Black Lives Matter. Support the Equal Justice Initiative.

Source file src/runtime/mgcmark.go

Documentation: runtime

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Garbage collector: marking and scanning
     6  
     7  package runtime
     8  
     9  import (
    10  	"runtime/internal/atomic"
    11  	"runtime/internal/sys"
    12  	"unsafe"
    13  )
    14  
    15  const (
    16  	fixedRootFinalizers = iota
    17  	fixedRootFreeGStacks
    18  	fixedRootCount
    19  
    20  	// rootBlockBytes is the number of bytes to scan per data or
    21  	// BSS root.
    22  	rootBlockBytes = 256 << 10
    23  
    24  	// maxObletBytes is the maximum bytes of an object to scan at
    25  	// once. Larger objects will be split up into "oblets" of at
    26  	// most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
    27  	// scan preemption at ~100 µs.
    28  	//
    29  	// This must be > _MaxSmallSize so that the object base is the
    30  	// span base.
    31  	maxObletBytes = 128 << 10
    32  
    33  	// drainCheckThreshold specifies how many units of work to do
    34  	// between self-preemption checks in gcDrain. Assuming a scan
    35  	// rate of 1 MB/ms, this is ~100 µs. Lower values have higher
    36  	// overhead in the scan loop (the scheduler check may perform
    37  	// a syscall, so its overhead is nontrivial). Higher values
    38  	// make the system less responsive to incoming work.
    39  	drainCheckThreshold = 100000
    40  
    41  	// pagesPerSpanRoot indicates how many pages to scan from a span root
    42  	// at a time. Used by special root marking.
    43  	//
    44  	// Higher values improve throughput by increasing locality, but
    45  	// increase the minimum latency of a marking operation.
    46  	//
    47  	// Must be a multiple of the pageInUse bitmap element size and
    48  	// must also evenly divide pagesPerArena.
    49  	pagesPerSpanRoot = 512
    50  )
    51  
    52  // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
    53  // some miscellany) and initializes scanning-related state.
    54  //
    55  // The world must be stopped.
    56  func gcMarkRootPrepare() {
    57  	assertWorldStopped()
    58  
    59  	// Compute how many data and BSS root blocks there are.
    60  	nBlocks := func(bytes uintptr) int {
    61  		return int(divRoundUp(bytes, rootBlockBytes))
    62  	}
    63  
    64  	work.nDataRoots = 0
    65  	work.nBSSRoots = 0
    66  
    67  	// Scan globals.
    68  	for _, datap := range activeModules() {
    69  		nDataRoots := nBlocks(datap.edata - datap.data)
    70  		if nDataRoots > work.nDataRoots {
    71  			work.nDataRoots = nDataRoots
    72  		}
    73  	}
    74  
    75  	for _, datap := range activeModules() {
    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.nStackRoots = int(atomic.Loaduintptr(&allglen))
   104  
   105  	work.markrootNext = 0
   106  	work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
   107  
   108  	// Calculate base indexes of each root type
   109  	work.baseData = uint32(fixedRootCount)
   110  	work.baseBSS = work.baseData + uint32(work.nDataRoots)
   111  	work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
   112  	work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
   113  	work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
   114  }
   115  
   116  // gcMarkRootCheck checks that all roots have been scanned. It is
   117  // purely for debugging.
   118  func gcMarkRootCheck() {
   119  	if work.markrootNext < work.markrootJobs {
   120  		print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
   121  		throw("left over markroot jobs")
   122  	}
   123  
   124  	// Check that stacks have been scanned.
   125  	//
   126  	// We only check the first nStackRoots Gs that we should have scanned.
   127  	// Since we don't care about newer Gs (see comment in
   128  	// gcMarkRootPrepare), no locking is required.
   129  	i := 0
   130  	forEachGRace(func(gp *g) {
   131  		if i >= work.nStackRoots {
   132  			return
   133  		}
   134  
   135  		if !gp.gcscandone {
   136  			println("gp", gp, "goid", gp.goid,
   137  				"status", readgstatus(gp),
   138  				"gcscandone", gp.gcscandone)
   139  			throw("scan missed a g")
   140  		}
   141  
   142  		i++
   143  	})
   144  }
   145  
   146  // ptrmask for an allocation containing a single pointer.
   147  var oneptrmask = [...]uint8{1}
   148  
   149  // markroot scans the i'th root.
   150  //
   151  // Preemption must be disabled (because this uses a gcWork).
   152  //
   153  // nowritebarrier is only advisory here.
   154  //
   155  //go:nowritebarrier
   156  func markroot(gcw *gcWork, i uint32) {
   157  	// Note: if you add a case here, please also update heapdump.go:dumproots.
   158  	switch {
   159  	case work.baseData <= i && i < work.baseBSS:
   160  		for _, datap := range activeModules() {
   161  			markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
   162  		}
   163  
   164  	case work.baseBSS <= i && i < work.baseSpans:
   165  		for _, datap := range activeModules() {
   166  			markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
   167  		}
   168  
   169  	case i == fixedRootFinalizers:
   170  		for fb := allfin; fb != nil; fb = fb.alllink {
   171  			cnt := uintptr(atomic.Load(&fb.cnt))
   172  			scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
   173  		}
   174  
   175  	case i == fixedRootFreeGStacks:
   176  		// Switch to the system stack so we can call
   177  		// stackfree.
   178  		systemstack(markrootFreeGStacks)
   179  
   180  	case work.baseSpans <= i && i < work.baseStacks:
   181  		// mark mspan.specials
   182  		markrootSpans(gcw, int(i-work.baseSpans))
   183  
   184  	default:
   185  		// the rest is scanning goroutine stacks
   186  		var gp *g
   187  		if work.baseStacks <= i && i < work.baseEnd {
   188  			// N.B. Atomic read of allglen in gcMarkRootPrepare
   189  			// acts as a barrier to ensure that allgs must be large
   190  			// enough to contain all relevant Gs.
   191  			gp = allgs[i-work.baseStacks]
   192  		} else {
   193  			throw("markroot: bad index")
   194  		}
   195  
   196  		// remember when we've first observed the G blocked
   197  		// needed only to output in traceback
   198  		status := readgstatus(gp) // We are not in a scan state
   199  		if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
   200  			gp.waitsince = work.tstart
   201  		}
   202  
   203  		// scanstack must be done on the system stack in case
   204  		// we're trying to scan our own stack.
   205  		systemstack(func() {
   206  			// If this is a self-scan, put the user G in
   207  			// _Gwaiting to prevent self-deadlock. It may
   208  			// already be in _Gwaiting if this is a mark
   209  			// worker or we're in mark termination.
   210  			userG := getg().m.curg
   211  			selfScan := gp == userG && readgstatus(userG) == _Grunning
   212  			if selfScan {
   213  				casgstatus(userG, _Grunning, _Gwaiting)
   214  				userG.waitreason = waitReasonGarbageCollectionScan
   215  			}
   216  
   217  			// TODO: suspendG blocks (and spins) until gp
   218  			// stops, which may take a while for
   219  			// running goroutines. Consider doing this in
   220  			// two phases where the first is non-blocking:
   221  			// we scan the stacks we can and ask running
   222  			// goroutines to scan themselves; and the
   223  			// second blocks.
