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  	"runtime/internal/atomic"
    12  	"runtime/internal/sys"
    13  	"unsafe"
    14  )
    15  
    16  const (
    17  	fixedRootFinalizers = iota
    18  	fixedRootFreeGStacks
    19  	fixedRootCount
    20  
    21  	// rootBlockBytes is the number of bytes to scan per data or
    22  	// BSS root.
    23  	rootBlockBytes = 256 << 10
    24  
    25  	// maxObletBytes is the maximum bytes of an object to scan at
    26  	// once. Larger objects will be split up into "oblets" of at
    27  	// most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
    28  	// scan preemption at ~100 µs.
    29  	//
    30  	// This must be > _MaxSmallSize so that the object base is the
    31  	// span base.
    32  	maxObletBytes = 128 << 10
    33  
    34  	// drainCheckThreshold specifies how many units of work to do
    35  	// between self-preemption checks in gcDrain. Assuming a scan
    36  	// rate of 1 MB/ms, this is ~100 µs. Lower values have higher
    37  	// overhead in the scan loop (the scheduler check may perform
    38  	// a syscall, so its overhead is nontrivial). Higher values
    39  	// make the system less responsive to incoming work.
    40  	drainCheckThreshold = 100000
    41  
    42  	// pagesPerSpanRoot indicates how many pages to scan from a span root
    43  	// at a time. Used by special root marking.
    44  	//
    45  	// Higher values improve throughput by increasing locality, but
    46  	// increase the minimum latency of a marking operation.
    47  	//
    48  	// Must be a multiple of the pageInUse bitmap element size and
    49  	// must also evenly divide pagesPerArena.
    50  	pagesPerSpanRoot = 512
    51  )
    52  
    53  // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
    54  // some miscellany) and initializes scanning-related state.
    55  //
    56  // The world must be stopped.
    57  func gcMarkRootPrepare() {
    58  	assertWorldStopped()
    59  
    60  	// Compute how many data and BSS root blocks there are.
    61  	nBlocks := func(bytes uintptr) int {
    62  		return int(divRoundUp(bytes, rootBlockBytes))
    63  	}
    64  
    65  	work.nDataRoots = 0
    66  	work.nBSSRoots = 0
    67  
    68  	// Scan globals.
    69  	for _, datap := range activeModules() {
    70  		nDataRoots := nBlocks(datap.edata - datap.data)
    71  		if nDataRoots > work.nDataRoots {
    72  			work.nDataRoots = nDataRoots
    73  		}
    74  	}
    75  
    76  	for _, datap := range activeModules() {
    77  		nBSSRoots := nBlocks(datap.ebss - datap.bss)
    78  		if nBSSRoots > work.nBSSRoots {
    79  			work.nBSSRoots = nBSSRoots
    80  		}
    81  	}
    82  
    83  	// Scan span roots for finalizer specials.
    84  	//
    85  	// We depend on addfinalizer to mark objects that get
    86  	// finalizers after root marking.
    87  	//
    88  	// We're going to scan the whole heap (that was available at the time the
    89  	// mark phase started, i.e. markArenas) for in-use spans which have specials.
    90  	//
    91  	// Break up the work into arenas, and further into chunks.
    92  	//
    93  	// Snapshot allArenas as markArenas. This snapshot is safe because allArenas
    94  	// is append-only.
    95  	mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
    96  	work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
    97  
    98  	// Scan stacks.
    99  	//
   100  	// Gs may be created after this point, but it's okay that we
   101  	// ignore them because they begin life without any roots, so
   102  	// there's nothing to scan, and any roots they create during
   103  	// the concurrent phase will be caught by the write barrier.
   104  	work.stackRoots = allGsSnapshot()
   105  	work.nStackRoots = len(work.stackRoots)
   106  
   107  	work.markrootNext = 0
   108  	work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
   109  
   110  	// Calculate base indexes of each root type
   111  	work.baseData = uint32(fixedRootCount)
   112  	work.baseBSS = work.baseData + uint32(work.nDataRoots)
   113  	work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
   114  	work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
   115  	work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
   116  }
   117  
   118  // gcMarkRootCheck checks that all roots have been scanned. It is
   119  // purely for debugging.
   120  func gcMarkRootCheck() {
   121  	if work.markrootNext < work.markrootJobs {
   122  		print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
   123  		throw("left over markroot jobs")
   124  	}
   125  
   126  	// Check that stacks have been scanned.
   127  	//
   128  	// We only check the first nStackRoots Gs that we should have scanned.
   129  	// Since we don't care about newer Gs (see comment in
   130  	// gcMarkRootPrepare), no locking is required.
   131  	i := 0
   132  	forEachGRace(func(gp *g) {
   133  		if i >= work.nStackRoots {
   134  			return
   135  		}
   136  
   137  		if !gp.gcscandone {
   138  			println("gp", gp, "goid", gp.goid,
   139  				"status", readgstatus(gp),
   140  				"gcscandone", gp.gcscandone)
   141  			throw("scan missed a g")
   142  		}
   143  
   144  		i++
   145  	})
   146  }
   147  
   148  // ptrmask for an allocation containing a single pointer.
   149  var oneptrmask = [...]uint8{1}
   150  
   151  // markroot scans the i'th root.
   152  //
   153  // Preemption must be disabled (because this uses a gcWork).
   154  //
   155  // Returns the amount of GC work credit produced by the operation.
   156  // If flushBgCredit is true, then that credit is also flushed
   157  // to the background credit pool.
   158  //
   159  // nowritebarrier is only advisory here.
   160  //
   161  //go:nowritebarrier
   162  func markroot(gcw *gcWork, i uint32, flushBgCredit bool) int64 {
   163  	// Note: if you add a case here, please also update heapdump.go:dumproots.
   164  	var workDone int64
   165  	var workCounter *atomic.Int64
   166  	switch {
   167  	case work.baseData <= i && i < work.baseBSS:
   168  		workCounter = &gcController.globalsScanWork
   169  		for _, datap := range activeModules() {
   170  			workDone += markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
   171  		}
   172  
   173  	case work.baseBSS <= i && i < work.baseSpans:
   174  		workCounter = &gcController.globalsScanWork
   175  		for _, datap := range activeModules() {
   176  			workDone += markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
   177  		}
   178  
   179  	case i == fixedRootFinalizers:
   180  		for fb := allfin; fb != nil; fb = fb.alllink {
   181  			cnt := uintptr(atomic.Load(&fb.cnt))
   182  			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
   183  		}
   184  
   185  	case i == fixedRootFreeGStacks:
   186  		// Switch to the system stack so we can call
   187  		// stackfree.
   188  		systemstack(markrootFreeGStacks)
   189  
   190  	case work.baseSpans <= i && i < work.baseStacks:
   191  		// mark mspan.specials
   192  		markrootSpans(gcw, int(i-work.baseSpans))
   193  
   194  	default:
   195  		// the rest is scanning goroutine stacks
   196  		workCounter = &gcController.stackScanWork
   197  		if i < work.baseStacks || work.baseEnd <= i {
   198  			printlock()
   199  			print("runtime: markroot index ", i, " not in stack roots range [", work.baseStacks, ", ", work.baseEnd, ")\n")
   200  			throw("markroot: bad index")
   201  		}
   202  		gp := work.stackRoots[i-work.baseStacks]
   203  
   204  		// remember when we've first observed the G blocked
   205  		// needed only to output in traceback
   206  		status := readgstatus(gp) // We are not in a scan state
   207  		if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
   208  			gp.waitsince = work.tstart
   209  		}
   210  
   211  		// scanstack must be done on the system stack in case
   212  		// we're trying to scan our own stack.
