the end of the address space,
   269  	// then b0 + rootBlockBytes might wrap around.
   270  	// These tests are written to avoid any possible overflow.
   271  	off := uintptr(shard) * rootBlockBytes
   272  	if off >= n0 {
   273  		return 0
   274  	}
   275  	b := b0 + off
   276  	ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
   277  	n := uintptr(rootBlockBytes)
   278  	if off+n > n0 {
   279  		n = n0 - off
   280  	}
   281  
   282  	// Scan this shard.
   283  	scanblock(b, n, ptrmask, gcw, nil)
   284  	return int64(n)
   285  }
   286  
   287  // markrootFreeGStacks frees stacks of dead Gs.
   288  //
   289  // This does not free stacks of dead Gs cached on Ps, but having a few
   290  // cached stacks around isn't a problem.
   291  func markrootFreeGStacks() {
   292  	// Take list of dead Gs with stacks.
   293  	lock(&sched.gFree.lock)
   294  	list := sched.gFree.stack
   295  	sched.gFree.stack = gList{}
   296  	unlock(&sched.gFree.lock)
   297  	if list.empty() {
   298  		return
   299  	}
   300  
   301  	// Free stacks.
   302  	q := gQueue{list.head, list.head}
   303  	for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
   304  		stackfree(gp.stack)
   305  		gp.stack.lo = 0
   306  		gp.stack.hi = 0
   307  		// Manipulate the queue directly since the Gs are
   308  		// already all linked the right way.
   309  		q.tail.set(gp)
   310  	}
   311  
   312  	// Put Gs back on the free list.
   313  	lock(&sched.gFree.lock)
   314  	sched.gFree.noStack.pushAll(q)
   315  	unlock(&sched.gFree.lock)
   316  }
   317  
   318  // markrootSpans marks roots for one shard of markArenas.
   319  //
   320  //go:nowritebarrier
   321  func markrootSpans(gcw *gcWork, shard int) {
   322  	// Objects with finalizers have two GC-related invariants:
   323  	//
   324  	// 1) Everything reachable from the object must be marked.
   325  	// This ensures that when we pass the object to its finalizer,
   326  	// everything the finalizer can reach will be retained.
   327  	//
   328  	// 2) Finalizer specials (which are not in the garbage
   329  	// collected heap) are roots. In practice, this means the fn
   330  	// field must be scanned.
   331  	//
   332  	// Objects with weak handles have only one invariant related
   333  	// to this function: weak handle specials (which are not in the
   334  	// garbage collected heap) are roots. In practice, this means
   335  	// the handle field must be scanned. Note that the value the
   336  	// handle pointer referenced does *not* need to be scanned. See
   337  	// the definition of specialWeakHandle for details.
   338  	sg := mheap_.sweepgen
   339  
   340  	// Find the arena and page index into that arena for this shard.
   341  	ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
   342  	ha := mheap_.arenas[ai.l1()][ai.l2()]
   343  	arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
   344  
   345  	// Construct slice of bitmap which we'll iterate over.
   346  	specialsbits := ha.pageSpecials[arenaPage/8:]
   347  	specialsbits = specialsbits[:pagesPerSpanRoot/8]
   348  	for i := range specialsbits {
   349  		// Find set bits, which correspond to spans with specials.
   350  		specials := atomic.Load8(&specialsbits[i])
   351  		if specials == 0 {
   352  			continue
   353  		}
   354  		for j := uint(0); j < 8; j++ {
   355  			if specials&(1<<j) == 0 {
   356  				continue
   357  			}
   358  			// Find the span for this bit.
   359  			//
   360  			// This value is guaranteed to be non-nil because having
   361  			// specials implies that the span is in-use, and since we're
   362  			// currently marking we can be sure that we don't have to worry
   363  			// about the span being freed and re-used.
   364  			s := ha.spans[arenaPage+uint(i)*8+j]
   365  
   366  			// The state must be mSpanInUse if the specials bit is set, so
   367  			// sanity check that.
   368  			if state := s.state.get(); state != mSpanInUse {
   369  				print("s.state = ", state, "\n")
   370  				throw("non in-use span found with specials bit set")
   371  			}
   372  			// Check that this span was swept (it may be cached or uncached).
   373  			if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
   374  				// sweepgen was updated (+2) during non-checkmark GC pass
   375  				print("sweep ", s.sweepgen, " ", sg, "\n")
   376  				throw("gc: unswept span")
   377  			}
   378  
   379  			// Lock the specials to prevent a special from being
   380  			// removed from the list while we're traversing it.
   381  			lock(&s.speciallock)
   382  			for sp := s.specials; sp != nil; sp = sp.next {
   383  				switch sp.kind {
   384  				case _KindSpecialFinalizer:
   385  					// don't mark finalized object, but scan it so we
   386  					// retain everything it points to.
   387  					spf := (*specialfinalizer)(unsafe.Pointer(sp))
   388  					// A finalizer can be set for an inner byte of an object, find object beginning.
   389  					p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
   390  
   391  					// Mark everything that can be reached from
   392  					// the object (but *not* the object itself or
   393  					// we'll never collect it).
   394  					if !s.spanclass.noscan() {
   395  						scanobject(p, gcw)
   396  					}
   397  
   398  					// The special itself is a root.
   399  					scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
   400  				case _KindSpecialWeakHandle:
   401  					// The special itself is a root.
   402  					spw := (*specialWeakHandle)(unsafe.Pointer(sp))
   403  					scanblock(uintptr(unsafe.Pointer(&spw.handle)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
   404  				}
   405  			}
   406  			unlock(&s.speciallock)
   407  		}
   408  	}
   409  }
   410  
   411  // gcAssistAlloc performs GC work to make gp's assist debt positive.
   412  // gp must be the calling user goroutine.
   413  //
   414  // This must be called with preemption enabled.
   415  func gcAssistAlloc(gp *g) {
   416  	// Don't assist in non-preemptible contexts. These are
   417  	// generally fragile and won't allow the assist to block.
   418  	if getg() == gp.m.g0 {
   419  		return
   420  	}
   421  	if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
   422  		return
   423  	}
   424  
   425  	// This extremely verbose boolean indicates whether we've
   426  	// entered mark assist from the perspective of the tracer.
   427  	//
   428  	// In the tracer, this is just before we call gcAssistAlloc1
   429  	// *regardless* of whether tracing is enabled. This is because
   430  	// the tracer allows for tracing to begin (and advance
   431  	// generations) in the middle of a GC mark phase, so we need to
   432  	// record some state so that the tracer can pick it up to ensure
   433  	// a consistent trace result.
   434  	//
   435  	// TODO(mknyszek): Hide the details of inMarkAssist in tracer
   436  	// functions and simplify all the state tracking. This is a lot.
   437  	enteredMarkAssistForTracing := false
   438  retry:
   439  	if gcCPULimiter.limiting() {
   440  		// If the CPU limiter is enabled, intentionally don't
   441  		// assist to reduce the amount of CPU time spent in the GC.
