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/abi"
    13  	"internal/cpu"
    14  	"internal/goarch"
    15  	"internal/goexperiment"
    16  	"internal/runtime/atomic"
    17  	"internal/runtime/gc"
    18  	"internal/runtime/sys"
    19  	"unsafe"
    20  )
    21  
    22  const (
    23  	// minPhysPageSize is a lower-bound on the physical page size. The
    24  	// true physical page size may be larger than this. In contrast,
    25  	// sys.PhysPageSize is an upper-bound on the physical page size.
    26  	minPhysPageSize = 4096
    27  
    28  	// maxPhysPageSize is the maximum page size the runtime supports.
    29  	maxPhysPageSize = 512 << 10
    30  
    31  	// maxPhysHugePageSize sets an upper-bound on the maximum huge page size
    32  	// that the runtime supports.
    33  	maxPhysHugePageSize = pallocChunkBytes
    34  
    35  	// pagesPerReclaimerChunk indicates how many pages to scan from the
    36  	// pageInUse bitmap at a time. Used by the page reclaimer.
    37  	//
    38  	// Higher values reduce contention on scanning indexes (such as
    39  	// h.reclaimIndex), but increase the minimum latency of the
    40  	// operation.
    41  	//
    42  	// The time required to scan this many pages can vary a lot depending
    43  	// on how many spans are actually freed. Experimentally, it can
    44  	// scan for pages at ~300 GB/ms on a 2.6GHz Core i7, but can only
    45  	// free spans at ~32 MB/ms. Using 512 pages bounds this at
    46  	// roughly 100µs.
    47  	//
    48  	// Must be a multiple of the pageInUse bitmap element size and
    49  	// must also evenly divide pagesPerArena.
    50  	pagesPerReclaimerChunk = 512
    51  
    52  	// physPageAlignedStacks indicates whether stack allocations must be
    53  	// physical page aligned. This is a requirement for MAP_STACK on
    54  	// OpenBSD.
    55  	physPageAlignedStacks = GOOS == "openbsd"
    56  )
    57  
    58  // Main malloc heap.
    59  // The heap itself is the "free" and "scav" treaps,
    60  // but all the other global data is here too.
    61  //
    62  // mheap must not be heap-allocated because it contains mSpanLists,
    63  // which must not be heap-allocated.
    64  type mheap struct {
    65  	_ sys.NotInHeap
    66  
    67  	// lock must only be acquired on the system stack, otherwise a g
    68  	// could self-deadlock if its stack grows with the lock held.
    69  	lock mutex
    70  
    71  	pages pageAlloc // page allocation data structure
    72  
    73  	sweepgen uint32 // sweep generation, see comment in mspan; written during STW
    74  
    75  	// allspans is a slice of all mspans ever created. Each mspan
    76  	// appears exactly once.
    77  	//
    78  	// The memory for allspans is manually managed and can be
    79  	// reallocated and move as the heap grows.
    80  	//
    81  	// In general, allspans is protected by mheap_.lock, which
    82  	// prevents concurrent access as well as freeing the backing
    83  	// store. Accesses during STW might not hold the lock, but
    84  	// must ensure that allocation cannot happen around the
    85  	// access (since that may free the backing store).
    86  	allspans []*mspan // all spans out there
    87  
    88  	// Proportional sweep
    89  	//
    90  	// These parameters represent a linear function from gcController.heapLive
    91  	// to page sweep count. The proportional sweep system works to
    92  	// stay in the black by keeping the current page sweep count
    93  	// above this line at the current gcController.heapLive.
    94  	//
    95  	// The line has slope sweepPagesPerByte and passes through a
    96  	// basis point at (sweepHeapLiveBasis, pagesSweptBasis). At
    97  	// any given time, the system is at (gcController.heapLive,
    98  	// pagesSwept) in this space.
    99  	//
   100  	// It is important that the line pass through a point we
   101  	// control rather than simply starting at a 0,0 origin
   102  	// because that lets us adjust sweep pacing at any time while
   103  	// accounting for current progress. If we could only adjust
   104  	// the slope, it would create a discontinuity in debt if any
   105  	// progress has already been made.
   106  	pagesInUse         atomic.Uintptr // pages of spans in stats mSpanInUse
   107  	pagesSwept         atomic.Uint64  // pages swept this cycle
   108  	pagesSweptBasis    atomic.Uint64  // pagesSwept to use as the origin of the sweep ratio
   109  	sweepHeapLiveBasis uint64         // value of gcController.heapLive to use as the origin of sweep ratio; written with lock, read without
   110  	sweepPagesPerByte  float64        // proportional sweep ratio; written with lock, read without
   111  
   112  	// Page reclaimer state
   113  
   114  	// reclaimIndex is the page index in heapArenas of next page to
   115  	// reclaim. Specifically, it refers to page (i %
   116  	// pagesPerArena) of arena heapArenas[i / pagesPerArena].
   117  	//
   118  	// If this is >= 1<<63, the page reclaimer is done scanning
   119  	// the page marks.
   120  	reclaimIndex atomic.Uint64
   121  
   122  	// reclaimCredit is spare credit for extra pages swept. Since
   123  	// the page reclaimer works in large chunks, it may reclaim
   124  	// more than requested. Any spare pages released go to this
   125  	// credit pool.
   126  	reclaimCredit atomic.Uintptr
   127  
   128  	_ cpu.CacheLinePad // prevents false-sharing between arenas and preceding variables
   129  
   130  	// arenas is the heap arena map. It points to the metadata for
   131  	// the heap for every arena frame of the entire usable virtual
   132  	// address space.
   133  	//
   134  	// Use arenaIndex to compute indexes into this array.
   135  	//
   136  	// For regions of the address space that are not backed by the
   137  	// Go heap, the arena map contains nil.
   138  	//
   139  	// Modifications are protected by mheap_.lock. Reads can be
   140  	// performed without locking; however, a given entry can
   141  	// transition from nil to non-nil at any time when the lock
   142  	// isn't held. (Entries never transitions back to nil.)
   143  	//
   144  	// In general, this is a two-level mapping consisting of an L1
   145  	// map and possibly many L2 maps. This saves space when there
   146  	// are a huge number of arena frames. However, on many
   147  	// platforms (even 64-bit), arenaL1Bits is 0, making this
   148  	// effectively a single-level map. In this case, arenas[0]
   149  	// will never be nil.
   150  	arenas [1 << arenaL1Bits]*[1 << arenaL2Bits]*heapArena
   151  
   152  	// arenasHugePages indicates whether arenas' L2 entries are eligible
   153  	// to be backed by huge pages.
   154  	arenasHugePages bool
   155  
   156  	// heapArenaAlloc is pre-reserved space for allocating heapArena
   157  	// objects. This is only used on 32-bit, where we pre-reserve
   158  	// this space to avoid interleaving it with the heap itself.
   159  	heapArenaAlloc linearAlloc
   160  
   161  	// arenaHints is a list of addresses at which to attempt to
   162  	// add more heap arenas. This is initially populated with a
   163  	// set of general hint addresses, and grown with the bounds of
   164  	// actual heap arena ranges.
   165  	arenaHints *arenaHint
   166  
   167  	// arena is a pre-reserved space for allocating heap arenas
   168  	// (the actual arenas). This is only used on 32-bit.
   169  	arena linearAlloc
   170  
   171  	// heapArenas is the arenaIndex of every mapped arena mapped for the heap.
   172  	// This can be used to iterate through the heap address space.
   173  	//
   174  	// Access is protected by mheap_.lock. However, since this is
   175  	// append-only and old backing arrays are never freed, it is
   176  	// safe to acquire mheap_.lock, copy the slice header, and
   177  	// then release mheap_.lock.
   178  	heapArenas []arenaIdx
   179  
   180  	// userArenaArenas is the arenaIndex of every mapped arena mapped for
   181  	// user arenas.
   182  	//
   183  	// Access is protected by mheap_.lock. However, since this is
   184  	// append-only and old backing arrays are never freed, it is
   185  	// safe to acquire mheap_.lock, copy the slice header, and
   186  	// then release mheap_.lock.
   187  	userArenaArenas []arenaIdx
   188  
   189  	// sweepArenas is a snapshot of heapArenas taken at the
   190  	// beginning of the sweep cycle. This can be read safely by
   191  	// simply blocking GC (by disabling preemption).
   192  	sweepArenas []arenaIdx
   193  
   194  	// markArenas is a snapshot of heapArenas taken at the beginning
   195  	// of the mark cycle. Because heapArenas is append-only, neither
   196  	// this slice nor its contents will change during the mark, so
   197  	// it can be read safely.
   198  	markArenas []arenaIdx
   199  
   200  	// curArena is the arena that the heap is currently growing
   201  	// into. This should always be physPageSize-aligned.
   202  	curArena struct {
   203  		base, end uintptr
   204  	}
   205  
   206  	// central free lists for small size classes.
   207  	// the padding makes sure that the mcentrals are
   208  	// spaced CacheLinePadSize bytes apart, so that each mcentral.lock
   209  	// gets its own cache line.
   210  	// central is indexed by spanClass.
   211  	central [numSpanClasses]struct {
   212  		mcentral mcentral
   213  		pad      [(cpu.CacheLinePadSize - unsafe.Sizeof(mcentral{})%cpu.CacheLinePadSize) % cpu.CacheLinePadSize]byte
   214  	}
   215  
   216  	spanalloc                  fixalloc // allocator for span*
   217  	cachealloc                 fixalloc // allocator for mcache*
   218  	specialfinalizeralloc      fixalloc // allocator for specialfinalizer*
   219  	specialCleanupAlloc        fixalloc // allocator for specialCleanup*
   220  	specialCheckFinalizerAlloc fixalloc // allocator for specialCheckFinalizer*
   221  	specialTinyBlockAlloc      fixalloc // allocator for specialTinyBlock*
   222  	specialprofilealloc        fixalloc // allocator for specialprofile*
   223  	specialReachableAlloc      fixalloc // allocator for specialReachable
   224  	specialPinCounterAlloc     fixalloc // allocator for specialPinCounter
   225  	specialWeakHandleAlloc     fixalloc // allocator for specialWeakHandle
   226  	specialBubbleAlloc         fixalloc // allocator for specialBubble
   227  	speciallock                mutex    // lock for special record allocators.
   228  	arenaHintAlloc             fixalloc // allocator for arenaHints
   229  
   230  	// User arena state.
   231  	//
   232  	// Protected by mheap_.lock.
   233  	userArena struct {
   234  		// arenaHints is a list of addresses at which to attempt to
   235  		// add more heap arenas for user arena chunks. This is initially
   236  		// populated with a set of general hint addresses, and grown with
   237  		// the bounds of actual heap arena ranges.
   238  		arenaHints *arenaHint
   239  
   240  		// quarantineList is a list of user arena spans that have been set to fault, but
   241  		// are waiting for all pointers into them to go away. Sweeping handles
   242  		// identifying when this is true, and moves the span to the ready list.
   243  		quarantineList mSpanList
   244  
   245  		// readyList is a list of empty user arena spans that are ready for reuse.
   246  		readyList mSpanList
   247  	}
   248  
   249  	// cleanupID is a counter which is incremented each time a cleanup special is added
   250  	// to a span. It's used to create globally unique identifiers for individual cleanup.
   251  	// cleanupID is protected by mheap_.speciallock. It must only be incremented while holding
   252  	// the lock. ID 0 is reserved. Users should increment first, then read the value.
   253  	cleanupID uint64
   254  
   255  	_ cpu.CacheLinePad
   256  
   257  	immortalWeakHandles immortalWeakHandleMap
   258  
   259  	unused *specialfinalizer // never set, just here to force the specialfinalizer type into DWARF
   260  }
   261  
   262  var mheap_ mheap
   263  
   264  // A heapArena stores metadata for a heap arena. heapArenas are stored
   265  // outside of the Go heap and accessed via the mheap_.arenas index.
   266  type heapArena struct {
   267  	_ sys.NotInHeap
   268  
   269  	// spans maps from virtual address page ID within this arena to *mspan.
   270  	// For allocated spans, their pages map to the span itself.
   271  	// For free spans, only the lowest and highest pages map to the span itself.
   272  	// Internal pages map to an arbitrary span.
   273  	// For pages that have never been allocated, spans entries are nil.
   274  	//
   275  	// Modifications are protected by mheap.lock. Reads can be
   276  	// performed without locking, but ONLY from indexes that are
   277  	// known to contain in-use or stack spans. This means there
   278  	// must not be a safe-point between establishing that an
   279  	// address is live and looking it up in the spans array.
   280  	spans [pagesPerArena]*mspan
   281  
   282  	// pageInUse is a bitmap that indicates which spans are in
   283  	// state mSpanInUse. This bitmap is indexed by page number,
   284  	// but only the bit corresponding to the first page in each
   285  	// span is used.
   286  	//
   287  	// Reads and writes are atomic.
   288  	pageInUse [pagesPerArena / 8]uint8
   289  
   290  	// pageMarks is a bitmap that indicates which spans have any
   291  	// marked objects on them. Like pageInUse, only the bit
   292  	// corresponding to the first page in each span is used.
   293  	//
   294  	// Writes are done atomically during marking. Reads are
   295  	// non-atomic and lock-free since they only occur during
   296  	// sweeping (and hence never race with writes).
   297  	//
   298  	// This is used to quickly find whole spans that can be freed.