   213  		systemstack(func() {
   214  			// If this is a self-scan, put the user G in
   215  			// _Gwaiting to prevent self-deadlock. It may
   216  			// already be in _Gwaiting if this is a mark
   217  			// worker or we're in mark termination.
   218  			userG := getg().m.curg
   219  			selfScan := gp == userG && readgstatus(userG) == _Grunning
   220  			if selfScan {
   221  				casGToWaiting(userG, _Grunning, waitReasonGarbageCollectionScan)
   222  			}
   223  
   224  			// TODO: suspendG blocks (and spins) until gp
   225  			// stops, which may take a while for
   226  			// running goroutines. Consider doing this in
   227  			// two phases where the first is non-blocking:
   228  			// we scan the stacks we can and ask running
   229  			// goroutines to scan themselves; and the
   230  			// second blocks.
   231  			stopped := suspendG(gp)
   232  			if stopped.dead {
   233  				gp.gcscandone = true
   234  				return
   235  			}
   236  			if gp.gcscandone {
   237  				throw("g already scanned")
   238  			}
   239  			workDone += scanstack(gp, gcw)
   240  			gp.gcscandone = true
   241  			resumeG(stopped)
   242  
   243  			if selfScan {
   244  				casgstatus(userG, _Gwaiting, _Grunning)
   245  			}
   246  		})
   247  	}
   248  	if workCounter != nil && workDone != 0 {
   249  		workCounter.Add(workDone)
   250  		if flushBgCredit {
   251  			gcFlushBgCredit(workDone)
   252  		}
   253  	}
   254  	return workDone
   255  }
   256  
   257  // markrootBlock scans the shard'th shard of the block of memory [b0,
   258  // b0+n0), with the given pointer mask.
   259  //
   260  // Returns the amount of work done.
   261  //
   262  //go:nowritebarrier
   263  func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
   264  	if rootBlockBytes%(8*goarch.PtrSize) != 0 {
   265  		// This is necessary to pick byte offsets in ptrmask0.
   266  		throw("rootBlockBytes must be a multiple of 8*ptrSize")
   267  	}
   268  
   269  	// Note that if b0 is toward the end of the address space,
   270  	// then b0 + rootBlockBytes might wrap around.
   271  	// These tests are written to avoid any possible overflow.
   272  	off := uintptr(shard) * rootBlockBytes
   273  	if off >= n0 {
   274  		return 0
   275  	}
   276  	b := b0 + off
   277  	ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
   278  	n := uintptr(rootBlockBytes)
   279  	if off+n > n0 {
   280  		n = n0 - off
   281  	}
   282  
   283  	// Scan this shard.
   284  	scanblock(b, n, ptrmask, gcw, nil)
   285  	return int64(n)
   286  }
   287  
   288  // markrootFreeGStacks frees stacks of dead Gs.
   289  //
   290  // This does not free stacks of dead Gs cached on Ps, but having a few
   291  // cached stacks around isn't a problem.
   292  func markrootFreeGStacks() {
   293  	// Take list of dead Gs with stacks.
   294  	lock(&sched.gFree.lock)
   295  	list := sched.gFree.stack
   296  	sched.gFree.stack = gList{}
   297  	unlock(&sched.gFree.lock)
   298  	if list.empty() {
   299  		return
   300  	}
   301  
   302  	// Free stacks.
   303  	q := gQueue{list.head, list.head}
   304  	for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
   305  		stackfree(gp.stack)
   306  		gp.stack.lo = 0
   307  		gp.stack.hi = 0
   308  		// Manipulate the queue directly since the Gs are
   309  		// already all linked the right way.
   310  		q.tail.set(gp)
   311  	}
   312  
   313  	// Put Gs back on the free list.
   314  	lock(&sched.gFree.lock)
   315  	sched.gFree.noStack.pushAll(q)
   316  	unlock(&sched.gFree.lock)
   317  }
   318  
   319  // markrootSpans marks roots for one shard of markArenas.
   320  //
   321  //go:nowritebarrier
   322  func markrootSpans(gcw *gcWork, shard int) {
   323  	// Objects with finalizers have two GC-related invariants:
   324  	//
   325  	// 1) Everything reachable from the object must be marked.
   326  	// This ensures that when we pass the object to its finalizer,
   327  	// everything the finalizer can reach will be retained.
   328  	//
   329  	// 2) Finalizer specials (which are not in the garbage
   330  	// collected heap) are roots. In practice, this means the fn
   331  	// field must be scanned.
   332  	sg := mheap_.sweepgen
   333  
   334  	// Find the arena and page index into that arena for this shard.
   335  	ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
   336  	ha := mheap_.arenas[ai.l1()][ai.l2()]
   337  	arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
   338  
   339  	// Construct slice of bitmap which we'll iterate over.
   340  	specialsbits := ha.pageSpecials[arenaPage/8:]
   341  	specialsbits = specialsbits[:pagesPerSpanRoot/8]
   342  	for i := range specialsbits {
   343  		// Find set bits, which correspond to spans with specials.
   344  		specials := atomic.Load8(&specialsbits[i])
   345  		if specials == 0 {
   346  			continue
   347  		}
   348  		for j := uint(0); j < 8; j++ {
   349  			if specials&(1<<j) == 0 {
   350  				continue
   351  			}
   352  			// Find the span for this bit.
   353  			//
   354  			// This value is guaranteed to be non-nil because having
   355  			// specials implies that the span is in-use, and since we're
   356  			// currently marking we can be sure that we don't have to worry
   357  			// about the span being freed and re-used.
   358  			s := ha.spans[arenaPage+uint(i)*8+j]
   359  
   360  			// The state must be mSpanInUse if the specials bit is set, so
   361  			// sanity check that.
   362  			if state := s.state.get(); state != mSpanInUse {
   363  				print("s.state = ", state, "\n")
   364  				throw("non in-use span found with specials bit set")
   365  			}
   366  			// Check that this span was swept (it may be cached or uncached).
   367  			if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
   368  				// sweepgen was updated (+2) during non-checkmark GC pass
   369  				print("sweep ", s.sweepgen, " ", sg, "\n")
   370  				throw("gc: unswept span")
   371  			}
   372  
   373  			// Lock the specials to prevent a special from being
   374  			// removed from the list while we're traversing it.
   375  			lock(&s.speciallock)
   376  			for sp := s.specials; sp != nil; sp = sp.next {
   377  				if sp.kind != _KindSpecialFinalizer {
   378  					continue
   379  				}
   380  				// don't mark finalized object, but scan it so we
   381  				// retain everything it points to.
   382  				spf := (*specialfinalizer)(unsafe.Pointer(sp))
   383  				// A finalizer can be set for an inner byte of an object, find object beginning.
   384  				p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
   385  
   386  				// Mark everything that can be reached from
   387  				// the object (but *not* the object itself or
   388  				// we'll never collect it).
   389  				if !s.spanclass.noscan() {
   390  					scanobject(p, gcw)
   391  				}
   392  
   393  				// The special itself is a root.
   394  				scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
   395  			}
   396  			unlock(&s.speciallock)
   397  		}
   398  	}
   399  }
   400  
   401  // gcAssistAlloc performs GC work to make gp's assist debt positive.
   402  // gp must be the calling user goroutine.
   403  //
   404  // This must be called with preemption enabled.