   442  		if enteredMarkAssistForTracing {
   443  			trace := traceAcquire()
   444  			if trace.ok() {
   445  				trace.GCMarkAssistDone()
   446  				// Set this *after* we trace the end to make sure
   447  				// that we emit an in-progress event if this is
   448  				// the first event for the goroutine in the trace
   449  				// or trace generation. Also, do this between
   450  				// acquire/release because this is part of the
   451  				// goroutine's trace state, and it must be atomic
   452  				// with respect to the tracer.
   453  				gp.inMarkAssist = false
   454  				traceRelease(trace)
   455  			} else {
   456  				// This state is tracked even if tracing isn't enabled.
   457  				// It's only used by the new tracer.
   458  				// See the comment on enteredMarkAssistForTracing.
   459  				gp.inMarkAssist = false
   460  			}
   461  		}
   462  		return
   463  	}
   464  	// Compute the amount of scan work we need to do to make the
   465  	// balance positive. When the required amount of work is low,
   466  	// we over-assist to build up credit for future allocations
   467  	// and amortize the cost of assisting.
   468  	assistWorkPerByte := gcController.assistWorkPerByte.Load()
   469  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   470  	debtBytes := -gp.gcAssistBytes
   471  	scanWork := int64(assistWorkPerByte * float64(debtBytes))
   472  	if scanWork < gcOverAssistWork {
   473  		scanWork = gcOverAssistWork
   474  		debtBytes = int64(assistBytesPerWork * float64(scanWork))
   475  	}
   476  
   477  	// Steal as much credit as we can from the background GC's
   478  	// scan credit. This is racy and may drop the background
   479  	// credit below 0 if two mutators steal at the same time. This
   480  	// will just cause steals to fail until credit is accumulated
   481  	// again, so in the long run it doesn't really matter, but we
   482  	// do have to handle the negative credit case.
   483  	bgScanCredit := gcController.bgScanCredit.Load()
   484  	stolen := int64(0)
   485  	if bgScanCredit > 0 {
   486  		if bgScanCredit < scanWork {
   487  			stolen = bgScanCredit
   488  			gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
   489  		} else {
   490  			stolen = scanWork
   491  			gp.gcAssistBytes += debtBytes
   492  		}
   493  		gcController.bgScanCredit.Add(-stolen)
   494  
   495  		scanWork -= stolen
   496  
   497  		if scanWork == 0 {
   498  			// We were able to steal all of the credit we
   499  			// needed.
   500  			if enteredMarkAssistForTracing {
   501  				trace := traceAcquire()
   502  				if trace.ok() {
   503  					trace.GCMarkAssistDone()
   504  					// Set this *after* we trace the end to make sure
   505  					// that we emit an in-progress event if this is
   506  					// the first event for the goroutine in the trace
   507  					// or trace generation. Also, do this between
   508  					// acquire/release because this is part of the
   509  					// goroutine's trace state, and it must be atomic
   510  					// with respect to the tracer.
   511  					gp.inMarkAssist = false
   512  					traceRelease(trace)
   513  				} else {
   514  					// This state is tracked even if tracing isn't enabled.
   515  					// It's only used by the new tracer.
   516  					// See the comment on enteredMarkAssistForTracing.
   517  					gp.inMarkAssist = false
   518  				}
   519  			}
   520  			return
   521  		}
   522  	}
   523  	if !enteredMarkAssistForTracing {
   524  		trace := traceAcquire()
   525  		if trace.ok() {
   526  			trace.GCMarkAssistStart()
   527  			// Set this *after* we trace the start, otherwise we may
   528  			// emit an in-progress event for an assist we're about to start.
   529  			gp.inMarkAssist = true
   530  			traceRelease(trace)
   531  		} else {
   532  			gp.inMarkAssist = true
   533  		}
   534  		// In the new tracer, set enter mark assist tracing if we
   535  		// ever pass this point, because we must manage inMarkAssist
   536  		// correctly.
   537  		//
   538  		// See the comment on enteredMarkAssistForTracing.
   539  		enteredMarkAssistForTracing = true
   540  	}
   541  
   542  	// Perform assist work
   543  	systemstack(func() {
   544  		gcAssistAlloc1(gp, scanWork)
   545  		// The user stack may have moved, so this can't touch
   546  		// anything on it until it returns from systemstack.
   547  	})
   548  
   549  	completed := gp.param != nil
   550  	gp.param = nil
   551  	if completed {
   552  		gcMarkDone()
   553  	}
   554  
   555  	if gp.gcAssistBytes < 0 {
   556  		// We were unable steal enough credit or perform
   557  		// enough work to pay off the assist debt. We need to
   558  		// do one of these before letting the mutator allocate
   559  		// more to prevent over-allocation.
   560  		//
   561  		// If this is because we were preempted, reschedule
   562  		// and try some more.
   563  		if gp.preempt {
   564  			Gosched()
   565  			goto retry
   566  		}
   567  
   568  		// Add this G to an assist queue and park. When the GC
   569  		// has more background credit, it will satisfy queued
   570  		// assists before flushing to the global credit pool.
   571  		//
   572  		// Note that this does *not* get woken up when more
   573  		// work is added to the work list. The theory is that
   574  		// there wasn't enough work to do anyway, so we might
   575  		// as well let background marking take care of the
   576  		// work that is available.
   577  		if !gcParkAssist() {
   578  			goto retry
   579  		}
   580  
   581  		// At this point either background GC has satisfied
   582  		// this G's assist debt, or the GC cycle is over.
   583  	}
   584  	if enteredMarkAssistForTracing {
   585  		trace := traceAcquire()
   586  		if trace.ok() {
   587  			trace.GCMarkAssistDone()
   588  			// Set this *after* we trace the end to make sure
   589  			// that we emit an in-progress event if this is
   590  			// the first event for the goroutine in the trace
   591  			// or trace generation. Also, do this between
   592  			// acquire/release because this is part of the
   593  			// goroutine's trace state, and it must be atomic
   594  			// with respect to the tracer.
   595  			gp.inMarkAssist = false
   596  			traceRelease(trace)
   597  		} else {
   598  			// This state is tracked even if tracing isn't enabled.
   599  			// It's only used by the new tracer.
   600  			// See the comment on enteredMarkAssistForTracing.
   601  			gp.inMarkAssist = false
   602  		}
   603  	}
   604  }
   605  
   606  // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
   607  // stack. This is a separate function to make it easier to see that
   608  // we're not capturing anything from the user stack, since the user
   609  // stack may move while we're in this function.
   610  //
   611  // gcAssistAlloc1 indicates whether this assist completed the mark
   612  // phase by setting gp.param to non-nil. This can't be communicated on
   613  // the stack since it may move.
   614  //
   615  //go:systemstack
   616  func gcAssistAlloc1(gp *g, scanWork int64) {
   617  	// Clear the flag indicating that this assist completed the
   618  	// mark phase.
   619  	gp.param = nil
   620  
   621  	if atomic.Load(&gcBlackenEnabled) == 0 {
   622  		// The gcBlackenEnabled check in malloc races with the
   623  		// store that clears it but an atomic check in every malloc
   624  		// would be a performance hit.
   625  		// Instead we recheck it here on the non-preemptible system
   626  		// stack to determine if we should perform an assist.
