Source file src/runtime/mheap.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  // Page heap.
     6  //
     7  // See malloc.go for overview.
     8  
     9  package runtime
    10  
    11  import (
    12  	"internal/cpu"
    13  	"internal/goarch"
    14  	"internal/runtime/atomic"
    15  	"runtime/internal/sys"
    16  	"unsafe"
    17  )
    18  
    19  const (
    20  	// minPhysPageSize is a lower-bound on the physical page size. The
    21  	// true physical page size may be larger than this. In contrast,
    22  	// sys.PhysPageSize is an upper-bound on the physical page size.
    23  	minPhysPageSize = 4096
    24  
    25  	// maxPhysPageSize is the maximum page size the runtime supports.
    26  	maxPhysPageSize = 512 << 10
    27  
    28  	// maxPhysHugePageSize sets an upper-bound on the maximum huge page size
    29  	// that the runtime supports.
    30  	maxPhysHugePageSize = pallocChunkBytes
    31  
    32  	// pagesPerReclaimerChunk indicates how many pages to scan from the
    33  	// pageInUse bitmap at a time. Used by the page reclaimer.
    34  	//
    35  	// Higher values reduce contention on scanning indexes (such as
    36  	// h.reclaimIndex), but increase the minimum latency of the
    37  	// operation.
    38  	//
    39  	// The time required to scan this many pages can vary a lot depending
    40  	// on how many spans are actually freed. Experimentally, it can
    41  	// scan for pages at ~300 GB/ms on a 2.6GHz Core i7, but can only
    42  	// free spans at ~32 MB/ms. Using 512 pages bounds this at
    43  	// roughly 100µs.
    44  	//
    45  	// Must be a multiple of the pageInUse bitmap element size and
    46  	// must also evenly divide pagesPerArena.
    47  	pagesPerReclaimerChunk = 512
    48  
    49  	// physPageAlignedStacks indicates whether stack allocations must be
    50  	// physical page aligned. This is a requirement for MAP_STACK on
    51  	// OpenBSD.
    52  	physPageAlignedStacks = GOOS == "openbsd"
    53  )
    54  
    55  // Main malloc heap.
    56  // The heap itself is the "free" and "scav" treaps,
    57  // but all the other global data is here too.
    58  //
    59  // mheap must not be heap-allocated because it contains mSpanLists,
    60  // which must not be heap-allocated.
    61  type mheap struct {
    62  	_ sys.NotInHeap
    63  
    64  	// lock must only be acquired on the system stack, otherwise a g
    65  	// could self-deadlock if its stack grows with the lock held.
    66  	lock mutex
    67  
    68  	pages pageAlloc // page allocation data structure
    69  
    70  	sweepgen uint32 // sweep generation, see comment in mspan; written during STW
    71  
    72  	// allspans is a slice of all mspans ever created. Each mspan
    73  	// appears exactly once.
    74  	//
    75  	// The memory for allspans is manually managed and can be
    76  	// reallocated and move as the heap grows.
    77  	//
    78  	// In general, allspans is protected by mheap_.lock, which
    79  	// prevents concurrent access as well as freeing the backing
    80  	// store. Accesses during STW might not hold the lock, but
    81  	// must ensure that allocation cannot happen around the
    82  	// access (since that may free the backing store).
    83  	allspans []*mspan // all spans out there
    84  
    85  	// Proportional sweep
    86  	//
    87  	// These parameters represent a linear function from gcController.heapLive
    88  	// to page sweep count. The proportional sweep system works to
    89  	// stay in the black by keeping the current page sweep count
    90  	// above this line at the current gcController.heapLive.
    91  	//
    92  	// The line has slope sweepPagesPerByte and passes through a
    93  	// basis point at (sweepHeapLiveBasis, pagesSweptBasis). At
    94  	// any given time, the system is at (gcController.heapLive,
    95  	// pagesSwept) in this space.
    96  	//
    97  	// It is important that the line pass through a point we
    98  	// control rather than simply starting at a 0,0 origin
    99  	// because that lets us adjust sweep pacing at any time while
   100  	// accounting for current progress. If we could only adjust
   101  	// the slope, it would create a discontinuity in debt if any
   102  	// progress has already been made.
   103  	pagesInUse         atomic.Uintptr // pages of spans in stats mSpanInUse
   104  	pagesSwept         atomic.Uint64  // pages swept this cycle
   105  	pagesSweptBasis    atomic.Uint64  // pagesSwept to use as the origin of the sweep ratio
   106  	sweepHeapLiveBasis uint64         // value of gcController.heapLive to use as the origin of sweep ratio; written with lock, read without
   107  	sweepPagesPerByte  float64        // proportional sweep ratio; written with lock, read without
   108  
   109  	// Page reclaimer state
   110  
   111  	// reclaimIndex is the page index in allArenas of next page to
   112  	// reclaim. Specifically, it refers to page (i %
   113  	// pagesPerArena) of arena allArenas[i / pagesPerArena].
   114  	//
   115  	// If this is >= 1<<63, the page reclaimer is done scanning
   116  	// the page marks.
   117  	reclaimIndex atomic.Uint64
   118  
   119  	// reclaimCredit is spare credit for extra pages swept. Since
   120  	// the page reclaimer works in large chunks, it may reclaim
   121  	// more than requested. Any spare pages released go to this
   122  	// credit pool.
   123  	reclaimCredit atomic.Uintptr
   124  
   125  	_ cpu.CacheLinePad // prevents false-sharing between arenas and preceding variables
   126  
   127  	// arenas is the heap arena map. It points to the metadata for
   128  	// the heap for every arena frame of the entire usable virtual
   129  	// address space.
   130  	//
   131  	// Use arenaIndex to compute indexes into this array.
   132  	//
   133  	// For regions of the address space that are not backed by the
   134  	// Go heap, the arena map contains nil.
   135  	//
   136  	// Modifications are protected by mheap_.lock. Reads can be
   137  	// performed without locking; however, a given entry can
   138  	// transition from nil to non-nil at any time when the lock
   139  	// isn't held. (Entries never transitions back to nil.)
   140  	//
   141  	// In general, this is a two-level mapping consisting of an L1
   142  	// map and possibly many L2 maps. This saves space when there
   143  	// are a huge number of arena frames. However, on many
   144  	// platforms (even 64-bit), arenaL1Bits is 0, making this
   145  	// effectively a single-level map. In this case, arenas[0]
   146  	// will never be nil.
   147  	arenas [1 << arenaL1Bits]*[1 << arenaL2Bits]*heapArena
   148  
   149  	// arenasHugePages indicates whether arenas' L2 entries are eligible
   150  	// to be backed by huge pages.
   151  	arenasHugePages bool
   152  
   153  	// heapArenaAlloc is pre-reserved space for allocating heapArena
   154  	// objects. This is only used on 32-bit, where we pre-reserve
   155  	// this space to avoid interleaving it with the heap itself.
   156  	heapArenaAlloc linearAlloc
   157  
   158  	// arenaHints is a list of addresses at which to attempt to
   159  	// add more heap arenas. This is initially populated with a
   160  	// set of general hint addresses, and grown with the bounds of
   161  	// actual heap arena ranges.
   162  	arenaHints *arenaHint
   163  
   164  	// arena is a pre-reserved space for allocating heap arenas
   165  	// (the actual arenas). This is only used on 32-bit.
   166  	arena linearAlloc
   167  
   168  	// allArenas is the arenaIndex of every mapped arena. This can
   169  	// be used to iterate through the address space.
   170  	//
   171  	// Access is protected by mheap_.lock. However, since this is
   172  	// append-only and old backing arrays are never freed, it is
   173  	// safe to acquire mheap_.lock, copy the slice header, and
   174  	// then release mheap_.lock.
   175  	allArenas []arenaIdx
   176  
   177  	// sweepArenas is a snapshot of allArenas taken at the
   178  	// beginning of the sweep cycle. This can be read safely by
   179  	// simply blocking GC (by disabling preemption).
   180  	sweepArenas []arenaIdx
   181  
   182  	// markArenas is a snapshot of allArenas taken at the beginning
   183  	// of the mark cycle. Because allArenas is append-only, neither
   184  	// this slice nor its contents will change during the mark, so
   185  	// it can be read safely.
   186  	markArenas []arenaIdx
   187  
   188  	// curArena is the arena that the heap is currently growing
   189  	// into. This should always be physPageSize-aligned.
   190  	curArena struct {
   191  		base, end uintptr
   192  	}
   193  
   194  	// central free lists for small size classes.
   195  	// the padding makes sure that the mcentrals are
   196  	// spaced CacheLinePadSize bytes apart, so that each mcentral.lock
   197  	// gets its own cache line.
   198  	// central is indexed by spanClass.
   199  	central [numSpanClasses]struct {
   200  		mcentral mcentral
   201  		pad      [(cpu.CacheLinePadSize - unsafe.Sizeof(mcentral{})%cpu.CacheLinePadSize) % cpu.CacheLinePadSize]byte
   202  	}
   203  
   204  	spanalloc              fixalloc // allocator for span*
   205  	cachealloc             fixalloc // allocator for mcache*
   206  	specialfinalizeralloc  fixalloc // allocator for specialfinalizer*
   207  	specialprofilealloc    fixalloc // allocator for specialprofile*
   208  	specialReachableAlloc  fixalloc // allocator for specialReachable
   209  	specialPinCounterAlloc fixalloc // allocator for specialPinCounter
   210  	specialWeakHandleAlloc fixalloc // allocator for specialWeakHandle
   211  	speciallock            mutex    // lock for special record allocators.
   212  	arenaHintAlloc         fixalloc // allocator for arenaHints
   213  
   214  	// User arena state.
   215  	//
   216  	// Protected by mheap_.lock.
   217  	userArena struct {
   218  		// arenaHints is a list of addresses at which to attempt to
   219  		// add more heap arenas for user arena chunks. This is initially
   220  		// populated with a set of general hint addresses, and grown with
   221  		// the bounds of actual heap arena ranges.
   222  		arenaHints *arenaHint
   223  
   224  		// quarantineList is a list of user arena spans that have been set to fault, but
   225  		// are waiting for all pointers into them to go away. Sweeping handles
   226  		// identifying when this is true, and moves the span to the ready list.
   227  		quarantineList mSpanList
   228  
   229  		// readyList is a list of empty user arena spans that are ready for reuse.
   230  		readyList mSpanList
   231  	}
   232  
   233  	unused *specialfinalizer // never set, just here to force the specialfinalizer type into DWARF
   234  }
   235  
   236  var mheap_ mheap
   237  
   238  // A heapArena stores metadata for a heap arena. heapArenas are stored
   239  // outside of the Go heap and accessed via the mheap_.arenas index.
   240  type heapArena struct {
   241  	_ sys.NotInHeap
   242  
   243  	// spans maps from virtual address page ID within this arena to *mspan.
   244  	// For allocated spans, their pages map to the span itself.
   245  	// For free spans, only the lowest and highest pages map to the span itself.
   246  	// Internal pages map to an arbitrary span.
   247  	// For pages that have never been allocated, spans entries are nil.
   248  	//
   249  	// Modifications are protected by mheap.lock. Reads can be
   250  	// performed without locking, but ONLY from indexes that are
   251  	// known to contain in-use or stack spans. This means there
   252  	// must not be a safe-point between establishing that an
   253  	// address is live and looking it up in the spans array.
   254  	spans [pagesPerArena]*mspan
   255  
   256  	// pageInUse is a bitmap that indicates which spans are in
   257  	// state mSpanInUse. This bitmap is indexed by page number,
   258  	// but only the bit corresponding to the first page in each
   259  	// span is used.
   260  	//
   261  	// Reads and writes are atomic.
   262  	pageInUse [pagesPerArena / 8]uint8
   263  
   264  	// pageMarks is a bitmap that indicates which spans have any
   265  	// marked objects on them. Like pageInUse, only the bit
   266  	// corresponding to the first page in each span is used.
   267  	//
   268  	// Writes are done atomically during marking. Reads are
   269  	// non-atomic and lock-free since they only occur during
   270  	// sweeping (and hence never race with writes).
   271  	//
   272  	// This is used to quickly find whole spans that can be freed.
   273  	//
   274  	// TODO(austin): It would be nice if this was uint64 for
   275  	// faster scanning, but we don't have 64-bit atomic bit
   276  	// operations.
   277  	pageMarks [pagesPerArena / 8]uint8
   278  
   279  	// pageSpecials is a bitmap that indicates which spans have
   280  	// specials (finalizers or other). Like pageInUse, only the bit
   281  	// corresponding to the first page in each span is used.
   282  	//
   283  	// Writes are done atomically whenever a special is added to
   284  	// a span and whenever the last special is removed from a span.
   285  	// Reads are done atomically to find spans containing specials
   286  	// during marking.
   287  	pageSpecials [pagesPerArena / 8]uint8
   288  
   289  	// checkmarks stores the debug.gccheckmark state. It is only
   290  	// used if debug.gccheckmark > 0.
   291  	checkmarks *checkmarksMap
   292  
   293  	// zeroedBase marks the first byte of the first page in this
   294  	// arena which hasn't been used yet and is therefore already
   295  	// zero. zeroedBase is relative to the arena base.
   296  	// Increases monotonically until it hits heapArenaBytes.
   297  	//
   298  	// This field is sufficient to determine if an allocation
   299  	// needs to be zeroed because the page allocator follows an
   300  	// address-ordered first-fit policy.
   301  	//
   302  	// Read atomically and written with an atomic CAS.
   303  	zeroedBase uintptr
   304  }
   305  
   306  // arenaHint is a hint for where to grow the heap arenas. See
   307  // mheap_.arenaHints.