   299  	//
   300  	// TODO(austin): It would be nice if this was uint64 for
   301  	// faster scanning, but we don't have 64-bit atomic bit
   302  	// operations.
   303  	pageMarks [pagesPerArena / 8]uint8
   304  
   305  	// pageSpecials is a bitmap that indicates which spans have
   306  	// specials (finalizers or other). Like pageInUse, only the bit
   307  	// corresponding to the first page in each span is used.
   308  	//
   309  	// Writes are done atomically whenever a special is added to
   310  	// a span and whenever the last special is removed from a span.
   311  	// Reads are done atomically to find spans containing specials
   312  	// during marking.
   313  	pageSpecials [pagesPerArena / 8]uint8
   314  
   315  	// pageUseSpanInlineMarkBits is a bitmap where each bit corresponds
   316  	// to a span, as only spans one page in size can have inline mark bits.
   317  	// The bit indicates that the span has a spanInlineMarkBits struct
   318  	// stored directly at the top end of the span's memory.
   319  	pageUseSpanInlineMarkBits [pagesPerArena / 8]uint8
   320  
   321  	// checkmarks stores the debug.gccheckmark state. It is only
   322  	// used if debug.gccheckmark > 0 or debug.checkfinalizers > 0.
   323  	checkmarks *checkmarksMap
   324  
   325  	// zeroedBase marks the first byte of the first page in this
   326  	// arena which hasn't been used yet and is therefore already
   327  	// zero. zeroedBase is relative to the arena base.
   328  	// Increases monotonically until it hits heapArenaBytes.
   329  	//
   330  	// This field is sufficient to determine if an allocation
   331  	// needs to be zeroed because the page allocator follows an
   332  	// address-ordered first-fit policy.
   333  	//
   334  	// Read atomically and written with an atomic CAS.
   335  	zeroedBase uintptr
   336  }
   337  
   338  // arenaHint is a hint for where to grow the heap arenas. See
   339  // mheap_.arenaHints.
   340  type arenaHint struct {
   341  	_    sys.NotInHeap
   342  	addr uintptr
   343  	down bool
   344  	next *arenaHint
   345  }
   346  
   347  // An mspan is a run of pages.
   348  //
   349  // When a mspan is in the heap free treap, state == mSpanFree
   350  // and heapmap(s->start) == span, heapmap(s->start+s->npages-1) == span.
   351  // If the mspan is in the heap scav treap, then in addition to the
   352  // above scavenged == true. scavenged == false in all other cases.
   353  //
   354  // When a mspan is allocated, state == mSpanInUse or mSpanManual
   355  // and heapmap(i) == span for all s->start <= i < s->start+s->npages.
   356  
   357  // Every mspan is in one doubly-linked list, either in the mheap's
   358  // busy list or one of the mcentral's span lists.
   359  
   360  // An mspan representing actual memory has state mSpanInUse,
   361  // mSpanManual, or mSpanFree. Transitions between these states are
   362  // constrained as follows:
   363  //
   364  //   - A span may transition from free to in-use or manual during any GC
   365  //     phase.
   366  //
   367  //   - During sweeping (gcphase == _GCoff), a span may transition from
   368  //     in-use to free (as a result of sweeping) or manual to free (as a
   369  //     result of stacks being freed).
   370  //
   371  //   - During GC (gcphase != _GCoff), a span *must not* transition from
   372  //     manual or in-use to free. Because concurrent GC may read a pointer
   373  //     and then look up its span, the span state must be monotonic.
   374  //
   375  // Setting mspan.state to mSpanInUse or mSpanManual must be done
   376  // atomically and only after all other span fields are valid.
   377  // Likewise, if inspecting a span is contingent on it being
   378  // mSpanInUse, the state should be loaded atomically and checked
   379  // before depending on other fields. This allows the garbage collector
   380  // to safely deal with potentially invalid pointers, since resolving
   381  // such pointers may race with a span being allocated.
   382  type mSpanState uint8
   383  
   384  const (
   385  	mSpanDead   mSpanState = iota
   386  	mSpanInUse             // allocated for garbage collected heap
   387  	mSpanManual            // allocated for manual management (e.g., stack allocator)
   388  )
   389  
   390  // mSpanStateNames are the names of the span states, indexed by
   391  // mSpanState.
   392  var mSpanStateNames = []string{
   393  	"mSpanDead",
   394  	"mSpanInUse",
   395  	"mSpanManual",
   396  }
   397  
   398  // mSpanStateBox holds an atomic.Uint8 to provide atomic operations on
   399  // an mSpanState. This is a separate type to disallow accidental comparison
   400  // or assignment with mSpanState.
   401  type mSpanStateBox struct {
   402  	s atomic.Uint8
   403  }
   404  
   405  // It is nosplit to match get, below.
   406  
   407  //go:nosplit
   408  func (b *mSpanStateBox) set(s mSpanState) {
   409  	b.s.Store(uint8(s))
   410  }
   411  
   412  // It is nosplit because it's called indirectly by typedmemclr,
   413  // which must not be preempted.
   414  
   415  //go:nosplit
   416  func (b *mSpanStateBox) get() mSpanState {
   417  	return mSpanState(b.s.Load())
   418  }
   419  
   420  type mspan struct {
   421  	_    sys.NotInHeap
   422  	next *mspan     // next span in list, or nil if none
   423  	prev *mspan     // previous span in list, or nil if none
   424  	list *mSpanList // For debugging.
   425  
   426  	startAddr uintptr // address of first byte of span aka s.base()
   427  	npages    uintptr // number of pages in span
   428  
   429  	manualFreeList gclinkptr // list of free objects in mSpanManual spans
   430  
   431  	// freeindex is the slot index between 0 and nelems at which to begin scanning
   432  	// for the next free object in this span.
   433  	// Each allocation scans allocBits starting at freeindex until it encounters a 0
   434  	// indicating a free object. freeindex is then adjusted so that subsequent scans begin
   435  	// just past the newly discovered free object.
   436  	//
   437  	// If freeindex == nelems, this span has no free objects.
   438  	//
   439  	// allocBits is a bitmap of objects in this span.
   440  	// If n >= freeindex and allocBits[n/8] & (1<<(n%8)) is 0
   441  	// then object n is free;
   442  	// otherwise, object n is allocated. Bits starting at nelems are
   443  	// undefined and should never be referenced.
   444  	//
   445  	// Object n starts at address n*elemsize + (start << pageShift).
   446  	freeindex uint16
   447  	// TODO: Look up nelems from sizeclass and remove this field if it
   448  	// helps performance.
   449  	nelems uint16 // number of object in the span.
   450  	// freeIndexForScan is like freeindex, except that freeindex is
   451  	// used by the allocator whereas freeIndexForScan is used by the
   452  	// GC scanner. They are two fields so that the GC sees the object
   453  	// is allocated only when the object and the heap bits are
   454  	// initialized (see also the assignment of freeIndexForScan in
   455  	// mallocgc, and issue 54596).
   456  	freeIndexForScan uint16
   457  
   458  	// Temporary storage for the object index that caused this span to
   459  	// be queued for scanning.
   460  	//
   461  	// Used only with goexperiment.GreenTeaGC.
   462  	scanIdx uint16
   463  
   464  	// Cache of the allocBits at freeindex. allocCache is shifted
   465  	// such that the lowest bit corresponds to the bit freeindex.
   466  	// allocCache holds the complement of allocBits, thus allowing
   467  	// ctz (count trailing zero) to use it directly.
   468  	// allocCache may contain bits beyond s.nelems; the caller must ignore
   469  	// these.
   470  	allocCache uint64
   471  
   472  	// allocBits and gcmarkBits hold pointers to a span's mark and
   473  	// allocation bits. The pointers are 8 byte aligned.
   474  	// There are three arenas where this data is held.
   475  	// free: Dirty arenas that are no longer accessed
   476  	//       and can be reused.
   477  	// next: Holds information to be used in the next GC cycle.
   478  	// current: Information being used during this GC cycle.
   479  	// previous: Information being used during the last GC cycle.
   480  	// A new GC cycle starts with the call to finishsweep_m.
   481  	// finishsweep_m moves the previous arena to the free arena,
   482  	// the current arena to the previous arena, and
   483  	// the next arena to the current arena.
   484  	// The next arena is populated as the spans request
   485  	// memory to hold gcmarkBits for the next GC cycle as well
   486  	// as allocBits for newly allocated spans.
   487  	//
   488  	// The pointer arithmetic is done "by hand" instead of using
   489  	// arrays to avoid bounds checks along critical performance
   490  	// paths.
   491  	// The sweep will free the old allocBits and set allocBits to the
   492  	// gcmarkBits. The gcmarkBits are replaced with a fresh zeroed
   493  	// out memory.
   494  	allocBits  *gcBits
   495  	gcmarkBits *gcBits
   496  	pinnerBits *gcBits // bitmap for pinned objects; accessed atomically
   497  
   498  	// sweep generation:
   499  	// if sweepgen == h->sweepgen - 2, the span needs sweeping
   500  	// if sweepgen == h->sweepgen - 1, the span is currently being swept
   501  	// if sweepgen == h->sweepgen, the span is swept and ready to use
   502  	// if sweepgen == h->sweepgen + 1, the span was cached before sweep began and is still cached, and needs sweeping
   503  	// if sweepgen == h->sweepgen + 3, the span was swept and then cached and is still cached
   504  	// h->sweepgen is incremented by 2 after every GC
   505  
   506  	sweepgen              uint32
   507  	divMul                uint32        // for divide by elemsize
   508  	allocCount            uint16        // number of allocated objects
   509  	spanclass             spanClass     // size class and noscan (uint8)
   510  	state                 mSpanStateBox // mSpanInUse etc; accessed atomically (get/set methods)
   511  	needzero              uint8         // needs to be zeroed before allocation
   512  	isUserArenaChunk      bool          // whether or not this span represents a user arena
   513  	allocCountBeforeCache uint16        // a copy of allocCount that is stored just before this span is cached
   514  	elemsize              uintptr       // computed from sizeclass or from npages
   515  	limit                 uintptr       // end of data in span
   516  	speciallock           mutex         // guards specials list and changes to pinnerBits
   517  	specials              *special      // linked list of special records sorted by offset.
   518  	userArenaChunkFree    addrRange     // interval for managing chunk allocation
   519  	largeType             *_type        // malloc header for large objects.
   520  }
   521  
   522  func (s *mspan) base() uintptr {
   523  	return s.startAddr
   524  }
   525  
   526  func (s *mspan) layout() (size, n, total uintptr) {
   527  	total = s.npages << gc.PageShift
   528  	size = s.elemsize
   529  	if size > 0 {
   530  		n = total / size
   531  	}
   532  	return
   533  }
   534  
   535  // recordspan adds a newly allocated span to h.allspans.
   536  //
   537  // This only happens the first time a span is allocated from
   538  // mheap.spanalloc (it is not called when a span is reused).
   539  //
   540  // Write barriers are disallowed here because it can be called from
   541  // gcWork when allocating new workbufs. However, because it's an
   542  // indirect call from the fixalloc initializer, the compiler can't see
   543  // this.
   544  //
   545  // The heap lock must be held.
   546  //
   547  //go:nowritebarrierrec
   548  func recordspan(vh unsafe.Pointer, p unsafe.Pointer) {
   549  	h := (*mheap)(vh)
   550  	s := (*mspan)(p)
   551  
   552  	assertLockHeld(&h.lock)
   553  
   554  	if len(h.allspans) >= cap(h.allspans) {
   555  		n := 64 * 1024 / goarch.PtrSize
   556  		if n < cap(h.allspans)*3/2 {
   557  			n = cap(h.allspans) * 3 / 2
   558  		}
   559  		var new []*mspan
   560  		sp := (*slice)(unsafe.Pointer(&new))
   561  		sp.array = sysAlloc(uintptr(n)*goarch.PtrSize, &memstats.other_sys, "allspans array")
   562  		if sp.array == nil {
   563  			throw("runtime: cannot allocate memory")
   564  		}
   565  		sp.len = len(h.allspans)
   566  		sp.cap = n
   567  		if len(h.allspans) > 0 {
   568  			copy(new, h.allspans)
   569  		}
   570  		oldAllspans := h.allspans
   571  		*(*notInHeapSlice)(unsafe.Pointer(&h.allspans)) = *(*notInHeapSlice)(unsafe.Pointer(&new))
   572  		if len(oldAllspans) != 0 {
   573  			sysFree(unsafe.Pointer(&oldAllspans[0]), uintptr(cap(oldAllspans))*unsafe.Sizeof(oldAllspans[0]), &memstats.other_sys)
   574  		}
   575  	}
   576  	h.allspans = h.allspans[:len(h.allspans)+1]
   577  	h.allspans[len(h.allspans)-1] = s
   578  }
   579  
   580  // A spanClass represents the size class and noscan-ness of a span.
   581  //
   582  // Each size class has a noscan spanClass and a scan spanClass. The
   583  // noscan spanClass contains only noscan objects, which do not contain
   584  // pointers and thus do not need to be scanned by the garbage
   585  // collector.