   627  
   628  		// GC is done, so ignore any remaining debt.
   629  		gp.gcAssistBytes = 0
   630  		return
   631  	}
   632  	// Track time spent in this assist. Since we're on the
   633  	// system stack, this is non-preemptible, so we can
   634  	// just measure start and end time.
   635  	//
   636  	// Limiter event tracking might be disabled if we end up here
   637  	// while on a mark worker.
   638  	startTime := nanotime()
   639  	trackLimiterEvent := gp.m.p.ptr().limiterEvent.start(limiterEventMarkAssist, startTime)
   640  
   641  	decnwait := atomic.Xadd(&work.nwait, -1)
   642  	if decnwait == work.nproc {
   643  		println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
   644  		throw("nwait > work.nprocs")
   645  	}
   646  
   647  	// gcDrainN requires the caller to be preemptible.
   648  	casGToWaitingForGC(gp, _Grunning, waitReasonGCAssistMarking)
   649  
   650  	// drain own cached work first in the hopes that it
   651  	// will be more cache friendly.
   652  	gcw := &getg().m.p.ptr().gcw
   653  	workDone := gcDrainN(gcw, scanWork)
   654  
   655  	casgstatus(gp, _Gwaiting, _Grunning)
   656  
   657  	// Record that we did this much scan work.
   658  	//
   659  	// Back out the number of bytes of assist credit that
   660  	// this scan work counts for. The "1+" is a poor man's
   661  	// round-up, to ensure this adds credit even if
   662  	// assistBytesPerWork is very low.
   663  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   664  	gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
   665  
   666  	// If this is the last worker and we ran out of work,
   667  	// signal a completion point.
   668  	incnwait := atomic.Xadd(&work.nwait, +1)
   669  	if incnwait > work.nproc {
   670  		println("runtime: work.nwait=", incnwait,
   671  			"work.nproc=", work.nproc)
   672  		throw("work.nwait > work.nproc")
   673  	}
   674  
   675  	if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
   676  		// This has reached a background completion point. Set
   677  		// gp.param to a non-nil value to indicate this. It
   678  		// doesn't matter what we set it to (it just has to be
   679  		// a valid pointer).
   680  		gp.param = unsafe.Pointer(gp)
   681  	}
   682  	now := nanotime()
   683  	duration := now - startTime
   684  	pp := gp.m.p.ptr()
   685  	pp.gcAssistTime += duration
   686  	if trackLimiterEvent {
   687  		pp.limiterEvent.stop(limiterEventMarkAssist, now)
   688  	}
   689  	if pp.gcAssistTime > gcAssistTimeSlack {
   690  		gcController.assistTime.Add(pp.gcAssistTime)
   691  		gcCPULimiter.update(now)
   692  		pp.gcAssistTime = 0
   693  	}
   694  }
   695  
   696  // gcWakeAllAssists wakes all currently blocked assists. This is used
   697  // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
   698  // new assists from going to sleep after this point.
   699  func gcWakeAllAssists() {
   700  	lock(&work.assistQueue.lock)
   701  	list := work.assistQueue.q.popList()
   702  	injectglist(&list)
   703  	unlock(&work.assistQueue.lock)
   704  }
   705  
   706  // gcParkAssist puts the current goroutine on the assist queue and parks.
   707  //
   708  // gcParkAssist reports whether the assist is now satisfied. If it
   709  // returns false, the caller must retry the assist.
   710  func gcParkAssist() bool {
   711  	lock(&work.assistQueue.lock)
   712  	// If the GC cycle finished while we were getting the lock,
   713  	// exit the assist. The cycle can't finish while we hold the
   714  	// lock.
   715  	if atomic.Load(&gcBlackenEnabled) == 0 {
   716  		unlock(&work.assistQueue.lock)
   717  		return true
   718  	}
   719  
   720  	gp := getg()
   721  	oldList := work.assistQueue.q
   722  	work.assistQueue.q.pushBack(gp)
   723  
   724  	// Recheck for background credit now that this G is in
   725  	// the queue, but can still back out. This avoids a
   726  	// race in case background marking has flushed more
   727  	// credit since we checked above.
   728  	if gcController.bgScanCredit.Load() > 0 {
   729  		work.assistQueue.q = oldList
   730  		if oldList.tail != 0 {
   731  			oldList.tail.ptr().schedlink.set(nil)
   732  		}
   733  		unlock(&work.assistQueue.lock)
   734  		return false
   735  	}
   736  	// Park.
   737  	goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceBlockGCMarkAssist, 2)
   738  	return true
   739  }
   740  
   741  // gcFlushBgCredit flushes scanWork units of background scan work
   742  // credit. This first satisfies blocked assists on the
   743  // work.assistQueue and then flushes any remaining credit to
   744  // gcController.bgScanCredit.
   745  //
   746  // Write barriers are disallowed because this is used by gcDrain after
   747  // it has ensured that all work is drained and this must preserve that
   748  // condition.
   749  //
   750  //go:nowritebarrierrec
   751  func gcFlushBgCredit(scanWork int64) {
   752  	if work.assistQueue.q.empty() {
   753  		// Fast path; there are no blocked assists. There's a
   754  		// small window here where an assist may add itself to
   755  		// the blocked queue and park. If that happens, we'll
   756  		// just get it on the next flush.
   757  		gcController.bgScanCredit.Add(scanWork)
   758  		return
   759  	}
   760  
   761  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   762  	scanBytes := int64(float64(scanWork) * assistBytesPerWork)
   763  
   764  	lock(&work.assistQueue.lock)
   765  	for !work.assistQueue.q.empty() && scanBytes > 0 {
   766  		gp := work.assistQueue.q.pop()
   767  		// Note that gp.gcAssistBytes is negative because gp
   768  		// is in debt. Think carefully about the signs below.
   769  		if scanBytes+gp.gcAssistBytes >= 0 {
   770  			// Satisfy this entire assist debt.
   771  			scanBytes += gp.gcAssistBytes
   772  			gp.gcAssistBytes = 0
   773  			// It's important that we *not* put gp in
   774  			// runnext. Otherwise, it's possible for user
   775  			// code to exploit the GC worker's high
   776  			// scheduler priority to get itself always run
   777  			// before other goroutines and always in the
   778  			// fresh quantum started by GC.
   779  			ready(gp, 0, false)
   780  		} else {
   781  			// Partially satisfy this assist.
   782  			gp.gcAssistBytes += scanBytes
   783  			scanBytes = 0
   784  			// As a heuristic, we move this assist to the
   785  			// back of the queue so that large assists
   786  			// can't clog up the assist queue and
   787  			// substantially delay small assists.
   788  			work.assistQueue.q.pushBack(gp)
   789  			break
   790  		}
   791  	}
   792  
   793  	if scanBytes > 0 {
   794  		// Convert from scan bytes back to work.
   795  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
   796  		scanWork = int64(float64(scanBytes) * assistWorkPerByte)
   797  		gcController.bgScanCredit.Add(scanWork)
   798  	}
   799  	unlock(&work.assistQueue.lock)
   800  }
   801  
   802  // scanstack scans gp's stack, greying all pointers found on the stack.