   405  func gcAssistAlloc(gp *g) {
   406  	// Don't assist in non-preemptible contexts. These are
   407  	// generally fragile and won't allow the assist to block.
   408  	if getg() == gp.m.g0 {
   409  		return
   410  	}
   411  	if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
   412  		return
   413  	}
   414  
   415  	traced := false
   416  retry:
   417  	if go119MemoryLimitSupport && gcCPULimiter.limiting() {
   418  		// If the CPU limiter is enabled, intentionally don't
   419  		// assist to reduce the amount of CPU time spent in the GC.
   420  		if traced {
   421  			traceGCMarkAssistDone()
   422  		}
   423  		return
   424  	}
   425  	// Compute the amount of scan work we need to do to make the
   426  	// balance positive. When the required amount of work is low,
   427  	// we over-assist to build up credit for future allocations
   428  	// and amortize the cost of assisting.
   429  	assistWorkPerByte := gcController.assistWorkPerByte.Load()
   430  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   431  	debtBytes := -gp.gcAssistBytes
   432  	scanWork := int64(assistWorkPerByte * float64(debtBytes))
   433  	if scanWork < gcOverAssistWork {
   434  		scanWork = gcOverAssistWork
   435  		debtBytes = int64(assistBytesPerWork * float64(scanWork))
   436  	}
   437  
   438  	// Steal as much credit as we can from the background GC's
   439  	// scan credit. This is racy and may drop the background
   440  	// credit below 0 if two mutators steal at the same time. This
   441  	// will just cause steals to fail until credit is accumulated
   442  	// again, so in the long run it doesn't really matter, but we
   443  	// do have to handle the negative credit case.
   444  	bgScanCredit := gcController.bgScanCredit.Load()
   445  	stolen := int64(0)
   446  	if bgScanCredit > 0 {
   447  		if bgScanCredit < scanWork {
   448  			stolen = bgScanCredit
   449  			gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
   450  		} else {
   451  			stolen = scanWork
   452  			gp.gcAssistBytes += debtBytes
   453  		}
   454  		gcController.bgScanCredit.Add(-stolen)
   455  
   456  		scanWork -= stolen
   457  
   458  		if scanWork == 0 {
   459  			// We were able to steal all of the credit we
   460  			// needed.
   461  			if traced {
   462  				traceGCMarkAssistDone()
   463  			}
   464  			return
   465  		}
   466  	}
   467  
   468  	if trace.enabled && !traced {
   469  		traced = true
   470  		traceGCMarkAssistStart()
   471  	}
   472  
   473  	// Perform assist work
   474  	systemstack(func() {
   475  		gcAssistAlloc1(gp, scanWork)
   476  		// The user stack may have moved, so this can't touch
   477  		// anything on it until it returns from systemstack.
   478  	})
   479  
   480  	completed := gp.param != nil
   481  	gp.param = nil
   482  	if completed {
   483  		gcMarkDone()
   484  	}
   485  
   486  	if gp.gcAssistBytes < 0 {
   487  		// We were unable steal enough credit or perform
   488  		// enough work to pay off the assist debt. We need to
   489  		// do one of these before letting the mutator allocate
   490  		// more to prevent over-allocation.
   491  		//
   492  		// If this is because we were preempted, reschedule
   493  		// and try some more.
   494  		if gp.preempt {
   495  			Gosched()
   496  			goto retry
   497  		}
   498  
   499  		// Add this G to an assist queue and park. When the GC
   500  		// has more background credit, it will satisfy queued
   501  		// assists before flushing to the global credit pool.
   502  		//
   503  		// Note that this does *not* get woken up when more
   504  		// work is added to the work list. The theory is that
   505  		// there wasn't enough work to do anyway, so we might
   506  		// as well let background marking take care of the
   507  		// work that is available.
   508  		if !gcParkAssist() {
   509  			goto retry
   510  		}
   511  
   512  		// At this point either background GC has satisfied
   513  		// this G's assist debt, or the GC cycle is over.
   514  	}
   515  	if traced {
   516  		traceGCMarkAssistDone()
   517  	}
   518  }
   519  
   520  // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
   521  // stack. This is a separate function to make it easier to see that
   522  // we're not capturing anything from the user stack, since the user
   523  // stack may move while we're in this function.
   524  //
   525  // gcAssistAlloc1 indicates whether this assist completed the mark
   526  // phase by setting gp.param to non-nil. This can't be communicated on
   527  // the stack since it may move.
   528  //
   529  //go:systemstack
   530  func gcAssistAlloc1(gp *g, scanWork int64) {
   531  	// Clear the flag indicating that this assist completed the
   532  	// mark phase.
   533  	gp.param = nil
   534  
   535  	if atomic.Load(&gcBlackenEnabled) == 0 {
   536  		// The gcBlackenEnabled check in malloc races with the
   537  		// store that clears it but an atomic check in every malloc
   538  		// would be a performance hit.
   539  		// Instead we recheck it here on the non-preemptable system
   540  		// stack to determine if we should perform an assist.
   541  
   542  		// GC is done, so ignore any remaining debt.
   543  		gp.gcAssistBytes = 0
   544  		return
   545  	}
   546  	// Track time spent in this assist. Since we're on the
   547  	// system stack, this is non-preemptible, so we can
   548  	// just measure start and end time.
   549  	//
   550  	// Limiter event tracking might be disabled if we end up here
   551  	// while on a mark worker.
   552  	startTime := nanotime()
   553  	trackLimiterEvent := gp.m.p.ptr().limiterEvent.start(limiterEventMarkAssist, startTime)
   554  
   555  	decnwait := atomic.Xadd(&work.nwait, -1)
   556  	if decnwait == work.nproc {
   557  		println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
   558  		throw("nwait > work.nprocs")
   559  	}
   560  
   561  	// gcDrainN requires the caller to be preemptible.
   562  	casGToWaiting(gp, _Grunning, waitReasonGCAssistMarking)
   563  
   564  	// drain own cached work first in the hopes that it
   565  	// will be more cache friendly.
   566  	gcw := &getg().m.p.ptr().gcw
   567  	workDone := gcDrainN(gcw, scanWork)
   568  
   569  	casgstatus(gp, _Gwaiting, _Grunning)
   570  
   571  	// Record that we did this much scan work.
   572  	//
   573  	// Back out the number of bytes of assist credit that
   574  	// this scan work counts for. The "1+" is a poor man's
   575  	// round-up, to ensure this adds credit even if
   576  	// assistBytesPerWork is very low.
   577  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   578  	gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
   579  
   580  	// If this is the last worker and we ran out of work,
   581  	// signal a completion point.
   582  	incnwait := atomic.Xadd(&work.nwait, +1)
   583  	if incnwait > work.nproc {
   584  		println("runtime: work.nwait=", incnwait,
   585  			"work.nproc=", work.nproc)
   586  		throw("work.nwait > work.nproc")
   587  	}
   588  
   589  	if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
   590  		// This has reached a background completion point. Set
   591  		// gp.param to a non-nil value to indicate this. It
   592  		// doesn't matter what we set it to (it just has to be
   593  		// a valid pointer).
   594  		gp.param = unsafe.Pointer(gp)
   595  	}
   596  	now := nanotime()
   597  	duration := now - startTime
   598  	pp := gp.m.p.ptr()
   599  	pp.gcAssistTime += duration
   600  	if trackLimiterEvent {
   601  		pp.limiterEvent.stop(limiterEventMarkAssist, now)
   602  	}
   603  	if pp.gcAssistTime > gcAssistTimeSlack {
   604  		gcController.assistTime.Add(pp.gcAssistTime)
   605  		gcCPULimiter.update(now)
   606  		pp.gcAssistTime = 0
   607  	}
   608  }
   609  
   610  // gcWakeAllAssists wakes all currently blocked assists. This is used
   611  // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
   612  // new assists from going to sleep after this point.