   586  type spanClass uint8
   587  
   588  const (
   589  	numSpanClasses = gc.NumSizeClasses << 1
   590  	tinySpanClass  = spanClass(tinySizeClass<<1 | 1)
   591  )
   592  
   593  func makeSpanClass(sizeclass uint8, noscan bool) spanClass {
   594  	return spanClass(sizeclass<<1) | spanClass(bool2int(noscan))
   595  }
   596  
   597  //go:nosplit
   598  func (sc spanClass) sizeclass() int8 {
   599  	return int8(sc >> 1)
   600  }
   601  
   602  //go:nosplit
   603  func (sc spanClass) noscan() bool {
   604  	return sc&1 != 0
   605  }
   606  
   607  // arenaIndex returns the index into mheap_.arenas of the arena
   608  // containing metadata for p. This index combines of an index into the
   609  // L1 map and an index into the L2 map and should be used as
   610  // mheap_.arenas[ai.l1()][ai.l2()].
   611  //
   612  // If p is outside the range of valid heap addresses, either l1() or
   613  // l2() will be out of bounds.
   614  //
   615  // It is nosplit because it's called by spanOf and several other
   616  // nosplit functions.
   617  //
   618  //go:nosplit
   619  func arenaIndex(p uintptr) arenaIdx {
   620  	return arenaIdx((p - arenaBaseOffset) / heapArenaBytes)
   621  }
   622  
   623  // arenaBase returns the low address of the region covered by heap
   624  // arena i.
   625  func arenaBase(i arenaIdx) uintptr {
   626  	return uintptr(i)*heapArenaBytes + arenaBaseOffset
   627  }
   628  
   629  type arenaIdx uint
   630  
   631  // l1 returns the "l1" portion of an arenaIdx.
   632  //
   633  // Marked nosplit because it's called by spanOf and other nosplit
   634  // functions.
   635  //
   636  //go:nosplit
   637  func (i arenaIdx) l1() uint {
   638  	if arenaL1Bits == 0 {
   639  		// Let the compiler optimize this away if there's no
   640  		// L1 map.
   641  		return 0
   642  	} else {
   643  		return uint(i) >> arenaL1Shift
   644  	}
   645  }
   646  
   647  // l2 returns the "l2" portion of an arenaIdx.
   648  //
   649  // Marked nosplit because it's called by spanOf and other nosplit funcs.
   650  // functions.
   651  //
   652  //go:nosplit
   653  func (i arenaIdx) l2() uint {
   654  	if arenaL1Bits == 0 {
   655  		return uint(i)
   656  	} else {
   657  		return uint(i) & (1<<arenaL2Bits - 1)
   658  	}
   659  }
   660  
   661  // inheap reports whether b is a pointer into a (potentially dead) heap object.
   662  // It returns false for pointers into mSpanManual spans.
   663  // Non-preemptible because it is used by write barriers.
   664  //
   665  //go:nowritebarrier
   666  //go:nosplit
   667  func inheap(b uintptr) bool {
   668  	return spanOfHeap(b) != nil
   669  }
   670  
   671  // inHeapOrStack is a variant of inheap that returns true for pointers
   672  // into any allocated heap span.
   673  //
   674  //go:nowritebarrier
   675  //go:nosplit
   676  func inHeapOrStack(b uintptr) bool {
   677  	s := spanOf(b)
   678  	if s == nil || b < s.base() {
   679  		return false
   680  	}
   681  	switch s.state.get() {
   682  	case mSpanInUse, mSpanManual:
   683  		return b < s.limit
   684  	default:
   685  		return false
   686  	}
   687  }
   688  
   689  // spanOf returns the span of p. If p does not point into the heap
   690  // arena or no span has ever contained p, spanOf returns nil.
   691  //
   692  // If p does not point to allocated memory, this may return a non-nil
   693  // span that does *not* contain p. If this is a possibility, the
   694  // caller should either call spanOfHeap or check the span bounds
   695  // explicitly.
   696  //
   697  // Must be nosplit because it has callers that are nosplit.
   698  //
   699  //go:nosplit
   700  func spanOf(p uintptr) *mspan {
   701  	// This function looks big, but we use a lot of constant
   702  	// folding around arenaL1Bits to get it under the inlining
   703  	// budget. Also, many of the checks here are safety checks
   704  	// that Go needs to do anyway, so the generated code is quite
   705  	// short.
   706  	ri := arenaIndex(p)
   707  	if arenaL1Bits == 0 {
   708  		// If there's no L1, then ri.l1() can't be out of bounds but ri.l2() can.
   709  		if ri.l2() >= uint(len(mheap_.arenas[0])) {
   710  			return nil
   711  		}
   712  	} else {
   713  		// If there's an L1, then ri.l1() can be out of bounds but ri.l2() can't.
   714  		if ri.l1() >= uint(len(mheap_.arenas)) {
   715  			return nil
   716  		}
   717  	}
   718  	l2 := mheap_.arenas[ri.l1()]
   719  	if arenaL1Bits != 0 && l2 == nil { // Should never happen if there's no L1.
   720  		return nil
   721  	}
   722  	ha := l2[ri.l2()]
   723  	if ha == nil {
   724  		return nil
   725  	}
   726  	return ha.spans[(p/pageSize)%pagesPerArena]
   727  }
   728  
   729  // spanOfUnchecked is equivalent to spanOf, but the caller must ensure
   730  // that p points into an allocated heap arena.
   731  //
   732  // Must be nosplit because it has callers that are nosplit.
   733  //
   734  //go:nosplit
   735  func spanOfUnchecked(p uintptr) *mspan {
   736  	ai := arenaIndex(p)
   737  	return mheap_.arenas[ai.l1()][ai.l2()].spans[(p/pageSize)%pagesPerArena]
   738  }
   739  
   740  // spanOfHeap is like spanOf, but returns nil if p does not point to a
   741  // heap object.
   742  //
   743  // Must be nosplit because it has callers that are nosplit.
   744  //
   745  //go:nosplit
   746  func spanOfHeap(p uintptr) *mspan {
   747  	s := spanOf(p)
   748  	// s is nil if it's never been allocated. Otherwise, we check
   749  	// its state first because we don't trust this pointer, so we
   750  	// have to synchronize with span initialization. Then, it's
   751  	// still possible we picked up a stale span pointer, so we
   752  	// have to check the span's bounds.
   753  	if s == nil || s.state.get() != mSpanInUse || p < s.base() || p >= s.limit {
   754  		return nil
   755  	}
   756  	return s
   757  }
   758  
   759  // pageIndexOf returns the arena, page index, and page mask for pointer p.
   760  // The caller must ensure p is in the heap.
   761  func pageIndexOf(p uintptr) (arena *heapArena, pageIdx uintptr, pageMask uint8) {
   762  	ai := arenaIndex(p)
   763  	arena = mheap_.arenas[ai.l1()][ai.l2()]
   764  	pageIdx = ((p / pageSize) / 8) % uintptr(len(arena.pageInUse))
   765  	pageMask = byte(1 << ((p / pageSize) % 8))
   766  	return
   767  }
   768  
   769  // heapArenaOf returns the heap arena for p, if one exists.
   770  func heapArenaOf(p uintptr) *heapArena {
   771  	ri := arenaIndex(p)
   772  	if arenaL1Bits == 0 {
   773  		// If there's no L1, then ri.l1() can't be out of bounds but ri.l2() can.
   774  		if ri.l2() >= uint(len(mheap_.arenas[0])) {
   775  			return nil
   776  		}
   777  	} else {
   778  		// If there's an L1, then ri.l1() can be out of bounds but ri.l2() can't.
   779  		if ri.l1() >= uint(len(mheap_.arenas)) {
   780  			return nil
   781  		}
   782  	}
   783  	l2 := mheap_.arenas[ri.l1()]
   784  	if arenaL1Bits != 0 && l2 == nil { // Should never happen if there's no L1.
   785  		return nil
   786  	}
   787  	return l2[ri.l2()]
   788  }
   789  
   790  // Initialize the heap.
   791  func (h *mheap) init() {
   792  	lockInit(&h.lock, lockRankMheap)
   793  	lockInit(&h.speciallock, lockRankMheapSpecial)
   794  
   795  	h.spanalloc.init(unsafe.Sizeof(mspan{}), recordspan, unsafe.Pointer(h), &memstats.mspan_sys)
   796  	h.cachealloc.init(unsafe.Sizeof(mcache{}), nil, nil, &memstats.mcache_sys)
   797  	h.specialfinalizeralloc.init(unsafe.Sizeof(specialfinalizer{}), nil, nil, &memstats.other_sys)
   798  	h.specialCleanupAlloc.init(unsafe.Sizeof(specialCleanup{}), nil, nil, &memstats.other_sys)
   799  	h.specialCheckFinalizerAlloc.init(unsafe.Sizeof(specialCheckFinalizer{}), nil, nil, &memstats.other_sys)
   800  	h.specialTinyBlockAlloc.init(unsafe.Sizeof(specialTinyBlock{}), nil, nil, &memstats.other_sys)
   801  	h.specialprofilealloc.init(unsafe.Sizeof(specialprofile{}), nil, nil, &memstats.other_sys)
   802  	h.specialReachableAlloc.init(unsafe.Sizeof(specialReachable{}), nil, nil, &memstats.other_sys)
   803  	h.specialPinCounterAlloc.init(unsafe.Sizeof(specialPinCounter{}), nil, nil, &memstats.other_sys)
   804  	h.specialWeakHandleAlloc.init(unsafe.Sizeof(specialWeakHandle{}), nil, nil, &memstats.gcMiscSys)
   805  	h.specialBubbleAlloc.init(unsafe.Sizeof(specialBubble{}), nil, nil, &memstats.other_sys)
   806  	h.arenaHintAlloc.init(unsafe.Sizeof(arenaHint{}), nil, nil, &memstats.other_sys)
   807  
   808  	// Don't zero mspan allocations. Background sweeping can
   809  	// inspect a span concurrently with allocating it, so it's
   810  	// important that the span's sweepgen survive across freeing
   811  	// and re-allocating a span to prevent background sweeping
   812  	// from improperly cas'ing it from 0.
   813  	//
   814  	// This is safe because mspan contains no heap pointers.
   815  	h.spanalloc.zero = false
   816  
   817  	// h->mapcache needs no init
   818  
   819  	for i := range h.central {
   820  		h.central[i].mcentral.init(spanClass(i))
   821  	}
   822  
   823  	h.pages.init(&h.lock, &memstats.gcMiscSys, false)
   824  
   825  	xRegInitAlloc()
   826  }
   827  
   828  // reclaim sweeps and reclaims at least npage pages into the heap.
   829  // It is called before allocating npage pages to keep growth in check.
   830  //
   831  // reclaim implements the page-reclaimer half of the sweeper.
   832  //
   833  // h.lock must NOT be held.
   834  func (h *mheap) reclaim(npage uintptr) {
   835  	// TODO(austin): Half of the time spent freeing spans is in
   836  	// locking/unlocking the heap (even with low contention). We
   837  	// could make the slow path here several times faster by
   838  	// batching heap frees.
   839  
   840  	// Bail early if there's no more reclaim work.
   841  	if h.reclaimIndex.Load() >= 1<<63 {
   842  		return
   843  	}
   844  
   845  	// Disable preemption so the GC can't start while we're
   846  	// sweeping, so we can read h.sweepArenas, and so
   847  	// traceGCSweepStart/Done pair on the P.
   848  	mp := acquirem()
   849  
   850  	trace := traceAcquire()
   851  	if trace.ok() {
   852  		trace.GCSweepStart()
   853  		traceRelease(trace)
   854  	}
   855  
   856  	arenas := h.sweepArenas
   857  	locked := false
   858  	for npage > 0 {
   859  		// Pull from accumulated credit first.
   860  		if credit := h.reclaimCredit.Load(); credit > 0 {
   861  			take := credit
   862  			if take > npage {
   863  				// Take only what we need.
   864  				take = npage
   865  			}
   866  			if h.reclaimCredit.CompareAndSwap(credit, credit-take) {
   867  				npage -= take
   868  			}
   869  			continue
   870  		}
   871  
   872  		// Claim a chunk of work.
   873  		idx := uintptr(h.reclaimIndex.Add(pagesPerReclaimerChunk) - pagesPerReclaimerChunk)
   874  		if idx/pagesPerArena >= uintptr(len(arenas)) {
   875  			// Page reclaiming is done.
   876  			h.reclaimIndex.Store(1 << 63)
   877  			break
   878  		}
   879  
   880  		if !locked {
   881  			// Lock the heap for reclaimChunk.
   882  			lock(&h.lock)
   883  			locked = true
   884  		}
   885  
   886  		// Scan this chunk.
   887  		nfound := h.reclaimChunk(arenas, idx, pagesPerReclaimerChunk)
   888  		if nfound <= npage {
   889  			npage -= nfound
   890  		} else {
   891  			// Put spare pages toward global credit.
   892  			h.reclaimCredit.Add(nfound - npage)
   893  			npage = 0
   894  		}
   895  	}
   896  	if locked {
   897  		unlock(&h.lock)
   898  	}
   899  
   900  	trace = traceAcquire()
   901  	if trace.ok() {
   902  		trace.GCSweepDone()
   903  		traceRelease(trace)
   904  	}
   905  	releasem(mp)
   906  }
   907  
   908  // reclaimChunk sweeps unmarked spans that start at page indexes [pageIdx, pageIdx+n).
   909  // It returns the number of pages returned to the heap.
   910  //
   911  // h.lock must be held and the caller must be non-preemptible. Note: h.lock may be
   912  // temporarily unlocked and re-locked in order to do sweeping or if tracing is
   913  // enabled.