   803  //
   804  // Returns the amount of scan work performed, but doesn't update
   805  // gcController.stackScanWork or flush any credit. Any background credit produced
   806  // by this function should be flushed by its caller. scanstack itself can't
   807  // safely flush because it may result in trying to wake up a goroutine that
   808  // was just scanned, resulting in a self-deadlock.
   809  //
   810  // scanstack will also shrink the stack if it is safe to do so. If it
   811  // is not, it schedules a stack shrink for the next synchronous safe
   812  // point.
   813  //
   814  // scanstack is marked go:systemstack because it must not be preempted
   815  // while using a workbuf.
   816  //
   817  //go:nowritebarrier
   818  //go:systemstack
   819  func scanstack(gp *g, gcw *gcWork) int64 {
   820  	if readgstatus(gp)&_Gscan == 0 {
   821  		print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
   822  		throw("scanstack - bad status")
   823  	}
   824  
   825  	switch readgstatus(gp) &^ _Gscan {
   826  	default:
   827  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   828  		throw("mark - bad status")
   829  	case _Gdead:
   830  		return 0
   831  	case _Grunning:
   832  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   833  		throw("scanstack: goroutine not stopped")
   834  	case _Grunnable, _Gsyscall, _Gwaiting:
   835  		// ok
   836  	}
   837  
   838  	if gp == getg() {
   839  		throw("can't scan our own stack")
   840  	}
   841  
   842  	// scannedSize is the amount of work we'll be reporting.
   843  	//
   844  	// It is less than the allocated size (which is hi-lo).
   845  	var sp uintptr
   846  	if gp.syscallsp != 0 {
   847  		sp = gp.syscallsp // If in a system call this is the stack pointer (gp.sched.sp can be 0 in this case on Windows).
   848  	} else {
   849  		sp = gp.sched.sp
   850  	}
   851  	scannedSize := gp.stack.hi - sp
   852  
   853  	// Keep statistics for initial stack size calculation.
   854  	// Note that this accumulates the scanned size, not the allocated size.
   855  	p := getg().m.p.ptr()
   856  	p.scannedStackSize += uint64(scannedSize)
   857  	p.scannedStacks++
   858  
   859  	if isShrinkStackSafe(gp) {
   860  		// Shrink the stack if not much of it is being used.
   861  		shrinkstack(gp)
   862  	} else {
   863  		// Otherwise, shrink the stack at the next sync safe point.
   864  		gp.preemptShrink = true
   865  	}
   866  
   867  	var state stackScanState
   868  	state.stack = gp.stack
   869  
   870  	if stackTraceDebug {
   871  		println("stack trace goroutine", gp.goid)
   872  	}
   873  
   874  	if debugScanConservative && gp.asyncSafePoint {
   875  		print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
   876  	}
   877  
   878  	// Scan the saved context register. This is effectively a live
   879  	// register that gets moved back and forth between the
   880  	// register and sched.ctxt without a write barrier.
   881  	if gp.sched.ctxt != nil {
   882  		scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   883  	}
   884  
   885  	// Scan the stack. Accumulate a list of stack objects.
   886  	var u unwinder
   887  	for u.init(gp, 0); u.valid(); u.next() {
   888  		scanframeworker(&u.frame, &state, gcw)
   889  	}
   890  
   891  	// Find additional pointers that point into the stack from the heap.
   892  	// Currently this includes defers and panics. See also function copystack.
   893  
   894  	// Find and trace other pointers in defer records.
   895  	for d := gp._defer; d != nil; d = d.link {
   896  		if d.fn != nil {
   897  			// Scan the func value, which could be a stack allocated closure.
   898  			// See issue 30453.
   899  			scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   900  		}
   901  		if d.link != nil {
   902  			// The link field of a stack-allocated defer record might point
   903  			// to a heap-allocated defer record. Keep that heap record live.
   904  			scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   905  		}
   906  		// Retain defers records themselves.
   907  		// Defer records might not be reachable from the G through regular heap
   908  		// tracing because the defer linked list might weave between the stack and the heap.
   909  		if d.heap {
   910  			scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   911  		}
   912  	}
   913  	if gp._panic != nil {
   914  		// Panics are always stack allocated.
   915  		state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
   916  	}
   917  
   918  	// Find and scan all reachable stack objects.
   919  	//
   920  	// The state's pointer queue prioritizes precise pointers over
   921  	// conservative pointers so that we'll prefer scanning stack
   922  	// objects precisely.
   923  	state.buildIndex()
   924  	for {
   925  		p, conservative := state.getPtr()
   926  		if p == 0 {
   927  			break
   928  		}
   929  		obj := state.findObject(p)
   930  		if obj == nil {
   931  			continue
   932  		}
   933  		r := obj.r
   934  		if r == nil {
   935  			// We've already scanned this object.
   936  			continue
   937  		}
   938  		obj.setRecord(nil) // Don't scan it again.
   939  		if stackTraceDebug {
   940  			printlock()
   941  			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
   942  			if conservative {
   943  				print(" (conservative)")
   944  			}
   945  			println()
   946  			printunlock()
   947  		}
   948  		gcdata := r.gcdata()
   949  		var s *mspan
   950  		if r.useGCProg() {
   951  			// This path is pretty unlikely, an object large enough
   952  			// to have a GC program allocated on the stack.
   953  			// We need some space to unpack the program into a straight
   954  			// bitmask, which we allocate/free here.
   955  			// TODO: it would be nice if there were a way to run a GC
   956  			// program without having to store all its bits. We'd have
   957  			// to change from a Lempel-Ziv style program to something else.
   958  			// Or we can forbid putting objects on stacks if they require
   959  			// a gc program (see issue 27447).
   960  			s = materializeGCProg(r.ptrdata(), gcdata)
   961  			gcdata = (*byte)(unsafe.Pointer(s.startAddr))
   962  		}
   963  
   964  		b := state.stack.lo + uintptr(obj.off)
   965  		if conservative {
   966  			scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
   967  		} else {
   968  			scanblock(b, r.ptrdata(), gcdata, gcw, &state)
   969  		}
   970  
   971  		if s != nil {
   972  			dematerializeGCProg(s)
   973  		}
   974  	}
   975  
   976  	// Deallocate object buffers.
   977  	// (Pointer buffers were all deallocated in the loop above.)
   978  	for state.head != nil {
   979  		x := state.head
   980  		state.head = x.next
   981  		if stackTraceDebug {
   982  			for i := 0; i < x.nobj; i++ {
   983  				obj := &x.obj[i]
   984  				if obj.r == nil { // reachable
   985  					continue
   986  				}
   987  				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
   988  				// Note: not necessarily really dead - only reachable-from-ptr dead.
   989  			}
   990  		}
   991  		x.nobj = 0
   992  		putempty((*workbuf)(unsafe.Pointer(x)))
   993  	}
   994  	if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
   995  		throw("remaining pointer buffers")
   996  	}
   997  	return int64(scannedSize)
   998  }
   999  
  1000  // Scan a stack frame: local variables and function arguments/results.