   224  			stopped := suspendG(gp)
   225  			if stopped.dead {
   226  				gp.gcscandone = true
   227  				return
   228  			}
   229  			if gp.gcscandone {
   230  				throw("g already scanned")
   231  			}
   232  			scanstack(gp, gcw)
   233  			gp.gcscandone = true
   234  			resumeG(stopped)
   235  
   236  			if selfScan {
   237  				casgstatus(userG, _Gwaiting, _Grunning)
   238  			}
   239  		})
   240  	}
   241  }
   242  
   243  // markrootBlock scans the shard'th shard of the block of memory [b0,
   244  // b0+n0), with the given pointer mask.
   245  //
   246  //go:nowritebarrier
   247  func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
   248  	if rootBlockBytes%(8*sys.PtrSize) != 0 {
   249  		// This is necessary to pick byte offsets in ptrmask0.
   250  		throw("rootBlockBytes must be a multiple of 8*ptrSize")
   251  	}
   252  
   253  	// Note that if b0 is toward the end of the address space,
   254  	// then b0 + rootBlockBytes might wrap around.
   255  	// These tests are written to avoid any possible overflow.
   256  	off := uintptr(shard) * rootBlockBytes
   257  	if off >= n0 {
   258  		return
   259  	}
   260  	b := b0 + off
   261  	ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*sys.PtrSize))))
   262  	n := uintptr(rootBlockBytes)
   263  	if off+n > n0 {
   264  		n = n0 - off
   265  	}
   266  
   267  	// Scan this shard.
   268  	scanblock(b, n, ptrmask, gcw, nil)
   269  }
   270  
   271  // markrootFreeGStacks frees stacks of dead Gs.
   272  //
   273  // This does not free stacks of dead Gs cached on Ps, but having a few
   274  // cached stacks around isn't a problem.
   275  func markrootFreeGStacks() {
   276  	// Take list of dead Gs with stacks.
   277  	lock(&sched.gFree.lock)
   278  	list := sched.gFree.stack
   279  	sched.gFree.stack = gList{}
   280  	unlock(&sched.gFree.lock)
   281  	if list.empty() {
   282  		return
   283  	}
   284  
   285  	// Free stacks.
   286  	q := gQueue{list.head, list.head}
   287  	for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
   288  		stackfree(gp.stack)
   289  		gp.stack.lo = 0
   290  		gp.stack.hi = 0
   291  		// Manipulate the queue directly since the Gs are
   292  		// already all linked the right way.
   293  		q.tail.set(gp)
   294  	}
   295  
   296  	// Put Gs back on the free list.
   297  	lock(&sched.gFree.lock)
   298  	sched.gFree.noStack.pushAll(q)
   299  	unlock(&sched.gFree.lock)
   300  }
   301  
   302  // markrootSpans marks roots for one shard of markArenas.
   303  //
   304  //go:nowritebarrier
   305  func markrootSpans(gcw *gcWork, shard int) {
   306  	// Objects with finalizers have two GC-related invariants:
   307  	//
   308  	// 1) Everything reachable from the object must be marked.
   309  	// This ensures that when we pass the object to its finalizer,
   310  	// everything the finalizer can reach will be retained.
   311  	//
   312  	// 2) Finalizer specials (which are not in the garbage
   313  	// collected heap) are roots. In practice, this means the fn
   314  	// field must be scanned.
   315  	sg := mheap_.sweepgen
   316  
   317  	// Find the arena and page index into that arena for this shard.
   318  	ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
   319  	ha := mheap_.arenas[ai.l1()][ai.l2()]
   320  	arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
   321  
   322  	// Construct slice of bitmap which we'll iterate over.
   323  	specialsbits := ha.pageSpecials[arenaPage/8:]
   324  	specialsbits = specialsbits[:pagesPerSpanRoot/8]
   325  	for i := range specialsbits {
   326  		// Find set bits, which correspond to spans with specials.
   327  		specials := atomic.Load8(&specialsbits[i])
   328  		if specials == 0 {
   329  			continue
   330  		}
   331  		for j := uint(0); j < 8; j++ {
   332  			if specials&(1<<j) == 0 {
   333  				continue
   334  			}
   335  			// Find the span for this bit.
   336  			//
   337  			// This value is guaranteed to be non-nil because having
   338  			// specials implies that the span is in-use, and since we're
   339  			// currently marking we can be sure that we don't have to worry
   340  			// about the span being freed and re-used.
   341  			s := ha.spans[arenaPage+uint(i)*8+j]
   342  
   343  			// The state must be mSpanInUse if the specials bit is set, so
   344  			// sanity check that.
   345  			if state := s.state.get(); state != mSpanInUse {
   346  				print("s.state = ", state, "\n")
   347  				throw("non in-use span found with specials bit set")
   348  			}
   349  			// Check that this span was swept (it may be cached or uncached).
   350  			if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
   351  				// sweepgen was updated (+2) during non-checkmark GC pass
   352  				print("sweep ", s.sweepgen, " ", sg, "\n")
   353  				throw("gc: unswept span")
   354  			}
   355  
   356  			// Lock the specials to prevent a special from being
   357  			// removed from the list while we're traversing it.
   358  			lock(&s.speciallock)
   359  			for sp := s.specials; sp != nil; sp = sp.next {
   360  				if sp.kind != _KindSpecialFinalizer {
   361  					continue
   362  				}
   363  				// don't mark finalized object, but scan it so we
   364  				// retain everything it points to.
   365  				spf := (*specialfinalizer)(unsafe.Pointer(sp))
   366  				// A finalizer can be set for an inner byte of an object, find object beginning.
   367  				p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
   368  
   369  				// Mark everything that can be reached from
   370  				// the object (but *not* the object itself or
   371  				// we'll never collect it).
   372  				scanobject(p, gcw)
   373  
   374  				// The special itself is a root.
   375  				scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw, nil)
   376  			}
   377  			unlock(&s.speciallock)
   378  		}
   379  	}
   380  }
   381  
   382  // gcAssistAlloc performs GC work to make gp's assist debt positive.
   383  // gp must be the calling user gorountine.
   384  //
   385  // This must be called with preemption enabled.
   386  func gcAssistAlloc(gp *g) {
   387  	// Don't assist in non-preemptible contexts. These are
   388  	// generally fragile and won't allow the assist to block.
   389  	if getg() == gp.m.g0 {
   390  		return
   391  	}
   392  	if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
   393  		return
   394  	}
   395  
   396  	traced := false
   397  retry:
   398  	// Compute the amount of scan work we need to do to make the
   399  	// balance positive. When the required amount of work is low,
   400  	// we over-assist to build up credit for future allocations
   401  	// and amortize the cost of assisting.