   914  func (h *mheap) reclaimChunk(arenas []arenaIdx, pageIdx, n uintptr) uintptr {
   915  	// The heap lock must be held because this accesses the
   916  	// heapArena.spans arrays using potentially non-live pointers.
   917  	// In particular, if a span were freed and merged concurrently
   918  	// with this probing heapArena.spans, it would be possible to
   919  	// observe arbitrary, stale span pointers.
   920  	assertLockHeld(&h.lock)
   921  
   922  	n0 := n
   923  	var nFreed uintptr
   924  	sl := sweep.active.begin()
   925  	if !sl.valid {
   926  		return 0
   927  	}
   928  	for n > 0 {
   929  		ai := arenas[pageIdx/pagesPerArena]
   930  		ha := h.arenas[ai.l1()][ai.l2()]
   931  
   932  		// Get a chunk of the bitmap to work on.
   933  		arenaPage := uint(pageIdx % pagesPerArena)
   934  		inUse := ha.pageInUse[arenaPage/8:]
   935  		marked := ha.pageMarks[arenaPage/8:]
   936  		if uintptr(len(inUse)) > n/8 {
   937  			inUse = inUse[:n/8]
   938  			marked = marked[:n/8]
   939  		}
   940  
   941  		// Scan this bitmap chunk for spans that are in-use
   942  		// but have no marked objects on them.
   943  		for i := range inUse {
   944  			inUseUnmarked := atomic.Load8(&inUse[i]) &^ marked[i]
   945  			if inUseUnmarked == 0 {
   946  				continue
   947  			}
   948  
   949  			for j := uint(0); j < 8; j++ {
   950  				if inUseUnmarked&(1<<j) != 0 {
   951  					s := ha.spans[arenaPage+uint(i)*8+j]
   952  					if s, ok := sl.tryAcquire(s); ok {
   953  						npages := s.npages
   954  						unlock(&h.lock)
   955  						if s.sweep(false) {
   956  							nFreed += npages
   957  						}
   958  						lock(&h.lock)
   959  						// Reload inUse. It's possible nearby
   960  						// spans were freed when we dropped the
   961  						// lock and we don't want to get stale
   962  						// pointers from the spans array.
   963  						inUseUnmarked = atomic.Load8(&inUse[i]) &^ marked[i]
   964  					}
   965  				}
   966  			}
   967  		}
   968  
   969  		// Advance.
   970  		pageIdx += uintptr(len(inUse) * 8)
   971  		n -= uintptr(len(inUse) * 8)
   972  	}
   973  	sweep.active.end(sl)
   974  	trace := traceAcquire()
   975  	if trace.ok() {
   976  		unlock(&h.lock)
   977  		// Account for pages scanned but not reclaimed.
   978  		trace.GCSweepSpan((n0 - nFreed) * pageSize)
   979  		traceRelease(trace)
   980  		lock(&h.lock)
   981  	}
   982  
   983  	assertLockHeld(&h.lock) // Must be locked on return.
   984  	return nFreed
   985  }
   986  
   987  // spanAllocType represents the type of allocation to make, or
   988  // the type of allocation to be freed.
   989  type spanAllocType uint8
   990  
   991  const (
   992  	spanAllocHeap    spanAllocType = iota // heap span
   993  	spanAllocStack                        // stack span
   994  	spanAllocWorkBuf                      // work buf span
   995  )
   996  
   997  // manual returns true if the span allocation is manually managed.
   998  func (s spanAllocType) manual() bool {
   999  	return s != spanAllocHeap
  1000  }
  1001  
  1002  // alloc allocates a new span of npage pages from the GC'd heap.
  1003  //
  1004  // spanclass indicates the span's size class and scannability.
  1005  //
  1006  // Returns a span that has been fully initialized. span.needzero indicates
  1007  // whether the span has been zeroed. Note that it may not be.
  1008  func (h *mheap) alloc(npages uintptr, spanclass spanClass) *mspan {
  1009  	// Don't do any operations that lock the heap on the G stack.
  1010  	// It might trigger stack growth, and the stack growth code needs
  1011  	// to be able to allocate heap.
  1012  	var s *mspan
  1013  	systemstack(func() {
  1014  		// To prevent excessive heap growth, before allocating n pages
  1015  		// we need to sweep and reclaim at least n pages.
  1016  		if !isSweepDone() {
  1017  			h.reclaim(npages)
  1018  		}
  1019  		s = h.allocSpan(npages, spanAllocHeap, spanclass)
  1020  	})
  1021  	return s
  1022  }
  1023  
  1024  // allocManual allocates a manually-managed span of npage pages.
  1025  // allocManual returns nil if allocation fails.
  1026  //
  1027  // allocManual adds the bytes used to *stat, which should be a
  1028  // memstats in-use field. Unlike allocations in the GC'd heap, the
  1029  // allocation does *not* count toward heapInUse.
  1030  //
  1031  // The memory backing the returned span may not be zeroed if
  1032  // span.needzero is set.
  1033  //
  1034  // allocManual must be called on the system stack because it may
  1035  // acquire the heap lock via allocSpan. See mheap for details.
  1036  //
  1037  // If new code is written to call allocManual, do NOT use an
  1038  // existing spanAllocType value and instead declare a new one.
  1039  //
  1040  //go:systemstack
  1041  func (h *mheap) allocManual(npages uintptr, typ spanAllocType) *mspan {
  1042  	if !typ.manual() {
  1043  		throw("manual span allocation called with non-manually-managed type")
  1044  	}
  1045  	return h.allocSpan(npages, typ, 0)
  1046  }
  1047  
  1048  // setSpans modifies the span map so [spanOf(base), spanOf(base+npage*pageSize))
  1049  // is s.
  1050  func (h *mheap) setSpans(base, npage uintptr, s *mspan) {
  1051  	p := base / pageSize
  1052  	ai := arenaIndex(base)
  1053  	ha := h.arenas[ai.l1()][ai.l2()]
  1054  	for n := uintptr(0); n < npage; n++ {
  1055  		i := (p + n) % pagesPerArena
  1056  		if i == 0 {
  1057  			ai = arenaIndex(base + n*pageSize)
  1058  			ha = h.arenas[ai.l1()][ai.l2()]
  1059  		}
  1060  		ha.spans[i] = s
  1061  	}
  1062  }
  1063  
  1064  // allocNeedsZero checks if the region of address space [base, base+npage*pageSize),
  1065  // assumed to be allocated, needs to be zeroed, updating heap arena metadata for
  1066  // future allocations.
  1067  //
  1068  // This must be called each time pages are allocated from the heap, even if the page
  1069  // allocator can otherwise prove the memory it's allocating is already zero because
  1070  // they're fresh from the operating system. It updates heapArena metadata that is
  1071  // critical for future page allocations.
  1072  //
  1073  // There are no locking constraints on this method.
  1074  func (h *mheap) allocNeedsZero(base, npage uintptr) (needZero bool) {
  1075  	for npage > 0 {
  1076  		ai := arenaIndex(base)
  1077  		ha := h.arenas[ai.l1()][ai.l2()]
  1078  
  1079  		zeroedBase := atomic.Loaduintptr(&ha.zeroedBase)
  1080  		arenaBase := base % heapArenaBytes
  1081  		if arenaBase < zeroedBase {
  1082  			// We extended into the non-zeroed part of the
  1083  			// arena, so this region needs to be zeroed before use.
  1084  			//
  1085  			// zeroedBase is monotonically increasing, so if we see this now then
  1086  			// we can be sure we need to zero this memory region.
  1087  			//
  1088  			// We still need to update zeroedBase for this arena, and
  1089  			// potentially more arenas.
  1090  			needZero = true
  1091  		}
  1092  		// We may observe arenaBase > zeroedBase if we're racing with one or more
  1093  		// allocations which are acquiring memory directly before us in the address
  1094  		// space. But, because we know no one else is acquiring *this* memory, it's
  1095  		// still safe to not zero.
  1096  
  1097  		// Compute how far into the arena we extend into, capped
  1098  		// at heapArenaBytes.
  1099  		arenaLimit := arenaBase + npage*pageSize
  1100  		if arenaLimit > heapArenaBytes {
  1101  			arenaLimit = heapArenaBytes
  1102  		}
  1103  		// Increase ha.zeroedBase so it's >= arenaLimit.
  1104  		// We may be racing with other updates.
  1105  		for arenaLimit > zeroedBase {
  1106  			if atomic.Casuintptr(&ha.zeroedBase, zeroedBase, arenaLimit) {
  1107  				break
  1108  			}
  1109  			zeroedBase = atomic.Loaduintptr(&ha.zeroedBase)
  1110  			// Double check basic conditions of zeroedBase.
  1111  			if zeroedBase <= arenaLimit && zeroedBase > arenaBase {
  1112  				// The zeroedBase moved into the space we were trying to
  1113  				// claim. That's very bad, and indicates someone allocated
  1114  				// the same region we did.
  1115  				throw("potentially overlapping in-use allocations detected")
  1116  			}
  1117  		}
  1118  
  1119  		// Move base forward and subtract from npage to move into
  1120  		// the next arena, or finish.
  1121  		base += arenaLimit - arenaBase
  1122  		npage -= (arenaLimit - arenaBase) / pageSize
  1123  	}
  1124  	return
  1125  }
  1126  
  1127  // tryAllocMSpan attempts to allocate an mspan object from
  1128  // the P-local cache, but may fail.
  1129  //
  1130  // h.lock need not be held.
  1131  //
  1132  // This caller must ensure that its P won't change underneath
  1133  // it during this function. Currently to ensure that we enforce
  1134  // that the function is run on the system stack, because that's
  1135  // the only place it is used now. In the future, this requirement
  1136  // may be relaxed if its use is necessary elsewhere.
  1137  //
  1138  //go:systemstack
  1139  func (h *mheap) tryAllocMSpan() *mspan {
  1140  	pp := getg().m.p.ptr()
  1141  	// If we don't have a p or the cache is empty, we can't do
  1142  	// anything here.
  1143  	if pp == nil || pp.mspancache.len == 0 {
  1144  		return nil
  1145  	}
  1146  	// Pull off the last entry in the cache.
  1147  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1148  	pp.mspancache.len--
  1149  	return s
  1150  }
  1151  
  1152  // allocMSpanLocked allocates an mspan object.
  1153  //
  1154  // h.lock must be held.
  1155  //
  1156  // allocMSpanLocked must be called on the system stack because
  1157  // its caller holds the heap lock. See mheap for details.
  1158  // Running on the system stack also ensures that we won't
  1159  // switch Ps during this function. See tryAllocMSpan for details.
  1160  //
  1161  //go:systemstack
  1162  func (h *mheap) allocMSpanLocked() *mspan {
  1163  	assertLockHeld(&h.lock)
  1164  
  1165  	pp := getg().m.p.ptr()
  1166  	if pp == nil {
  1167  		// We don't have a p so just do the normal thing.
  1168  		return (*mspan)(h.spanalloc.alloc())
  1169  	}
  1170  	// Refill the cache if necessary.
  1171  	if pp.mspancache.len == 0 {
  1172  		const refillCount = len(pp.mspancache.buf) / 2
  1173  		for i := 0; i < refillCount; i++ {
  1174  			pp.mspancache.buf[i] = (*mspan)(h.spanalloc.alloc())
  1175  		}
  1176  		pp.mspancache.len = refillCount
  1177  	}
  1178  	// Pull off the last entry in the cache.
  1179  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1180  	pp.mspancache.len--
  1181  	return s
  1182  }
  1183  
  1184  // freeMSpanLocked free an mspan object.
  1185  //
  1186  // h.lock must be held.
  1187  //
  1188  // freeMSpanLocked must be called on the system stack because
  1189  // its caller holds the heap lock. See mheap for details.
  1190  // Running on the system stack also ensures that we won't
  1191  // switch Ps during this function. See tryAllocMSpan for details.
  1192  //
  1193  //go:systemstack
  1194  func (h *mheap) freeMSpanLocked(s *mspan) {
  1195  	assertLockHeld(&h.lock)
  1196  
  1197  	pp := getg().m.p.ptr()
  1198  	// First try to free the mspan directly to the cache.
  1199  	if pp != nil && pp.mspancache.len < len(pp.mspancache.buf) {
  1200  		pp.mspancache.buf[pp.mspancache.len] = s
  1201  		pp.mspancache.len++
  1202  		return
  1203  	}
  1204  	// Failing that (or if we don't have a p), just free it to
  1205  	// the heap.
  1206  	h.spanalloc.free(unsafe.Pointer(s))
  1207  }
  1208  
  1209  // allocSpan allocates an mspan which owns npages worth of memory.
  1210  //
  1211  // If typ.manual() == false, allocSpan allocates a heap span of class spanclass
  1212  // and updates heap accounting. If manual == true, allocSpan allocates a
  1213  // manually-managed span (spanclass is ignored), and the caller is
  1214  // responsible for any accounting related to its use of the span. Either
  1215  // way, allocSpan will atomically add the bytes in the newly allocated
  1216  // span to *sysStat.
  1217  //
  1218  // The returned span is fully initialized.
  1219  //
  1220  // h.lock must not be held.
  1221  //
  1222  // allocSpan must be called on the system stack both because it acquires
  1223  // the heap lock and because it must block GC transitions.
  1224  //
  1225  //go:systemstack
  1226  func (h *mheap) allocSpan(npages uintptr, typ spanAllocType, spanclass spanClass) (s *mspan) {
  1227  	// Function-global state.