  1001  //
  1002  //go:nowritebarrier
  1003  func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
  1004  	if _DebugGC > 1 && frame.continpc != 0 {
  1005  		print("scanframe ", funcname(frame.fn), "\n")
  1006  	}
  1007  
  1008  	isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == abi.FuncID_asyncPreempt
  1009  	isDebugCall := frame.fn.valid() && frame.fn.funcID == abi.FuncID_debugCallV2
  1010  	if state.conservative || isAsyncPreempt || isDebugCall {
  1011  		if debugScanConservative {
  1012  			println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
  1013  		}
  1014  
  1015  		// Conservatively scan the frame. Unlike the precise
  1016  		// case, this includes the outgoing argument space
  1017  		// since we may have stopped while this function was
  1018  		// setting up a call.
  1019  		//
  1020  		// TODO: We could narrow this down if the compiler
  1021  		// produced a single map per function of stack slots
  1022  		// and registers that ever contain a pointer.
  1023  		if frame.varp != 0 {
  1024  			size := frame.varp - frame.sp
  1025  			if size > 0 {
  1026  				scanConservative(frame.sp, size, nil, gcw, state)
  1027  			}
  1028  		}
  1029  
  1030  		// Scan arguments to this frame.
  1031  		if n := frame.argBytes(); n != 0 {
  1032  			// TODO: We could pass the entry argument map
  1033  			// to narrow this down further.
  1034  			scanConservative(frame.argp, n, nil, gcw, state)
  1035  		}
  1036  
  1037  		if isAsyncPreempt || isDebugCall {
  1038  			// This function's frame contained the
  1039  			// registers for the asynchronously stopped
  1040  			// parent frame. Scan the parent
  1041  			// conservatively.
  1042  			state.conservative = true
  1043  		} else {
  1044  			// We only wanted to scan those two frames
  1045  			// conservatively. Clear the flag for future
  1046  			// frames.
  1047  			state.conservative = false
  1048  		}
  1049  		return
  1050  	}
  1051  
  1052  	locals, args, objs := frame.getStackMap(false)
  1053  
  1054  	// Scan local variables if stack frame has been allocated.
  1055  	if locals.n > 0 {
  1056  		size := uintptr(locals.n) * goarch.PtrSize
  1057  		scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
  1058  	}
  1059  
  1060  	// Scan arguments.
  1061  	if args.n > 0 {
  1062  		scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
  1063  	}
  1064  
  1065  	// Add all stack objects to the stack object list.
  1066  	if frame.varp != 0 {
  1067  		// varp is 0 for defers, where there are no locals.
  1068  		// In that case, there can't be a pointer to its args, either.
  1069  		// (And all args would be scanned above anyway.)
  1070  		for i := range objs {
  1071  			obj := &objs[i]
  1072  			off := obj.off
  1073  			base := frame.varp // locals base pointer
  1074  			if off >= 0 {
  1075  				base = frame.argp // arguments and return values base pointer
  1076  			}
  1077  			ptr := base + uintptr(off)
  1078  			if ptr < frame.sp {
  1079  				// object hasn't been allocated in the frame yet.
  1080  				continue
  1081  			}
  1082  			if stackTraceDebug {
  1083  				println("stkobj at", hex(ptr), "of size", obj.size)
  1084  			}
  1085  			state.addObject(ptr, obj)
  1086  		}
  1087  	}
  1088  }
  1089  
  1090  type gcDrainFlags int
  1091  
  1092  const (
  1093  	gcDrainUntilPreempt gcDrainFlags = 1 << iota
  1094  	gcDrainFlushBgCredit
  1095  	gcDrainIdle
  1096  	gcDrainFractional
  1097  )
  1098  
  1099  // gcDrainMarkWorkerIdle is a wrapper for gcDrain that exists to better account
  1100  // mark time in profiles.
  1101  func gcDrainMarkWorkerIdle(gcw *gcWork) {
  1102  	gcDrain(gcw, gcDrainIdle|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1103  }
  1104  
  1105  // gcDrainMarkWorkerDedicated is a wrapper for gcDrain that exists to better account
  1106  // mark time in profiles.
  1107  func gcDrainMarkWorkerDedicated(gcw *gcWork, untilPreempt bool) {
  1108  	flags := gcDrainFlushBgCredit
  1109  	if untilPreempt {
  1110  		flags |= gcDrainUntilPreempt
  1111  	}
  1112  	gcDrain(gcw, flags)
  1113  }
  1114  
  1115  // gcDrainMarkWorkerFractional is a wrapper for gcDrain that exists to better account
  1116  // mark time in profiles.
  1117  func gcDrainMarkWorkerFractional(gcw *gcWork) {
  1118  	gcDrain(gcw, gcDrainFractional|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1119  }
  1120  
  1121  // gcDrain scans roots and objects in work buffers, blackening grey
  1122  // objects until it is unable to get more work. It may return before
  1123  // GC is done; it's the caller's responsibility to balance work from
  1124  // other Ps.
  1125  //
  1126  // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
  1127  // is set.
  1128  //
  1129  // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
  1130  // to do.
  1131  //
  1132  // If flags&gcDrainFractional != 0, gcDrain self-preempts when
  1133  // pollFractionalWorkerExit() returns true. This implies
  1134  // gcDrainNoBlock.
  1135  //
  1136  // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
  1137  // credit to gcController.bgScanCredit every gcCreditSlack units of
  1138  // scan work.
  1139  //
  1140  // gcDrain will always return if there is a pending STW or forEachP.
  1141  //
  1142  // Disabling write barriers is necessary to ensure that after we've
  1143  // confirmed that we've drained gcw, that we don't accidentally end
  1144  // up flipping that condition by immediately adding work in the form
  1145  // of a write barrier buffer flush.
  1146  //
  1147  // Don't set nowritebarrierrec because it's safe for some callees to
  1148  // have write barriers enabled.
  1149  //
  1150  //go:nowritebarrier
  1151  func gcDrain(gcw *gcWork, flags gcDrainFlags) {
  1152  	if !writeBarrier.enabled {
  1153  		throw("gcDrain phase incorrect")
  1154  	}
  1155  
  1156  	// N.B. We must be running in a non-preemptible context, so it's
  1157  	// safe to hold a reference to our P here.
  1158  	gp := getg().m.curg
  1159  	pp := gp.m.p.ptr()
  1160  	preemptible := flags&gcDrainUntilPreempt != 0
  1161  	flushBgCredit := flags&gcDrainFlushBgCredit != 0
  1162  	idle := flags&gcDrainIdle != 0
  1163  
  1164  	initScanWork := gcw.heapScanWork
  1165  
  1166  	// checkWork is the scan work before performing the next
  1167  	// self-preempt check.
  1168  	checkWork := int64(1<<63 - 1)
  1169  	var check func() bool
  1170  	if flags&(gcDrainIdle|gcDrainFractional) != 0 {
  1171  		checkWork = initScanWork + drainCheckThreshold
  1172  		if idle {
  1173  			check = pollWork
  1174  		} else if flags&gcDrainFractional != 0 {
  1175  			check = pollFractionalWorkerExit
  1176  		}
  1177  	}
  1178  
  1179  	// Drain root marking jobs.