   613  func gcWakeAllAssists() {
   614  	lock(&work.assistQueue.lock)
   615  	list := work.assistQueue.q.popList()
   616  	injectglist(&list)
   617  	unlock(&work.assistQueue.lock)
   618  }
   619  
   620  // gcParkAssist puts the current goroutine on the assist queue and parks.
   621  //
   622  // gcParkAssist reports whether the assist is now satisfied. If it
   623  // returns false, the caller must retry the assist.
   624  func gcParkAssist() bool {
   625  	lock(&work.assistQueue.lock)
   626  	// If the GC cycle finished while we were getting the lock,
   627  	// exit the assist. The cycle can't finish while we hold the
   628  	// lock.
   629  	if atomic.Load(&gcBlackenEnabled) == 0 {
   630  		unlock(&work.assistQueue.lock)
   631  		return true
   632  	}
   633  
   634  	gp := getg()
   635  	oldList := work.assistQueue.q
   636  	work.assistQueue.q.pushBack(gp)
   637  
   638  	// Recheck for background credit now that this G is in
   639  	// the queue, but can still back out. This avoids a
   640  	// race in case background marking has flushed more
   641  	// credit since we checked above.
   642  	if gcController.bgScanCredit.Load() > 0 {
   643  		work.assistQueue.q = oldList
   644  		if oldList.tail != 0 {
   645  			oldList.tail.ptr().schedlink.set(nil)
   646  		}
   647  		unlock(&work.assistQueue.lock)
   648  		return false
   649  	}
   650  	// Park.
   651  	goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
   652  	return true
   653  }
   654  
   655  // gcFlushBgCredit flushes scanWork units of background scan work
   656  // credit. This first satisfies blocked assists on the
   657  // work.assistQueue and then flushes any remaining credit to
   658  // gcController.bgScanCredit.
   659  //
   660  // Write barriers are disallowed because this is used by gcDrain after
   661  // it has ensured that all work is drained and this must preserve that
   662  // condition.
   663  //
   664  //go:nowritebarrierrec
   665  func gcFlushBgCredit(scanWork int64) {
   666  	if work.assistQueue.q.empty() {
   667  		// Fast path; there are no blocked assists. There's a
   668  		// small window here where an assist may add itself to
   669  		// the blocked queue and park. If that happens, we'll
   670  		// just get it on the next flush.
   671  		gcController.bgScanCredit.Add(scanWork)
   672  		return
   673  	}
   674  
   675  	assistBytesPerWork := gcController.assistBytesPerWork.Load()
   676  	scanBytes := int64(float64(scanWork) * assistBytesPerWork)
   677  
   678  	lock(&work.assistQueue.lock)
   679  	for !work.assistQueue.q.empty() && scanBytes > 0 {
   680  		gp := work.assistQueue.q.pop()
   681  		// Note that gp.gcAssistBytes is negative because gp
   682  		// is in debt. Think carefully about the signs below.
   683  		if scanBytes+gp.gcAssistBytes >= 0 {
   684  			// Satisfy this entire assist debt.
   685  			scanBytes += gp.gcAssistBytes
   686  			gp.gcAssistBytes = 0
   687  			// It's important that we *not* put gp in
   688  			// runnext. Otherwise, it's possible for user
   689  			// code to exploit the GC worker's high
   690  			// scheduler priority to get itself always run
   691  			// before other goroutines and always in the
   692  			// fresh quantum started by GC.
   693  			ready(gp, 0, false)
   694  		} else {
   695  			// Partially satisfy this assist.
   696  			gp.gcAssistBytes += scanBytes
   697  			scanBytes = 0
   698  			// As a heuristic, we move this assist to the
   699  			// back of the queue so that large assists
   700  			// can't clog up the assist queue and
   701  			// substantially delay small assists.
   702  			work.assistQueue.q.pushBack(gp)
   703  			break
   704  		}
   705  	}
   706  
   707  	if scanBytes > 0 {
   708  		// Convert from scan bytes back to work.
   709  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
   710  		scanWork = int64(float64(scanBytes) * assistWorkPerByte)
   711  		gcController.bgScanCredit.Add(scanWork)
   712  	}
   713  	unlock(&work.assistQueue.lock)
   714  }
   715  
   716  // scanstack scans gp's stack, greying all pointers found on the stack.
   717  //
   718  // Returns the amount of scan work performed, but doesn't update
   719  // gcController.stackScanWork or flush any credit. Any background credit produced
   720  // by this function should be flushed by its caller. scanstack itself can't
   721  // safely flush because it may result in trying to wake up a goroutine that
   722  // was just scanned, resulting in a self-deadlock.
   723  //
   724  // scanstack will also shrink the stack if it is safe to do so. If it
   725  // is not, it schedules a stack shrink for the next synchronous safe
   726  // point.
   727  //
   728  // scanstack is marked go:systemstack because it must not be preempted
   729  // while using a workbuf.
   730  //
   731  //go:nowritebarrier
   732  //go:systemstack
   733  func scanstack(gp *g, gcw *gcWork) int64 {
   734  	if readgstatus(gp)&_Gscan == 0 {
   735  		print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
   736  		throw("scanstack - bad status")
   737  	}
   738  
   739  	switch readgstatus(gp) &^ _Gscan {
   740  	default:
   741  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   742  		throw("mark - bad status")
   743  	case _Gdead:
   744  		return 0
   745  	case _Grunning:
   746  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   747  		throw("scanstack: goroutine not stopped")
   748  	case _Grunnable, _Gsyscall, _Gwaiting:
   749  		// ok
   750  	}
   751  
   752  	if gp == getg() {
   753  		throw("can't scan our own stack")
   754  	}
   755  
   756  	// scannedSize is the amount of work we'll be reporting.
   757  	//
   758  	// It is less than the allocated size (which is hi-lo).
   759  	var sp uintptr
   760  	if gp.syscallsp != 0 {
   761  		sp = gp.syscallsp // If in a system call this is the stack pointer (gp.sched.sp can be 0 in this case on Windows).
   762  	} else {
   763  		sp = gp.sched.sp
   764  	}
   765  	scannedSize := gp.stack.hi - sp
   766  
   767  	// Keep statistics for initial stack size calculation.
   768  	// Note that this accumulates the scanned size, not the allocated size.
   769  	p := getg().m.p.ptr()
   770  	p.scannedStackSize += uint64(scannedSize)
   771  	p.scannedStacks++
   772  
   773  	if isShrinkStackSafe(gp) {
   774  		// Shrink the stack if not much of it is being used.
   775  		shrinkstack(gp)
   776  	} else {
   777  		// Otherwise, shrink the stack at the next sync safe point.
   778  		gp.preemptShrink = true
   779  	}
   780  
   781  	var state stackScanState
   782  	state.stack = gp.stack
   783  
   784  	if stackTraceDebug {
   785  		println("stack trace goroutine", gp.goid)
   786  	}
   787  
   788  	if debugScanConservative && gp.asyncSafePoint {
   789  		print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
   790  	}
   791  
   792  	// Scan the saved context register. This is effectively a live
   793  	// register that gets moved back and forth between the
   794  	// register and sched.ctxt without a write barrier.