  1228  	gp := getg()
  1229  	base, scav := uintptr(0), uintptr(0)
  1230  	growth := uintptr(0)
  1231  
  1232  	// On some platforms we need to provide physical page aligned stack
  1233  	// allocations. Where the page size is less than the physical page
  1234  	// size, we already manage to do this by default.
  1235  	needPhysPageAlign := physPageAlignedStacks && typ == spanAllocStack && pageSize < physPageSize
  1236  
  1237  	// If the allocation is small enough, try the page cache!
  1238  	// The page cache does not support aligned allocations, so we cannot use
  1239  	// it if we need to provide a physical page aligned stack allocation.
  1240  	pp := gp.m.p.ptr()
  1241  	if !needPhysPageAlign && pp != nil && npages < pageCachePages/4 {
  1242  		c := &pp.pcache
  1243  
  1244  		// If the cache is empty, refill it.
  1245  		if c.empty() {
  1246  			lock(&h.lock)
  1247  			*c = h.pages.allocToCache()
  1248  			unlock(&h.lock)
  1249  		}
  1250  
  1251  		// Try to allocate from the cache.
  1252  		base, scav = c.alloc(npages)
  1253  		if base != 0 {
  1254  			s = h.tryAllocMSpan()
  1255  			if s != nil {
  1256  				goto HaveSpan
  1257  			}
  1258  			// We have a base but no mspan, so we need
  1259  			// to lock the heap.
  1260  		}
  1261  	}
  1262  
  1263  	// For one reason or another, we couldn't get the
  1264  	// whole job done without the heap lock.
  1265  	lock(&h.lock)
  1266  
  1267  	if needPhysPageAlign {
  1268  		// Overallocate by a physical page to allow for later alignment.
  1269  		extraPages := physPageSize / pageSize
  1270  
  1271  		// Find a big enough region first, but then only allocate the
  1272  		// aligned portion. We can't just allocate and then free the
  1273  		// edges because we need to account for scavenged memory, and
  1274  		// that's difficult with alloc.
  1275  		//
  1276  		// Note that we skip updates to searchAddr here. It's OK if
  1277  		// it's stale and higher than normal; it'll operate correctly,
  1278  		// just come with a performance cost.
  1279  		base, _ = h.pages.find(npages + extraPages)
  1280  		if base == 0 {
  1281  			var ok bool
  1282  			growth, ok = h.grow(npages + extraPages)
  1283  			if !ok {
  1284  				unlock(&h.lock)
  1285  				return nil
  1286  			}
  1287  			base, _ = h.pages.find(npages + extraPages)
  1288  			if base == 0 {
  1289  				throw("grew heap, but no adequate free space found")
  1290  			}
  1291  		}
  1292  		base = alignUp(base, physPageSize)
  1293  		scav = h.pages.allocRange(base, npages)
  1294  	}
  1295  
  1296  	if base == 0 {
  1297  		// Try to acquire a base address.
  1298  		base, scav = h.pages.alloc(npages)
  1299  		if base == 0 {
  1300  			var ok bool
  1301  			growth, ok = h.grow(npages)
  1302  			if !ok {
  1303  				unlock(&h.lock)
  1304  				return nil
  1305  			}
  1306  			base, scav = h.pages.alloc(npages)
  1307  			if base == 0 {
  1308  				throw("grew heap, but no adequate free space found")
  1309  			}
  1310  		}
  1311  	}
  1312  	if s == nil {
  1313  		// We failed to get an mspan earlier, so grab
  1314  		// one now that we have the heap lock.
  1315  		s = h.allocMSpanLocked()
  1316  	}
  1317  	unlock(&h.lock)
  1318  
  1319  HaveSpan:
  1320  	// Decide if we need to scavenge in response to what we just allocated.
  1321  	// Specifically, we track the maximum amount of memory to scavenge of all
  1322  	// the alternatives below, assuming that the maximum satisfies *all*
  1323  	// conditions we check (e.g. if we need to scavenge X to satisfy the
  1324  	// memory limit and Y to satisfy heap-growth scavenging, and Y > X, then
  1325  	// it's fine to pick Y, because the memory limit is still satisfied).
  1326  	//
  1327  	// It's fine to do this after allocating because we expect any scavenged
  1328  	// pages not to get touched until we return. Simultaneously, it's important
  1329  	// to do this before calling sysUsed because that may commit address space.
  1330  	bytesToScavenge := uintptr(0)
  1331  	forceScavenge := false
  1332  	if limit := gcController.memoryLimit.Load(); !gcCPULimiter.limiting() {
  1333  		// Assist with scavenging to maintain the memory limit by the amount
  1334  		// that we expect to page in.
  1335  		inuse := gcController.mappedReady.Load()
  1336  		// Be careful about overflow, especially with uintptrs. Even on 32-bit platforms
  1337  		// someone can set a really big memory limit that isn't math.MaxInt64.
  1338  		if uint64(scav)+inuse > uint64(limit) {
  1339  			bytesToScavenge = uintptr(uint64(scav) + inuse - uint64(limit))
  1340  			forceScavenge = true
  1341  		}
  1342  	}
  1343  	if goal := scavenge.gcPercentGoal.Load(); goal != ^uint64(0) && growth > 0 {
  1344  		// We just caused a heap growth, so scavenge down what will soon be used.
  1345  		// By scavenging inline we deal with the failure to allocate out of
  1346  		// memory fragments by scavenging the memory fragments that are least
  1347  		// likely to be re-used.
  1348  		//
  1349  		// Only bother with this because we're not using a memory limit. We don't
  1350  		// care about heap growths as long as we're under the memory limit, and the
  1351  		// previous check for scaving already handles that.
  1352  		if retained := heapRetained(); retained+uint64(growth) > goal {
  1353  			// The scavenging algorithm requires the heap lock to be dropped so it
  1354  			// can acquire it only sparingly. This is a potentially expensive operation
  1355  			// so it frees up other goroutines to allocate in the meanwhile. In fact,
  1356  			// they can make use of the growth we just created.
  1357  			todo := growth
  1358  			if overage := uintptr(retained + uint64(growth) - goal); todo > overage {
  1359  				todo = overage
  1360  			}
  1361  			if todo > bytesToScavenge {
  1362  				bytesToScavenge = todo
  1363  			}
  1364  		}
  1365  	}
  1366  	// There are a few very limited circumstances where we won't have a P here.
  1367  	// It's OK to simply skip scavenging in these cases. Something else will notice
  1368  	// and pick up the tab.
  1369  	var now int64
  1370  	if pp != nil && bytesToScavenge > 0 {
  1371  		// Measure how long we spent scavenging and add that measurement to the assist
  1372  		// time so we can track it for the GC CPU limiter.
  1373  		//
  1374  		// Limiter event tracking might be disabled if we end up here
  1375  		// while on a mark worker.
  1376  		start := nanotime()
  1377  		track := pp.limiterEvent.start(limiterEventScavengeAssist, start)
  1378  
  1379  		// Scavenge, but back out if the limiter turns on.
  1380  		released := h.pages.scavenge(bytesToScavenge, func() bool {
  1381  			return gcCPULimiter.limiting()
  1382  		}, forceScavenge)
  1383  
  1384  		mheap_.pages.scav.releasedEager.Add(released)
  1385  
  1386  		// Finish up accounting.
  1387  		now = nanotime()
  1388  		if track {
  1389  			pp.limiterEvent.stop(limiterEventScavengeAssist, now)
  1390  		}
  1391  		scavenge.assistTime.Add(now - start)
  1392  	}
  1393  
  1394  	// Initialize the span.
  1395  	h.initSpan(s, typ, spanclass, base, npages)
  1396  
  1397  	if valgrindenabled {
  1398  		valgrindMempoolMalloc(unsafe.Pointer(arenaBase(arenaIndex(base))), unsafe.Pointer(base), npages*pageSize)
  1399  	}
  1400  
  1401  	// Commit and account for any scavenged memory that the span now owns.
  1402  	nbytes := npages * pageSize
  1403  	if scav != 0 {
  1404  		// sysUsed all the pages that are actually available
  1405  		// in the span since some of them might be scavenged.
  1406  		sysUsed(unsafe.Pointer(base), nbytes, scav)
  1407  		gcController.heapReleased.add(-int64(scav))
  1408  	}
  1409  	// Update stats.
  1410  	gcController.heapFree.add(-int64(nbytes - scav))
  1411  	if typ == spanAllocHeap {
  1412  		gcController.heapInUse.add(int64(nbytes))
  1413  	}
  1414  	// Update consistent stats.
  1415  	stats := memstats.heapStats.acquire()
  1416  	atomic.Xaddint64(&stats.committed, int64(scav))
  1417  	atomic.Xaddint64(&stats.released, -int64(scav))
  1418  	switch typ {
  1419  	case spanAllocHeap:
  1420  		atomic.Xaddint64(&stats.inHeap, int64(nbytes))
  1421  	case spanAllocStack:
  1422  		atomic.Xaddint64(&stats.inStacks, int64(nbytes))
  1423  	case spanAllocWorkBuf:
  1424  		atomic.Xaddint64(&stats.inWorkBufs, int64(nbytes))
  1425  	}
  1426  	memstats.heapStats.release()
  1427  
  1428  	// Trace the span alloc.
  1429  	if traceAllocFreeEnabled() {
  1430  		trace := traceAcquire()
  1431  		if trace.ok() {
  1432  			trace.SpanAlloc(s)
  1433  			traceRelease(trace)
  1434  		}
  1435  	}
  1436  	return s
  1437  }
  1438  
  1439  // initSpan initializes a blank span s which will represent the range
  1440  // [base, base+npages*pageSize). typ is the type of span being allocated.
  1441  func (h *mheap) initSpan(s *mspan, typ spanAllocType, spanclass spanClass, base, npages uintptr) {
  1442  	// At this point, both s != nil and base != 0, and the heap
  1443  	// lock is no longer held. Initialize the span.
  1444  	s.init(base, npages)
  1445  	if h.allocNeedsZero(base, npages) {
  1446  		s.needzero = 1
  1447  	}
  1448  	nbytes := npages * pageSize
  1449  	if typ.manual() {
  1450  		s.manualFreeList = 0
  1451  		s.nelems = 0
  1452  		s.state.set(mSpanManual)
  1453  	} else {
  1454  		// We must set span properties before the span is published anywhere
  1455  		// since we're not holding the heap lock.
  1456  		s.spanclass = spanclass
  1457  		if sizeclass := spanclass.sizeclass(); sizeclass == 0 {
  1458  			s.elemsize = nbytes
  1459  			s.nelems = 1
  1460  			s.divMul = 0
  1461  		} else {
  1462  			s.elemsize = uintptr(gc.SizeClassToSize[sizeclass])
  1463  			if goexperiment.GreenTeaGC {
  1464  				var reserve uintptr
  1465  				if gcUsesSpanInlineMarkBits(s.elemsize) {
  1466  					// Reserve space for the inline mark bits.
  1467  					reserve += unsafe.Sizeof(spanInlineMarkBits{})
  1468  				}
  1469  				if heapBitsInSpan(s.elemsize) && !s.spanclass.noscan() {
  1470  					// Reserve space for the pointer/scan bitmap at the end.
  1471  					reserve += nbytes / goarch.PtrSize / 8
  1472  				}
  1473  				s.nelems = uint16((nbytes - reserve) / s.elemsize)
  1474  			} else {
  1475  				if !s.spanclass.noscan() && heapBitsInSpan(s.elemsize) {
  1476  					// Reserve space for the pointer/scan bitmap at the end.
  1477  					s.nelems = uint16((nbytes - (nbytes / goarch.PtrSize / 8)) / s.elemsize)
  1478  				} else {
  1479  					s.nelems = uint16(nbytes / s.elemsize)
  1480  				}
  1481  			}
  1482  			s.divMul = gc.SizeClassToDivMagic[sizeclass]
  1483  		}
  1484  
  1485  		// Initialize mark and allocation structures.
  1486  		s.freeindex = 0
  1487  		s.freeIndexForScan = 0
  1488  		s.allocCache = ^uint64(0) // all 1s indicating all free.
  1489  		s.gcmarkBits = newMarkBits(uintptr(s.nelems))
  1490  		s.allocBits = newAllocBits(uintptr(s.nelems))
  1491  
  1492  		// Adjust s.limit down to the object-containing part of the span.
  1493  		s.limit = s.base() + s.elemsize*uintptr(s.nelems)
  1494  
  1495  		// It's safe to access h.sweepgen without the heap lock because it's
  1496  		// only ever updated with the world stopped and we run on the
  1497  		// systemstack which blocks a STW transition.
  1498  		atomic.Store(&s.sweepgen, h.sweepgen)
  1499  
  1500  		// Now that the span is filled in, set its state. This
  1501  		// is a publication barrier for the other fields in
  1502  		// the span. While valid pointers into this span
  1503  		// should never be visible until the span is returned,
  1504  		// if the garbage collector finds an invalid pointer,
  1505  		// access to the span may race with initialization of
  1506  		// the span. We resolve this race by atomically
  1507  		// setting the state after the span is fully
  1508  		// initialized, and atomically checking the state in
  1509  		// any situation where a pointer is suspect.
  1510  		s.state.set(mSpanInUse)
  1511  	}
  1512  
  1513  	// Publish the span in various locations.