   402  	assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
   403  	assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
   404  	debtBytes := -gp.gcAssistBytes
   405  	scanWork := int64(assistWorkPerByte * float64(debtBytes))
   406  	if scanWork < gcOverAssistWork {
   407  		scanWork = gcOverAssistWork
   408  		debtBytes = int64(assistBytesPerWork * float64(scanWork))
   409  	}
   410  
   411  	// Steal as much credit as we can from the background GC's
   412  	// scan credit. This is racy and may drop the background
   413  	// credit below 0 if two mutators steal at the same time. This
   414  	// will just cause steals to fail until credit is accumulated
   415  	// again, so in the long run it doesn't really matter, but we
   416  	// do have to handle the negative credit case.
   417  	bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
   418  	stolen := int64(0)
   419  	if bgScanCredit > 0 {
   420  		if bgScanCredit < scanWork {
   421  			stolen = bgScanCredit
   422  			gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
   423  		} else {
   424  			stolen = scanWork
   425  			gp.gcAssistBytes += debtBytes
   426  		}
   427  		atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
   428  
   429  		scanWork -= stolen
   430  
   431  		if scanWork == 0 {
   432  			// We were able to steal all of the credit we
   433  			// needed.
   434  			if traced {
   435  				traceGCMarkAssistDone()
   436  			}
   437  			return
   438  		}
   439  	}
   440  
   441  	if trace.enabled && !traced {
   442  		traced = true
   443  		traceGCMarkAssistStart()
   444  	}
   445  
   446  	// Perform assist work
   447  	systemstack(func() {
   448  		gcAssistAlloc1(gp, scanWork)
   449  		// The user stack may have moved, so this can't touch
   450  		// anything on it until it returns from systemstack.
   451  	})
   452  
   453  	completed := gp.param != nil
   454  	gp.param = nil
   455  	if completed {
   456  		gcMarkDone()
   457  	}
   458  
   459  	if gp.gcAssistBytes < 0 {
   460  		// We were unable steal enough credit or perform
   461  		// enough work to pay off the assist debt. We need to
   462  		// do one of these before letting the mutator allocate
   463  		// more to prevent over-allocation.
   464  		//
   465  		// If this is because we were preempted, reschedule
   466  		// and try some more.
   467  		if gp.preempt {
   468  			Gosched()
   469  			goto retry
   470  		}
   471  
   472  		// Add this G to an assist queue and park. When the GC
   473  		// has more background credit, it will satisfy queued
   474  		// assists before flushing to the global credit pool.
   475  		//
   476  		// Note that this does *not* get woken up when more
   477  		// work is added to the work list. The theory is that
   478  		// there wasn't enough work to do anyway, so we might
   479  		// as well let background marking take care of the
   480  		// work that is available.
   481  		if !gcParkAssist() {
   482  			goto retry
   483  		}
   484  
   485  		// At this point either background GC has satisfied
   486  		// this G's assist debt, or the GC cycle is over.
   487  	}
   488  	if traced {
   489  		traceGCMarkAssistDone()
   490  	}
   491  }
   492  
   493  // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
   494  // stack. This is a separate function to make it easier to see that
   495  // we're not capturing anything from the user stack, since the user
   496  // stack may move while we're in this function.
   497  //
   498  // gcAssistAlloc1 indicates whether this assist completed the mark
   499  // phase by setting gp.param to non-nil. This can't be communicated on
   500  // the stack since it may move.
   501  //
   502  //go:systemstack
   503  func gcAssistAlloc1(gp *g, scanWork int64) {
   504  	// Clear the flag indicating that this assist completed the
   505  	// mark phase.
   506  	gp.param = nil
   507  
   508  	if atomic.Load(&gcBlackenEnabled) == 0 {
   509  		// The gcBlackenEnabled check in malloc races with the
   510  		// store that clears it but an atomic check in every malloc
   511  		// would be a performance hit.
   512  		// Instead we recheck it here on the non-preemptable system
   513  		// stack to determine if we should perform an assist.
   514  
   515  		// GC is done, so ignore any remaining debt.
   516  		gp.gcAssistBytes = 0
   517  		return
   518  	}
   519  	// Track time spent in this assist. Since we're on the
   520  	// system stack, this is non-preemptible, so we can
   521  	// just measure start and end time.
   522  	startTime := nanotime()
   523  
   524  	decnwait := atomic.Xadd(&work.nwait, -1)
   525  	if decnwait == work.nproc {
   526  		println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
   527  		throw("nwait > work.nprocs")
   528  	}
   529  
   530  	// gcDrainN requires the caller to be preemptible.
   531  	casgstatus(gp, _Grunning, _Gwaiting)
   532  	gp.waitreason = waitReasonGCAssistMarking
   533  
   534  	// drain own cached work first in the hopes that it
   535  	// will be more cache friendly.
   536  	gcw := &getg().m.p.ptr().gcw
   537  	workDone := gcDrainN(gcw, scanWork)
   538  
   539  	casgstatus(gp, _Gwaiting, _Grunning)
   540  
   541  	// Record that we did this much scan work.
   542  	//
   543  	// Back out the number of bytes of assist credit that
   544  	// this scan work counts for. The "1+" is a poor man's
   545  	// round-up, to ensure this adds credit even if
   546  	// assistBytesPerWork is very low.
   547  	assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
   548  	gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
   549  
   550  	// If this is the last worker and we ran out of work,
   551  	// signal a completion point.
   552  	incnwait := atomic.Xadd(&work.nwait, +1)
   553  	if incnwait > work.nproc {
   554  		println("runtime: work.nwait=", incnwait,
   555  			"work.nproc=", work.nproc)
   556  		throw("work.nwait > work.nproc")
   557  	}
   558  
   559  	if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
   560  		// This has reached a background completion point. Set
   561  		// gp.param to a non-nil value to indicate this. It
   562  		// doesn't matter what we set it to (it just has to be
   563  		// a valid pointer).
   564  		gp.param = unsafe.Pointer(gp)
   565  	}
   566  	duration := nanotime() - startTime
   567  	_p_ := gp.m.p.ptr()
   568  	_p_.gcAssistTime += duration
   569  	if _p_.gcAssistTime > gcAssistTimeSlack {
   570  		atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
   571  		_p_.gcAssistTime = 0
   572  	}
   573  }
   574  
   575  // gcWakeAllAssists wakes all currently blocked assists. This is used
   576  // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
   577  // new assists from going to sleep after this point.
   578  func gcWakeAllAssists() {
   579  	lock(&work.assistQueue.lock)
   580  	list := work.assistQueue.q.popList()
   581  	injectglist(&list)
   582  	unlock(&work.assistQueue.lock)
   583  }
   584  
   585  // gcParkAssist puts the current goroutine on the assist queue and parks.
   586  //
   587  // gcParkAssist reports whether the assist is now satisfied. If it
   588  // returns false, the caller must retry the assist.
   589  //
   590  //go:nowritebarrier
   591  func gcParkAssist() bool {
   592  	lock(&work.assistQueue.lock)
   593  	// If the GC cycle finished while we were getting the lock,
   594  	// exit the assist. The cycle can't finish while we hold the
   595  	// lock.