   795  	if gp.sched.ctxt != nil {
   796  		scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   797  	}
   798  
   799  	// Scan the stack. Accumulate a list of stack objects.
   800  	scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
   801  		scanframeworker(frame, &state, gcw)
   802  		return true
   803  	}
   804  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
   805  
   806  	// Find additional pointers that point into the stack from the heap.
   807  	// Currently this includes defers and panics. See also function copystack.
   808  
   809  	// Find and trace other pointers in defer records.
   810  	for d := gp._defer; d != nil; d = d.link {
   811  		if d.fn != nil {
   812  			// Scan the func value, which could be a stack allocated closure.
   813  			// See issue 30453.
   814  			scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   815  		}
   816  		if d.link != nil {
   817  			// The link field of a stack-allocated defer record might point
   818  			// to a heap-allocated defer record. Keep that heap record live.
   819  			scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   820  		}
   821  		// Retain defers records themselves.
   822  		// Defer records might not be reachable from the G through regular heap
   823  		// tracing because the defer linked list might weave between the stack and the heap.
   824  		if d.heap {
   825  			scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
   826  		}
   827  	}
   828  	if gp._panic != nil {
   829  		// Panics are always stack allocated.
   830  		state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
   831  	}
   832  
   833  	// Find and scan all reachable stack objects.
   834  	//
   835  	// The state's pointer queue prioritizes precise pointers over
   836  	// conservative pointers so that we'll prefer scanning stack
   837  	// objects precisely.
   838  	state.buildIndex()
   839  	for {
   840  		p, conservative := state.getPtr()
   841  		if p == 0 {
   842  			break
   843  		}
   844  		obj := state.findObject(p)
   845  		if obj == nil {
   846  			continue
   847  		}
   848  		r := obj.r
   849  		if r == nil {
   850  			// We've already scanned this object.
   851  			continue
   852  		}
   853  		obj.setRecord(nil) // Don't scan it again.
   854  		if stackTraceDebug {
   855  			printlock()
   856  			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
   857  			if conservative {
   858  				print(" (conservative)")
   859  			}
   860  			println()
   861  			printunlock()
   862  		}
   863  		gcdata := r.gcdata()
   864  		var s *mspan
   865  		if r.useGCProg() {
   866  			// This path is pretty unlikely, an object large enough
   867  			// to have a GC program allocated on the stack.
   868  			// We need some space to unpack the program into a straight
   869  			// bitmask, which we allocate/free here.
   870  			// TODO: it would be nice if there were a way to run a GC
   871  			// program without having to store all its bits. We'd have
   872  			// to change from a Lempel-Ziv style program to something else.
   873  			// Or we can forbid putting objects on stacks if they require
   874  			// a gc program (see issue 27447).
   875  			s = materializeGCProg(r.ptrdata(), gcdata)
   876  			gcdata = (*byte)(unsafe.Pointer(s.startAddr))
   877  		}
   878  
   879  		b := state.stack.lo + uintptr(obj.off)
   880  		if conservative {
   881  			scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
   882  		} else {
   883  			scanblock(b, r.ptrdata(), gcdata, gcw, &state)
   884  		}
   885  
   886  		if s != nil {
   887  			dematerializeGCProg(s)
   888  		}
   889  	}
   890  
   891  	// Deallocate object buffers.
   892  	// (Pointer buffers were all deallocated in the loop above.)
   893  	for state.head != nil {
   894  		x := state.head
   895  		state.head = x.next
   896  		if stackTraceDebug {
   897  			for i := 0; i < x.nobj; i++ {
   898  				obj := &x.obj[i]
   899  				if obj.r == nil { // reachable
   900  					continue
   901  				}
   902  				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
   903  				// Note: not necessarily really dead - only reachable-from-ptr dead.
   904  			}
   905  		}
   906  		x.nobj = 0
   907  		putempty((*workbuf)(unsafe.Pointer(x)))
   908  	}
   909  	if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
   910  		throw("remaining pointer buffers")
   911  	}
   912  	return int64(scannedSize)
   913  }
   914  
   915  // Scan a stack frame: local variables and function arguments/results.
   916  //
   917  //go:nowritebarrier
   918  func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
   919  	if _DebugGC > 1 && frame.continpc != 0 {
   920  		print("scanframe ", funcname(frame.fn), "\n")
   921  	}
   922  
   923  	isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
   924  	isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
   925  	if state.conservative || isAsyncPreempt || isDebugCall {
   926  		if debugScanConservative {
   927  			println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
   928  		}
   929  
   930  		// Conservatively scan the frame. Unlike the precise
   931  		// case, this includes the outgoing argument space
   932  		// since we may have stopped while this function was
   933  		// setting up a call.
   934  		//
   935  		// TODO: We could narrow this down if the compiler
   936  		// produced a single map per function of stack slots
   937  		// and registers that ever contain a pointer.
   938  		if frame.varp != 0 {
   939  			size := frame.varp - frame.sp
   940  			if size > 0 {
   941  				scanConservative(frame.sp, size, nil, gcw, state)
   942  			}
   943  		}
   944  
   945  		// Scan arguments to this frame.
   946  		if n := frame.argBytes(); n != 0 {
   947  			// TODO: We could pass the entry argument map
   948  			// to narrow this down further.
   949  			scanConservative(frame.argp, n, nil, gcw, state)
   950  		}
   951  
   952  		if isAsyncPreempt || isDebugCall {
   953  			// This function's frame contained the
   954  			// registers for the asynchronously stopped
   955  			// parent frame. Scan the parent
   956  			// conservatively.
   957  			state.conservative = true
   958  		} else {
   959  			// We only wanted to scan those two frames
   960  			// conservatively. Clear the flag for future
   961  			// frames.
   962  			state.conservative = false
   963  		}
   964  		return
   965  	}
   966  
   967  	locals, args, objs := frame.getStackMap(&state.cache, false)
   968  
   969  	// Scan local variables if stack frame has been allocated.
   970  	if locals.n > 0 {
   971  		size := uintptr(locals.n) * goarch.PtrSize
   972  		scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
   973  	}
   974  
   975  	// Scan arguments.
   976  	if args.n > 0 {
   977  		scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
   978  	}
   979  
   980  	// Add all stack objects to the stack object list.
   981  	if frame.varp != 0 {
   982  		// varp is 0 for defers, where there are no locals.
   983  		// In that case, there can't be a pointer to its args, either.
   984  		// (And all args would be scanned above anyway.)
   985  		for i := range objs {
   986  			obj := &objs[i]
   987  			off := obj.off
   988  			base := frame.varp // locals base pointer
   989  			if off >= 0 {
   990  				base = frame.argp // arguments and return values base pointer
   991  			}
   992  			ptr := base + uintptr(off)
   993  			if ptr < frame.sp {
   994  				// object hasn't been allocated in the frame yet.
   995  				continue
   996  			}
   997  			if stackTraceDebug {
   998  				println("stkobj at", hex(ptr), "of size", obj.size)
   999  			}
  1000  			state.addObject(ptr, obj)
  1001  		}
  1002  	}
  1003  }
  1004  
  1005  type gcDrainFlags int
  1006  
  1007  const (
  1008  	gcDrainUntilPreempt gcDrainFlags = 1 << iota
  1009  	gcDrainFlushBgCredit
  1010  	gcDrainIdle
  1011  	gcDrainFractional
  1012  )
  1013  
  1014  // gcDrain scans roots and objects in work buffers, blackening grey
  1015  // objects until it is unable to get more work. It may return before
  1016  // GC is done; it's the caller's responsibility to balance work from
  1017  // other Ps.