  1180  	if work.markrootNext < work.markrootJobs {
  1181  		// Stop if we're preemptible, if someone wants to STW, or if
  1182  		// someone is calling forEachP.
  1183  		for !(gp.preempt && (preemptible || sched.gcwaiting.Load() || pp.runSafePointFn != 0)) {
  1184  			job := atomic.Xadd(&work.markrootNext, +1) - 1
  1185  			if job >= work.markrootJobs {
  1186  				break
  1187  			}
  1188  			markroot(gcw, job, flushBgCredit)
  1189  			if check != nil && check() {
  1190  				goto done
  1191  			}
  1192  		}
  1193  	}
  1194  
  1195  	// Drain heap marking jobs.
  1196  	//
  1197  	// Stop if we're preemptible, if someone wants to STW, or if
  1198  	// someone is calling forEachP.
  1199  	//
  1200  	// TODO(mknyszek): Consider always checking gp.preempt instead
  1201  	// of having the preempt flag, and making an exception for certain
  1202  	// mark workers in retake. That might be simpler than trying to
  1203  	// enumerate all the reasons why we might want to preempt, even
  1204  	// if we're supposed to be mostly non-preemptible.
  1205  	for !(gp.preempt && (preemptible || sched.gcwaiting.Load() || pp.runSafePointFn != 0)) {
  1206  		// Try to keep work available on the global queue. We used to
  1207  		// check if there were waiting workers, but it's better to
  1208  		// just keep work available than to make workers wait. In the
  1209  		// worst case, we'll do O(log(_WorkbufSize)) unnecessary
  1210  		// balances.
  1211  		if work.full == 0 {
  1212  			gcw.balance()
  1213  		}
  1214  
  1215  		b := gcw.tryGetFast()
  1216  		if b == 0 {
  1217  			b = gcw.tryGet()
  1218  			if b == 0 {
  1219  				// Flush the write barrier
  1220  				// buffer; this may create
  1221  				// more work.
  1222  				wbBufFlush()
  1223  				b = gcw.tryGet()
  1224  			}
  1225  		}
  1226  		if b == 0 {
  1227  			// Unable to get work.
  1228  			break
  1229  		}
  1230  		scanobject(b, gcw)
  1231  
  1232  		// Flush background scan work credit to the global
  1233  		// account if we've accumulated enough locally so
  1234  		// mutator assists can draw on it.
  1235  		if gcw.heapScanWork >= gcCreditSlack {
  1236  			gcController.heapScanWork.Add(gcw.heapScanWork)
  1237  			if flushBgCredit {
  1238  				gcFlushBgCredit(gcw.heapScanWork - initScanWork)
  1239  				initScanWork = 0
  1240  			}
  1241  			checkWork -= gcw.heapScanWork
  1242  			gcw.heapScanWork = 0
  1243  
  1244  			if checkWork <= 0 {
  1245  				checkWork += drainCheckThreshold
  1246  				if check != nil && check() {
  1247  					break
  1248  				}
  1249  			}
  1250  		}
  1251  	}
  1252  
  1253  done:
  1254  	// Flush remaining scan work credit.
  1255  	if gcw.heapScanWork > 0 {
  1256  		gcController.heapScanWork.Add(gcw.heapScanWork)
  1257  		if flushBgCredit {
  1258  			gcFlushBgCredit(gcw.heapScanWork - initScanWork)
  1259  		}
  1260  		gcw.heapScanWork = 0
  1261  	}
  1262  }
  1263  
  1264  // gcDrainN blackens grey objects until it has performed roughly
  1265  // scanWork units of scan work or the G is preempted. This is
  1266  // best-effort, so it may perform less work if it fails to get a work
  1267  // buffer. Otherwise, it will perform at least n units of work, but
  1268  // may perform more because scanning is always done in whole object
  1269  // increments. It returns the amount of scan work performed.
  1270  //
  1271  // The caller goroutine must be in a preemptible state (e.g.,
  1272  // _Gwaiting) to prevent deadlocks during stack scanning. As a
  1273  // consequence, this must be called on the system stack.
  1274  //
  1275  //go:nowritebarrier
  1276  //go:systemstack
  1277  func gcDrainN(gcw *gcWork, scanWork int64) int64 {
  1278  	if !writeBarrier.enabled {
  1279  		throw("gcDrainN phase incorrect")
  1280  	}
  1281  
  1282  	// There may already be scan work on the gcw, which we don't
  1283  	// want to claim was done by this call.
  1284  	workFlushed := -gcw.heapScanWork
  1285  
  1286  	// In addition to backing out because of a preemption, back out
  1287  	// if the GC CPU limiter is enabled.
  1288  	gp := getg().m.curg
  1289  	for !gp.preempt && !gcCPULimiter.limiting() && workFlushed+gcw.heapScanWork < scanWork {
  1290  		// See gcDrain comment.
  1291  		if work.full == 0 {
  1292  			gcw.balance()
  1293  		}
  1294  
  1295  		b := gcw.tryGetFast()
  1296  		if b == 0 {
  1297  			b = gcw.tryGet()
  1298  			if b == 0 {
  1299  				// Flush the write barrier buffer;
  1300  				// this may create more work.
  1301  				wbBufFlush()
  1302  				b = gcw.tryGet()
  1303  			}
  1304  		}
  1305  
  1306  		if b == 0 {
  1307  			// Try to do a root job.
  1308  			if work.markrootNext < work.markrootJobs {
  1309  				job := atomic.Xadd(&work.markrootNext, +1) - 1
  1310  				if job < work.markrootJobs {
  1311  					workFlushed += markroot(gcw, job, false)
  1312  					continue
  1313  				}
  1314  			}
  1315  			// No heap or root jobs.
  1316  			break
  1317  		}
  1318  
  1319  		scanobject(b, gcw)
  1320  
  1321  		// Flush background scan work credit.
  1322  		if gcw.heapScanWork >= gcCreditSlack {
  1323  			gcController.heapScanWork.Add(gcw.heapScanWork)
  1324  			workFlushed += gcw.heapScanWork
  1325  			gcw.heapScanWork = 0
  1326  		}
  1327  	}
  1328  
  1329  	// Unlike gcDrain, there's no need to flush remaining work
  1330  	// here because this never flushes to bgScanCredit and
  1331  	// gcw.dispose will flush any remaining work to scanWork.
  1332  
  1333  	return workFlushed + gcw.heapScanWork
  1334  }
  1335  
  1336  // scanblock scans b as scanobject would, but using an explicit
  1337  // pointer bitmap instead of the heap bitmap.
  1338  //
  1339  // This is used to scan non-heap roots, so it does not update
  1340  // gcw.bytesMarked or gcw.heapScanWork.
  1341  //
  1342  // If stk != nil, possible stack pointers are also reported to stk.putPtr.
  1343  //
  1344  //go:nowritebarrier
  1345  func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
  1346  	// Use local copies of original parameters, so that a stack trace
  1347  	// due to one of the throws below shows the original block
  1348  	// base and extent.
  1349  	b := b0
  1350  	n := n0
  1351  
  1352  	for i := uintptr(0); i < n; {
  1353  		// Find bits for the next word.