   596  	if atomic.Load(&gcBlackenEnabled) == 0 {
   597  		unlock(&work.assistQueue.lock)
   598  		return true
   599  	}
   600  
   601  	gp := getg()
   602  	oldList := work.assistQueue.q
   603  	work.assistQueue.q.pushBack(gp)
   604  
   605  	// Recheck for background credit now that this G is in
   606  	// the queue, but can still back out. This avoids a
   607  	// race in case background marking has flushed more
   608  	// credit since we checked above.
   609  	if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
   610  		work.assistQueue.q = oldList
   611  		if oldList.tail != 0 {
   612  			oldList.tail.ptr().schedlink.set(nil)
   613  		}
   614  		unlock(&work.assistQueue.lock)
   615  		return false
   616  	}
   617  	// Park.
   618  	goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
   619  	return true
   620  }
   621  
   622  // gcFlushBgCredit flushes scanWork units of background scan work
   623  // credit. This first satisfies blocked assists on the
   624  // work.assistQueue and then flushes any remaining credit to
   625  // gcController.bgScanCredit.
   626  //
   627  // Write barriers are disallowed because this is used by gcDrain after
   628  // it has ensured that all work is drained and this must preserve that
   629  // condition.
   630  //
   631  //go:nowritebarrierrec
   632  func gcFlushBgCredit(scanWork int64) {
   633  	if work.assistQueue.q.empty() {
   634  		// Fast path; there are no blocked assists. There's a
   635  		// small window here where an assist may add itself to
   636  		// the blocked queue and park. If that happens, we'll
   637  		// just get it on the next flush.
   638  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   639  		return
   640  	}
   641  
   642  	assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
   643  	scanBytes := int64(float64(scanWork) * assistBytesPerWork)
   644  
   645  	lock(&work.assistQueue.lock)
   646  	for !work.assistQueue.q.empty() && scanBytes > 0 {
   647  		gp := work.assistQueue.q.pop()
   648  		// Note that gp.gcAssistBytes is negative because gp
   649  		// is in debt. Think carefully about the signs below.
   650  		if scanBytes+gp.gcAssistBytes >= 0 {
   651  			// Satisfy this entire assist debt.
   652  			scanBytes += gp.gcAssistBytes
   653  			gp.gcAssistBytes = 0
   654  			// It's important that we *not* put gp in
   655  			// runnext. Otherwise, it's possible for user
   656  			// code to exploit the GC worker's high
   657  			// scheduler priority to get itself always run
   658  			// before other goroutines and always in the
   659  			// fresh quantum started by GC.
   660  			ready(gp, 0, false)
   661  		} else {
   662  			// Partially satisfy this assist.
   663  			gp.gcAssistBytes += scanBytes
   664  			scanBytes = 0
   665  			// As a heuristic, we move this assist to the
   666  			// back of the queue so that large assists
   667  			// can't clog up the assist queue and
   668  			// substantially delay small assists.
   669  			work.assistQueue.q.pushBack(gp)
   670  			break
   671  		}
   672  	}
   673  
   674  	if scanBytes > 0 {
   675  		// Convert from scan bytes back to work.
   676  		assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
   677  		scanWork = int64(float64(scanBytes) * assistWorkPerByte)
   678  		atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
   679  	}
   680  	unlock(&work.assistQueue.lock)
   681  }
   682  
   683  // scanstack scans gp's stack, greying all pointers found on the stack.
   684  //
   685  // scanstack will also shrink the stack if it is safe to do so. If it
   686  // is not, it schedules a stack shrink for the next synchronous safe
   687  // point.
   688  //
   689  // scanstack is marked go:systemstack because it must not be preempted
   690  // while using a workbuf.
   691  //
   692  //go:nowritebarrier
   693  //go:systemstack
   694  func scanstack(gp *g, gcw *gcWork) {
   695  	if readgstatus(gp)&_Gscan == 0 {
   696  		print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
   697  		throw("scanstack - bad status")
   698  	}
   699  
   700  	switch readgstatus(gp) &^ _Gscan {
   701  	default:
   702  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   703  		throw("mark - bad status")
   704  	case _Gdead:
   705  		return
   706  	case _Grunning:
   707  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   708  		throw("scanstack: goroutine not stopped")
   709  	case _Grunnable, _Gsyscall, _Gwaiting:
   710  		// ok
   711  	}
   712  
   713  	if gp == getg() {
   714  		throw("can't scan our own stack")
   715  	}
   716  
   717  	if isShrinkStackSafe(gp) {
   718  		// Shrink the stack if not much of it is being used.
   719  		shrinkstack(gp)
   720  	} else {
   721  		// Otherwise, shrink the stack at the next sync safe point.
   722  		gp.preemptShrink = true
   723  	}
   724  
   725  	var state stackScanState
   726  	state.stack = gp.stack
   727  
   728  	if stackTraceDebug {
   729  		println("stack trace goroutine", gp.goid)
   730  	}
   731  
   732  	if debugScanConservative && gp.asyncSafePoint {
   733  		print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
   734  	}
   735  
   736  	// Scan the saved context register. This is effectively a live
   737  	// register that gets moved back and forth between the
   738  	// register and sched.ctxt without a write barrier.
   739  	if gp.sched.ctxt != nil {
   740  		scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   741  	}
   742  
   743  	// Scan the stack. Accumulate a list of stack objects.
   744  	scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
   745  		scanframeworker(frame, &state, gcw)
   746  		return true
   747  	}
   748  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
   749  
   750  	// Find additional pointers that point into the stack from the heap.
   751  	// Currently this includes defers and panics. See also function copystack.
   752  
   753  	// Find and trace all defer arguments.
   754  	tracebackdefers(gp, scanframe, nil)
   755  
   756  	// Find and trace other pointers in defer records.
   757  	for d := gp._defer; d != nil; d = d.link {
   758  		if d.fn != nil {
   759  			// tracebackdefers above does not scan the func value, which could
   760  			// be a stack allocated closure. See issue 30453.
   761  			scanblock(uintptr(unsafe.Pointer(&d.fn)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   762  		}
   763  		if d.link != nil {
   764  			// The link field of a stack-allocated defer record might point
   765  			// to a heap-allocated defer record. Keep that heap record live.
   766  			scanblock(uintptr(unsafe.Pointer(&d.link)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   767  		}
   768  		// Retain defers records themselves.
   769  		// Defer records might not be reachable from the G through regular heap
   770  		// tracing because the defer linked list might weave between the stack and the heap.
   771  		if d.heap {
   772  			scanblock(uintptr(unsafe.Pointer(&d)), sys.PtrSize, &oneptrmask[0], gcw, &state)
   773  		}
   774  	}
   775  	if gp._panic != nil {
   776  		// Panics are always stack allocated.
   777  		state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
   778  	}
   779  
   780  	// Find and scan all reachable stack objects.
   781  	//
   782  	// The state's pointer queue prioritizes precise pointers over
   783  	// conservative pointers so that we'll prefer scanning stack
   784  	// objects precisely.