   308  type arenaHint struct {
   309  	_    sys.NotInHeap
   310  	addr uintptr
   311  	down bool
   312  	next *arenaHint
   313  }
   314  
   315  // An mspan is a run of pages.
   316  //
   317  // When a mspan is in the heap free treap, state == mSpanFree
   318  // and heapmap(s->start) == span, heapmap(s->start+s->npages-1) == span.
   319  // If the mspan is in the heap scav treap, then in addition to the
   320  // above scavenged == true. scavenged == false in all other cases.
   321  //
   322  // When a mspan is allocated, state == mSpanInUse or mSpanManual
   323  // and heapmap(i) == span for all s->start <= i < s->start+s->npages.
   324  
   325  // Every mspan is in one doubly-linked list, either in the mheap's
   326  // busy list or one of the mcentral's span lists.
   327  
   328  // An mspan representing actual memory has state mSpanInUse,
   329  // mSpanManual, or mSpanFree. Transitions between these states are
   330  // constrained as follows:
   331  //
   332  //   - A span may transition from free to in-use or manual during any GC
   333  //     phase.
   334  //
   335  //   - During sweeping (gcphase == _GCoff), a span may transition from
   336  //     in-use to free (as a result of sweeping) or manual to free (as a
   337  //     result of stacks being freed).
   338  //
   339  //   - During GC (gcphase != _GCoff), a span *must not* transition from
   340  //     manual or in-use to free. Because concurrent GC may read a pointer
   341  //     and then look up its span, the span state must be monotonic.
   342  //
   343  // Setting mspan.state to mSpanInUse or mSpanManual must be done
   344  // atomically and only after all other span fields are valid.
   345  // Likewise, if inspecting a span is contingent on it being
   346  // mSpanInUse, the state should be loaded atomically and checked
   347  // before depending on other fields. This allows the garbage collector
   348  // to safely deal with potentially invalid pointers, since resolving
   349  // such pointers may race with a span being allocated.
   350  type mSpanState uint8
   351  
   352  const (
   353  	mSpanDead   mSpanState = iota
   354  	mSpanInUse             // allocated for garbage collected heap
   355  	mSpanManual            // allocated for manual management (e.g., stack allocator)
   356  )
   357  
   358  // mSpanStateNames are the names of the span states, indexed by
   359  // mSpanState.
   360  var mSpanStateNames = []string{
   361  	"mSpanDead",
   362  	"mSpanInUse",
   363  	"mSpanManual",
   364  }
   365  
   366  // mSpanStateBox holds an atomic.Uint8 to provide atomic operations on
   367  // an mSpanState. This is a separate type to disallow accidental comparison
   368  // or assignment with mSpanState.
   369  type mSpanStateBox struct {
   370  	s atomic.Uint8
   371  }
   372  
   373  // It is nosplit to match get, below.
   374  
   375  //go:nosplit
   376  func (b *mSpanStateBox) set(s mSpanState) {
   377  	b.s.Store(uint8(s))
   378  }
   379  
   380  // It is nosplit because it's called indirectly by typedmemclr,
   381  // which must not be preempted.
   382  
   383  //go:nosplit
   384  func (b *mSpanStateBox) get() mSpanState {
   385  	return mSpanState(b.s.Load())
   386  }
   387  
   388  // mSpanList heads a linked list of spans.
   389  type mSpanList struct {
   390  	_     sys.NotInHeap
   391  	first *mspan // first span in list, or nil if none
   392  	last  *mspan // last span in list, or nil if none
   393  }
   394  
   395  type mspan struct {
   396  	_    sys.NotInHeap
   397  	next *mspan     // next span in list, or nil if none
   398  	prev *mspan     // previous span in list, or nil if none
   399  	list *mSpanList // For debugging.
   400  
   401  	startAddr uintptr // address of first byte of span aka s.base()
   402  	npages    uintptr // number of pages in span
   403  
   404  	manualFreeList gclinkptr // list of free objects in mSpanManual spans
   405  
   406  	// freeindex is the slot index between 0 and nelems at which to begin scanning
   407  	// for the next free object in this span.
   408  	// Each allocation scans allocBits starting at freeindex until it encounters a 0
   409  	// indicating a free object. freeindex is then adjusted so that subsequent scans begin
   410  	// just past the newly discovered free object.
   411  	//
   412  	// If freeindex == nelem, this span has no free objects.
   413  	//
   414  	// allocBits is a bitmap of objects in this span.
   415  	// If n >= freeindex and allocBits[n/8] & (1<<(n%8)) is 0
   416  	// then object n is free;
   417  	// otherwise, object n is allocated. Bits starting at nelem are
   418  	// undefined and should never be referenced.
   419  	//
   420  	// Object n starts at address n*elemsize + (start << pageShift).
   421  	freeindex uint16
   422  	// TODO: Look up nelems from sizeclass and remove this field if it
   423  	// helps performance.
   424  	nelems uint16 // number of object in the span.
   425  	// freeIndexForScan is like freeindex, except that freeindex is
   426  	// used by the allocator whereas freeIndexForScan is used by the
   427  	// GC scanner. They are two fields so that the GC sees the object
   428  	// is allocated only when the object and the heap bits are
   429  	// initialized (see also the assignment of freeIndexForScan in
   430  	// mallocgc, and issue 54596).
   431  	freeIndexForScan uint16
   432  
   433  	// Cache of the allocBits at freeindex. allocCache is shifted
   434  	// such that the lowest bit corresponds to the bit freeindex.
   435  	// allocCache holds the complement of allocBits, thus allowing
   436  	// ctz (count trailing zero) to use it directly.
   437  	// allocCache may contain bits beyond s.nelems; the caller must ignore
   438  	// these.
   439  	allocCache uint64
   440  
   441  	// allocBits and gcmarkBits hold pointers to a span's mark and
   442  	// allocation bits. The pointers are 8 byte aligned.
   443  	// There are three arenas where this data is held.
   444  	// free: Dirty arenas that are no longer accessed
   445  	//       and can be reused.
   446  	// next: Holds information to be used in the next GC cycle.
   447  	// current: Information being used during this GC cycle.
   448  	// previous: Information being used during the last GC cycle.
   449  	// A new GC cycle starts with the call to finishsweep_m.
   450  	// finishsweep_m moves the previous arena to the free arena,
   451  	// the current arena to the previous arena, and
   452  	// the next arena to the current arena.
   453  	// The next arena is populated as the spans request
   454  	// memory to hold gcmarkBits for the next GC cycle as well
   455  	// as allocBits for newly allocated spans.
   456  	//
   457  	// The pointer arithmetic is done "by hand" instead of using
   458  	// arrays to avoid bounds checks along critical performance
   459  	// paths.
   460  	// The sweep will free the old allocBits and set allocBits to the
   461  	// gcmarkBits. The gcmarkBits are replaced with a fresh zeroed
   462  	// out memory.
   463  	allocBits  *gcBits
   464  	gcmarkBits *gcBits
   465  	pinnerBits *gcBits // bitmap for pinned objects; accessed atomically
   466  
   467  	// sweep generation:
   468  	// if sweepgen == h->sweepgen - 2, the span needs sweeping
   469  	// if sweepgen == h->sweepgen - 1, the span is currently being swept
   470  	// if sweepgen == h->sweepgen, the span is swept and ready to use
   471  	// if sweepgen == h->sweepgen + 1, the span was cached before sweep began and is still cached, and needs sweeping
   472  	// if sweepgen == h->sweepgen + 3, the span was swept and then cached and is still cached
   473  	// h->sweepgen is incremented by 2 after every GC
   474  
   475  	sweepgen              uint32
   476  	divMul                uint32        // for divide by elemsize
   477  	allocCount            uint16        // number of allocated objects
   478  	spanclass             spanClass     // size class and noscan (uint8)
   479  	state                 mSpanStateBox // mSpanInUse etc; accessed atomically (get/set methods)
   480  	needzero              uint8         // needs to be zeroed before allocation
   481  	isUserArenaChunk      bool          // whether or not this span represents a user arena
   482  	allocCountBeforeCache uint16        // a copy of allocCount that is stored just before this span is cached
   483  	elemsize              uintptr       // computed from sizeclass or from npages
   484  	limit                 uintptr       // end of data in span
   485  	speciallock           mutex         // guards specials list and changes to pinnerBits
   486  	specials              *special      // linked list of special records sorted by offset.
   487  	userArenaChunkFree    addrRange     // interval for managing chunk allocation
   488  	largeType             *_type        // malloc header for large objects.
   489  }
   490  
   491  func (s *mspan) base() uintptr {
   492  	return s.startAddr
   493  }
   494  
   495  func (s *mspan) layout() (size, n, total uintptr) {
   496  	total = s.npages << _PageShift
   497  	size = s.elemsize
   498  	if size > 0 {
   499  		n = total / size
   500  	}
   501  	return
   502  }
   503  
   504  // recordspan adds a newly allocated span to h.allspans.
   505  //
   506  // This only happens the first time a span is allocated from
   507  // mheap.spanalloc (it is not called when a span is reused).
   508  //
   509  // Write barriers are disallowed here because it can be called from
   510  // gcWork when allocating new workbufs. However, because it's an
   511  // indirect call from the fixalloc initializer, the compiler can't see
   512  // this.
   513  //
   514  // The heap lock must be held.
   515  //
   516  //go:nowritebarrierrec
   517  func recordspan(vh unsafe.Pointer, p unsafe.Pointer) {
   518  	h := (*mheap)(vh)
   519  	s := (*mspan)(p)
   520  
   521  	assertLockHeld(&h.lock)
   522  
   523  	if len(h.allspans) >= cap(h.allspans) {
   524  		n := 64 * 1024 / goarch.PtrSize
   525  		if n < cap(h.allspans)*3/2 {
   526  			n = cap(h.allspans) * 3 / 2
   527  		}
   528  		var new []*mspan
   529  		sp := (*slice)(unsafe.Pointer(&new))
   530  		sp.array = sysAlloc(uintptr(n)*goarch.PtrSize, &memstats.other_sys)
   531  		if sp.array == nil {
   532  			throw("runtime: cannot allocate memory")
   533  		}
   534  		sp.len = len(h.allspans)
   535  		sp.cap = n
   536  		if len(h.allspans) > 0 {
   537  			copy(new, h.allspans)
   538  		}
   539  		oldAllspans := h.allspans
   540  		*(*notInHeapSlice)(unsafe.Pointer(&h.allspans)) = *(*notInHeapSlice)(unsafe.Pointer(&new))
   541  		if len(oldAllspans) != 0 {
   542  			sysFree(unsafe.Pointer(&oldAllspans[0]), uintptr(cap(oldAllspans))*unsafe.Sizeof(oldAllspans[0]), &memstats.other_sys)
   543  		}
   544  	}
   545  	h.allspans = h.allspans[:len(h.allspans)+1]
   546  	h.allspans[len(h.allspans)-1] = s
   547  }
   548  
   549  // A spanClass represents the size class and noscan-ness of a span.
   550  //
   551  // Each size class has a noscan spanClass and a scan spanClass. The
   552  // noscan spanClass contains only noscan objects, which do not contain
   553  // pointers and thus do not need to be scanned by the garbage
   554  // collector.
   555  type spanClass uint8
   556  
   557  const (
   558  	numSpanClasses = _NumSizeClasses << 1
   559  	tinySpanClass  = spanClass(tinySizeClass<<1 | 1)
   560  )
   561  
   562  func makeSpanClass(sizeclass uint8, noscan bool) spanClass {
   563  	return spanClass(sizeclass<<1) | spanClass(bool2int(noscan))
   564  }
   565  
   566  //go:nosplit
   567  func (sc spanClass) sizeclass() int8 {
   568  	return int8(sc >> 1)
   569  }
   570  
   571  //go:nosplit
   572  func (sc spanClass) noscan() bool {
   573  	return sc&1 != 0
   574  }
   575  
   576  // arenaIndex returns the index into mheap_.arenas of the arena
   577  // containing metadata for p. This index combines of an index into the
   578  // L1 map and an index into the L2 map and should be used as
   579  // mheap_.arenas[ai.l1()][ai.l2()].
   580  //
   581  // If p is outside the range of valid heap addresses, either l1() or
   582  // l2() will be out of bounds.
   583  //
   584  // It is nosplit because it's called by spanOf and several other
   585  // nosplit functions.
   586  //
   587  //go:nosplit
   588  func arenaIndex(p uintptr) arenaIdx {
   589  	return arenaIdx((p - arenaBaseOffset) / heapArenaBytes)
   590  }
   591  
   592  // arenaBase returns the low address of the region covered by heap
   593  // arena i.
   594  func arenaBase(i arenaIdx) uintptr {
   595  	return uintptr(i)*heapArenaBytes + arenaBaseOffset
   596  }
   597  
   598  type arenaIdx uint
   599  
   600  // l1 returns the "l1" portion of an arenaIdx.
   601  //
   602  // Marked nosplit because it's called by spanOf and other nosplit
   603  // functions.
   604  //
   605  //go:nosplit
   606  func (i arenaIdx) l1() uint {
   607  	if arenaL1Bits == 0 {
   608  		// Let the compiler optimize this away if there's no
   609  		// L1 map.
   610  		return 0
   611  	} else {
   612  		return uint(i) >> arenaL1Shift
   613  	}
   614  }
   615  
   616  // l2 returns the "l2" portion of an arenaIdx.
   617  //
   618  // Marked nosplit because it's called by spanOf and other nosplit funcs.
   619  // functions.