  1018  //
  1019  // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
  1020  // is set.
  1021  //
  1022  // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
  1023  // to do.
  1024  //
  1025  // If flags&gcDrainFractional != 0, gcDrain self-preempts when
  1026  // pollFractionalWorkerExit() returns true. This implies
  1027  // gcDrainNoBlock.
  1028  //
  1029  // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
  1030  // credit to gcController.bgScanCredit every gcCreditSlack units of
  1031  // scan work.
  1032  //
  1033  // gcDrain will always return if there is a pending STW.
  1034  //
  1035  //go:nowritebarrier
  1036  func gcDrain(gcw *gcWork, flags gcDrainFlags) {
  1037  	if !writeBarrier.needed {
  1038  		throw("gcDrain phase incorrect")
  1039  	}
  1040  
  1041  	gp := getg().m.curg
  1042  	preemptible := flags&gcDrainUntilPreempt != 0
  1043  	flushBgCredit := flags&gcDrainFlushBgCredit != 0
  1044  	idle := flags&gcDrainIdle != 0
  1045  
  1046  	initScanWork := gcw.heapScanWork
  1047  
  1048  	// checkWork is the scan work before performing the next
  1049  	// self-preempt check.
  1050  	checkWork := int64(1<<63 - 1)
  1051  	var check func() bool
  1052  	if flags&(gcDrainIdle|gcDrainFractional) != 0 {
  1053  		checkWork = initScanWork + drainCheckThreshold
  1054  		if idle {
  1055  			check = pollWork
  1056  		} else if flags&gcDrainFractional != 0 {
  1057  			check = pollFractionalWorkerExit
  1058  		}
  1059  	}
  1060  
  1061  	// Drain root marking jobs.
  1062  	if work.markrootNext < work.markrootJobs {
  1063  		// Stop if we're preemptible or if someone wants to STW.
  1064  		for !(gp.preempt && (preemptible || sched.gcwaiting.Load())) {
  1065  			job := atomic.Xadd(&work.markrootNext, +1) - 1
  1066  			if job >= work.markrootJobs {
  1067  				break
  1068  			}
  1069  			markroot(gcw, job, flushBgCredit)
  1070  			if check != nil && check() {
  1071  				goto done
  1072  			}
  1073  		}
  1074  	}
  1075  
  1076  	// Drain heap marking jobs.
  1077  	// Stop if we're preemptible or if someone wants to STW.
  1078  	for !(gp.preempt && (preemptible || sched.gcwaiting.Load())) {
  1079  		// Try to keep work available on the global queue. We used to
  1080  		// check if there were waiting workers, but it's better to
  1081  		// just keep work available than to make workers wait. In the
  1082  		// worst case, we'll do O(log(_WorkbufSize)) unnecessary
  1083  		// balances.
  1084  		if work.full == 0 {
  1085  			gcw.balance()
  1086  		}
  1087  
  1088  		b := gcw.tryGetFast()
  1089  		if b == 0 {
  1090  			b = gcw.tryGet()
  1091  			if b == 0 {
  1092  				// Flush the write barrier
  1093  				// buffer; this may create
  1094  				// more work.
  1095  				wbBufFlush(nil, 0)
  1096  				b = gcw.tryGet()
  1097  			}
  1098  		}
  1099  		if b == 0 {
  1100  			// Unable to get work.
  1101  			break
  1102  		}
  1103  		scanobject(b, gcw)
  1104  
  1105  		// Flush background scan work credit to the global
  1106  		// account if we've accumulated enough locally so
  1107  		// mutator assists can draw on it.
  1108  		if gcw.heapScanWork >= gcCreditSlack {
  1109  			gcController.heapScanWork.Add(gcw.heapScanWork)
  1110  			if flushBgCredit {
  1111  				gcFlushBgCredit(gcw.heapScanWork - initScanWork)
  1112  				initScanWork = 0
  1113  			}
  1114  			checkWork -= gcw.heapScanWork
  1115  			gcw.heapScanWork = 0
  1116  
  1117  			if checkWork <= 0 {
  1118  				checkWork += drainCheckThreshold
  1119  				if check != nil && check() {
  1120  					break
  1121  				}
  1122  			}
  1123  		}
  1124  	}
  1125  
  1126  done:
  1127  	// Flush remaining scan work credit.
  1128  	if gcw.heapScanWork > 0 {
  1129  		gcController.heapScanWork.Add(gcw.heapScanWork)
  1130  		if flushBgCredit {
  1131  			gcFlushBgCredit(gcw.heapScanWork - initScanWork)
  1132  		}
  1133  		gcw.heapScanWork = 0
  1134  	}
  1135  }
  1136  
  1137  // gcDrainN blackens grey objects until it has performed roughly
  1138  // scanWork units of scan work or the G is preempted. This is
  1139  // best-effort, so it may perform less work if it fails to get a work
  1140  // buffer. Otherwise, it will perform at least n units of work, but
  1141  // may perform more because scanning is always done in whole object
  1142  // increments. It returns the amount of scan work performed.
  1143  //
  1144  // The caller goroutine must be in a preemptible state (e.g.,
  1145  // _Gwaiting) to prevent deadlocks during stack scanning. As a
  1146  // consequence, this must be called on the system stack.
  1147  //
  1148  //go:nowritebarrier
  1149  //go:systemstack
  1150  func gcDrainN(gcw *gcWork, scanWork int64) int64 {
  1151  	if !writeBarrier.needed {
  1152  		throw("gcDrainN phase incorrect")
  1153  	}
  1154  
  1155  	// There may already be scan work on the gcw, which we don't
  1156  	// want to claim was done by this call.
  1157  	workFlushed := -gcw.heapScanWork
  1158  
  1159  	// In addition to backing out because of a preemption, back out
  1160  	// if the GC CPU limiter is enabled.
  1161  	gp := getg().m.curg
  1162  	for !gp.preempt && !gcCPULimiter.limiting() && workFlushed+gcw.heapScanWork < scanWork {
  1163  		// See gcDrain comment.
  1164  		if work.full == 0 {
  1165  			gcw.balance()
  1166  		}
  1167  
  1168  		b := gcw.tryGetFast()
  1169  		if b == 0 {
  1170  			b = gcw.tryGet()
  1171  			if b == 0 {
  1172  				// Flush the write barrier buffer;
  1173  				// this may create more work.
  1174  				wbBufFlush(nil, 0)
  1175  				b = gcw.tryGet()
  1176  			}
  1177  		}
  1178  
  1179  		if b == 0 {
  1180  			// Try to do a root job.
  1181  			if work.markrootNext < work.markrootJobs {
  1182  				job := atomic.Xadd(&work.markrootNext, +1) - 1
  1183  				if job < work.markrootJobs {
  1184  					workFlushed += markroot(gcw, job, false)
  1185  					continue
  1186  				}
  1187  			}
  1188  			// No heap or root jobs.
  1189  			break
  1190  		}
  1191  
  1192  		scanobject(b, gcw)
  1193  
  1194  		// Flush background scan work credit.
  1195  		if gcw.heapScanWork >= gcCreditSlack {
  1196  			gcController.heapScanWork.Add(gcw.heapScanWork)
  1197  			workFlushed += gcw.heapScanWork
  1198  			gcw.heapScanWork = 0
  1199  		}
  1200  	}
  1201  
  1202  	// Unlike gcDrain, there's no need to flush remaining work
  1203  	// here because this never flushes to bgScanCredit and
  1204  	// gcw.dispose will flush any remaining work to scanWork.