   785  	state.buildIndex()
   786  	for {
   787  		p, conservative := state.getPtr()
   788  		if p == 0 {
   789  			break
   790  		}
   791  		obj := state.findObject(p)
   792  		if obj == nil {
   793  			continue
   794  		}
   795  		r := obj.r
   796  		if r == nil {
   797  			// We've already scanned this object.
   798  			continue
   799  		}
   800  		obj.setRecord(nil) // Don't scan it again.
   801  		if stackTraceDebug {
   802  			printlock()
   803  			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
   804  			if conservative {
   805  				print(" (conservative)")
   806  			}
   807  			println()
   808  			printunlock()
   809  		}
   810  		gcdata := r.gcdata
   811  		var s *mspan
   812  		if r.useGCProg() {
   813  			// This path is pretty unlikely, an object large enough
   814  			// to have a GC program allocated on the stack.
   815  			// We need some space to unpack the program into a straight
   816  			// bitmask, which we allocate/free here.
   817  			// TODO: it would be nice if there were a way to run a GC
   818  			// program without having to store all its bits. We'd have
   819  			// to change from a Lempel-Ziv style program to something else.
   820  			// Or we can forbid putting objects on stacks if they require
   821  			// a gc program (see issue 27447).
   822  			s = materializeGCProg(r.ptrdata(), gcdata)
   823  			gcdata = (*byte)(unsafe.Pointer(s.startAddr))
   824  		}
   825  
   826  		b := state.stack.lo + uintptr(obj.off)
   827  		if conservative {
   828  			scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
   829  		} else {
   830  			scanblock(b, r.ptrdata(), gcdata, gcw, &state)
   831  		}
   832  
   833  		if s != nil {
   834  			dematerializeGCProg(s)
   835  		}
   836  	}
   837  
   838  	// Deallocate object buffers.
   839  	// (Pointer buffers were all deallocated in the loop above.)
   840  	for state.head != nil {
   841  		x := state.head
   842  		state.head = x.next
   843  		if stackTraceDebug {
   844  			for i := 0; i < x.nobj; i++ {
   845  				obj := &x.obj[i]
   846  				if obj.r == nil { // reachable
   847  					continue
   848  				}
   849  				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
   850  				// Note: not necessarily really dead - only reachable-from-ptr dead.
   851  			}
   852  		}
   853  		x.nobj = 0
   854  		putempty((*workbuf)(unsafe.Pointer(x)))
   855  	}
   856  	if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
   857  		throw("remaining pointer buffers")
   858  	}
   859  }
   860  
   861  // Scan a stack frame: local variables and function arguments/results.
   862  //go:nowritebarrier
   863  func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
   864  	if _DebugGC > 1 && frame.continpc != 0 {
   865  		print("scanframe ", funcname(frame.fn), "\n")
   866  	}
   867  
   868  	isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
   869  	isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
   870  	if state.conservative || isAsyncPreempt || isDebugCall {
   871  		if debugScanConservative {
   872  			println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
   873  		}
   874  
   875  		// Conservatively scan the frame. Unlike the precise
   876  		// case, this includes the outgoing argument space
   877  		// since we may have stopped while this function was
   878  		// setting up a call.
   879  		//
   880  		// TODO: We could narrow this down if the compiler
   881  		// produced a single map per function of stack slots
   882  		// and registers that ever contain a pointer.
   883  		if frame.varp != 0 {
   884  			size := frame.varp - frame.sp
   885  			if size > 0 {
   886  				scanConservative(frame.sp, size, nil, gcw, state)
   887  			}
   888  		}
   889  
   890  		// Scan arguments to this frame.
   891  		if frame.arglen != 0 {
   892  			// TODO: We could pass the entry argument map
   893  			// to narrow this down further.
   894  			scanConservative(frame.argp, frame.arglen, nil, gcw, state)
   895  		}
   896  
   897  		if isAsyncPreempt || isDebugCall {
   898  			// This function's frame contained the
   899  			// registers for the asynchronously stopped
   900  			// parent frame. Scan the parent
   901  			// conservatively.
   902  			state.conservative = true
   903  		} else {
   904  			// We only wanted to scan those two frames
   905  			// conservatively. Clear the flag for future
   906  			// frames.
   907  			state.conservative = false
   908  		}
   909  		return
   910  	}
   911  
   912  	locals, args, objs := getStackMap(frame, &state.cache, false)
   913  
   914  	// Scan local variables if stack frame has been allocated.
   915  	if locals.n > 0 {
   916  		size := uintptr(locals.n) * sys.PtrSize
   917  		scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
   918  	}
   919  
   920  	// Scan arguments.
   921  	if args.n > 0 {
   922  		scanblock(frame.argp, uintptr(args.n)*sys.PtrSize, args.bytedata, gcw, state)
   923  	}
   924  
   925  	// Add all stack objects to the stack object list.
   926  	if frame.varp != 0 {
   927  		// varp is 0 for defers, where there are no locals.
   928  		// In that case, there can't be a pointer to its args, either.
   929  		// (And all args would be scanned above anyway.)
   930  		for i, obj := range objs {
   931  			off := obj.off
   932  			base := frame.varp // locals base pointer
   933  			if off >= 0 {
   934  				base = frame.argp // arguments and return values base pointer
   935  			}
   936  			ptr := base + uintptr(off)
   937  			if ptr < frame.sp {
   938  				// object hasn't been allocated in the frame yet.
   939  				continue
   940  			}
   941  			if stackTraceDebug {
   942  				println("stkobj at", hex(ptr), "of size", obj.size)
   943  			}
   944  			state.addObject(ptr, &objs[i])
   945  		}
   946  	}
   947  }
   948  
   949  type gcDrainFlags int
   950  
   951  const (
   952  	gcDrainUntilPreempt gcDrainFlags = 1 << iota
   953  	gcDrainFlushBgCredit
   954  	gcDrainIdle
   955  	gcDrainFractional
   956  )
   957  
   958  // gcDrain scans roots and objects in work buffers, blackening grey
   959  // objects until it is unable to get more work. It may return before
   960  // GC is done; it's the caller's responsibility to balance work from
   961  // other Ps.
   962  //
   963  // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
   964  // is set.
   965  //
   966  // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
   967  // to do.
   968  //
   969  // If flags&gcDrainFractional != 0, gcDrain self-preempts when
   970  // pollFractionalWorkerExit() returns true. This implies
   971  // gcDrainNoBlock.
   972  //
   973  // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
   974  // credit to gcController.bgScanCredit every gcCreditSlack units of
   975  // scan work.
   976  //
   977  // gcDrain will always return if there is a pending STW.
   978  //
   979  //go:nowritebarrier
   980  func gcDrain(gcw *gcWork, flags gcDrainFlags) {
   981  	if !writeBarrier.needed {
   982  		throw("gcDrain phase incorrect")
   983  	}
   984  
   985  	gp := getg().m.curg
   986  	preemptible := flags&gcDrainUntilPreempt != 0
   987  	flushBgCredit := flags&gcDrainFlushBgCredit != 0
   988  	idle := flags&gcDrainIdle != 0
   989  
   990  	initScanWork := gcw.scanWork
   991  
   992  	// checkWork is the scan work before performing the next
   993  	// self-preempt check.