  1514  
  1515  	// This is safe to call without the lock held because the slots
  1516  	// related to this span will only ever be read or modified by
  1517  	// this thread until pointers into the span are published (and
  1518  	// we execute a publication barrier at the end of this function
  1519  	// before that happens) or pageInUse is updated.
  1520  	h.setSpans(s.base(), npages, s)
  1521  
  1522  	if !typ.manual() {
  1523  		// Mark in-use span in arena page bitmap.
  1524  		//
  1525  		// This publishes the span to the page sweeper, so
  1526  		// it's imperative that the span be completely initialized
  1527  		// prior to this line.
  1528  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1529  		atomic.Or8(&arena.pageInUse[pageIdx], pageMask)
  1530  
  1531  		// Mark packed span.
  1532  		if gcUsesSpanInlineMarkBits(s.elemsize) {
  1533  			atomic.Or8(&arena.pageUseSpanInlineMarkBits[pageIdx], pageMask)
  1534  		}
  1535  
  1536  		// Update related page sweeper stats.
  1537  		h.pagesInUse.Add(npages)
  1538  	}
  1539  
  1540  	// Make sure the newly allocated span will be observed
  1541  	// by the GC before pointers into the span are published.
  1542  	publicationBarrier()
  1543  }
  1544  
  1545  // Try to add at least npage pages of memory to the heap,
  1546  // returning how much the heap grew by and whether it worked.
  1547  //
  1548  // h.lock must be held.
  1549  func (h *mheap) grow(npage uintptr) (uintptr, bool) {
  1550  	assertLockHeld(&h.lock)
  1551  
  1552  	firstGrow := h.curArena.base == 0
  1553  
  1554  	// We must grow the heap in whole palloc chunks.
  1555  	// We call sysMap below but note that because we
  1556  	// round up to pallocChunkPages which is on the order
  1557  	// of MiB (generally >= to the huge page size) we
  1558  	// won't be calling it too much.
  1559  	ask := alignUp(npage, pallocChunkPages) * pageSize
  1560  
  1561  	totalGrowth := uintptr(0)
  1562  	// This may overflow because ask could be very large
  1563  	// and is otherwise unrelated to h.curArena.base.
  1564  	end := h.curArena.base + ask
  1565  	nBase := alignUp(end, physPageSize)
  1566  	if nBase > h.curArena.end || /* overflow */ end < h.curArena.base {
  1567  		// Not enough room in the current arena. Allocate more
  1568  		// arena space. This may not be contiguous with the
  1569  		// current arena, so we have to request the full ask.
  1570  		av, asize := h.sysAlloc(ask, &h.arenaHints, &h.heapArenas)
  1571  		if av == nil {
  1572  			inUse := gcController.heapFree.load() + gcController.heapReleased.load() + gcController.heapInUse.load()
  1573  			print("runtime: out of memory: cannot allocate ", ask, "-byte block (", inUse, " in use)\n")
  1574  			return 0, false
  1575  		}
  1576  
  1577  		if uintptr(av) == h.curArena.end {
  1578  			// The new space is contiguous with the old
  1579  			// space, so just extend the current space.
  1580  			h.curArena.end = uintptr(av) + asize
  1581  		} else {
  1582  			// The new space is discontiguous. Track what
  1583  			// remains of the current space and switch to
  1584  			// the new space. This should be rare.
  1585  			if size := h.curArena.end - h.curArena.base; size != 0 {
  1586  				// Transition this space from Reserved to Prepared and mark it
  1587  				// as released since we'll be able to start using it after updating
  1588  				// the page allocator and releasing the lock at any time.
  1589  				sysMap(unsafe.Pointer(h.curArena.base), size, &gcController.heapReleased, "heap")
  1590  				// Update stats.
  1591  				stats := memstats.heapStats.acquire()
  1592  				atomic.Xaddint64(&stats.released, int64(size))
  1593  				memstats.heapStats.release()
  1594  				// Update the page allocator's structures to make this
  1595  				// space ready for allocation.
  1596  				h.pages.grow(h.curArena.base, size)
  1597  				totalGrowth += size
  1598  			}
  1599  			// Switch to the new space.
  1600  			h.curArena.base = uintptr(av)
  1601  			h.curArena.end = uintptr(av) + asize
  1602  
  1603  			if firstGrow && randomizeHeapBase {
  1604  				// The top heapAddrBits-logHeapArenaBytes are randomized, we now
  1605  				// want to randomize the next
  1606  				// logHeapArenaBytes-log2(pallocChunkBytes) bits, making sure
  1607  				// h.curArena.base is aligned to pallocChunkBytes.
  1608  				bits := logHeapArenaBytes - logPallocChunkBytes
  1609  				offset := nextHeapRandBits(bits)
  1610  				h.curArena.base = alignDown(h.curArena.base|(offset<<logPallocChunkBytes), pallocChunkBytes)
  1611  			}
  1612  		}
  1613  
  1614  		// Recalculate nBase.
  1615  		// We know this won't overflow, because sysAlloc returned
  1616  		// a valid region starting at h.curArena.base which is at
  1617  		// least ask bytes in size.
  1618  		nBase = alignUp(h.curArena.base+ask, physPageSize)
  1619  	}
  1620  
  1621  	// Grow into the current arena.
  1622  	v := h.curArena.base
  1623  	h.curArena.base = nBase
  1624  
  1625  	// Transition the space we're going to use from Reserved to Prepared.
  1626  	//
  1627  	// The allocation is always aligned to the heap arena
  1628  	// size which is always > physPageSize, so its safe to
  1629  	// just add directly to heapReleased.
  1630  	sysMap(unsafe.Pointer(v), nBase-v, &gcController.heapReleased, "heap")
  1631  
  1632  	// The memory just allocated counts as both released
  1633  	// and idle, even though it's not yet backed by spans.
  1634  	stats := memstats.heapStats.acquire()
  1635  	atomic.Xaddint64(&stats.released, int64(nBase-v))
  1636  	memstats.heapStats.release()
  1637  
  1638  	// Update the page allocator's structures to make this
  1639  	// space ready for allocation.
  1640  	h.pages.grow(v, nBase-v)
  1641  	totalGrowth += nBase - v
  1642  
  1643  	if firstGrow && randomizeHeapBase {
  1644  		// The top heapAddrBits-log2(pallocChunkBytes) bits are now randomized,
  1645  		// we finally want to randomize the next
  1646  		// log2(pallocChunkBytes)-log2(pageSize) bits, while maintaining
  1647  		// alignment to pageSize. We do this by calculating a random number of
  1648  		// pages into the current arena, and marking them as allocated. The
  1649  		// address of the next available page becomes our fully randomized base
  1650  		// heap address.
  1651  		randOffset := nextHeapRandBits(logPallocChunkBytes)
  1652  		randNumPages := alignDown(randOffset, pageSize) / pageSize
  1653  		if randNumPages != 0 {
  1654  			h.pages.markRandomPaddingPages(v, randNumPages)
  1655  		}
  1656  	}
  1657  
  1658  	return totalGrowth, true
  1659  }
  1660  
  1661  // Free the span back into the heap.
  1662  func (h *mheap) freeSpan(s *mspan) {
  1663  	systemstack(func() {
  1664  		// Trace the span free.
  1665  		if traceAllocFreeEnabled() {
  1666  			trace := traceAcquire()
  1667  			if trace.ok() {
  1668  				trace.SpanFree(s)
  1669  				traceRelease(trace)
  1670  			}
  1671  		}
  1672  
  1673  		lock(&h.lock)
  1674  		if msanenabled {
  1675  			// Tell msan that this entire span is no longer in use.
  1676  			base := unsafe.Pointer(s.base())
  1677  			bytes := s.npages << gc.PageShift
  1678  			msanfree(base, bytes)
  1679  		}
  1680  		if asanenabled {
  1681  			// Tell asan that this entire span is no longer in use.
  1682  			base := unsafe.Pointer(s.base())
  1683  			bytes := s.npages << gc.PageShift
  1684  			asanpoison(base, bytes)
  1685  		}
  1686  		if valgrindenabled {
  1687  			base := s.base()
  1688  			valgrindMempoolFree(unsafe.Pointer(arenaBase(arenaIndex(base))), unsafe.Pointer(base))
  1689  		}
  1690  		h.freeSpanLocked(s, spanAllocHeap)
  1691  		unlock(&h.lock)
  1692  	})
  1693  }
  1694  
  1695  // freeManual frees a manually-managed span returned by allocManual.
  1696  // typ must be the same as the spanAllocType passed to the allocManual that
  1697  // allocated s.
  1698  //
  1699  // This must only be called when gcphase == _GCoff. See mSpanState for
  1700  // an explanation.
  1701  //
  1702  // freeManual must be called on the system stack because it acquires
  1703  // the heap lock. See mheap for details.
  1704  //
  1705  //go:systemstack
  1706  func (h *mheap) freeManual(s *mspan, typ spanAllocType) {
  1707  	// Trace the span free.
  1708  	if traceAllocFreeEnabled() {
  1709  		trace := traceAcquire()
  1710  		if trace.ok() {
  1711  			trace.SpanFree(s)
  1712  			traceRelease(trace)
  1713  		}
  1714  	}
  1715  
  1716  	s.needzero = 1
  1717  	lock(&h.lock)
  1718  	if valgrindenabled {
  1719  		base := s.base()
  1720  		valgrindMempoolFree(unsafe.Pointer(arenaBase(arenaIndex(base))), unsafe.Pointer(base))
  1721  	}
  1722  	h.freeSpanLocked(s, typ)
  1723  	unlock(&h.lock)
  1724  }
  1725  
  1726  func (h *mheap) freeSpanLocked(s *mspan, typ spanAllocType) {
  1727  	assertLockHeld(&h.lock)
  1728  
  1729  	switch s.state.get() {
  1730  	case mSpanManual:
  1731  		if s.allocCount != 0 {
  1732  			throw("mheap.freeSpanLocked - invalid stack free")
  1733  		}
  1734  	case mSpanInUse:
  1735  		if s.isUserArenaChunk {
  1736  			throw("mheap.freeSpanLocked - invalid free of user arena chunk")
  1737  		}
  1738  		if s.allocCount != 0 || s.sweepgen != h.sweepgen {
  1739  			print("mheap.freeSpanLocked - span ", s, " ptr ", hex(s.base()), " allocCount ", s.allocCount, " sweepgen ", s.sweepgen, "/", h.sweepgen, "\n")
  1740  			throw("mheap.freeSpanLocked - invalid free")
  1741  		}
  1742  		h.pagesInUse.Add(-s.npages)
  1743  
  1744  		// Clear in-use bit in arena page bitmap.
  1745  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1746  		atomic.And8(&arena.pageInUse[pageIdx], ^pageMask)
  1747  
  1748  		// Clear small heap span bit if necessary.
  1749  		if gcUsesSpanInlineMarkBits(s.elemsize) {
  1750  			atomic.And8(&arena.pageUseSpanInlineMarkBits[pageIdx], ^pageMask)
  1751  		}
  1752  	default:
  1753  		throw("mheap.freeSpanLocked - invalid span state")
  1754  	}
  1755  
  1756  	// Update stats.
  1757  	//
  1758  	// Mirrors the code in allocSpan.
  1759  	nbytes := s.npages * pageSize
  1760  	gcController.heapFree.add(int64(nbytes))
  1761  	if typ == spanAllocHeap {
  1762  		gcController.heapInUse.add(-int64(nbytes))
  1763  	}
  1764  	// Update consistent stats.
  1765  	stats := memstats.heapStats.acquire()
  1766  	switch typ {
  1767  	case spanAllocHeap:
  1768  		atomic.Xaddint64(&stats.inHeap, -int64(nbytes))
  1769  	case spanAllocStack:
  1770  		atomic.Xaddint64(&stats.inStacks, -int64(nbytes))
  1771  	case spanAllocWorkBuf:
  1772  		atomic.Xaddint64(&stats.inWorkBufs, -int64(nbytes))
  1773  	}
  1774  	memstats.heapStats.release()
  1775  
  1776  	// Mark the space as free.
  1777  	h.pages.free(s.base(), s.npages)
  1778  
  1779  	// Free the span structure. We no longer have a use for it.
  1780  	s.state.set(mSpanDead)
  1781  	h.freeMSpanLocked(s)
  1782  }
  1783  
  1784  // scavengeAll acquires the heap lock (blocking any additional
  1785  // manipulation of the page allocator) and iterates over the whole
  1786  // heap, scavenging every free page available.
  1787  //
  1788  // Must run on the system stack because it acquires the heap lock.
  1789  //
  1790  //go:systemstack
  1791  func (h *mheap) scavengeAll() {
  1792  	// Disallow malloc or panic while holding the heap lock. We do
  1793  	// this here because this is a non-mallocgc entry-point to
  1794  	// the mheap API.
  1795  	gp := getg()
  1796  	gp.m.mallocing++
  1797  
  1798  	// Force scavenge everything.
  1799  	released := h.pages.scavenge(^uintptr(0), nil, true)
  1800  
  1801  	gp.m.mallocing--
  1802  
  1803  	if debug.scavtrace > 0 {
  1804  		printScavTrace(0, released, true)
  1805  	}
  1806  }
  1807  
  1808  //go:linkname runtime_debug_freeOSMemory runtime/debug.freeOSMemory
  1809  func runtime_debug_freeOSMemory() {
  1810  	GC()
  1811  	systemstack(func() { mheap_.scavengeAll() })
  1812  }
  1813  
  1814  // Initialize a new span with the given start and npages.