  1205  
  1206  	return workFlushed + gcw.heapScanWork
  1207  }
  1208  
  1209  // scanblock scans b as scanobject would, but using an explicit
  1210  // pointer bitmap instead of the heap bitmap.
  1211  //
  1212  // This is used to scan non-heap roots, so it does not update
  1213  // gcw.bytesMarked or gcw.heapScanWork.
  1214  //
  1215  // If stk != nil, possible stack pointers are also reported to stk.putPtr.
  1216  //
  1217  //go:nowritebarrier
  1218  func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
  1219  	// Use local copies of original parameters, so that a stack trace
  1220  	// due to one of the throws below shows the original block
  1221  	// base and extent.
  1222  	b := b0
  1223  	n := n0
  1224  
  1225  	for i := uintptr(0); i < n; {
  1226  		// Find bits for the next word.
  1227  		bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
  1228  		if bits == 0 {
  1229  			i += goarch.PtrSize * 8
  1230  			continue
  1231  		}
  1232  		for j := 0; j < 8 && i < n; j++ {
  1233  			if bits&1 != 0 {
  1234  				// Same work as in scanobject; see comments there.
  1235  				p := *(*uintptr)(unsafe.Pointer(b + i))
  1236  				if p != 0 {
  1237  					if obj, span, objIndex := findObject(p, b, i); obj != 0 {
  1238  						greyobject(obj, b, i, span, gcw, objIndex)
  1239  					} else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
  1240  						stk.putPtr(p, false)
  1241  					}
  1242  				}
  1243  			}
  1244  			bits >>= 1
  1245  			i += goarch.PtrSize
  1246  		}
  1247  	}
  1248  }
  1249  
  1250  // scanobject scans the object starting at b, adding pointers to gcw.
  1251  // b must point to the beginning of a heap object or an oblet.
  1252  // scanobject consults the GC bitmap for the pointer mask and the
  1253  // spans for the size of the object.
  1254  //
  1255  //go:nowritebarrier
  1256  func scanobject(b uintptr, gcw *gcWork) {
  1257  	// Prefetch object before we scan it.
  1258  	//
  1259  	// This will overlap fetching the beginning of the object with initial
  1260  	// setup before we start scanning the object.
  1261  	sys.Prefetch(b)
  1262  
  1263  	// Find the bits for b and the size of the object at b.
  1264  	//
  1265  	// b is either the beginning of an object, in which case this
  1266  	// is the size of the object to scan, or it points to an
  1267  	// oblet, in which case we compute the size to scan below.
  1268  	s := spanOfUnchecked(b)
  1269  	n := s.elemsize
  1270  	if n == 0 {
  1271  		throw("scanobject n == 0")
  1272  	}
  1273  	if s.spanclass.noscan() {
  1274  		// Correctness-wise this is ok, but it's inefficient
  1275  		// if noscan objects reach here.
  1276  		throw("scanobject of a noscan object")
  1277  	}
  1278  
  1279  	if n > maxObletBytes {
  1280  		// Large object. Break into oblets for better
  1281  		// parallelism and lower latency.
  1282  		if b == s.base() {
  1283  			// Enqueue the other oblets to scan later.
  1284  			// Some oblets may be in b's scalar tail, but
  1285  			// these will be marked as "no more pointers",
  1286  			// so we'll drop out immediately when we go to
  1287  			// scan those.
  1288  			for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
  1289  				if !gcw.putFast(oblet) {
  1290  					gcw.put(oblet)
  1291  				}
  1292  			}
  1293  		}
  1294  
  1295  		// Compute the size of the oblet. Since this object
  1296  		// must be a large object, s.base() is the beginning
  1297  		// of the object.
  1298  		n = s.base() + s.elemsize - b
  1299  		if n > maxObletBytes {
  1300  			n = maxObletBytes
  1301  		}
  1302  	}
  1303  
  1304  	hbits := heapBitsForAddr(b, n)
  1305  	var scanSize uintptr
  1306  	for {
  1307  		var addr uintptr
  1308  		if hbits, addr = hbits.nextFast(); addr == 0 {
  1309  			if hbits, addr = hbits.next(); addr == 0 {
  1310  				break
  1311  			}
  1312  		}
  1313  
  1314  		// Keep track of farthest pointer we found, so we can
  1315  		// update heapScanWork. TODO: is there a better metric,
  1316  		// now that we can skip scalar portions pretty efficiently?
  1317  		scanSize = addr - b + goarch.PtrSize
  1318  
  1319  		// Work here is duplicated in scanblock and above.
  1320  		// If you make changes here, make changes there too.
  1321  		obj := *(*uintptr)(unsafe.Pointer(addr))
  1322  
  1323  		// At this point we have extracted the next potential pointer.
  1324  		// Quickly filter out nil and pointers back to the current object.
  1325  		if obj != 0 && obj-b >= n {
  1326  			// Test if obj points into the Go heap and, if so,
  1327  			// mark the object.
  1328  			//
  1329  			// Note that it's possible for findObject to
  1330  			// fail if obj points to a just-allocated heap
  1331  			// object because of a race with growing the
  1332  			// heap. In this case, we know the object was
  1333  			// just allocated and hence will be marked by
  1334  			// allocation itself.
  1335  			if obj, span, objIndex := findObject(obj, b, addr-b); obj != 0 {
  1336  				greyobject(obj, b, addr-b, span, gcw, objIndex)
  1337  			}
  1338  		}
  1339  	}
  1340  	gcw.bytesMarked += uint64(n)
  1341  	gcw.heapScanWork += int64(scanSize)
  1342  }
  1343  
  1344  // scanConservative scans block [b, b+n) conservatively, treating any
  1345  // pointer-like value in the block as a pointer.
  1346  //
  1347  // If ptrmask != nil, only words that are marked in ptrmask are
  1348  // considered as potential pointers.
  1349  //
  1350  // If state != nil, it's assumed that [b, b+n) is a block in the stack
  1351  // and may contain pointers to stack objects.
  1352  func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
  1353  	if debugScanConservative {
  1354  		printlock()
  1355  		print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
  1356  		hexdumpWords(b, b+n, func(p uintptr) byte {
  1357  			if ptrmask != nil {
  1358  				word := (p - b) / goarch.PtrSize
  1359  				bits := *addb(ptrmask, word/8)
  1360  				if (bits>>(word%8))&1 == 0 {
  1361  					return '$'
  1362  				}
  1363  			}
  1364  
  1365  			val := *(*uintptr)(unsafe.Pointer(p))
  1366  			if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1367  				return '@'
  1368  			}
  1369  
  1370  			span := spanOfHeap(val)
  1371  			if span == nil {
  1372  				return ' '
  1373  			}
  1374  			idx := span.objIndex(val)
  1375  			if span.isFree(idx) {
  1376  				return ' '
  1377  			}
  1378  			return '*'
  1379  		})
  1380  		printunlock()
  1381  	}
  1382  
  1383  	for i := uintptr(0); i < n; i += goarch.PtrSize {
  1384  		if ptrmask != nil {
  1385  			word := i / goarch.PtrSize
  1386  			bits := *addb(ptrmask, word/8)
  1387  			if bits == 0 {
  1388  				// Skip 8 words (the loop increment will do the 8th)
  1389  				//
  1390  				// This must be the first time we've
  1391  				// seen this word of ptrmask, so i
  1392  				// must be 8-word-aligned, but check
  1393  				// our reasoning just in case.