   994  	checkWork := int64(1<<63 - 1)
   995  	var check func() bool
   996  	if flags&(gcDrainIdle|gcDrainFractional) != 0 {
   997  		checkWork = initScanWork + drainCheckThreshold
   998  		if idle {
   999  			check = pollWork
  1000  		} else if flags&gcDrainFractional != 0 {
  1001  			check = pollFractionalWorkerExit
  1002  		}
  1003  	}
  1004  
  1005  	// Drain root marking jobs.
  1006  	if work.markrootNext < work.markrootJobs {
  1007  		// Stop if we're preemptible or if someone wants to STW.
  1008  		for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
  1009  			job := atomic.Xadd(&work.markrootNext, +1) - 1
  1010  			if job >= work.markrootJobs {
  1011  				break
  1012  			}
  1013  			markroot(gcw, job)
  1014  			if check != nil && check() {
  1015  				goto done
  1016  			}
  1017  		}
  1018  	}
  1019  
  1020  	// Drain heap marking jobs.
  1021  	// Stop if we're preemptible or if someone wants to STW.
  1022  	for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
  1023  		// Try to keep work available on the global queue. We used to
  1024  		// check if there were waiting workers, but it's better to
  1025  		// just keep work available than to make workers wait. In the
  1026  		// worst case, we'll do O(log(_WorkbufSize)) unnecessary
  1027  		// balances.
  1028  		if work.full == 0 {
  1029  			gcw.balance()
  1030  		}
  1031  
  1032  		b := gcw.tryGetFast()
  1033  		if b == 0 {
  1034  			b = gcw.tryGet()
  1035  			if b == 0 {
  1036  				// Flush the write barrier
  1037  				// buffer; this may create
  1038  				// more work.
  1039  				wbBufFlush(nil, 0)
  1040  				b = gcw.tryGet()
  1041  			}
  1042  		}
  1043  		if b == 0 {
  1044  			// Unable to get work.
  1045  			break
  1046  		}
  1047  		scanobject(b, gcw)
  1048  
  1049  		// Flush background scan work credit to the global
  1050  		// account if we've accumulated enough locally so
  1051  		// mutator assists can draw on it.
  1052  		if gcw.scanWork >= gcCreditSlack {
  1053  			atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1054  			if flushBgCredit {
  1055  				gcFlushBgCredit(gcw.scanWork - initScanWork)
  1056  				initScanWork = 0
  1057  			}
  1058  			checkWork -= gcw.scanWork
  1059  			gcw.scanWork = 0
  1060  
  1061  			if checkWork <= 0 {
  1062  				checkWork += drainCheckThreshold
  1063  				if check != nil && check() {
  1064  					break
  1065  				}
  1066  			}
  1067  		}
  1068  	}
  1069  
  1070  done:
  1071  	// Flush remaining scan work credit.
  1072  	if gcw.scanWork > 0 {
  1073  		atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1074  		if flushBgCredit {
  1075  			gcFlushBgCredit(gcw.scanWork - initScanWork)
  1076  		}
  1077  		gcw.scanWork = 0
  1078  	}
  1079  }
  1080  
  1081  // gcDrainN blackens grey objects until it has performed roughly
  1082  // scanWork units of scan work or the G is preempted. This is
  1083  // best-effort, so it may perform less work if it fails to get a work
  1084  // buffer. Otherwise, it will perform at least n units of work, but
  1085  // may perform more because scanning is always done in whole object
  1086  // increments. It returns the amount of scan work performed.
  1087  //
  1088  // The caller goroutine must be in a preemptible state (e.g.,
  1089  // _Gwaiting) to prevent deadlocks during stack scanning. As a
  1090  // consequence, this must be called on the system stack.
  1091  //
  1092  //go:nowritebarrier
  1093  //go:systemstack
  1094  func gcDrainN(gcw *gcWork, scanWork int64) int64 {
  1095  	if !writeBarrier.needed {
  1096  		throw("gcDrainN phase incorrect")
  1097  	}
  1098  
  1099  	// There may already be scan work on the gcw, which we don't
  1100  	// want to claim was done by this call.
  1101  	workFlushed := -gcw.scanWork
  1102  
  1103  	gp := getg().m.curg
  1104  	for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
  1105  		// See gcDrain comment.
  1106  		if work.full == 0 {
  1107  			gcw.balance()
  1108  		}
  1109  
  1110  		// This might be a good place to add prefetch code...
  1111  		// if(wbuf.nobj > 4) {
  1112  		//         PREFETCH(wbuf->obj[wbuf.nobj - 3];
  1113  		//  }
  1114  		//
  1115  		b := gcw.tryGetFast()
  1116  		if b == 0 {
  1117  			b = gcw.tryGet()
  1118  			if b == 0 {
  1119  				// Flush the write barrier buffer;
  1120  				// this may create more work.
  1121  				wbBufFlush(nil, 0)
  1122  				b = gcw.tryGet()
  1123  			}
  1124  		}
  1125  
  1126  		if b == 0 {
  1127  			// Try to do a root job.
  1128  			//
  1129  			// TODO: Assists should get credit for this
  1130  			// work.
  1131  			if work.markrootNext < work.markrootJobs {
  1132  				job := atomic.Xadd(&work.markrootNext, +1) - 1
  1133  				if job < work.markrootJobs {
  1134  					markroot(gcw, job)
  1135  					continue
  1136  				}
  1137  			}
  1138  			// No heap or root jobs.
  1139  			break
  1140  		}
  1141  		scanobject(b, gcw)
  1142  
  1143  		// Flush background scan work credit.
  1144  		if gcw.scanWork >= gcCreditSlack {
  1145  			atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
  1146  			workFlushed += gcw.scanWork
  1147  			gcw.scanWork = 0
  1148  		}
  1149  	}
  1150  
  1151  	// Unlike gcDrain, there's no need to flush remaining work
  1152  	// here because this never flushes to bgScanCredit and
  1153  	// gcw.dispose will flush any remaining work to scanWork.
  1154  
  1155  	return workFlushed + gcw.scanWork
  1156  }
  1157  
  1158  // scanblock scans b as scanobject would, but using an explicit
  1159  // pointer bitmap instead of the heap bitmap.
  1160  //
  1161  // This is used to scan non-heap roots, so it does not update
  1162  // gcw.bytesMarked or gcw.scanWork.
  1163  //
  1164  // If stk != nil, possible stack pointers are also reported to stk.putPtr.
  1165  //go:nowritebarrier
  1166  func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
  1167  	// Use local copies of original parameters, so that a stack trace
  1168  	// due to one of the throws below shows the original block
  1169  	// base and extent.
  1170  	b := b0
  1171  	n := n0
  1172  
  1173  	for i := uintptr(0); i < n; {
  1174  		// Find bits for the next word.