  1354  		bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
  1355  		if bits == 0 {
  1356  			i += goarch.PtrSize * 8
  1357  			continue
  1358  		}
  1359  		for j := 0; j < 8 && i < n; j++ {
  1360  			if bits&1 != 0 {
  1361  				// Same work as in scanobject; see comments there.
  1362  				p := *(*uintptr)(unsafe.Pointer(b + i))
  1363  				if p != 0 {
  1364  					if obj, span, objIndex := findObject(p, b, i); obj != 0 {
  1365  						greyobject(obj, b, i, span, gcw, objIndex)
  1366  					} else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
  1367  						stk.putPtr(p, false)
  1368  					}
  1369  				}
  1370  			}
  1371  			bits >>= 1
  1372  			i += goarch.PtrSize
  1373  		}
  1374  	}
  1375  }
  1376  
  1377  // scanobject scans the object starting at b, adding pointers to gcw.
  1378  // b must point to the beginning of a heap object or an oblet.
  1379  // scanobject consults the GC bitmap for the pointer mask and the
  1380  // spans for the size of the object.
  1381  //
  1382  //go:nowritebarrier
  1383  func scanobject(b uintptr, gcw *gcWork) {
  1384  	// Prefetch object before we scan it.
  1385  	//
  1386  	// This will overlap fetching the beginning of the object with initial
  1387  	// setup before we start scanning the object.
  1388  	sys.Prefetch(b)
  1389  
  1390  	// Find the bits for b and the size of the object at b.
  1391  	//
  1392  	// b is either the beginning of an object, in which case this
  1393  	// is the size of the object to scan, or it points to an
  1394  	// oblet, in which case we compute the size to scan below.
  1395  	s := spanOfUnchecked(b)
  1396  	n := s.elemsize
  1397  	if n == 0 {
  1398  		throw("scanobject n == 0")
  1399  	}
  1400  	if s.spanclass.noscan() {
  1401  		// Correctness-wise this is ok, but it's inefficient
  1402  		// if noscan objects reach here.
  1403  		throw("scanobject of a noscan object")
  1404  	}
  1405  
  1406  	var tp typePointers
  1407  	if n > maxObletBytes {
  1408  		// Large object. Break into oblets for better
  1409  		// parallelism and lower latency.
  1410  		if b == s.base() {
  1411  			// Enqueue the other oblets to scan later.
  1412  			// Some oblets may be in b's scalar tail, but
  1413  			// these will be marked as "no more pointers",
  1414  			// so we'll drop out immediately when we go to
  1415  			// scan those.
  1416  			for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
  1417  				if !gcw.putFast(oblet) {
  1418  					gcw.put(oblet)
  1419  				}
  1420  			}
  1421  		}
  1422  
  1423  		// Compute the size of the oblet. Since this object
  1424  		// must be a large object, s.base() is the beginning
  1425  		// of the object.
  1426  		n = s.base() + s.elemsize - b
  1427  		n = min(n, maxObletBytes)
  1428  		tp = s.typePointersOfUnchecked(s.base())
  1429  		tp = tp.fastForward(b-tp.addr, b+n)
  1430  	} else {
  1431  		tp = s.typePointersOfUnchecked(b)
  1432  	}
  1433  
  1434  	var scanSize uintptr
  1435  	for {
  1436  		var addr uintptr
  1437  		if tp, addr = tp.nextFast(); addr == 0 {
  1438  			if tp, addr = tp.next(b + n); addr == 0 {
  1439  				break
  1440  			}
  1441  		}
  1442  
  1443  		// Keep track of farthest pointer we found, so we can
  1444  		// update heapScanWork. TODO: is there a better metric,
  1445  		// now that we can skip scalar portions pretty efficiently?
  1446  		scanSize = addr - b + goarch.PtrSize
  1447  
  1448  		// Work here is duplicated in scanblock and above.
  1449  		// If you make changes here, make changes there too.
  1450  		obj := *(*uintptr)(unsafe.Pointer(addr))
  1451  
  1452  		// At this point we have extracted the next potential pointer.
  1453  		// Quickly filter out nil and pointers back to the current object.
  1454  		if obj != 0 && obj-b >= n {
  1455  			// Test if obj points into the Go heap and, if so,
  1456  			// mark the object.
  1457  			//
  1458  			// Note that it's possible for findObject to
  1459  			// fail if obj points to a just-allocated heap
  1460  			// object because of a race with growing the
  1461  			// heap. In this case, we know the object was
  1462  			// just allocated and hence will be marked by
  1463  			// allocation itself.
  1464  			if obj, span, objIndex := findObject(obj, b, addr-b); obj != 0 {
  1465  				greyobject(obj, b, addr-b, span, gcw, objIndex)
  1466  			}
  1467  		}
  1468  	}
  1469  	gcw.bytesMarked += uint64(n)
  1470  	gcw.heapScanWork += int64(scanSize)
  1471  }
  1472  
  1473  // scanConservative scans block [b, b+n) conservatively, treating any
  1474  // pointer-like value in the block as a pointer.
  1475  //
  1476  // If ptrmask != nil, only words that are marked in ptrmask are
  1477  // considered as potential pointers.
  1478  //
  1479  // If state != nil, it's assumed that [b, b+n) is a block in the stack
  1480  // and may contain pointers to stack objects.
  1481  func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
  1482  	if debugScanConservative {
  1483  		printlock()
  1484  		print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
  1485  		hexdumpWords(b, b+n, func(p uintptr) byte {
  1486  			if ptrmask != nil {
  1487  				word := (p - b) / goarch.PtrSize
  1488  				bits := *addb(ptrmask, word/8)
  1489  				if (bits>>(word%8))&1 == 0 {
  1490  					return '$'
  1491  				}
  1492  			}
  1493  
  1494  			val := *(*uintptr)(unsafe.Pointer(p))
  1495  			if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1496  				return '@'
  1497  			}
  1498  
  1499  			span := spanOfHeap(val)
  1500  			if span == nil {
  1501  				return ' '
  1502  			}
  1503  			idx := span.objIndex(val)
  1504  			if span.isFree(idx) {
  1505  				return ' '
  1506  			}
  1507  			return '*'
  1508  		})
  1509  		printunlock()
  1510  	}
  1511  
  1512  	for i := uintptr(0); i < n; i += goarch.PtrSize {
  1513  		if ptrmask != nil {
  1514  			word := i / goarch.PtrSize
  1515  			bits := *addb(ptrmask, word/8)
  1516  			if bits == 0 {
  1517  				// Skip 8 words (the loop increment will do the 8th)
  1518  				//
  1519  				// This must be the first time we've
  1520  				// seen this word of ptrmask, so i
  1521  				// must be 8-word-aligned, but check
  1522  				// our reasoning just in case.
  1523  				if i%(goarch.PtrSize*8) != 0 {
  1524  					throw("misaligned mask")
  1525  				}
  1526  				i += goarch.PtrSize*8 - goarch.PtrSize
  1527  				continue
  1528  			}
  1529  			if (bits>>(word%8))&1 == 0 {
  1530  				continue
  1531  			}
  1532  		}
  1533  
  1534  		val := *(*uintptr)(unsafe.Pointer(b + i))
  1535  
  1536  		// Check if val points into the stack.