  1815  func (span *mspan) init(base uintptr, npages uintptr) {
  1816  	// span is *not* zeroed.
  1817  	span.next = nil
  1818  	span.prev = nil
  1819  	span.list = nil
  1820  	span.startAddr = base
  1821  	span.npages = npages
  1822  	span.limit = base + npages*gc.PageSize // see go.dev/issue/74288; adjusted later for heap spans
  1823  	span.allocCount = 0
  1824  	span.spanclass = 0
  1825  	span.elemsize = 0
  1826  	span.speciallock.key = 0
  1827  	span.specials = nil
  1828  	span.needzero = 0
  1829  	span.freeindex = 0
  1830  	span.freeIndexForScan = 0
  1831  	span.allocBits = nil
  1832  	span.gcmarkBits = nil
  1833  	span.pinnerBits = nil
  1834  	span.state.set(mSpanDead)
  1835  	lockInit(&span.speciallock, lockRankMspanSpecial)
  1836  }
  1837  
  1838  func (span *mspan) inList() bool {
  1839  	return span.list != nil
  1840  }
  1841  
  1842  // mSpanList heads a linked list of spans.
  1843  type mSpanList struct {
  1844  	_     sys.NotInHeap
  1845  	first *mspan // first span in list, or nil if none
  1846  	last  *mspan // last span in list, or nil if none
  1847  }
  1848  
  1849  // Initialize an empty doubly-linked list.
  1850  func (list *mSpanList) init() {
  1851  	list.first = nil
  1852  	list.last = nil
  1853  }
  1854  
  1855  func (list *mSpanList) remove(span *mspan) {
  1856  	if span.list != list {
  1857  		print("runtime: failed mSpanList.remove span.npages=", span.npages,
  1858  			" span=", span, " prev=", span.prev, " span.list=", span.list, " list=", list, "\n")
  1859  		throw("mSpanList.remove")
  1860  	}
  1861  	if list.first == span {
  1862  		list.first = span.next
  1863  	} else {
  1864  		span.prev.next = span.next
  1865  	}
  1866  	if list.last == span {
  1867  		list.last = span.prev
  1868  	} else {
  1869  		span.next.prev = span.prev
  1870  	}
  1871  	span.next = nil
  1872  	span.prev = nil
  1873  	span.list = nil
  1874  }
  1875  
  1876  func (list *mSpanList) isEmpty() bool {
  1877  	return list.first == nil
  1878  }
  1879  
  1880  func (list *mSpanList) insert(span *mspan) {
  1881  	if span.next != nil || span.prev != nil || span.list != nil {
  1882  		println("runtime: failed mSpanList.insert", span, span.next, span.prev, span.list)
  1883  		throw("mSpanList.insert")
  1884  	}
  1885  	span.next = list.first
  1886  	if list.first != nil {
  1887  		// The list contains at least one span; link it in.
  1888  		// The last span in the list doesn't change.
  1889  		list.first.prev = span
  1890  	} else {
  1891  		// The list contains no spans, so this is also the last span.
  1892  		list.last = span
  1893  	}
  1894  	list.first = span
  1895  	span.list = list
  1896  }
  1897  
  1898  func (list *mSpanList) insertBack(span *mspan) {
  1899  	if span.next != nil || span.prev != nil || span.list != nil {
  1900  		println("runtime: failed mSpanList.insertBack", span, span.next, span.prev, span.list)
  1901  		throw("mSpanList.insertBack")
  1902  	}
  1903  	span.prev = list.last
  1904  	if list.last != nil {
  1905  		// The list contains at least one span.
  1906  		list.last.next = span
  1907  	} else {
  1908  		// The list contains no spans, so this is also the first span.
  1909  		list.first = span
  1910  	}
  1911  	list.last = span
  1912  	span.list = list
  1913  }
  1914  
  1915  // takeAll removes all spans from other and inserts them at the front
  1916  // of list.
  1917  func (list *mSpanList) takeAll(other *mSpanList) {
  1918  	if other.isEmpty() {
  1919  		return
  1920  	}
  1921  
  1922  	// Reparent everything in other to list.
  1923  	for s := other.first; s != nil; s = s.next {
  1924  		s.list = list
  1925  	}
  1926  
  1927  	// Concatenate the lists.
  1928  	if list.isEmpty() {
  1929  		*list = *other
  1930  	} else {
  1931  		// Neither list is empty. Put other before list.
  1932  		other.last.next = list.first
  1933  		list.first.prev = other.last
  1934  		list.first = other.first
  1935  	}
  1936  
  1937  	other.first, other.last = nil, nil
  1938  }
  1939  
  1940  // mSpanQueue is like an mSpanList but is FIFO instead of LIFO and may
  1941  // be allocated on the stack. (mSpanList can be visible from the mspan
  1942  // itself, so it is marked as not-in-heap).
  1943  type mSpanQueue struct {
  1944  	head, tail *mspan
  1945  	n          int
  1946  }
  1947  
  1948  // push adds s to the end of the queue.
  1949  func (q *mSpanQueue) push(s *mspan) {
  1950  	if s.next != nil {
  1951  		throw("span already on list")
  1952  	}
  1953  	if q.tail == nil {
  1954  		q.tail, q.head = s, s
  1955  	} else {
  1956  		q.tail.next = s
  1957  		q.tail = s
  1958  	}
  1959  	q.n++
  1960  }
  1961  
  1962  // pop removes a span from the head of the queue, if any.
  1963  func (q *mSpanQueue) pop() *mspan {
  1964  	if q.head == nil {
  1965  		return nil
  1966  	}
  1967  	s := q.head
  1968  	q.head = s.next
  1969  	s.next = nil
  1970  	if q.head == nil {
  1971  		q.tail = nil
  1972  	}
  1973  	q.n--
  1974  	return s
  1975  }
  1976  
  1977  // takeAll removes all the spans from q2 and adds them to the end of q1, in order.
  1978  func (q1 *mSpanQueue) takeAll(q2 *mSpanQueue) {
  1979  	if q2.head == nil {
  1980  		return
  1981  	}
  1982  	if q1.head == nil {
  1983  		*q1 = *q2
  1984  	} else {
  1985  		q1.tail.next = q2.head
  1986  		q1.tail = q2.tail
  1987  		q1.n += q2.n
  1988  	}
  1989  	q2.tail = nil
  1990  	q2.head = nil
  1991  	q2.n = 0
  1992  }
  1993  
  1994  // popN removes n spans from the head of the queue and returns them as a new queue.
  1995  func (q *mSpanQueue) popN(n int) mSpanQueue {
  1996  	var newQ mSpanQueue
  1997  	if n <= 0 {
  1998  		return newQ
  1999  	}
  2000  	if n >= q.n {
  2001  		newQ = *q
  2002  		q.tail = nil
  2003  		q.head = nil
  2004  		q.n = 0
  2005  		return newQ
  2006  	}
  2007  	s := q.head
  2008  	for range n - 1 {
  2009  		s = s.next
  2010  	}
  2011  	q.n -= n
  2012  	newQ.head = q.head
  2013  	newQ.tail = s
  2014  	newQ.n = n
  2015  	q.head = s.next
  2016  	s.next = nil
  2017  	return newQ
  2018  }
  2019  
  2020  const (
  2021  	// _KindSpecialTinyBlock indicates that a given allocation is a tiny block.
  2022  	// Ordered before KindSpecialFinalizer and KindSpecialCleanup so that it
  2023  	// always appears first in the specials list.
  2024  	// Used only if debug.checkfinalizers != 0.
  2025  	_KindSpecialTinyBlock = 1
  2026  	// _KindSpecialFinalizer is for tracking finalizers.
  2027  	_KindSpecialFinalizer = 2
  2028  	// _KindSpecialWeakHandle is used for creating weak pointers.
  2029  	_KindSpecialWeakHandle = 3
  2030  	// _KindSpecialProfile is for memory profiling.
  2031  	_KindSpecialProfile = 4
  2032  	// _KindSpecialReachable is a special used for tracking
  2033  	// reachability during testing.
  2034  	_KindSpecialReachable = 5
  2035  	// _KindSpecialPinCounter is a special used for objects that are pinned
  2036  	// multiple times
  2037  	_KindSpecialPinCounter = 6
  2038  	// _KindSpecialCleanup is for tracking cleanups.
  2039  	_KindSpecialCleanup = 7
  2040  	// _KindSpecialCheckFinalizer adds additional context to a finalizer or cleanup.
  2041  	// Used only if debug.checkfinalizers != 0.
  2042  	_KindSpecialCheckFinalizer = 8
  2043  	// _KindSpecialBubble is used to associate objects with synctest bubbles.
  2044  	_KindSpecialBubble = 9
  2045  )
  2046  
  2047  type special struct {
  2048  	_      sys.NotInHeap
  2049  	next   *special // linked list in span
  2050  	offset uintptr  // span offset of object
  2051  	kind   byte     // kind of special
  2052  }
  2053  
  2054  // spanHasSpecials marks a span as having specials in the arena bitmap.
  2055  func spanHasSpecials(s *mspan) {
  2056  	arenaPage := (s.base() / pageSize) % pagesPerArena
  2057  	ai := arenaIndex(s.base())
  2058  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  2059  	atomic.Or8(&ha.pageSpecials[arenaPage/8], uint8(1)<<(arenaPage%8))
  2060  }
  2061  
  2062  // spanHasNoSpecials marks a span as having no specials in the arena bitmap.
  2063  func spanHasNoSpecials(s *mspan) {
  2064  	arenaPage := (s.base() / pageSize) % pagesPerArena
  2065  	ai := arenaIndex(s.base())
  2066  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  2067  	atomic.And8(&ha.pageSpecials[arenaPage/8], ^(uint8(1) << (arenaPage % 8)))
  2068  }
  2069  
  2070  // addspecial adds the special record s to the list of special records for
  2071  // the object p. All fields of s should be filled in except for
  2072  // offset & next, which this routine will fill in.
  2073  // Returns true if the special was successfully added, false otherwise.
  2074  // (The add will fail only if a record with the same p and s->kind
  2075  // already exists unless force is set to true.)
  2076  func addspecial(p unsafe.Pointer, s *special, force bool) bool {
  2077  	span := spanOfHeap(uintptr(p))
  2078  	if span == nil {
  2079  		throw("addspecial on invalid pointer")
  2080  	}
  2081  
  2082  	// Ensure that the span is swept.
  2083  	// Sweeping accesses the specials list w/o locks, so we have
  2084  	// to synchronize with it. And it's just much safer.
  2085  	mp := acquirem()
  2086  	span.ensureSwept()
  2087  
  2088  	offset := uintptr(p) - span.base()
  2089  	kind := s.kind
  2090  
  2091  	lock(&span.speciallock)
  2092  
  2093  	// Find splice point, check for existing record.
  2094  	iter, exists := span.specialFindSplicePoint(offset, kind)
  2095  	if !exists || force {
  2096  		// Splice in record, fill in offset.
  2097  		s.offset = offset
  2098  		s.next = *iter
  2099  		*iter = s
  2100  		spanHasSpecials(span)
  2101  	}
  2102  
  2103  	unlock(&span.speciallock)
  2104  	releasem(mp)
  2105  	// We're converting p to a uintptr and looking it up, and we
  2106  	// don't want it to die and get swept while we're doing so.
  2107  	KeepAlive(p)
  2108  	return !exists || force // already exists or addition was forced
  2109  }
  2110  
  2111  // Removes the Special record of the given kind for the object p.
  2112  // Returns the record if the record existed, nil otherwise.
  2113  // The caller must FixAlloc_Free the result.
  2114  func removespecial(p unsafe.Pointer, kind uint8) *special {
  2115  	span := spanOfHeap(uintptr(p))
  2116  	if span == nil {
  2117  		throw("removespecial on invalid pointer")
  2118  	}
  2119  
  2120  	// Ensure that the span is swept.
  2121  	// Sweeping accesses the specials list w/o locks, so we have
  2122  	// to synchronize with it. And it's just much safer.
  2123  	mp := acquirem()
  2124  	span.ensureSwept()
  2125  
  2126  	offset := uintptr(p) - span.base()
  2127  
  2128  	var result *special
  2129  	lock(&span.speciallock)
  2130  
  2131  	iter, exists := span.specialFindSplicePoint(offset, kind)
  2132  	if exists {
  2133  		s := *iter
  2134  		*iter = s.next
  2135  		result = s
  2136  	}
  2137  	if span.specials == nil {
  2138  		spanHasNoSpecials(span)
  2139  	}
  2140  	unlock(&span.speciallock)
  2141  	releasem(mp)
  2142  	return result
  2143  }
  2144  
  2145  // Find a splice point in the sorted list and check for an already existing
  2146  // record. Returns a pointer to the next-reference in the list predecessor.
  2147  // Returns true, if the referenced item is an exact match.
  2148  func (span *mspan) specialFindSplicePoint(offset uintptr, kind byte) (**special, bool) {
  2149  	// Find splice point, check for existing record.