   620  //
   621  //go:nosplit
   622  func (i arenaIdx) l2() uint {
   623  	if arenaL1Bits == 0 {
   624  		return uint(i)
   625  	} else {
   626  		return uint(i) & (1<<arenaL2Bits - 1)
   627  	}
   628  }
   629  
   630  // inheap reports whether b is a pointer into a (potentially dead) heap object.
   631  // It returns false for pointers into mSpanManual spans.
   632  // Non-preemptible because it is used by write barriers.
   633  //
   634  //go:nowritebarrier
   635  //go:nosplit
   636  func inheap(b uintptr) bool {
   637  	return spanOfHeap(b) != nil
   638  }
   639  
   640  // inHeapOrStack is a variant of inheap that returns true for pointers
   641  // into any allocated heap span.
   642  //
   643  //go:nowritebarrier
   644  //go:nosplit
   645  func inHeapOrStack(b uintptr) bool {
   646  	s := spanOf(b)
   647  	if s == nil || b < s.base() {
   648  		return false
   649  	}
   650  	switch s.state.get() {
   651  	case mSpanInUse, mSpanManual:
   652  		return b < s.limit
   653  	default:
   654  		return false
   655  	}
   656  }
   657  
   658  // spanOf returns the span of p. If p does not point into the heap
   659  // arena or no span has ever contained p, spanOf returns nil.
   660  //
   661  // If p does not point to allocated memory, this may return a non-nil
   662  // span that does *not* contain p. If this is a possibility, the
   663  // caller should either call spanOfHeap or check the span bounds
   664  // explicitly.
   665  //
   666  // Must be nosplit because it has callers that are nosplit.
   667  //
   668  //go:nosplit
   669  func spanOf(p uintptr) *mspan {
   670  	// This function looks big, but we use a lot of constant
   671  	// folding around arenaL1Bits to get it under the inlining
   672  	// budget. Also, many of the checks here are safety checks
   673  	// that Go needs to do anyway, so the generated code is quite
   674  	// short.
   675  	ri := arenaIndex(p)
   676  	if arenaL1Bits == 0 {
   677  		// If there's no L1, then ri.l1() can't be out of bounds but ri.l2() can.
   678  		if ri.l2() >= uint(len(mheap_.arenas[0])) {
   679  			return nil
   680  		}
   681  	} else {
   682  		// If there's an L1, then ri.l1() can be out of bounds but ri.l2() can't.
   683  		if ri.l1() >= uint(len(mheap_.arenas)) {
   684  			return nil
   685  		}
   686  	}
   687  	l2 := mheap_.arenas[ri.l1()]
   688  	if arenaL1Bits != 0 && l2 == nil { // Should never happen if there's no L1.
   689  		return nil
   690  	}
   691  	ha := l2[ri.l2()]
   692  	if ha == nil {
   693  		return nil
   694  	}
   695  	return ha.spans[(p/pageSize)%pagesPerArena]
   696  }
   697  
   698  // spanOfUnchecked is equivalent to spanOf, but the caller must ensure
   699  // that p points into an allocated heap arena.
   700  //
   701  // Must be nosplit because it has callers that are nosplit.
   702  //
   703  //go:nosplit
   704  func spanOfUnchecked(p uintptr) *mspan {
   705  	ai := arenaIndex(p)
   706  	return mheap_.arenas[ai.l1()][ai.l2()].spans[(p/pageSize)%pagesPerArena]
   707  }
   708  
   709  // spanOfHeap is like spanOf, but returns nil if p does not point to a
   710  // heap object.
   711  //
   712  // Must be nosplit because it has callers that are nosplit.
   713  //
   714  //go:nosplit
   715  func spanOfHeap(p uintptr) *mspan {
   716  	s := spanOf(p)
   717  	// s is nil if it's never been allocated. Otherwise, we check
   718  	// its state first because we don't trust this pointer, so we
   719  	// have to synchronize with span initialization. Then, it's
   720  	// still possible we picked up a stale span pointer, so we
   721  	// have to check the span's bounds.
   722  	if s == nil || s.state.get() != mSpanInUse || p < s.base() || p >= s.limit {
   723  		return nil
   724  	}
   725  	return s
   726  }
   727  
   728  // pageIndexOf returns the arena, page index, and page mask for pointer p.
   729  // The caller must ensure p is in the heap.
   730  func pageIndexOf(p uintptr) (arena *heapArena, pageIdx uintptr, pageMask uint8) {
   731  	ai := arenaIndex(p)
   732  	arena = mheap_.arenas[ai.l1()][ai.l2()]
   733  	pageIdx = ((p / pageSize) / 8) % uintptr(len(arena.pageInUse))
   734  	pageMask = byte(1 << ((p / pageSize) % 8))
   735  	return
   736  }
   737  
   738  // Initialize the heap.
   739  func (h *mheap) init() {
   740  	lockInit(&h.lock, lockRankMheap)
   741  	lockInit(&h.speciallock, lockRankMheapSpecial)
   742  
   743  	h.spanalloc.init(unsafe.Sizeof(mspan{}), recordspan, unsafe.Pointer(h), &memstats.mspan_sys)
   744  	h.cachealloc.init(unsafe.Sizeof(mcache{}), nil, nil, &memstats.mcache_sys)
   745  	h.specialfinalizeralloc.init(unsafe.Sizeof(specialfinalizer{}), nil, nil, &memstats.other_sys)
   746  	h.specialprofilealloc.init(unsafe.Sizeof(specialprofile{}), nil, nil, &memstats.other_sys)
   747  	h.specialReachableAlloc.init(unsafe.Sizeof(specialReachable{}), nil, nil, &memstats.other_sys)
   748  	h.specialPinCounterAlloc.init(unsafe.Sizeof(specialPinCounter{}), nil, nil, &memstats.other_sys)
   749  	h.specialWeakHandleAlloc.init(unsafe.Sizeof(specialWeakHandle{}), nil, nil, &memstats.gcMiscSys)
   750  	h.arenaHintAlloc.init(unsafe.Sizeof(arenaHint{}), nil, nil, &memstats.other_sys)
   751  
   752  	// Don't zero mspan allocations. Background sweeping can
   753  	// inspect a span concurrently with allocating it, so it's
   754  	// important that the span's sweepgen survive across freeing
   755  	// and re-allocating a span to prevent background sweeping
   756  	// from improperly cas'ing it from 0.
   757  	//
   758  	// This is safe because mspan contains no heap pointers.
   759  	h.spanalloc.zero = false
   760  
   761  	// h->mapcache needs no init
   762  
   763  	for i := range h.central {
   764  		h.central[i].mcentral.init(spanClass(i))
   765  	}
   766  
   767  	h.pages.init(&h.lock, &memstats.gcMiscSys, false)
   768  }
   769  
   770  // reclaim sweeps and reclaims at least npage pages into the heap.
   771  // It is called before allocating npage pages to keep growth in check.
   772  //
   773  // reclaim implements the page-reclaimer half of the sweeper.
   774  //
   775  // h.lock must NOT be held.
   776  func (h *mheap) reclaim(npage uintptr) {
   777  	// TODO(austin): Half of the time spent freeing spans is in
   778  	// locking/unlocking the heap (even with low contention). We
   779  	// could make the slow path here several times faster by
   780  	// batching heap frees.
   781  
   782  	// Bail early if there's no more reclaim work.
   783  	if h.reclaimIndex.Load() >= 1<<63 {
   784  		return
   785  	}
   786  
   787  	// Disable preemption so the GC can't start while we're
   788  	// sweeping, so we can read h.sweepArenas, and so
   789  	// traceGCSweepStart/Done pair on the P.
   790  	mp := acquirem()
   791  
   792  	trace := traceAcquire()
   793  	if trace.ok() {
   794  		trace.GCSweepStart()
   795  		traceRelease(trace)
   796  	}
   797  
   798  	arenas := h.sweepArenas
   799  	locked := false
   800  	for npage > 0 {
   801  		// Pull from accumulated credit first.
   802  		if credit := h.reclaimCredit.Load(); credit > 0 {
   803  			take := credit
   804  			if take > npage {
   805  				// Take only what we need.
   806  				take = npage
   807  			}
   808  			if h.reclaimCredit.CompareAndSwap(credit, credit-take) {
   809  				npage -= take
   810  			}
   811  			continue
   812  		}
   813  
   814  		// Claim a chunk of work.
   815  		idx := uintptr(h.reclaimIndex.Add(pagesPerReclaimerChunk) - pagesPerReclaimerChunk)
   816  		if idx/pagesPerArena >= uintptr(len(arenas)) {
   817  			// Page reclaiming is done.
   818  			h.reclaimIndex.Store(1 << 63)
   819  			break
   820  		}
   821  
   822  		if !locked {
   823  			// Lock the heap for reclaimChunk.
   824  			lock(&h.lock)
   825  			locked = true
   826  		}
   827  
   828  		// Scan this chunk.
   829  		nfound := h.reclaimChunk(arenas, idx, pagesPerReclaimerChunk)
   830  		if nfound <= npage {
   831  			npage -= nfound
   832  		} else {
   833  			// Put spare pages toward global credit.
   834  			h.reclaimCredit.Add(nfound - npage)
   835  			npage = 0
   836  		}
   837  	}
   838  	if locked {
   839  		unlock(&h.lock)
   840  	}
   841  
   842  	trace = traceAcquire()
   843  	if trace.ok() {
   844  		trace.GCSweepDone()
   845  		traceRelease(trace)
   846  	}
   847  	releasem(mp)
   848  }
   849  
   850  // reclaimChunk sweeps unmarked spans that start at page indexes [pageIdx, pageIdx+n).
   851  // It returns the number of pages returned to the heap.
   852  //
   853  // h.lock must be held and the caller must be non-preemptible. Note: h.lock may be
   854  // temporarily unlocked and re-locked in order to do sweeping or if tracing is
   855  // enabled.
   856  func (h *mheap) reclaimChunk(arenas []arenaIdx, pageIdx, n uintptr) uintptr {
   857  	// The heap lock must be held because this accesses the
   858  	// heapArena.spans arrays using potentially non-live pointers.
   859  	// In particular, if a span were freed and merged concurrently
   860  	// with this probing heapArena.spans, it would be possible to
   861  	// observe arbitrary, stale span pointers.
   862  	assertLockHeld(&h.lock)
   863  
   864  	n0 := n
   865  	var nFreed uintptr
   866  	sl := sweep.active.begin()
   867  	if !sl.valid {
   868  		return 0
   869  	}
   870  	for n > 0 {
   871  		ai := arenas[pageIdx/pagesPerArena]
   872  		ha := h.arenas[ai.l1()][ai.l2()]
   873  
   874  		// Get a chunk of the bitmap to work on.
   875  		arenaPage := uint(pageIdx % pagesPerArena)
   876  		inUse := ha.pageInUse[arenaPage/8:]
   877  		marked := ha.pageMarks[arenaPage/8:]
   878  		if uintptr(len(inUse)) > n/8 {
   879  			inUse = inUse[:n/8]
   880  			marked = marked[:n/8]
   881  		}
   882  
   883  		// Scan this bitmap chunk for spans that are in-use
   884  		// but have no marked objects on them.
   885  		for i := range inUse {
   886  			inUseUnmarked := atomic.Load8(&inUse[i]) &^ marked[i]
   887  			if inUseUnmarked == 0 {
   888  				continue
   889  			}
   890  
   891  			for j := uint(0); j < 8; j++ {
   892  				if inUseUnmarked&(1<<j) != 0 {
   893  					s := ha.spans[arenaPage+uint(i)*8+j]
   894  					if s, ok := sl.tryAcquire(s); ok {
   895  						npages := s.npages
   896  						unlock(&h.lock)
   897  						if s.sweep(false) {
   898  							nFreed += npages
   899  						}
   900  						lock(&h.lock)
   901  						// Reload inUse. It's possible nearby
   902  						// spans were freed when we dropped the
   903  						// lock and we don't want to get stale
   904  						// pointers from the spans array.
   905  						inUseUnmarked = atomic.Load8(&inUse[i]) &^ marked[i]
   906  					}
   907  				}
   908  			}
   909  		}
   910  
   911  		// Advance.
   912  		pageIdx += uintptr(len(inUse) * 8)
   913  		n -= uintptr(len(inUse) * 8)
   914  	}
   915  	sweep.active.end(sl)
   916  	trace := traceAcquire()
   917  	if trace.ok() {
   918  		unlock(&h.lock)
   919  		// Account for pages scanned but not reclaimed.
   920  		trace.GCSweepSpan((n0 - nFreed) * pageSize)
   921  		traceRelease(trace)
   922  		lock(&h.lock)
   923  	}
   924  
   925  	assertLockHeld(&h.lock) // Must be locked on return.
   926  	return nFreed
   927  }
   928  
   929  // spanAllocType represents the type of allocation to make, or
   930  // the type of allocation to be freed.
   931  type spanAllocType uint8
   932  
   933  const (
   934  	spanAllocHeap          spanAllocType = iota // heap span
   935  	spanAllocStack                              // stack span
   936  	spanAllocPtrScalarBits                      // unrolled GC prog bitmap span
   937  	spanAllocWorkBuf                            // work buf span
   938  )
   939  
   940  // manual returns true if the span allocation is manually managed.
   941  func (s spanAllocType) manual() bool {
   942  	return s != spanAllocHeap
   943  }
   944  
   945  // alloc allocates a new span of npage pages from the GC'd heap.
   946  //
   947  // spanclass indicates the span's size class and scannability.
   948  //
   949  // Returns a span that has been fully initialized. span.needzero indicates
   950  // whether the span has been zeroed. Note that it may not be.