  1537  		if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1538  			// val may point to a stack object. This
  1539  			// object may be dead from last cycle and
  1540  			// hence may contain pointers to unallocated
  1541  			// objects, but unlike heap objects we can't
  1542  			// tell if it's already dead. Hence, if all
  1543  			// pointers to this object are from
  1544  			// conservative scanning, we have to scan it
  1545  			// defensively, too.
  1546  			state.putPtr(val, true)
  1547  			continue
  1548  		}
  1549  
  1550  		// Check if val points to a heap span.
  1551  		span := spanOfHeap(val)
  1552  		if span == nil {
  1553  			continue
  1554  		}
  1555  
  1556  		// Check if val points to an allocated object.
  1557  		idx := span.objIndex(val)
  1558  		if span.isFree(idx) {
  1559  			continue
  1560  		}
  1561  
  1562  		// val points to an allocated object. Mark it.
  1563  		obj := span.base() + idx*span.elemsize
  1564  		greyobject(obj, b, i, span, gcw, idx)
  1565  	}
  1566  }
  1567  
  1568  // Shade the object if it isn't already.
  1569  // The object is not nil and known to be in the heap.
  1570  // Preemption must be disabled.
  1571  //
  1572  //go:nowritebarrier
  1573  func shade(b uintptr) {
  1574  	if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
  1575  		gcw := &getg().m.p.ptr().gcw
  1576  		greyobject(obj, 0, 0, span, gcw, objIndex)
  1577  	}
  1578  }
  1579  
  1580  // obj is the start of an object with mark mbits.
  1581  // If it isn't already marked, mark it and enqueue into gcw.
  1582  // base and off are for debugging only and could be removed.
  1583  //
  1584  // See also wbBufFlush1, which partially duplicates this logic.
  1585  //
  1586  //go:nowritebarrierrec
  1587  func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
  1588  	// obj should be start of allocation, and so must be at least pointer-aligned.
  1589  	if obj&(goarch.PtrSize-1) != 0 {
  1590  		throw("greyobject: obj not pointer-aligned")
  1591  	}
  1592  	mbits := span.markBitsForIndex(objIndex)
  1593  
  1594  	if useCheckmark {
  1595  		if setCheckmark(obj, base, off, mbits) {
  1596  			// Already marked.
  1597  			return
  1598  		}
  1599  	} else {
  1600  		if debug.gccheckmark > 0 && span.isFree(objIndex) {
  1601  			print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
  1602  			gcDumpObject("base", base, off)
  1603  			gcDumpObject("obj", obj, ^uintptr(0))
  1604  			getg().m.traceback = 2
  1605  			throw("marking free object")
  1606  		}
  1607  
  1608  		// If marked we have nothing to do.
  1609  		if mbits.isMarked() {
  1610  			return
  1611  		}
  1612  		mbits.setMarked()
  1613  
  1614  		// Mark span.
  1615  		arena, pageIdx, pageMask := pageIndexOf(span.base())
  1616  		if arena.pageMarks[pageIdx]&pageMask == 0 {
  1617  			atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1618  		}
  1619  
  1620  		// If this is a noscan object, fast-track it to black
  1621  		// instead of greying it.
  1622  		if span.spanclass.noscan() {
  1623  			gcw.bytesMarked += uint64(span.elemsize)
  1624  			return
  1625  		}
  1626  	}
  1627  
  1628  	// We're adding obj to P's local workbuf, so it's likely
  1629  	// this object will be processed soon by the same P.
  1630  	// Even if the workbuf gets flushed, there will likely still be
  1631  	// some benefit on platforms with inclusive shared caches.
  1632  	sys.Prefetch(obj)
  1633  	// Queue the obj for scanning.
  1634  	if !gcw.putFast(obj) {
  1635  		gcw.put(obj)
  1636  	}
  1637  }
  1638  
  1639  // gcDumpObject dumps the contents of obj for debugging and marks the
  1640  // field at byte offset off in obj.
  1641  func gcDumpObject(label string, obj, off uintptr) {
  1642  	s := spanOf(obj)
  1643  	print(label, "=", hex(obj))
  1644  	if s == nil {
  1645  		print(" s=nil\n")
  1646  		return
  1647  	}
  1648  	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
  1649  	if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
  1650  		print(mSpanStateNames[state], "\n")
  1651  	} else {
  1652  		print("unknown(", state, ")\n")
  1653  	}
  1654  
  1655  	skipped := false
  1656  	size := s.elemsize
  1657  	if s.state.get() == mSpanManual && size == 0 {
  1658  		// We're printing something from a stack frame. We
  1659  		// don't know how big it is, so just show up to an
  1660  		// including off.
  1661  		size = off + goarch.PtrSize
  1662  	}
  1663  	for i := uintptr(0); i < size; i += goarch.PtrSize {
  1664  		// For big objects, just print the beginning (because
  1665  		// that usually hints at the object's type) and the
  1666  		// fields around off.
  1667  		if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
  1668  			skipped = true
  1669  			continue
  1670  		}
  1671  		if skipped {
  1672  			print(" ...\n")
  1673  			skipped = false
  1674  		}
  1675  		print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
  1676  		if i == off {
  1677  			print(" <==")
  1678  		}
  1679  		print("\n")
  1680  	}
  1681  	if skipped {
  1682  		print(" ...\n")
  1683  	}
  1684  }
  1685  
  1686  // gcmarknewobject marks a newly allocated object black. obj must
  1687  // not contain any non-nil pointers.
  1688  //
  1689  // This is nosplit so it can manipulate a gcWork without preemption.
  1690  //
  1691  //go:nowritebarrier
  1692  //go:nosplit
  1693  func gcmarknewobject(span *mspan, obj uintptr) {
  1694  	if useCheckmark { // The world should be stopped so this should not happen.
  1695  		throw("gcmarknewobject called while doing checkmark")
  1696  	}
  1697  
  1698  	// Mark object.
  1699  	objIndex := span.objIndex(obj)
  1700  	span.markBitsForIndex(objIndex).setMarked()
  1701  
  1702  	// Mark span.
  1703  	arena, pageIdx, pageMask := pageIndexOf(span.base())
  1704  	if arena.pageMarks[pageIdx]&pageMask == 0 {
  1705  		atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1706  	}
  1707  
  1708  	gcw := &getg().m.p.ptr().gcw
  1709  	gcw.bytesMarked += uint64(span.elemsize)
  1710  }
  1711  
  1712  // gcMarkTinyAllocs greys all active tiny alloc blocks.
  1713  //
  1714  // The world must be stopped.
  1715  func gcMarkTinyAllocs() {
  1716  	assertWorldStopped()
  1717  
  1718  	for _, p := range allp {
  1719  		c := p.mcache
  1720  		if c == nil || c.tiny == 0 {
  1721  			continue
  1722  		}
  1723  		_, span, objIndex := findObject(c.tiny, 0, 0)
  1724  		gcw := &p.gcw
  1725  		greyobject(c.tiny, 0, 0, span, gcw, objIndex)
  1726  	}
  1727  }
  1728  

View as plain text