  1175  		bits := uint32(*addb(ptrmask, i/(sys.PtrSize*8)))
  1176  		if bits == 0 {
  1177  			i += sys.PtrSize * 8
  1178  			continue
  1179  		}
  1180  		for j := 0; j < 8 && i < n; j++ {
  1181  			if bits&1 != 0 {
  1182  				// Same work as in scanobject; see comments there.
  1183  				p := *(*uintptr)(unsafe.Pointer(b + i))
  1184  				if p != 0 {
  1185  					if obj, span, objIndex := findObject(p, b, i); obj != 0 {
  1186  						greyobject(obj, b, i, span, gcw, objIndex)
  1187  					} else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
  1188  						stk.putPtr(p, false)
  1189  					}
  1190  				}
  1191  			}
  1192  			bits >>= 1
  1193  			i += sys.PtrSize
  1194  		}
  1195  	}
  1196  }
  1197  
  1198  // scanobject scans the object starting at b, adding pointers to gcw.
  1199  // b must point to the beginning of a heap object or an oblet.
  1200  // scanobject consults the GC bitmap for the pointer mask and the
  1201  // spans for the size of the object.
  1202  //
  1203  //go:nowritebarrier
  1204  func scanobject(b uintptr, gcw *gcWork) {
  1205  	// Find the bits for b and the size of the object at b.
  1206  	//
  1207  	// b is either the beginning of an object, in which case this
  1208  	// is the size of the object to scan, or it points to an
  1209  	// oblet, in which case we compute the size to scan below.
  1210  	hbits := heapBitsForAddr(b)
  1211  	s := spanOfUnchecked(b)
  1212  	n := s.elemsize
  1213  	if n == 0 {
  1214  		throw("scanobject n == 0")
  1215  	}
  1216  
  1217  	if n > maxObletBytes {
  1218  		// Large object. Break into oblets for better
  1219  		// parallelism and lower latency.
  1220  		if b == s.base() {
  1221  			// It's possible this is a noscan object (not
  1222  			// from greyobject, but from other code
  1223  			// paths), in which case we must *not* enqueue
  1224  			// oblets since their bitmaps will be
  1225  			// uninitialized.
  1226  			if s.spanclass.noscan() {
  1227  				// Bypass the whole scan.
  1228  				gcw.bytesMarked += uint64(n)
  1229  				return
  1230  			}
  1231  
  1232  			// Enqueue the other oblets to scan later.
  1233  			// Some oblets may be in b's scalar tail, but
  1234  			// these will be marked as "no more pointers",
  1235  			// so we'll drop out immediately when we go to
  1236  			// scan those.
  1237  			for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
  1238  				if !gcw.putFast(oblet) {
  1239  					gcw.put(oblet)
  1240  				}
  1241  			}
  1242  		}
  1243  
  1244  		// Compute the size of the oblet. Since this object
  1245  		// must be a large object, s.base() is the beginning
  1246  		// of the object.
  1247  		n = s.base() + s.elemsize - b
  1248  		if n > maxObletBytes {
  1249  			n = maxObletBytes
  1250  		}
  1251  	}
  1252  
  1253  	var i uintptr
  1254  	for i = 0; i < n; i, hbits = i+sys.PtrSize, hbits.next() {
  1255  		// Load bits once. See CL 22712 and issue 16973 for discussion.
  1256  		bits := hbits.bits()
  1257  		if bits&bitScan == 0 {
  1258  			break // no more pointers in this object
  1259  		}
  1260  		if bits&bitPointer == 0 {
  1261  			continue // not a pointer
  1262  		}
  1263  
  1264  		// Work here is duplicated in scanblock and above.
  1265  		// If you make changes here, make changes there too.
  1266  		obj := *(*uintptr)(unsafe.Pointer(b + i))
  1267  
  1268  		// At this point we have extracted the next potential pointer.
  1269  		// Quickly filter out nil and pointers back to the current object.
  1270  		if obj != 0 && obj-b >= n {
  1271  			// Test if obj points into the Go heap and, if so,
  1272  			// mark the object.
  1273  			//
  1274  			// Note that it's possible for findObject to
  1275  			// fail if obj points to a just-allocated heap
  1276  			// object because of a race with growing the
  1277  			// heap. In this case, we know the object was
  1278  			// just allocated and hence will be marked by
  1279  			// allocation itself.
  1280  			if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
  1281  				greyobject(obj, b, i, span, gcw, objIndex)
  1282  			}
  1283  		}
  1284  	}
  1285  	gcw.bytesMarked += uint64(n)
  1286  	gcw.scanWork += int64(i)
  1287  }
  1288  
  1289  // scanConservative scans block [b, b+n) conservatively, treating any
  1290  // pointer-like value in the block as a pointer.
  1291  //
  1292  // If ptrmask != nil, only words that are marked in ptrmask are
  1293  // considered as potential pointers.
  1294  //
  1295  // If state != nil, it's assumed that [b, b+n) is a block in the stack
  1296  // and may contain pointers to stack objects.
  1297  func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
  1298  	if debugScanConservative {
  1299  		printlock()
  1300  		print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
  1301  		hexdumpWords(b, b+n, func(p uintptr) byte {
  1302  			if ptrmask != nil {
  1303  				word := (p - b) / sys.PtrSize
  1304  				bits := *addb(ptrmask, word/8)
  1305  				if (bits>>(word%8))&1 == 0 {
  1306  					return '$'
  1307  				}
  1308  			}
  1309  
  1310  			val := *(*uintptr)(unsafe.Pointer(p))
  1311  			if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1312  				return '@'
  1313  			}
  1314  
  1315  			span := spanOfHeap(val)
  1316  			if span == nil {
  1317  				return ' '
  1318  			}
  1319  			idx := span.objIndex(val)
  1320  			if span.isFree(idx) {
  1321  				return ' '
  1322  			}
  1323  			return '*'
  1324  		})
  1325  		printunlock()
  1326  	}
  1327  
  1328  	for i := uintptr(0); i < n; i += sys.PtrSize {
  1329  		if ptrmask != nil {
  1330  			word := i / sys.PtrSize
  1331  			bits := *addb(ptrmask, word/8)
  1332  			if bits == 0 {
  1333  				// Skip 8 words (the loop increment will do the 8th)
  1334  				//
  1335  				// This must be the first time we've
  1336  				// seen this word of ptrmask, so i
  1337  				// must be 8-word-aligned, but check
  1338  				// our reasoning just in case.
  1339  				if i%(sys.PtrSize*8) != 0 {
  1340  					throw("misaligned mask")
  1341  				}
  1342  				i += sys.PtrSize*8 - sys.PtrSize
  1343  				continue
  1344  			}
  1345  			if (bits>>(word%8))&1 == 0 {
  1346  				continue
  1347  			}
  1348  		}
  1349  
  1350  		val := *(*uintptr)(unsafe.Pointer(b + i))
  1351  
  1352  		// Check if val points into the stack.