   951  func (h *mheap) alloc(npages uintptr, spanclass spanClass) *mspan {
   952  	// Don't do any operations that lock the heap on the G stack.
   953  	// It might trigger stack growth, and the stack growth code needs
   954  	// to be able to allocate heap.
   955  	var s *mspan
   956  	systemstack(func() {
   957  		// To prevent excessive heap growth, before allocating n pages
   958  		// we need to sweep and reclaim at least n pages.
   959  		if !isSweepDone() {
   960  			h.reclaim(npages)
   961  		}
   962  		s = h.allocSpan(npages, spanAllocHeap, spanclass)
   963  	})
   964  	return s
   965  }
   966  
   967  // allocManual allocates a manually-managed span of npage pages.
   968  // allocManual returns nil if allocation fails.
   969  //
   970  // allocManual adds the bytes used to *stat, which should be a
   971  // memstats in-use field. Unlike allocations in the GC'd heap, the
   972  // allocation does *not* count toward heapInUse.
   973  //
   974  // The memory backing the returned span may not be zeroed if
   975  // span.needzero is set.
   976  //
   977  // allocManual must be called on the system stack because it may
   978  // acquire the heap lock via allocSpan. See mheap for details.
   979  //
   980  // If new code is written to call allocManual, do NOT use an
   981  // existing spanAllocType value and instead declare a new one.
   982  //
   983  //go:systemstack
   984  func (h *mheap) allocManual(npages uintptr, typ spanAllocType) *mspan {
   985  	if !typ.manual() {
   986  		throw("manual span allocation called with non-manually-managed type")
   987  	}
   988  	return h.allocSpan(npages, typ, 0)
   989  }
   990  
   991  // setSpans modifies the span map so [spanOf(base), spanOf(base+npage*pageSize))
   992  // is s.
   993  func (h *mheap) setSpans(base, npage uintptr, s *mspan) {
   994  	p := base / pageSize
   995  	ai := arenaIndex(base)
   996  	ha := h.arenas[ai.l1()][ai.l2()]
   997  	for n := uintptr(0); n < npage; n++ {
   998  		i := (p + n) % pagesPerArena
   999  		if i == 0 {
  1000  			ai = arenaIndex(base + n*pageSize)
  1001  			ha = h.arenas[ai.l1()][ai.l2()]
  1002  		}
  1003  		ha.spans[i] = s
  1004  	}
  1005  }
  1006  
  1007  // allocNeedsZero checks if the region of address space [base, base+npage*pageSize),
  1008  // assumed to be allocated, needs to be zeroed, updating heap arena metadata for
  1009  // future allocations.
  1010  //
  1011  // This must be called each time pages are allocated from the heap, even if the page
  1012  // allocator can otherwise prove the memory it's allocating is already zero because
  1013  // they're fresh from the operating system. It updates heapArena metadata that is
  1014  // critical for future page allocations.
  1015  //
  1016  // There are no locking constraints on this method.
  1017  func (h *mheap) allocNeedsZero(base, npage uintptr) (needZero bool) {
  1018  	for npage > 0 {
  1019  		ai := arenaIndex(base)
  1020  		ha := h.arenas[ai.l1()][ai.l2()]
  1021  
  1022  		zeroedBase := atomic.Loaduintptr(&ha.zeroedBase)
  1023  		arenaBase := base % heapArenaBytes
  1024  		if arenaBase < zeroedBase {
  1025  			// We extended into the non-zeroed part of the
  1026  			// arena, so this region needs to be zeroed before use.
  1027  			//
  1028  			// zeroedBase is monotonically increasing, so if we see this now then
  1029  			// we can be sure we need to zero this memory region.
  1030  			//
  1031  			// We still need to update zeroedBase for this arena, and
  1032  			// potentially more arenas.
  1033  			needZero = true
  1034  		}
  1035  		// We may observe arenaBase > zeroedBase if we're racing with one or more
  1036  		// allocations which are acquiring memory directly before us in the address
  1037  		// space. But, because we know no one else is acquiring *this* memory, it's
  1038  		// still safe to not zero.
  1039  
  1040  		// Compute how far into the arena we extend into, capped
  1041  		// at heapArenaBytes.
  1042  		arenaLimit := arenaBase + npage*pageSize
  1043  		if arenaLimit > heapArenaBytes {
  1044  			arenaLimit = heapArenaBytes
  1045  		}
  1046  		// Increase ha.zeroedBase so it's >= arenaLimit.
  1047  		// We may be racing with other updates.
  1048  		for arenaLimit > zeroedBase {
  1049  			if atomic.Casuintptr(&ha.zeroedBase, zeroedBase, arenaLimit) {
  1050  				break
  1051  			}
  1052  			zeroedBase = atomic.Loaduintptr(&ha.zeroedBase)
  1053  			// Double check basic conditions of zeroedBase.
  1054  			if zeroedBase <= arenaLimit && zeroedBase > arenaBase {
  1055  				// The zeroedBase moved into the space we were trying to
  1056  				// claim. That's very bad, and indicates someone allocated
  1057  				// the same region we did.
  1058  				throw("potentially overlapping in-use allocations detected")
  1059  			}
  1060  		}
  1061  
  1062  		// Move base forward and subtract from npage to move into
  1063  		// the next arena, or finish.
  1064  		base += arenaLimit - arenaBase
  1065  		npage -= (arenaLimit - arenaBase) / pageSize
  1066  	}
  1067  	return
  1068  }
  1069  
  1070  // tryAllocMSpan attempts to allocate an mspan object from
  1071  // the P-local cache, but may fail.
  1072  //
  1073  // h.lock need not be held.
  1074  //
  1075  // This caller must ensure that its P won't change underneath
  1076  // it during this function. Currently to ensure that we enforce
  1077  // that the function is run on the system stack, because that's
  1078  // the only place it is used now. In the future, this requirement
  1079  // may be relaxed if its use is necessary elsewhere.
  1080  //
  1081  //go:systemstack
  1082  func (h *mheap) tryAllocMSpan() *mspan {
  1083  	pp := getg().m.p.ptr()
  1084  	// If we don't have a p or the cache is empty, we can't do
  1085  	// anything here.
  1086  	if pp == nil || pp.mspancache.len == 0 {
  1087  		return nil
  1088  	}
  1089  	// Pull off the last entry in the cache.
  1090  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1091  	pp.mspancache.len--
  1092  	return s
  1093  }
  1094  
  1095  // allocMSpanLocked allocates an mspan object.
  1096  //
  1097  // h.lock must be held.
  1098  //
  1099  // allocMSpanLocked must be called on the system stack because
  1100  // its caller holds the heap lock. See mheap for details.
  1101  // Running on the system stack also ensures that we won't
  1102  // switch Ps during this function. See tryAllocMSpan for details.
  1103  //
  1104  //go:systemstack
  1105  func (h *mheap) allocMSpanLocked() *mspan {
  1106  	assertLockHeld(&h.lock)
  1107  
  1108  	pp := getg().m.p.ptr()
  1109  	if pp == nil {
  1110  		// We don't have a p so just do the normal thing.
  1111  		return (*mspan)(h.spanalloc.alloc())
  1112  	}
  1113  	// Refill the cache if necessary.
  1114  	if pp.mspancache.len == 0 {
  1115  		const refillCount = len(pp.mspancache.buf) / 2
  1116  		for i := 0; i < refillCount; i++ {
  1117  			pp.mspancache.buf[i] = (*mspan)(h.spanalloc.alloc())
  1118  		}
  1119  		pp.mspancache.len = refillCount
  1120  	}
  1121  	// Pull off the last entry in the cache.
  1122  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1123  	pp.mspancache.len--
  1124  	return s
  1125  }
  1126  
  1127  // freeMSpanLocked free an mspan object.
  1128  //
  1129  // h.lock must be held.
  1130  //
  1131  // freeMSpanLocked must be called on the system stack because
  1132  // its caller holds the heap lock. See mheap for details.
  1133  // Running on the system stack also ensures that we won't
  1134  // switch Ps during this function. See tryAllocMSpan for details.
  1135  //
  1136  //go:systemstack
  1137  func (h *mheap) freeMSpanLocked(s *mspan) {
  1138  	assertLockHeld(&h.lock)
  1139  
  1140  	pp := getg().m.p.ptr()
  1141  	// First try to free the mspan directly to the cache.
  1142  	if pp != nil && pp.mspancache.len < len(pp.mspancache.buf) {
  1143  		pp.mspancache.buf[pp.mspancache.len] = s
  1144  		pp.mspancache.len++
  1145  		return
  1146  	}
  1147  	// Failing that (or if we don't have a p), just free it to
  1148  	// the heap.
  1149  	h.spanalloc.free(unsafe.Pointer(s))
  1150  }
  1151  
  1152  // allocSpan allocates an mspan which owns npages worth of memory.
  1153  //
  1154  // If typ.manual() == false, allocSpan allocates a heap span of class spanclass
  1155  // and updates heap accounting. If manual == true, allocSpan allocates a
  1156  // manually-managed span (spanclass is ignored), and the caller is
  1157  // responsible for any accounting related to its use of the span. Either
  1158  // way, allocSpan will atomically add the bytes in the newly allocated
  1159  // span to *sysStat.
  1160  //
  1161  // The returned span is fully initialized.
  1162  //
  1163  // h.lock must not be held.
  1164  //
  1165  // allocSpan must be called on the system stack both because it acquires
  1166  // the heap lock and because it must block GC transitions.
  1167  //
  1168  //go:systemstack
  1169  func (h *mheap) allocSpan(npages uintptr, typ spanAllocType, spanclass spanClass) (s *mspan) {
  1170  	// Function-global state.
  1171  	gp := getg()
  1172  	base, scav := uintptr(0), uintptr(0)
  1173  	growth := uintptr(0)
  1174  
  1175  	// On some platforms we need to provide physical page aligned stack
  1176  	// allocations. Where the page size is less than the physical page
  1177  	// size, we already manage to do this by default.
  1178  	needPhysPageAlign := physPageAlignedStacks && typ == spanAllocStack && pageSize < physPageSize
  1179  
  1180  	// If the allocation is small enough, try the page cache!
  1181  	// The page cache does not support aligned allocations, so we cannot use
  1182  	// it if we need to provide a physical page aligned stack allocation.
  1183  	pp := gp.m.p.ptr()
  1184  	if !needPhysPageAlign && pp != nil && npages < pageCachePages/4 {
  1185  		c := &pp.pcache
  1186  
  1187  		// If the cache is empty, refill it.
  1188  		if c.empty() {
  1189  			lock(&h.lock)
  1190  			*c = h.pages.allocToCache()
  1191  			unlock(&h.lock)
  1192  		}
  1193  
  1194  		// Try to allocate from the cache.
  1195  		base, scav = c.alloc(npages)
  1196  		if base != 0 {
  1197  			s = h.tryAllocMSpan()
  1198  			if s != nil {
  1199  				goto HaveSpan
  1200  			}
  1201  			// We have a base but no mspan, so we need
  1202  			// to lock the heap.
  1203  		}
  1204  	}
  1205  
  1206  	// For one reason or another, we couldn't get the
  1207  	// whole job done without the heap lock.
  1208  	lock(&h.lock)
  1209  
  1210  	if needPhysPageAlign {
  1211  		// Overallocate by a physical page to allow for later alignment.
  1212  		extraPages := physPageSize / pageSize
  1213  
  1214  		// Find a big enough region first, but then only allocate the
  1215  		// aligned portion. We can't just allocate and then free the
  1216  		// edges because we need to account for scavenged memory, and
  1217  		// that's difficult with alloc.
  1218  		//
  1219  		// Note that we skip updates to searchAddr here. It's OK if
  1220  		// it's stale and higher than normal; it'll operate correctly,
  1221  		// just come with a performance cost.
  1222  		base, _ = h.pages.find(npages + extraPages)
  1223  		if base == 0 {
  1224  			var ok bool
  1225  			growth, ok = h.grow(npages + extraPages)
  1226  			if !ok {
  1227  				unlock(&h.lock)
  1228  				return nil
  1229  			}
  1230  			base, _ = h.pages.find(npages + extraPages)
  1231  			if base == 0 {
  1232  				throw("grew heap, but no adequate free space found")
  1233  			}
  1234  		}
  1235  		base = alignUp(base, physPageSize)
  1236  		scav = h.pages.allocRange(base, npages)
  1237  	}
  1238  
  1239  	if base == 0 {
  1240  		// Try to acquire a base address.
  1241  		base, scav = h.pages.alloc(npages)
  1242  		if base == 0 {
  1243  			var ok bool
  1244  			growth, ok = h.grow(npages)
  1245  			if !ok {
  1246  				unlock(&h.lock)
  1247  				return nil
  1248  			}
  1249  			base, scav = h.pages.alloc(npages)
  1250  			if base == 0 {
  1251  				throw("grew heap, but no adequate free space found")
  1252  			}
  1253  		}
  1254  	}
  1255  	if s == nil {
  1256  		// We failed to get an mspan earlier, so grab
  1257  		// one now that we have the heap lock.
  1258  		s = h.allocMSpanLocked()
  1259  	}
  1260  	unlock(&h.lock)
  1261  
  1262  HaveSpan:
  1263  	// Decide if we need to scavenge in response to what we just allocated.
  1264  	// Specifically, we track the maximum amount of memory to scavenge of all
  1265  	// the alternatives below, assuming that the maximum satisfies *all*
  1266  	// conditions we check (e.g. if we need to scavenge X to satisfy the
  1267  	// memory limit and Y to satisfy heap-growth scavenging, and Y > X, then
  1268  	// it's fine to pick Y, because the memory limit is still satisfied).