  1394  				if i%(goarch.PtrSize*8) != 0 {
  1395  					throw("misaligned mask")
  1396  				}
  1397  				i += goarch.PtrSize*8 - goarch.PtrSize
  1398  				continue
  1399  			}
  1400  			if (bits>>(word%8))&1 == 0 {
  1401  				continue
  1402  			}
  1403  		}
  1404  
  1405  		val := *(*uintptr)(unsafe.Pointer(b + i))
  1406  
  1407  		// Check if val points into the stack.
  1408  		if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1409  			// val may point to a stack object. This
  1410  			// object may be dead from last cycle and
  1411  			// hence may contain pointers to unallocated
  1412  			// objects, but unlike heap objects we can't
  1413  			// tell if it's already dead. Hence, if all
  1414  			// pointers to this object are from
  1415  			// conservative scanning, we have to scan it
  1416  			// defensively, too.
  1417  			state.putPtr(val, true)
  1418  			continue
  1419  		}
  1420  
  1421  		// Check if val points to a heap span.
  1422  		span := spanOfHeap(val)
  1423  		if span == nil {
  1424  			continue
  1425  		}
  1426  
  1427  		// Check if val points to an allocated object.
  1428  		idx := span.objIndex(val)
  1429  		if span.isFree(idx) {
  1430  			continue
  1431  		}
  1432  
  1433  		// val points to an allocated object. Mark it.
  1434  		obj := span.base() + idx*span.elemsize
  1435  		greyobject(obj, b, i, span, gcw, idx)
  1436  	}
  1437  }
  1438  
  1439  // Shade the object if it isn't already.
  1440  // The object is not nil and known to be in the heap.
  1441  // Preemption must be disabled.
  1442  //
  1443  //go:nowritebarrier
  1444  func shade(b uintptr) {
  1445  	if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
  1446  		gcw := &getg().m.p.ptr().gcw
  1447  		greyobject(obj, 0, 0, span, gcw, objIndex)
  1448  	}
  1449  }
  1450  
  1451  // obj is the start of an object with mark mbits.
  1452  // If it isn't already marked, mark it and enqueue into gcw.
  1453  // base and off are for debugging only and could be removed.
  1454  //
  1455  // See also wbBufFlush1, which partially duplicates this logic.
  1456  //
  1457  //go:nowritebarrierrec
  1458  func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
  1459  	// obj should be start of allocation, and so must be at least pointer-aligned.
  1460  	if obj&(goarch.PtrSize-1) != 0 {
  1461  		throw("greyobject: obj not pointer-aligned")
  1462  	}
  1463  	mbits := span.markBitsForIndex(objIndex)
  1464  
  1465  	if useCheckmark {
  1466  		if setCheckmark(obj, base, off, mbits) {
  1467  			// Already marked.
  1468  			return
  1469  		}
  1470  	} else {
  1471  		if debug.gccheckmark > 0 && span.isFree(objIndex) {
  1472  			print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
  1473  			gcDumpObject("base", base, off)
  1474  			gcDumpObject("obj", obj, ^uintptr(0))
  1475  			getg().m.traceback = 2
  1476  			throw("marking free object")
  1477  		}
  1478  
  1479  		// If marked we have nothing to do.
  1480  		if mbits.isMarked() {
  1481  			return
  1482  		}
  1483  		mbits.setMarked()
  1484  
  1485  		// Mark span.
  1486  		arena, pageIdx, pageMask := pageIndexOf(span.base())
  1487  		if arena.pageMarks[pageIdx]&pageMask == 0 {
  1488  			atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1489  		}
  1490  
  1491  		// If this is a noscan object, fast-track it to black
  1492  		// instead of greying it.
  1493  		if span.spanclass.noscan() {
  1494  			gcw.bytesMarked += uint64(span.elemsize)
  1495  			return
  1496  		}
  1497  	}
  1498  
  1499  	// We're adding obj to P's local workbuf, so it's likely
  1500  	// this object will be processed soon by the same P.
  1501  	// Even if the workbuf gets flushed, there will likely still be
  1502  	// some benefit on platforms with inclusive shared caches.
  1503  	sys.Prefetch(obj)
  1504  	// Queue the obj for scanning.
  1505  	if !gcw.putFast(obj) {
  1506  		gcw.put(obj)
  1507  	}
  1508  }
  1509  
  1510  // gcDumpObject dumps the contents of obj for debugging and marks the
  1511  // field at byte offset off in obj.
  1512  func gcDumpObject(label string, obj, off uintptr) {
  1513  	s := spanOf(obj)
  1514  	print(label, "=", hex(obj))
  1515  	if s == nil {
  1516  		print(" s=nil\n")
  1517  		return
  1518  	}
  1519  	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
  1520  	if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
  1521  		print(mSpanStateNames[state], "\n")
  1522  	} else {
  1523  		print("unknown(", state, ")\n")
  1524  	}
  1525  
  1526  	skipped := false
  1527  	size := s.elemsize
  1528  	if s.state.get() == mSpanManual && size == 0 {
  1529  		// We're printing something from a stack frame. We
  1530  		// don't know how big it is, so just show up to an
  1531  		// including off.
  1532  		size = off + goarch.PtrSize
  1533  	}
  1534  	for i := uintptr(0); i < size; i += goarch.PtrSize {
  1535  		// For big objects, just print the beginning (because
  1536  		// that usually hints at the object's type) and the
  1537  		// fields around off.
  1538  		if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
  1539  			skipped = true
  1540  			continue
  1541  		}
  1542  		if skipped {
  1543  			print(" ...\n")
  1544  			skipped = false
  1545  		}
  1546  		print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
  1547  		if i == off {
  1548  			print(" <==")
  1549  		}
  1550  		print("\n")
  1551  	}
  1552  	if skipped {
  1553  		print(" ...\n")
  1554  	}
  1555  }
  1556  
  1557  // gcmarknewobject marks a newly allocated object black. obj must
  1558  // not contain any non-nil pointers.
  1559  //
  1560  // This is nosplit so it can manipulate a gcWork without preemption.
  1561  //
  1562  //go:nowritebarrier
  1563  //go:nosplit
  1564  func gcmarknewobject(span *mspan, obj, size uintptr) {
  1565  	if useCheckmark { // The world should be stopped so this should not happen.
  1566  		throw("gcmarknewobject called while doing checkmark")
  1567  	}
  1568  
  1569  	// Mark object.
  1570  	objIndex := span.objIndex(obj)
  1571  	span.markBitsForIndex(objIndex).setMarked()
  1572  
  1573  	// Mark span.
  1574  	arena, pageIdx, pageMask := pageIndexOf(span.base())
  1575  	if arena.pageMarks[pageIdx]&pageMask == 0 {
  1576  		atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1577  	}
  1578  
  1579  	gcw := &getg().m.p.ptr().gcw
  1580  	gcw.bytesMarked += uint64(size)
  1581  }
  1582  
  1583  // gcMarkTinyAllocs greys all active tiny alloc blocks.
  1584  //
  1585  // The world must be stopped.
  1586  func gcMarkTinyAllocs() {
  1587  	assertWorldStopped()
  1588  
  1589  	for _, p := range allp {
  1590  		c := p.mcache
  1591  		if c == nil || c.tiny == 0 {
  1592  			continue
  1593  		}
  1594  		_, span, objIndex := findObject(c.tiny, 0, 0)
  1595  		gcw := &p.gcw
  1596  		greyobject(c.tiny, 0, 0, span, gcw, objIndex)
  1597  	}
  1598  }
  1599  

View as plain text