  1353  		if state != nil && state.stack.lo <= val && val < state.stack.hi {
  1354  			// val may point to a stack object. This
  1355  			// object may be dead from last cycle and
  1356  			// hence may contain pointers to unallocated
  1357  			// objects, but unlike heap objects we can't
  1358  			// tell if it's already dead. Hence, if all
  1359  			// pointers to this object are from
  1360  			// conservative scanning, we have to scan it
  1361  			// defensively, too.
  1362  			state.putPtr(val, true)
  1363  			continue
  1364  		}
  1365  
  1366  		// Check if val points to a heap span.
  1367  		span := spanOfHeap(val)
  1368  		if span == nil {
  1369  			continue
  1370  		}
  1371  
  1372  		// Check if val points to an allocated object.
  1373  		idx := span.objIndex(val)
  1374  		if span.isFree(idx) {
  1375  			continue
  1376  		}
  1377  
  1378  		// val points to an allocated object. Mark it.
  1379  		obj := span.base() + idx*span.elemsize
  1380  		greyobject(obj, b, i, span, gcw, idx)
  1381  	}
  1382  }
  1383  
  1384  // Shade the object if it isn't already.
  1385  // The object is not nil and known to be in the heap.
  1386  // Preemption must be disabled.
  1387  //go:nowritebarrier
  1388  func shade(b uintptr) {
  1389  	if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
  1390  		gcw := &getg().m.p.ptr().gcw
  1391  		greyobject(obj, 0, 0, span, gcw, objIndex)
  1392  	}
  1393  }
  1394  
  1395  // obj is the start of an object with mark mbits.
  1396  // If it isn't already marked, mark it and enqueue into gcw.
  1397  // base and off are for debugging only and could be removed.
  1398  //
  1399  // See also wbBufFlush1, which partially duplicates this logic.
  1400  //
  1401  //go:nowritebarrierrec
  1402  func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
  1403  	// obj should be start of allocation, and so must be at least pointer-aligned.
  1404  	if obj&(sys.PtrSize-1) != 0 {
  1405  		throw("greyobject: obj not pointer-aligned")
  1406  	}
  1407  	mbits := span.markBitsForIndex(objIndex)
  1408  
  1409  	if useCheckmark {
  1410  		if setCheckmark(obj, base, off, mbits) {
  1411  			// Already marked.
  1412  			return
  1413  		}
  1414  	} else {
  1415  		if debug.gccheckmark > 0 && span.isFree(objIndex) {
  1416  			print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
  1417  			gcDumpObject("base", base, off)
  1418  			gcDumpObject("obj", obj, ^uintptr(0))
  1419  			getg().m.traceback = 2
  1420  			throw("marking free object")
  1421  		}
  1422  
  1423  		// If marked we have nothing to do.
  1424  		if mbits.isMarked() {
  1425  			return
  1426  		}
  1427  		mbits.setMarked()
  1428  
  1429  		// Mark span.
  1430  		arena, pageIdx, pageMask := pageIndexOf(span.base())
  1431  		if arena.pageMarks[pageIdx]&pageMask == 0 {
  1432  			atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1433  		}
  1434  
  1435  		// If this is a noscan object, fast-track it to black
  1436  		// instead of greying it.
  1437  		if span.spanclass.noscan() {
  1438  			gcw.bytesMarked += uint64(span.elemsize)
  1439  			return
  1440  		}
  1441  	}
  1442  
  1443  	// Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
  1444  	// seems like a nice optimization that can be added back in.
  1445  	// There needs to be time between the PREFETCH and the use.
  1446  	// Previously we put the obj in an 8 element buffer that is drained at a rate
  1447  	// to give the PREFETCH time to do its work.
  1448  	// Use of PREFETCHNTA might be more appropriate than PREFETCH
  1449  	if !gcw.putFast(obj) {
  1450  		gcw.put(obj)
  1451  	}
  1452  }
  1453  
  1454  // gcDumpObject dumps the contents of obj for debugging and marks the
  1455  // field at byte offset off in obj.
  1456  func gcDumpObject(label string, obj, off uintptr) {
  1457  	s := spanOf(obj)
  1458  	print(label, "=", hex(obj))
  1459  	if s == nil {
  1460  		print(" s=nil\n")
  1461  		return
  1462  	}
  1463  	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
  1464  	if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
  1465  		print(mSpanStateNames[state], "\n")
  1466  	} else {
  1467  		print("unknown(", state, ")\n")
  1468  	}
  1469  
  1470  	skipped := false
  1471  	size := s.elemsize
  1472  	if s.state.get() == mSpanManual && size == 0 {
  1473  		// We're printing something from a stack frame. We
  1474  		// don't know how big it is, so just show up to an
  1475  		// including off.
  1476  		size = off + sys.PtrSize
  1477  	}
  1478  	for i := uintptr(0); i < size; i += sys.PtrSize {
  1479  		// For big objects, just print the beginning (because
  1480  		// that usually hints at the object's type) and the
  1481  		// fields around off.
  1482  		if !(i < 128*sys.PtrSize || off-16*sys.PtrSize < i && i < off+16*sys.PtrSize) {
  1483  			skipped = true
  1484  			continue
  1485  		}
  1486  		if skipped {
  1487  			print(" ...\n")
  1488  			skipped = false
  1489  		}
  1490  		print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
  1491  		if i == off {
  1492  			print(" <==")
  1493  		}
  1494  		print("\n")
  1495  	}
  1496  	if skipped {
  1497  		print(" ...\n")
  1498  	}
  1499  }
  1500  
  1501  // gcmarknewobject marks a newly allocated object black. obj must
  1502  // not contain any non-nil pointers.
  1503  //
  1504  // This is nosplit so it can manipulate a gcWork without preemption.
  1505  //
  1506  //go:nowritebarrier
  1507  //go:nosplit
  1508  func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
  1509  	if useCheckmark { // The world should be stopped so this should not happen.
  1510  		throw("gcmarknewobject called while doing checkmark")
  1511  	}
  1512  
  1513  	// Mark object.
  1514  	objIndex := span.objIndex(obj)
  1515  	span.markBitsForIndex(objIndex).setMarked()
  1516  
  1517  	// Mark span.
  1518  	arena, pageIdx, pageMask := pageIndexOf(span.base())
  1519  	if arena.pageMarks[pageIdx]&pageMask == 0 {
  1520  		atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
  1521  	}
  1522  
  1523  	gcw := &getg().m.p.ptr().gcw
  1524  	gcw.bytesMarked += uint64(size)
  1525  	gcw.scanWork += int64(scanSize)
  1526  }
  1527  
  1528  // gcMarkTinyAllocs greys all active tiny alloc blocks.
  1529  //
  1530  // The world must be stopped.
  1531  func gcMarkTinyAllocs() {
  1532  	assertWorldStopped()
  1533  
  1534  	for _, p := range allp {
  1535  		c := p.mcache
  1536  		if c == nil || c.tiny == 0 {
  1537  			continue
  1538  		}
  1539  		_, span, objIndex := findObject(c.tiny, 0, 0)
  1540  		gcw := &p.gcw
  1541  		greyobject(c.tiny, 0, 0, span, gcw, objIndex)
  1542  	}
  1543  }
  1544  

View as plain text