  1269  	//
  1270  	// It's fine to do this after allocating because we expect any scavenged
  1271  	// pages not to get touched until we return. Simultaneously, it's important
  1272  	// to do this before calling sysUsed because that may commit address space.
  1273  	bytesToScavenge := uintptr(0)
  1274  	forceScavenge := false
  1275  	if limit := gcController.memoryLimit.Load(); !gcCPULimiter.limiting() {
  1276  		// Assist with scavenging to maintain the memory limit by the amount
  1277  		// that we expect to page in.
  1278  		inuse := gcController.mappedReady.Load()
  1279  		// Be careful about overflow, especially with uintptrs. Even on 32-bit platforms
  1280  		// someone can set a really big memory limit that isn't maxInt64.
  1281  		if uint64(scav)+inuse > uint64(limit) {
  1282  			bytesToScavenge = uintptr(uint64(scav) + inuse - uint64(limit))
  1283  			forceScavenge = true
  1284  		}
  1285  	}
  1286  	if goal := scavenge.gcPercentGoal.Load(); goal != ^uint64(0) && growth > 0 {
  1287  		// We just caused a heap growth, so scavenge down what will soon be used.
  1288  		// By scavenging inline we deal with the failure to allocate out of
  1289  		// memory fragments by scavenging the memory fragments that are least
  1290  		// likely to be re-used.
  1291  		//
  1292  		// Only bother with this because we're not using a memory limit. We don't
  1293  		// care about heap growths as long as we're under the memory limit, and the
  1294  		// previous check for scaving already handles that.
  1295  		if retained := heapRetained(); retained+uint64(growth) > goal {
  1296  			// The scavenging algorithm requires the heap lock to be dropped so it
  1297  			// can acquire it only sparingly. This is a potentially expensive operation
  1298  			// so it frees up other goroutines to allocate in the meanwhile. In fact,
  1299  			// they can make use of the growth we just created.
  1300  			todo := growth
  1301  			if overage := uintptr(retained + uint64(growth) - goal); todo > overage {
  1302  				todo = overage
  1303  			}
  1304  			if todo > bytesToScavenge {
  1305  				bytesToScavenge = todo
  1306  			}
  1307  		}
  1308  	}
  1309  	// There are a few very limited circumstances where we won't have a P here.
  1310  	// It's OK to simply skip scavenging in these cases. Something else will notice
  1311  	// and pick up the tab.
  1312  	var now int64
  1313  	if pp != nil && bytesToScavenge > 0 {
  1314  		// Measure how long we spent scavenging and add that measurement to the assist
  1315  		// time so we can track it for the GC CPU limiter.
  1316  		//
  1317  		// Limiter event tracking might be disabled if we end up here
  1318  		// while on a mark worker.
  1319  		start := nanotime()
  1320  		track := pp.limiterEvent.start(limiterEventScavengeAssist, start)
  1321  
  1322  		// Scavenge, but back out if the limiter turns on.
  1323  		released := h.pages.scavenge(bytesToScavenge, func() bool {
  1324  			return gcCPULimiter.limiting()
  1325  		}, forceScavenge)
  1326  
  1327  		mheap_.pages.scav.releasedEager.Add(released)
  1328  
  1329  		// Finish up accounting.
  1330  		now = nanotime()
  1331  		if track {
  1332  			pp.limiterEvent.stop(limiterEventScavengeAssist, now)
  1333  		}
  1334  		scavenge.assistTime.Add(now - start)
  1335  	}
  1336  
  1337  	// Initialize the span.
  1338  	h.initSpan(s, typ, spanclass, base, npages)
  1339  
  1340  	// Commit and account for any scavenged memory that the span now owns.
  1341  	nbytes := npages * pageSize
  1342  	if scav != 0 {
  1343  		// sysUsed all the pages that are actually available
  1344  		// in the span since some of them might be scavenged.
  1345  		sysUsed(unsafe.Pointer(base), nbytes, scav)
  1346  		gcController.heapReleased.add(-int64(scav))
  1347  	}
  1348  	// Update stats.
  1349  	gcController.heapFree.add(-int64(nbytes - scav))
  1350  	if typ == spanAllocHeap {
  1351  		gcController.heapInUse.add(int64(nbytes))
  1352  	}
  1353  	// Update consistent stats.
  1354  	stats := memstats.heapStats.acquire()
  1355  	atomic.Xaddint64(&stats.committed, int64(scav))
  1356  	atomic.Xaddint64(&stats.released, -int64(scav))
  1357  	switch typ {
  1358  	case spanAllocHeap:
  1359  		atomic.Xaddint64(&stats.inHeap, int64(nbytes))
  1360  	case spanAllocStack:
  1361  		atomic.Xaddint64(&stats.inStacks, int64(nbytes))
  1362  	case spanAllocPtrScalarBits:
  1363  		atomic.Xaddint64(&stats.inPtrScalarBits, int64(nbytes))
  1364  	case spanAllocWorkBuf:
  1365  		atomic.Xaddint64(&stats.inWorkBufs, int64(nbytes))
  1366  	}
  1367  	memstats.heapStats.release()
  1368  
  1369  	// Trace the span alloc.
  1370  	if traceAllocFreeEnabled() {
  1371  		trace := traceTryAcquire()
  1372  		if trace.ok() {
  1373  			trace.SpanAlloc(s)
  1374  			traceRelease(trace)
  1375  		}
  1376  	}
  1377  	return s
  1378  }
  1379  
  1380  // initSpan initializes a blank span s which will represent the range
  1381  // [base, base+npages*pageSize). typ is the type of span being allocated.
  1382  func (h *mheap) initSpan(s *mspan, typ spanAllocType, spanclass spanClass, base, npages uintptr) {
  1383  	// At this point, both s != nil and base != 0, and the heap
  1384  	// lock is no longer held. Initialize the span.
  1385  	s.init(base, npages)
  1386  	if h.allocNeedsZero(base, npages) {
  1387  		s.needzero = 1
  1388  	}
  1389  	nbytes := npages * pageSize
  1390  	if typ.manual() {
  1391  		s.manualFreeList = 0
  1392  		s.nelems = 0
  1393  		s.limit = s.base() + s.npages*pageSize
  1394  		s.state.set(mSpanManual)
  1395  	} else {
  1396  		// We must set span properties before the span is published anywhere
  1397  		// since we're not holding the heap lock.
  1398  		s.spanclass = spanclass
  1399  		if sizeclass := spanclass.sizeclass(); sizeclass == 0 {
  1400  			s.elemsize = nbytes
  1401  			s.nelems = 1
  1402  			s.divMul = 0
  1403  		} else {
  1404  			s.elemsize = uintptr(class_to_size[sizeclass])
  1405  			if !s.spanclass.noscan() && heapBitsInSpan(s.elemsize) {
  1406  				// Reserve space for the pointer/scan bitmap at the end.
  1407  				s.nelems = uint16((nbytes - (nbytes / goarch.PtrSize / 8)) / s.elemsize)
  1408  			} else {
  1409  				s.nelems = uint16(nbytes / s.elemsize)
  1410  			}
  1411  			s.divMul = class_to_divmagic[sizeclass]
  1412  		}
  1413  
  1414  		// Initialize mark and allocation structures.
  1415  		s.freeindex = 0
  1416  		s.freeIndexForScan = 0
  1417  		s.allocCache = ^uint64(0) // all 1s indicating all free.
  1418  		s.gcmarkBits = newMarkBits(uintptr(s.nelems))
  1419  		s.allocBits = newAllocBits(uintptr(s.nelems))
  1420  
  1421  		// It's safe to access h.sweepgen without the heap lock because it's
  1422  		// only ever updated with the world stopped and we run on the
  1423  		// systemstack which blocks a STW transition.
  1424  		atomic.Store(&s.sweepgen, h.sweepgen)
  1425  
  1426  		// Now that the span is filled in, set its state. This
  1427  		// is a publication barrier for the other fields in
  1428  		// the span. While valid pointers into this span
  1429  		// should never be visible until the span is returned,
  1430  		// if the garbage collector finds an invalid pointer,
  1431  		// access to the span may race with initialization of
  1432  		// the span. We resolve this race by atomically
  1433  		// setting the state after the span is fully
  1434  		// initialized, and atomically checking the state in
  1435  		// any situation where a pointer is suspect.
  1436  		s.state.set(mSpanInUse)
  1437  	}
  1438  
  1439  	// Publish the span in various locations.
  1440  
  1441  	// This is safe to call without the lock held because the slots
  1442  	// related to this span will only ever be read or modified by
  1443  	// this thread until pointers into the span are published (and
  1444  	// we execute a publication barrier at the end of this function
  1445  	// before that happens) or pageInUse is updated.
  1446  	h.setSpans(s.base(), npages, s)
  1447  
  1448  	if !typ.manual() {
  1449  		// Mark in-use span in arena page bitmap.
  1450  		//
  1451  		// This publishes the span to the page sweeper, so
  1452  		// it's imperative that the span be completely initialized
  1453  		// prior to this line.
  1454  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1455  		atomic.Or8(&arena.pageInUse[pageIdx], pageMask)
  1456  
  1457  		// Update related page sweeper stats.
  1458  		h.pagesInUse.Add(npages)
  1459  	}
  1460  
  1461  	// Make sure the newly allocated span will be observed
  1462  	// by the GC before pointers into the span are published.
  1463  	publicationBarrier()
  1464  }
  1465  
  1466  // Try to add at least npage pages of memory to the heap,
  1467  // returning how much the heap grew by and whether it worked.
  1468  //
  1469  // h.lock must be held.
  1470  func (h *mheap) grow(npage uintptr) (uintptr, bool) {
  1471  	assertLockHeld(&h.lock)
  1472  
  1473  	// We must grow the heap in whole palloc chunks.
  1474  	// We call sysMap below but note that because we
  1475  	// round up to pallocChunkPages which is on the order
  1476  	// of MiB (generally >= to the huge page size) we
  1477  	// won't be calling it too much.
  1478  	ask := alignUp(npage, pallocChunkPages) * pageSize
  1479  
  1480  	totalGrowth := uintptr(0)
  1481  	// This may overflow because ask could be very large
  1482  	// and is otherwise unrelated to h.curArena.base.
  1483  	end := h.curArena.base + ask
  1484  	nBase := alignUp(end, physPageSize)
  1485  	if nBase > h.curArena.end || /* overflow */ end < h.curArena.base {
  1486  		// Not enough room in the current arena. Allocate more
  1487  		// arena space. This may not be contiguous with the
  1488  		// current arena, so we have to request the full ask.
  1489  		av, asize := h.sysAlloc(ask, &h.arenaHints, true)
  1490  		if av == nil {
  1491  			inUse := gcController.heapFree.load() + gcController.heapReleased.load() + gcController.heapInUse.load()
  1492  			print("runtime: out of memory: cannot allocate ", ask, "-byte block (", inUse, " in use)\n")
  1493  			return 0, false
  1494  		}
  1495  
  1496  		if uintptr(av) == h.curArena.end {
  1497  			// The new space is contiguous with the old
  1498  			// space, so just extend the current space.
  1499  			h.curArena.end = uintptr(av) + asize
  1500  		} else {
  1501  			// The new space is discontiguous. Track what
  1502  			// remains of the current space and switch to
  1503  			// the new space. This should be rare.
  1504  			if size := h.curArena.end - h.curArena.base; size != 0 {
  1505  				// Transition this space from Reserved to Prepared and mark it
  1506  				// as released since we'll be able to start using it after updating
  1507  				// the page allocator and releasing the lock at any time.
  1508  				sysMap(unsafe.Pointer(h.curArena.base), size, &gcController.heapReleased)
  1509  				// Update stats.
  1510  				stats := memstats.heapStats.acquire()
  1511  				atomic.Xaddint64(&stats.released, int64(size))
  1512  				memstats.heapStats.release()
  1513  				// Update the page allocator's structures to make this
  1514  				// space ready for allocation.
  1515  				h.pages.grow(h.curArena.base, size)
  1516  				totalGrowth += size
  1517  			}
  1518  			// Switch to the new space.
  1519  			h.curArena.base = uintptr(av)
  1520  			h.curArena.end = uintptr(av) + asize
  1521  		}
  1522  
  1523  		// Recalculate nBase.
  1524  		// We know this won't overflow, because sysAlloc returned
  1525  		// a valid region starting at h.curArena.base which is at
  1526  		// least ask bytes in size.
  1527  		nBase = alignUp(h.curArena.base+ask, physPageSize)
  1528  	}
  1529  
  1530  	// Grow into the current arena.
  1531  	v := h.curArena.base
  1532  	h.curArena.base = nBase
  1533  
  1534  	// Transition the space we're going to use from Reserved to Prepared.
  1535  	//
  1536  	// The allocation is always aligned to the heap arena
  1537  	// size which is always > physPageSize, so its safe to
  1538  	// just add directly to heapReleased.
  1539  	sysMap(unsafe.Pointer(v), nBase-v, &gcController.heapReleased)
  1540  
  1541  	// The memory just allocated counts as both released
  1542  	// and idle, even though it's not yet backed by spans.
  1543  	stats := memstats.heapStats.acquire()
  1544  	atomic.Xaddint64(&stats.released, int64(nBase-v))
  1545  	memstats.heapStats.release()
  1546  
  1547  	// Update the page allocator's structures to make this
  1548  	// space ready for allocation.
  1549  	h.pages.grow(v, nBase-v)
  1550  	totalGrowth += nBase - v
  1551  	return totalGrowth, true
  1552  }
  1553  
  1554  // Free the span back into the heap.
  1555  func (h *mheap) freeSpan(s *mspan) {
  1556  	systemstack(func() {
  1557  		// Trace the span free.
  1558  		if traceAllocFreeEnabled() {
  1559  			trace := traceTryAcquire()
  1560  			if trace.ok() {
  1561  				trace.SpanFree(s)
  1562  				traceRelease(trace)
  1563  			}
  1564  		}
  1565  
  1566  		lock(&h.lock)
  1567  		if msanenabled {
  1568  			// Tell msan that this entire span is no longer in use.
  1569  			base := unsafe.Pointer(s.base())
  1570  			bytes := s.npages << _PageShift
  1571  			msanfree(base, bytes)
  1572  		}
  1573  		if asanenabled {
  1574  			// Tell asan that this entire span is no longer in use.
  1575  			base := unsafe.Pointer(s.base())
  1576  			bytes := s.npages << _PageShift
  1577  			asanpoison(base, bytes)
  1578  		}
  1579  		h.freeSpanLocked(s, spanAllocHeap)
  1580  		unlock(&h.lock)
  1581  	})
  1582  }
  1583  
  1584  // freeManual frees a manually-managed span returned by allocManual.
  1585  // typ must be the same as the spanAllocType passed to the allocManual that
  1586  // allocated s.
  1587  //
  1588  // This must only be called when gcphase == _GCoff. See mSpanState for
  1589  // an explanation.
  1590  //
  1591  // freeManual must be called on the system stack because it acquires
  1592  // the heap lock. See mheap for details.
  1593  //
  1594  //go:systemstack
  1595  func (h *mheap) freeManual(s *mspan, typ spanAllocType) {
  1596  	// Trace the span free.
  1597  	if traceAllocFreeEnabled() {
  1598  		trace := traceTryAcquire()
  1599  		if trace.ok() {
  1600  			trace.SpanFree(s)
  1601  			traceRelease(trace)
  1602  		}
  1603  	}
  1604  
  1605  	s.needzero = 1
  1606  	lock(&h.lock)
  1607  	h.freeSpanLocked(s, typ)
  1608  	unlock(&h.lock)
  1609  }
  1610  
  1611  func (h *mheap) freeSpanLocked(s *mspan, typ spanAllocType) {
  1612  	assertLockHeld(&h.lock)
  1613  
  1614  	switch s.state.get() {
  1615  	case mSpanManual:
  1616  		if s.allocCount != 0 {
  1617  			throw("mheap.freeSpanLocked - invalid stack free")
  1618  		}
  1619  	case mSpanInUse:
  1620  		if s.isUserArenaChunk {
  1621  			throw("mheap.freeSpanLocked - invalid free of user arena chunk")
  1622  		}
  1623  		if s.allocCount != 0 || s.sweepgen != h.sweepgen {
  1624  			print("mheap.freeSpanLocked - span ", s, " ptr ", hex(s.base()), " allocCount ", s.allocCount, " sweepgen ", s.sweepgen, "/", h.sweepgen, "\n")
  1625  			throw("mheap.freeSpanLocked - invalid free")
  1626  		}
  1627  		h.pagesInUse.Add(-s.npages)
  1628  
  1629  		// Clear in-use bit in arena page bitmap.
  1630  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1631  		atomic.And8(&arena.pageInUse[pageIdx], ^pageMask)
  1632  	default:
  1633  		throw("mheap.freeSpanLocked - invalid span state")
  1634  	}
  1635  
  1636  	// Update stats.
  1637  	//
  1638  	// Mirrors the code in allocSpan.
  1639  	nbytes := s.npages * pageSize
  1640  	gcController.heapFree.add(int64(nbytes))
  1641  	if typ == spanAllocHeap {
  1642  		gcController.heapInUse.add(-int64(nbytes))
  1643  	}
  1644  	// Update consistent stats.
  1645  	stats := memstats.heapStats.acquire()
  1646  	switch typ {
  1647  	case spanAllocHeap:
  1648  		atomic.Xaddint64(&stats.inHeap, -int64(nbytes))
  1649  	case spanAllocStack:
  1650  		atomic.Xaddint64(&stats.inStacks, -int64(nbytes))
  1651  	case spanAllocPtrScalarBits:
  1652  		atomic.Xaddint64(&stats.inPtrScalarBits, -int64(nbytes))
  1653  	case spanAllocWorkBuf:
  1654  		atomic.Xaddint64(&stats.inWorkBufs, -int64(nbytes))
  1655  	}
  1656  	memstats.heapStats.release()
  1657  
  1658  	// Mark the space as free.
  1659  	h.pages.free(s.base(), s.npages)
  1660  
  1661  	// Free the span structure. We no longer have a use for it.
  1662  	s.state.set(mSpanDead)
  1663  	h.freeMSpanLocked(s)
  1664  }
  1665  
  1666  // scavengeAll acquires the heap lock (blocking any additional
  1667  // manipulation of the page allocator) and iterates over the whole
  1668  // heap, scavenging every free page available.
  1669  //
  1670  // Must run on the system stack because it acquires the heap lock.
  1671  //
  1672  //go:systemstack
  1673  func (h *mheap) scavengeAll() {
  1674  	// Disallow malloc or panic while holding the heap lock. We do
  1675  	// this here because this is a non-mallocgc entry-point to
  1676  	// the mheap API.
  1677  	gp := getg()
  1678  	gp.m.mallocing++
  1679  
  1680  	// Force scavenge everything.
  1681  	released := h.pages.scavenge(^uintptr(0), nil, true)
  1682  
  1683  	gp.m.mallocing--
  1684  
  1685  	if debug.scavtrace > 0 {
  1686  		printScavTrace(0, released, true)
  1687  	}
  1688  }
  1689  
  1690  //go:linkname runtime_debug_freeOSMemory runtime/debug.freeOSMemory
  1691  func runtime_debug_freeOSMemory() {
  1692  	GC()
  1693  	systemstack(func() { mheap_.scavengeAll() })
  1694  }
  1695  
  1696  // Initialize a new span with the given start and npages.
  1697  func (span *mspan) init(base uintptr, npages uintptr) {
  1698  	// span is *not* zeroed.
  1699  	span.next = nil
  1700  	span.prev = nil
  1701  	span.list = nil
  1702  	span.startAddr = base
  1703  	span.npages = npages
  1704  	span.allocCount = 0
  1705  	span.spanclass = 0
  1706  	span.elemsize = 0
  1707  	span.speciallock.key = 0
  1708  	span.specials = nil
  1709  	span.needzero = 0
  1710  	span.freeindex = 0
  1711  	span.freeIndexForScan = 0
  1712  	span.allocBits = nil
  1713  	span.gcmarkBits = nil
  1714  	span.pinnerBits = nil
  1715  	span.state.set(mSpanDead)
  1716  	lockInit(&span.speciallock, lockRankMspanSpecial)
  1717  }
  1718  
  1719  func (span *mspan) inList() bool {
  1720  	return span.list != nil
  1721  }
  1722  
  1723  // Initialize an empty doubly-linked list.
  1724  func (list *mSpanList) init() {
  1725  	list.first = nil
  1726  	list.last = nil
  1727  }
  1728  
  1729  func (list *mSpanList) remove(span *mspan) {
  1730  	if span.list != list {
  1731  		print("runtime: failed mSpanList.remove span.npages=", span.npages,
  1732  			" span=", span, " prev=", span.prev, " span.list=", span.list, " list=", list, "\n")
  1733  		throw("mSpanList.remove")
  1734  	}
  1735  	if list.first == span {
  1736  		list.first = span.next
  1737  	} else {
  1738  		span.prev.next = span.next
  1739  	}
  1740  	if list.last == span {
  1741  		list.last = span.prev
  1742  	} else {
  1743  		span.next.prev = span.prev
  1744  	}
  1745  	span.next = nil
  1746  	span.prev = nil
  1747  	span.list = nil
  1748  }
  1749  
  1750  func (list *mSpanList) isEmpty() bool {
  1751  	return list.first == nil
  1752  }
  1753  
  1754  func (list *mSpanList) insert(span *mspan) {
  1755  	if span.next != nil || span.prev != nil || span.list != nil {
  1756  		println("runtime: failed mSpanList.insert", span, span.next, span.prev, span.list)
  1757  		throw("mSpanList.insert")
  1758  	}
  1759  	span.next = list.first
  1760  	if list.first != nil {
  1761  		// The list contains at least one span; link it in.
  1762  		// The last span in the list doesn't change.
  1763  		list.first.prev = span
  1764  	} else {
  1765  		// The list contains no spans, so this is also the last span.
  1766  		list.last = span
  1767  	}
  1768  	list.first = span
  1769  	span.list = list
  1770  }
  1771  
  1772  func (list *mSpanList) insertBack(span *mspan) {
  1773  	if span.next != nil || span.prev != nil || span.list != nil {
  1774  		println("runtime: failed mSpanList.insertBack", span, span.next, span.prev, span.list)
  1775  		throw("mSpanList.insertBack")
  1776  	}
  1777  	span.prev = list.last
  1778  	if list.last != nil {
  1779  		// The list contains at least one span.
  1780  		list.last.next = span
  1781  	} else {
  1782  		// The list contains no spans, so this is also the first span.
  1783  		list.first = span
  1784  	}
  1785  	list.last = span
  1786  	span.list = list
  1787  }
  1788  
  1789  // takeAll removes all spans from other and inserts them at the front
  1790  // of list.
  1791  func (list *mSpanList) takeAll(other *mSpanList) {
  1792  	if other.isEmpty() {
  1793  		return
  1794  	}
  1795  
  1796  	// Reparent everything in other to list.
  1797  	for s := other.first; s != nil; s = s.next {
  1798  		s.list = list
  1799  	}
  1800  
  1801  	// Concatenate the lists.
  1802  	if list.isEmpty() {
  1803  		*list = *other
  1804  	} else {
  1805  		// Neither list is empty. Put other before list.
  1806  		other.last.next = list.first
  1807  		list.first.prev = other.last
  1808  		list.first = other.first
  1809  	}
  1810  
  1811  	other.first, other.last = nil, nil
  1812  }
  1813  
  1814  const (
  1815  	// _KindSpecialFinalizer is for tracking finalizers.
  1816  	_KindSpecialFinalizer = 1
  1817  	// _KindSpecialWeakHandle is used for creating weak pointers.
  1818  	_KindSpecialWeakHandle = 2
  1819  	// _KindSpecialProfile is for memory profiling.
  1820  	_KindSpecialProfile = 3
  1821  	// _KindSpecialReachable is a special used for tracking
  1822  	// reachability during testing.
  1823  	_KindSpecialReachable = 4
  1824  	// _KindSpecialPinCounter is a special used for objects that are pinned
  1825  	// multiple times
  1826  	_KindSpecialPinCounter = 5
  1827  )
  1828  
  1829  type special struct {
  1830  	_      sys.NotInHeap
  1831  	next   *special // linked list in span
  1832  	offset uint16   // span offset of object
  1833  	kind   byte     // kind of special
  1834  }
  1835  
  1836  // spanHasSpecials marks a span as having specials in the arena bitmap.
  1837  func spanHasSpecials(s *mspan) {
  1838  	arenaPage := (s.base() / pageSize) % pagesPerArena
  1839  	ai := arenaIndex(s.base())
  1840  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  1841  	atomic.Or8(&ha.pageSpecials[arenaPage/8], uint8(1)<<(arenaPage%8))
  1842  }
  1843  
  1844  // spanHasNoSpecials marks a span as having no specials in the arena bitmap.
  1845  func spanHasNoSpecials(s *mspan) {
  1846  	arenaPage := (s.base() / pageSize) % pagesPerArena
  1847  	ai := arenaIndex(s.base())
  1848  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  1849  	atomic.And8(&ha.pageSpecials[arenaPage/8], ^(uint8(1) << (arenaPage % 8)))
  1850  }
  1851  
  1852  // Adds the special record s to the list of special records for
  1853  // the object p. All fields of s should be filled in except for
  1854  // offset & next, which this routine will fill in.
  1855  // Returns true if the special was successfully added, false otherwise.
  1856  // (The add will fail only if a record with the same p and s->kind
  1857  // already exists.)
  1858  func addspecial(p unsafe.Pointer, s *special) bool {
  1859  	span := spanOfHeap(uintptr(p))
  1860  	if span == nil {
  1861  		throw("addspecial on invalid pointer")
  1862  	}
  1863  
  1864  	// Ensure that the span is swept.
  1865  	// Sweeping accesses the specials list w/o locks, so we have
  1866  	// to synchronize with it. And it's just much safer.
  1867  	mp := acquirem()
  1868  	span.ensureSwept()
  1869  
  1870  	offset := uintptr(p) - span.base()
  1871  	kind := s.kind
  1872  
  1873  	lock(&span.speciallock)
  1874  
  1875  	// Find splice point, check for existing record.
  1876  	iter, exists := span.specialFindSplicePoint(offset, kind)
  1877  	if !exists {
  1878  		// Splice in record, fill in offset.
  1879  		s.offset = uint16(offset)
  1880  		s.next = *iter
  1881  		*iter = s
  1882  		spanHasSpecials(span)
  1883  	}
  1884  
  1885  	unlock(&span.speciallock)
  1886  	releasem(mp)
  1887  	return !exists // already exists
  1888  }
  1889  
  1890  // Removes the Special record of the given kind for the object p.
  1891  // Returns the record if the record existed, nil otherwise.
  1892  // The caller must FixAlloc_Free the result.
  1893  func removespecial(p unsafe.Pointer, kind uint8) *special {
  1894  	span := spanOfHeap(uintptr(p))
  1895  	if span == nil {
  1896  		throw("removespecial on invalid pointer")
  1897  	}
  1898  
  1899  	// Ensure that the span is swept.
  1900  	// Sweeping accesses the specials list w/o locks, so we have
  1901  	// to synchronize with it. And it's just much safer.
  1902  	mp := acquirem()
  1903  	span.ensureSwept()
  1904  
  1905  	offset := uintptr(p) - span.base()
  1906  
  1907  	var result *special
  1908  	lock(&span.speciallock)
  1909  
  1910  	iter, exists := span.specialFindSplicePoint(offset, kind)
  1911  	if exists {
  1912  		s := *iter
  1913  		*iter = s.next
  1914  		result = s
  1915  	}
  1916  	if span.specials == nil {
  1917  		spanHasNoSpecials(span)
  1918  	}
  1919  	unlock(&span.speciallock)
  1920  	releasem(mp)
  1921  	return result
  1922  }
  1923  
  1924  // Find a splice point in the sorted list and check for an already existing
  1925  // record. Returns a pointer to the next-reference in the list predecessor.
  1926  // Returns true, if the referenced item is an exact match.
  1927  func (span *mspan) specialFindSplicePoint(offset uintptr, kind byte) (**special, bool) {
  1928  	// Find splice point, check for existing record.
  1929  	iter := &span.specials
  1930  	found := false
  1931  	for {
  1932  		s := *iter
  1933  		if s == nil {
  1934  			break
  1935  		}
  1936  		if offset == uintptr(s.offset) && kind == s.kind {
  1937  			found = true
  1938  			break
  1939  		}
  1940  		if offset < uintptr(s.offset) || (offset == uintptr(s.offset) && kind < s.kind) {
  1941  			break
  1942  		}
  1943  		iter = &s.next
  1944  	}
  1945  	return iter, found
  1946  }
  1947  
  1948  // The described object has a finalizer set for it.
  1949  //
  1950  // specialfinalizer is allocated from non-GC'd memory, so any heap
  1951  // pointers must be specially handled.
  1952  type specialfinalizer struct {
  1953  	_       sys.NotInHeap
  1954  	special special
  1955  	fn      *funcval // May be a heap pointer.
  1956  	nret    uintptr
  1957  	fint    *_type   // May be a heap pointer, but always live.
  1958  	ot      *ptrtype // May be a heap pointer, but always live.
  1959  }
  1960  
  1961  // Adds a finalizer to the object p. Returns true if it succeeded.
  1962  func addfinalizer(p unsafe.Pointer, f *funcval, nret uintptr, fint *_type, ot *ptrtype) bool {
  1963  	lock(&mheap_.speciallock)
  1964  	s := (*specialfinalizer)(mheap_.specialfinalizeralloc.alloc())
  1965  	unlock(&mheap_.speciallock)
  1966  	s.special.kind = _KindSpecialFinalizer
  1967  	s.fn = f
  1968  	s.nret = nret
  1969  	s.fint = fint
  1970  	s.ot = ot
  1971  	if addspecial(p, &s.special) {
  1972  		// This is responsible for maintaining the same
  1973  		// GC-related invariants as markrootSpans in any
  1974  		// situation where it's possible that markrootSpans
  1975  		// has already run but mark termination hasn't yet.
  1976  		if gcphase != _GCoff {
  1977  			base, span, _ := findObject(uintptr(p), 0, 0)
  1978  			mp := acquirem()
  1979  			gcw := &mp.p.ptr().gcw
  1980  			// Mark everything reachable from the object
  1981  			// so it's retained for the finalizer.
  1982  			if !span.spanclass.noscan() {
  1983  				scanobject(base, gcw)
  1984  			}
  1985  			// Mark the finalizer itself, since the
  1986  			// special isn't part of the GC'd heap.
  1987  			scanblock(uintptr(unsafe.Pointer(&s.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  1988  			releasem(mp)
  1989  		}
  1990  		return true
  1991  	}
  1992  
  1993  	// There was an old finalizer
  1994  	lock(&mheap_.speciallock)
  1995  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  1996  	unlock(&mheap_.speciallock)
  1997  	return false
  1998  }
  1999  
  2000  // Removes the finalizer (if any) from the object p.
  2001  func removefinalizer(p unsafe.Pointer) {
  2002  	s := (*specialfinalizer)(unsafe.Pointer(removespecial(p, _KindSpecialFinalizer)))
  2003  	if s == nil {
  2004  		return // there wasn't a finalizer to remove
  2005  	}
  2006  	lock(&mheap_.speciallock)
  2007  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  2008  	unlock(&mheap_.speciallock)
  2009  }
  2010  
  2011  // The described object has a weak pointer.
  2012  //
  2013  // Weak pointers in the GC have the following invariants:
  2014  //
  2015  //   - Strong-to-weak conversions must ensure the strong pointer
  2016  //     remains live until the weak handle is installed. This ensures
  2017  //     that creating a weak pointer cannot fail.
  2018  //
  2019  //   - Weak-to-strong conversions require the weakly-referenced
  2020  //     object to be swept before the conversion may proceed. This
  2021  //     ensures that weak-to-strong conversions cannot resurrect
  2022  //     dead objects by sweeping them before that happens.
  2023  //
  2024  //   - Weak handles are unique and canonical for each byte offset into
  2025  //     an object that a strong pointer may point to, until an object
  2026  //     becomes unreachable.
  2027  //
  2028  //   - Weak handles contain nil as soon as an object becomes unreachable
  2029  //     the first time, before a finalizer makes it reachable again. New
  2030  //     weak handles created after resurrection are newly unique.
  2031  //
  2032  // specialWeakHandle is allocated from non-GC'd memory, so any heap
  2033  // pointers must be specially handled.
  2034  type specialWeakHandle struct {
  2035  	_       sys.NotInHeap
  2036  	special special
  2037  	// handle is a reference to the actual weak pointer.
  2038  	// It is always heap-allocated and must be explicitly kept
  2039  	// live so long as this special exists.
  2040  	handle *atomic.Uintptr
  2041  }
  2042  
  2043  //go:linkname internal_weak_runtime_registerWeakPointer internal/weak.runtime_registerWeakPointer
  2044  func internal_weak_runtime_registerWeakPointer(p unsafe.Pointer) unsafe.Pointer {
  2045  	return unsafe.Pointer(getOrAddWeakHandle(unsafe.Pointer(p)))
  2046  }
  2047  
  2048  //go:linkname internal_weak_runtime_makeStrongFromWeak internal/weak.runtime_makeStrongFromWeak
  2049  func internal_weak_runtime_makeStrongFromWeak(u unsafe.Pointer) unsafe.Pointer {
  2050  	handle := (*atomic.Uintptr)(u)
  2051  
  2052  	// Prevent preemption. We want to make sure that another GC cycle can't start
  2053  	// and that work.strongFromWeak.block can't change out from under us.
  2054  	mp := acquirem()
  2055  
  2056  	// Yield to the GC if necessary.
  2057  	if work.strongFromWeak.block {
  2058  		releasem(mp)
  2059  
  2060  		// Try to park and wait for mark termination.
  2061  		// N.B. gcParkStrongFromWeak calls acquirem before returning.
  2062  		mp = gcParkStrongFromWeak()
  2063  	}
  2064  
  2065  	p := handle.Load()
  2066  	if p == 0 {
  2067  		releasem(mp)
  2068  		return nil
  2069  	}
  2070  	// Be careful. p may or may not refer to valid memory anymore, as it could've been
  2071  	// swept and released already. It's always safe to ensure a span is swept, though,
  2072  	// even if it's just some random span.
  2073  	span := spanOfHeap(p)
  2074  	if span == nil {
  2075  		// The span probably got swept and released.
  2076  		releasem(mp)
  2077  		return nil
  2078  	}
  2079  	// Ensure the span is swept.
  2080  	span.ensureSwept()
  2081  
  2082  	// Now we can trust whatever we get from handle, so make a strong pointer.
  2083  	//
  2084  	// Even if we just swept some random span that doesn't contain this object, because
  2085  	// this object is long dead and its memory has since been reused, we'll just observe nil.
  2086  	ptr := unsafe.Pointer(handle.Load())
  2087  
  2088  	// This is responsible for maintaining the same GC-related
  2089  	// invariants as the Yuasa part of the write barrier. During
  2090  	// the mark phase, it's possible that we just created the only
  2091  	// valid pointer to the object pointed to by ptr. If it's only
  2092  	// ever referenced from our stack, and our stack is blackened
  2093  	// already, we could fail to mark it. So, mark it now.
  2094  	if gcphase != _GCoff {
  2095  		shade(uintptr(ptr))
  2096  	}
  2097  	releasem(mp)
  2098  
  2099  	// Explicitly keep ptr alive. This seems unnecessary since we return ptr,
  2100  	// but let's be explicit since it's important we keep ptr alive across the
  2101  	// call to shade.
  2102  	KeepAlive(ptr)
  2103  	return ptr
  2104  }
  2105  
  2106  // gcParkStrongFromWeak puts the current goroutine on the weak->strong queue and parks.
  2107  func gcParkStrongFromWeak() *m {
  2108  	// Prevent preemption as we check strongFromWeak, so it can't change out from under us.
  2109  	mp := acquirem()
  2110  
  2111  	for work.strongFromWeak.block {
  2112  		lock(&work.strongFromWeak.lock)
  2113  		releasem(mp) // N.B. Holding the lock prevents preemption.
  2114  
  2115  		// Queue ourselves up.
  2116  		work.strongFromWeak.q.pushBack(getg())
  2117  
  2118  		// Park.
  2119  		goparkunlock(&work.strongFromWeak.lock, waitReasonGCWeakToStrongWait, traceBlockGCWeakToStrongWait, 2)
  2120  
  2121  		// Re-acquire the current M since we're going to check the condition again.
  2122  		mp = acquirem()
  2123  
  2124  		// Re-check condition. We may have awoken in the next GC's mark termination phase.
  2125  	}
  2126  	return mp
  2127  }
  2128  
  2129  // gcWakeAllStrongFromWeak wakes all currently blocked weak->strong
  2130  // conversions. This is used at the end of a GC cycle.
  2131  //
  2132  // work.strongFromWeak.block must be false to prevent woken goroutines
  2133  // from immediately going back to sleep.
  2134  func gcWakeAllStrongFromWeak() {
  2135  	lock(&work.strongFromWeak.lock)
  2136  	list := work.strongFromWeak.q.popList()
  2137  	injectglist(&list)
  2138  	unlock(&work.strongFromWeak.lock)
  2139  }
  2140  
  2141  // Retrieves or creates a weak pointer handle for the object p.
  2142  func getOrAddWeakHandle(p unsafe.Pointer) *atomic.Uintptr {
  2143  	// First try to retrieve without allocating.
  2144  	if handle := getWeakHandle(p); handle != nil {
  2145  		// Keep p alive for the duration of the function to ensure
  2146  		// that it cannot die while we're trying to do this.
  2147  		KeepAlive(p)
  2148  		return handle
  2149  	}
  2150  
  2151  	lock(&mheap_.speciallock)
  2152  	s := (*specialWeakHandle)(mheap_.specialWeakHandleAlloc.alloc())
  2153  	unlock(&mheap_.speciallock)
  2154  
  2155  	handle := new(atomic.Uintptr)
  2156  	s.special.kind = _KindSpecialWeakHandle
  2157  	s.handle = handle
  2158  	handle.Store(uintptr(p))
  2159  	if addspecial(p, &s.special) {
  2160  		// This is responsible for maintaining the same
  2161  		// GC-related invariants as markrootSpans in any
  2162  		// situation where it's possible that markrootSpans
  2163  		// has already run but mark termination hasn't yet.
  2164  		if gcphase != _GCoff {
  2165  			mp := acquirem()
  2166  			gcw := &mp.p.ptr().gcw
  2167  			// Mark the weak handle itself, since the
  2168  			// special isn't part of the GC'd heap.
  2169  			scanblock(uintptr(unsafe.Pointer(&s.handle)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  2170  			releasem(mp)
  2171  		}
  2172  
  2173  		// Keep p alive for the duration of the function to ensure
  2174  		// that it cannot die while we're trying to do this.
  2175  		//
  2176  		// Same for handle, which is only stored in the special.
  2177  		// There's a window where it might die if we don't keep it
  2178  		// alive explicitly. Returning it here is probably good enough,
  2179  		// but let's be defensive and explicit. See #70455.
  2180  		KeepAlive(p)
  2181  		KeepAlive(handle)
  2182  		return handle
  2183  	}
  2184  
  2185  	// There was an existing handle. Free the special
  2186  	// and try again. We must succeed because we're explicitly
  2187  	// keeping p live until the end of this function. Either
  2188  	// we, or someone else, must have succeeded, because we can
  2189  	// only fail in the event of a race, and p will still be
  2190  	// be valid no matter how much time we spend here.
  2191  	lock(&mheap_.speciallock)
  2192  	mheap_.specialWeakHandleAlloc.free(unsafe.Pointer(s))
  2193  	unlock(&mheap_.speciallock)
  2194  
  2195  	handle = getWeakHandle(p)
  2196  	if handle == nil {
  2197  		throw("failed to get or create weak handle")
  2198  	}
  2199  
  2200  	// Keep p alive for the duration of the function to ensure
  2201  	// that it cannot die while we're trying to do this.
  2202  	//
  2203  	// Same for handle, just to be defensive.
  2204  	KeepAlive(p)
  2205  	KeepAlive(handle)
  2206  	return handle
  2207  }
  2208  
  2209  func getWeakHandle(p unsafe.Pointer) *atomic.Uintptr {
  2210  	span := spanOfHeap(uintptr(p))
  2211  	if span == nil {
  2212  		throw("getWeakHandle on invalid pointer")
  2213  	}
  2214  
  2215  	// Ensure that the span is swept.
  2216  	// Sweeping accesses the specials list w/o locks, so we have
  2217  	// to synchronize with it. And it's just much safer.
  2218  	mp := acquirem()
  2219  	span.ensureSwept()
  2220  
  2221  	offset := uintptr(p) - span.base()
  2222  
  2223  	lock(&span.speciallock)
  2224  
  2225  	// Find the existing record and return the handle if one exists.
  2226  	var handle *atomic.Uintptr
  2227  	iter, exists := span.specialFindSplicePoint(offset, _KindSpecialWeakHandle)
  2228  	if exists {
  2229  		handle = ((*specialWeakHandle)(unsafe.Pointer(*iter))).handle
  2230  	}
  2231  	unlock(&span.speciallock)
  2232  	releasem(mp)
  2233  
  2234  	// Keep p alive for the duration of the function to ensure
  2235  	// that it cannot die while we're trying to do this.
  2236  	KeepAlive(p)
  2237  	return handle
  2238  }
  2239  
  2240  // The described object is being heap profiled.
  2241  type specialprofile struct {
  2242  	_       sys.NotInHeap
  2243  	special special
  2244  	b       *bucket
  2245  }
  2246  
  2247  // Set the heap profile bucket associated with addr to b.
  2248  func setprofilebucket(p unsafe.Pointer, b *bucket) {
  2249  	lock(&mheap_.speciallock)
  2250  	s := (*specialprofile)(mheap_.specialprofilealloc.alloc())
  2251  	unlock(&mheap_.speciallock)
  2252  	s.special.kind = _KindSpecialProfile
  2253  	s.b = b
  2254  	if !addspecial(p, &s.special) {
  2255  		throw("setprofilebucket: profile already set")
  2256  	}
  2257  }
  2258  
  2259  // specialReachable tracks whether an object is reachable on the next
  2260  // GC cycle. This is used by testing.
  2261  type specialReachable struct {
  2262  	special   special
  2263  	done      bool
  2264  	reachable bool
  2265  }
  2266  
  2267  // specialPinCounter tracks whether an object is pinned multiple times.
  2268  type specialPinCounter struct {
  2269  	special special
  2270  	counter uintptr
  2271  }
  2272  
  2273  // specialsIter helps iterate over specials lists.
  2274  type specialsIter struct {
  2275  	pprev **special
  2276  	s     *special
  2277  }
  2278  
  2279  func newSpecialsIter(span *mspan) specialsIter {
  2280  	return specialsIter{&span.specials, span.specials}
  2281  }
  2282  
  2283  func (i *specialsIter) valid() bool {
  2284  	return i.s != nil
  2285  }
  2286  
  2287  func (i *specialsIter) next() {
  2288  	i.pprev = &i.s.next
  2289  	i.s = *i.pprev
  2290  }
  2291  
  2292  // unlinkAndNext removes the current special from the list and moves
  2293  // the iterator to the next special. It returns the unlinked special.
  2294  func (i *specialsIter) unlinkAndNext() *special {
  2295  	cur := i.s
  2296  	i.s = cur.next
  2297  	*i.pprev = i.s
  2298  	return cur
  2299  }
  2300  
  2301  // freeSpecial performs any cleanup on special s and deallocates it.
  2302  // s must already be unlinked from the specials list.
  2303  func freeSpecial(s *special, p unsafe.Pointer, size uintptr) {
  2304  	switch s.kind {
  2305  	case _KindSpecialFinalizer:
  2306  		sf := (*specialfinalizer)(unsafe.Pointer(s))
  2307  		queuefinalizer(p, sf.fn, sf.nret, sf.fint, sf.ot)
  2308  		lock(&mheap_.speciallock)
  2309  		mheap_.specialfinalizeralloc.free(unsafe.Pointer(sf))
  2310  		unlock(&mheap_.speciallock)
  2311  	case _KindSpecialWeakHandle:
  2312  		sw := (*specialWeakHandle)(unsafe.Pointer(s))
  2313  		sw.handle.Store(0)
  2314  		lock(&mheap_.speciallock)
  2315  		mheap_.specialWeakHandleAlloc.free(unsafe.Pointer(s))
  2316  		unlock(&mheap_.speciallock)
  2317  	case _KindSpecialProfile:
  2318  		sp := (*specialprofile)(unsafe.Pointer(s))
  2319  		mProf_Free(sp.b, size)
  2320  		lock(&mheap_.speciallock)
  2321  		mheap_.specialprofilealloc.free(unsafe.Pointer(sp))
  2322  		unlock(&mheap_.speciallock)
  2323  	case _KindSpecialReachable:
  2324  		sp := (*specialReachable)(unsafe.Pointer(s))
  2325  		sp.done = true
  2326  		// The creator frees these.
  2327  	case _KindSpecialPinCounter:
  2328  		lock(&mheap_.speciallock)
  2329  		mheap_.specialPinCounterAlloc.free(unsafe.Pointer(s))
  2330  		unlock(&mheap_.speciallock)
  2331  	default:
  2332  		throw("bad special kind")
  2333  		panic("not reached")
  2334  	}
  2335  }
  2336  
  2337  // gcBits is an alloc/mark bitmap. This is always used as gcBits.x.
  2338  type gcBits struct {
  2339  	_ sys.NotInHeap
  2340  	x uint8
  2341  }
  2342  
  2343  // bytep returns a pointer to the n'th byte of b.
  2344  func (b *gcBits) bytep(n uintptr) *uint8 {
  2345  	return addb(&b.x, n)
  2346  }
  2347  
  2348  // bitp returns a pointer to the byte containing bit n and a mask for
  2349  // selecting that bit from *bytep.
  2350  func (b *gcBits) bitp(n uintptr) (bytep *uint8, mask uint8) {
  2351  	return b.bytep(n / 8), 1 << (n % 8)
  2352  }
  2353  
  2354  const gcBitsChunkBytes = uintptr(64 << 10)
  2355  const gcBitsHeaderBytes = unsafe.Sizeof(gcBitsHeader{})
  2356  
  2357  type gcBitsHeader struct {
  2358  	free uintptr // free is the index into bits of the next free byte.
  2359  	next uintptr // *gcBits triggers recursive type bug. (issue 14620)
  2360  }
  2361  
  2362  type gcBitsArena struct {
  2363  	_ sys.NotInHeap
  2364  	// gcBitsHeader // side step recursive type bug (issue 14620) by including fields by hand.
  2365  	free uintptr // free is the index into bits of the next free byte; read/write atomically
  2366  	next *gcBitsArena
  2367  	bits [gcBitsChunkBytes - gcBitsHeaderBytes]gcBits
  2368  }
  2369  
  2370  var gcBitsArenas struct {
  2371  	lock     mutex
  2372  	free     *gcBitsArena
  2373  	next     *gcBitsArena // Read atomically. Write atomically under lock.
  2374  	current  *gcBitsArena
  2375  	previous *gcBitsArena
  2376  }
  2377  
  2378  // tryAlloc allocates from b or returns nil if b does not have enough room.
  2379  // This is safe to call concurrently.
  2380  func (b *gcBitsArena) tryAlloc(bytes uintptr) *gcBits {
  2381  	if b == nil || atomic.Loaduintptr(&b.free)+bytes > uintptr(len(b.bits)) {
  2382  		return nil
  2383  	}
  2384  	// Try to allocate from this block.
  2385  	end := atomic.Xadduintptr(&b.free, bytes)
  2386  	if end > uintptr(len(b.bits)) {
  2387  		return nil
  2388  	}
  2389  	// There was enough room.
  2390  	start := end - bytes
  2391  	return &b.bits[start]
  2392  }
  2393  
  2394  // newMarkBits returns a pointer to 8 byte aligned bytes
  2395  // to be used for a span's mark bits.
  2396  func newMarkBits(nelems uintptr) *gcBits {
  2397  	blocksNeeded := (nelems + 63) / 64
  2398  	bytesNeeded := blocksNeeded * 8
  2399  
  2400  	// Try directly allocating from the current head arena.
  2401  	head := (*gcBitsArena)(atomic.Loadp(unsafe.Pointer(&gcBitsArenas.next)))
  2402  	if p := head.tryAlloc(bytesNeeded); p != nil {
  2403  		return p
  2404  	}
  2405  
  2406  	// There's not enough room in the head arena. We may need to
  2407  	// allocate a new arena.
  2408  	lock(&gcBitsArenas.lock)
  2409  	// Try the head arena again, since it may have changed. Now
  2410  	// that we hold the lock, the list head can't change, but its
  2411  	// free position still can.
  2412  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2413  		unlock(&gcBitsArenas.lock)
  2414  		return p
  2415  	}
  2416  
  2417  	// Allocate a new arena. This may temporarily drop the lock.
  2418  	fresh := newArenaMayUnlock()
  2419  	// If newArenaMayUnlock dropped the lock, another thread may
  2420  	// have put a fresh arena on the "next" list. Try allocating
  2421  	// from next again.
  2422  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2423  		// Put fresh back on the free list.
  2424  		// TODO: Mark it "already zeroed"
  2425  		fresh.next = gcBitsArenas.free
  2426  		gcBitsArenas.free = fresh
  2427  		unlock(&gcBitsArenas.lock)
  2428  		return p
  2429  	}
  2430  
  2431  	// Allocate from the fresh arena. We haven't linked it in yet, so
  2432  	// this cannot race and is guaranteed to succeed.
  2433  	p := fresh.tryAlloc(bytesNeeded)
  2434  	if p == nil {
  2435  		throw("markBits overflow")
  2436  	}
  2437  
  2438  	// Add the fresh arena to the "next" list.
  2439  	fresh.next = gcBitsArenas.next
  2440  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), unsafe.Pointer(fresh))
  2441  
  2442  	unlock(&gcBitsArenas.lock)
  2443  	return p
  2444  }
  2445  
  2446  // newAllocBits returns a pointer to 8 byte aligned bytes
  2447  // to be used for this span's alloc bits.
  2448  // newAllocBits is used to provide newly initialized spans
  2449  // allocation bits. For spans not being initialized the
  2450  // mark bits are repurposed as allocation bits when
  2451  // the span is swept.
  2452  func newAllocBits(nelems uintptr) *gcBits {
  2453  	return newMarkBits(nelems)
  2454  }
  2455  
  2456  // nextMarkBitArenaEpoch establishes a new epoch for the arenas
  2457  // holding the mark bits. The arenas are named relative to the
  2458  // current GC cycle which is demarcated by the call to finishweep_m.
  2459  //
  2460  // All current spans have been swept.
  2461  // During that sweep each span allocated room for its gcmarkBits in
  2462  // gcBitsArenas.next block. gcBitsArenas.next becomes the gcBitsArenas.current
  2463  // where the GC will mark objects and after each span is swept these bits
  2464  // will be used to allocate objects.
  2465  // gcBitsArenas.current becomes gcBitsArenas.previous where the span's
  2466  // gcAllocBits live until all the spans have been swept during this GC cycle.
  2467  // The span's sweep extinguishes all the references to gcBitsArenas.previous
  2468  // by pointing gcAllocBits into the gcBitsArenas.current.
  2469  // The gcBitsArenas.previous is released to the gcBitsArenas.free list.
  2470  func nextMarkBitArenaEpoch() {
  2471  	lock(&gcBitsArenas.lock)
  2472  	if gcBitsArenas.previous != nil {
  2473  		if gcBitsArenas.free == nil {
  2474  			gcBitsArenas.free = gcBitsArenas.previous
  2475  		} else {
  2476  			// Find end of previous arenas.
  2477  			last := gcBitsArenas.previous
  2478  			for last = gcBitsArenas.previous; last.next != nil; last = last.next {
  2479  			}
  2480  			last.next = gcBitsArenas.free
  2481  			gcBitsArenas.free = gcBitsArenas.previous
  2482  		}
  2483  	}
  2484  	gcBitsArenas.previous = gcBitsArenas.current
  2485  	gcBitsArenas.current = gcBitsArenas.next
  2486  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), nil) // newMarkBits calls newArena when needed
  2487  	unlock(&gcBitsArenas.lock)
  2488  }
  2489  
  2490  // newArenaMayUnlock allocates and zeroes a gcBits arena.
  2491  // The caller must hold gcBitsArena.lock. This may temporarily release it.
  2492  func newArenaMayUnlock() *gcBitsArena {
  2493  	var result *gcBitsArena
  2494  	if gcBitsArenas.free == nil {
  2495  		unlock(&gcBitsArenas.lock)
  2496  		result = (*gcBitsArena)(sysAlloc(gcBitsChunkBytes, &memstats.gcMiscSys))
  2497  		if result == nil {
  2498  			throw("runtime: cannot allocate memory")
  2499  		}
  2500  		lock(&gcBitsArenas.lock)
  2501  	} else {
  2502  		result = gcBitsArenas.free
  2503  		gcBitsArenas.free = gcBitsArenas.free.next
  2504  		memclrNoHeapPointers(unsafe.Pointer(result), gcBitsChunkBytes)
  2505  	}
  2506  	result.next = nil
  2507  	// If result.bits is not 8 byte aligned adjust index so
  2508  	// that &result.bits[result.free] is 8 byte aligned.
  2509  	if unsafe.Offsetof(gcBitsArena{}.bits)&7 == 0 {
  2510  		result.free = 0
  2511  	} else {
  2512  		result.free = 8 - (uintptr(unsafe.Pointer(&result.bits[0])) & 7)
  2513  	}
  2514  	return result
  2515  }
  2516  

View as plain text