  2150  	iter := &span.specials
  2151  	found := false
  2152  	for {
  2153  		s := *iter
  2154  		if s == nil {
  2155  			break
  2156  		}
  2157  		if offset == s.offset && kind == s.kind {
  2158  			found = true
  2159  			break
  2160  		}
  2161  		if offset < s.offset || (offset == s.offset && kind < s.kind) {
  2162  			break
  2163  		}
  2164  		iter = &s.next
  2165  	}
  2166  	return iter, found
  2167  }
  2168  
  2169  // The described object has a finalizer set for it.
  2170  //
  2171  // specialfinalizer is allocated from non-GC'd memory, so any heap
  2172  // pointers must be specially handled.
  2173  type specialfinalizer struct {
  2174  	_       sys.NotInHeap
  2175  	special special
  2176  	fn      *funcval // May be a heap pointer.
  2177  	nret    uintptr
  2178  	fint    *_type   // May be a heap pointer, but always live.
  2179  	ot      *ptrtype // May be a heap pointer, but always live.
  2180  }
  2181  
  2182  // Adds a finalizer to the object p. Returns true if it succeeded.
  2183  func addfinalizer(p unsafe.Pointer, f *funcval, nret uintptr, fint *_type, ot *ptrtype) bool {
  2184  	lock(&mheap_.speciallock)
  2185  	s := (*specialfinalizer)(mheap_.specialfinalizeralloc.alloc())
  2186  	unlock(&mheap_.speciallock)
  2187  	s.special.kind = _KindSpecialFinalizer
  2188  	s.fn = f
  2189  	s.nret = nret
  2190  	s.fint = fint
  2191  	s.ot = ot
  2192  	if addspecial(p, &s.special, false) {
  2193  		// This is responsible for maintaining the same
  2194  		// GC-related invariants as markrootSpans in any
  2195  		// situation where it's possible that markrootSpans
  2196  		// has already run but mark termination hasn't yet.
  2197  		if gcphase != _GCoff {
  2198  			base, span, _ := findObject(uintptr(p), 0, 0)
  2199  			mp := acquirem()
  2200  			gcw := &mp.p.ptr().gcw
  2201  			// Mark everything reachable from the object
  2202  			// so it's retained for the finalizer.
  2203  			if !span.spanclass.noscan() {
  2204  				scanObject(base, gcw)
  2205  			}
  2206  			// Mark the finalizer itself, since the
  2207  			// special isn't part of the GC'd heap.
  2208  			scanblock(uintptr(unsafe.Pointer(&s.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  2209  			releasem(mp)
  2210  		}
  2211  		return true
  2212  	}
  2213  
  2214  	// There was an old finalizer
  2215  	lock(&mheap_.speciallock)
  2216  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  2217  	unlock(&mheap_.speciallock)
  2218  	return false
  2219  }
  2220  
  2221  // Removes the finalizer (if any) from the object p.
  2222  func removefinalizer(p unsafe.Pointer) {
  2223  	s := (*specialfinalizer)(unsafe.Pointer(removespecial(p, _KindSpecialFinalizer)))
  2224  	if s == nil {
  2225  		return // there wasn't a finalizer to remove
  2226  	}
  2227  	lock(&mheap_.speciallock)
  2228  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  2229  	unlock(&mheap_.speciallock)
  2230  }
  2231  
  2232  // The described object has a cleanup set for it.
  2233  type specialCleanup struct {
  2234  	_       sys.NotInHeap
  2235  	special special
  2236  	fn      *funcval
  2237  	// Globally unique ID for the cleanup, obtained from mheap_.cleanupID.
  2238  	id uint64
  2239  }
  2240  
  2241  // addCleanup attaches a cleanup function to the object. Multiple
  2242  // cleanups are allowed on an object, and even the same pointer.
  2243  // A cleanup id is returned which can be used to uniquely identify
  2244  // the cleanup.
  2245  func addCleanup(p unsafe.Pointer, f *funcval) uint64 {
  2246  	lock(&mheap_.speciallock)
  2247  	s := (*specialCleanup)(mheap_.specialCleanupAlloc.alloc())
  2248  	mheap_.cleanupID++ // Increment first. ID 0 is reserved.
  2249  	id := mheap_.cleanupID
  2250  	unlock(&mheap_.speciallock)
  2251  	s.special.kind = _KindSpecialCleanup
  2252  	s.fn = f
  2253  	s.id = id
  2254  
  2255  	mp := acquirem()
  2256  	addspecial(p, &s.special, true)
  2257  	// This is responsible for maintaining the same
  2258  	// GC-related invariants as markrootSpans in any
  2259  	// situation where it's possible that markrootSpans
  2260  	// has already run but mark termination hasn't yet.
  2261  	if gcphase != _GCoff {
  2262  		gcw := &mp.p.ptr().gcw
  2263  		// Mark the cleanup itself, since the
  2264  		// special isn't part of the GC'd heap.
  2265  		scanblock(uintptr(unsafe.Pointer(&s.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  2266  	}
  2267  	releasem(mp)
  2268  	// Keep f alive. There's a window in this function where it's
  2269  	// only reachable via the special while the special hasn't been
  2270  	// added to the specials list yet. This is similar to a bug
  2271  	// discovered for weak handles, see #70455.
  2272  	KeepAlive(f)
  2273  	return id
  2274  }
  2275  
  2276  // Always paired with a specialCleanup or specialfinalizer, adds context.
  2277  type specialCheckFinalizer struct {
  2278  	_         sys.NotInHeap
  2279  	special   special
  2280  	cleanupID uint64 // Needed to disambiguate cleanups.
  2281  	createPC  uintptr
  2282  	funcPC    uintptr
  2283  	ptrType   *_type
  2284  }
  2285  
  2286  // setFinalizerContext adds a specialCheckFinalizer to ptr. ptr must already have a
  2287  // finalizer special attached.
  2288  func setFinalizerContext(ptr unsafe.Pointer, ptrType *_type, createPC, funcPC uintptr) {
  2289  	setCleanupContext(ptr, ptrType, createPC, funcPC, 0)
  2290  }
  2291  
  2292  // setCleanupContext adds a specialCheckFinalizer to ptr. ptr must already have a
  2293  // finalizer or cleanup special attached. Pass 0 for the cleanupID to indicate
  2294  // a finalizer.
  2295  func setCleanupContext(ptr unsafe.Pointer, ptrType *_type, createPC, funcPC uintptr, cleanupID uint64) {
  2296  	lock(&mheap_.speciallock)
  2297  	s := (*specialCheckFinalizer)(mheap_.specialCheckFinalizerAlloc.alloc())
  2298  	unlock(&mheap_.speciallock)
  2299  	s.special.kind = _KindSpecialCheckFinalizer
  2300  	s.cleanupID = cleanupID
  2301  	s.createPC = createPC
  2302  	s.funcPC = funcPC
  2303  	s.ptrType = ptrType
  2304  
  2305  	mp := acquirem()
  2306  	addspecial(ptr, &s.special, true)
  2307  	releasem(mp)
  2308  	KeepAlive(ptr)
  2309  }
  2310  
  2311  func getCleanupContext(ptr uintptr, cleanupID uint64) *specialCheckFinalizer {
  2312  	assertWorldStopped()
  2313  
  2314  	span := spanOfHeap(ptr)
  2315  	if span == nil {
  2316  		return nil
  2317  	}
  2318  	var found *specialCheckFinalizer
  2319  	offset := ptr - span.base()
  2320  	iter, exists := span.specialFindSplicePoint(offset, _KindSpecialCheckFinalizer)
  2321  	if exists {
  2322  		for {
  2323  			s := *iter
  2324  			if s == nil {
  2325  				// Reached the end of the linked list. Stop searching at this point.
  2326  				break
  2327  			}
  2328  			if offset == s.offset && _KindSpecialCheckFinalizer == s.kind &&
  2329  				(*specialCheckFinalizer)(unsafe.Pointer(s)).cleanupID == cleanupID {
  2330  				// The special is a cleanup and contains a matching cleanup id.
  2331  				*iter = s.next
  2332  				found = (*specialCheckFinalizer)(unsafe.Pointer(s))
  2333  				break
  2334  			}
  2335  			if offset < s.offset || (offset == s.offset && _KindSpecialCheckFinalizer < s.kind) {
  2336  				// The special is outside the region specified for that kind of
  2337  				// special. The specials are sorted by kind.
  2338  				break
  2339  			}
  2340  			// Try the next special.
  2341  			iter = &s.next
  2342  		}
  2343  	}
  2344  	return found
  2345  }
  2346  
  2347  // clearFinalizerContext removes the specialCheckFinalizer for the given pointer, if any.
  2348  func clearFinalizerContext(ptr uintptr) {
  2349  	clearCleanupContext(ptr, 0)
  2350  }
  2351  
  2352  // clearFinalizerContext removes the specialCheckFinalizer for the given pointer and cleanup ID, if any.
  2353  func clearCleanupContext(ptr uintptr, cleanupID uint64) {
  2354  	// The following block removes the Special record of type cleanup for the object c.ptr.
  2355  	span := spanOfHeap(ptr)
  2356  	if span == nil {
  2357  		return
  2358  	}
  2359  	// Ensure that the span is swept.
  2360  	// Sweeping accesses the specials list w/o locks, so we have
  2361  	// to synchronize with it. And it's just much safer.
  2362  	mp := acquirem()
  2363  	span.ensureSwept()
  2364  
  2365  	offset := ptr - span.base()
  2366  
  2367  	var found *special
  2368  	lock(&span.speciallock)
  2369  
  2370  	iter, exists := span.specialFindSplicePoint(offset, _KindSpecialCheckFinalizer)
  2371  	if exists {
  2372  		for {
  2373  			s := *iter
  2374  			if s == nil {
  2375  				// Reached the end of the linked list. Stop searching at this point.
  2376  				break
  2377  			}
  2378  			if offset == s.offset && _KindSpecialCheckFinalizer == s.kind &&
  2379  				(*specialCheckFinalizer)(unsafe.Pointer(s)).cleanupID == cleanupID {
  2380  				// The special is a cleanup and contains a matching cleanup id.
  2381  				*iter = s.next
  2382  				found = s
  2383  				break
  2384  			}
  2385  			if offset < s.offset || (offset == s.offset && _KindSpecialCheckFinalizer < s.kind) {
  2386  				// The special is outside the region specified for that kind of
  2387  				// special. The specials are sorted by kind.
  2388  				break
  2389  			}
  2390  			// Try the next special.
  2391  			iter = &s.next
  2392  		}
  2393  	}
  2394  	if span.specials == nil {
  2395  		spanHasNoSpecials(span)
  2396  	}
  2397  	unlock(&span.speciallock)
  2398  	releasem(mp)
  2399  
  2400  	if found == nil {
  2401  		return
  2402  	}
  2403  	lock(&mheap_.speciallock)
  2404  	mheap_.specialCheckFinalizerAlloc.free(unsafe.Pointer(found))
  2405  	unlock(&mheap_.speciallock)
  2406  }
  2407  
  2408  // Indicates that an allocation is a tiny block.
  2409  // Used only if debug.checkfinalizers != 0.
  2410  type specialTinyBlock struct {
  2411  	_       sys.NotInHeap
  2412  	special special
  2413  }
  2414  
  2415  // setTinyBlockContext marks an allocation as a tiny block to diagnostics like
  2416  // checkfinalizer.
  2417  //
  2418  // A tiny block is only marked if it actually contains more than one distinct
  2419  // value, since we're using this for debugging.
  2420  func setTinyBlockContext(ptr unsafe.Pointer) {
  2421  	lock(&mheap_.speciallock)
  2422  	s := (*specialTinyBlock)(mheap_.specialTinyBlockAlloc.alloc())
  2423  	unlock(&mheap_.speciallock)
  2424  	s.special.kind = _KindSpecialTinyBlock
  2425  
  2426  	mp := acquirem()
  2427  	addspecial(ptr, &s.special, false)
  2428  	releasem(mp)
  2429  	KeepAlive(ptr)
  2430  }
  2431  
  2432  // inTinyBlock returns whether ptr is in a tiny alloc block, at one point grouped
  2433  // with other distinct values.
  2434  func inTinyBlock(ptr uintptr) bool {
  2435  	assertWorldStopped()
  2436  
  2437  	ptr = alignDown(ptr, maxTinySize)
  2438  	span := spanOfHeap(ptr)
  2439  	if span == nil {
  2440  		return false
  2441  	}
  2442  	offset := ptr - span.base()
  2443  	_, exists := span.specialFindSplicePoint(offset, _KindSpecialTinyBlock)
  2444  	return exists
  2445  }
  2446  
  2447  // The described object has a weak pointer.
  2448  //
  2449  // Weak pointers in the GC have the following invariants:
  2450  //
  2451  //   - Strong-to-weak conversions must ensure the strong pointer
  2452  //     remains live until the weak handle is installed. This ensures
  2453  //     that creating a weak pointer cannot fail.
  2454  //
  2455  //   - Weak-to-strong conversions require the weakly-referenced
  2456  //     object to be swept before the conversion may proceed. This
  2457  //     ensures that weak-to-strong conversions cannot resurrect
  2458  //     dead objects by sweeping them before that happens.
  2459  //
  2460  //   - Weak handles are unique and canonical for each byte offset into
  2461  //     an object that a strong pointer may point to, until an object
  2462  //     becomes unreachable.
  2463  //
  2464  //   - Weak handles contain nil as soon as an object becomes unreachable
  2465  //     the first time, before a finalizer makes it reachable again. New
  2466  //     weak handles created after resurrection are newly unique.
  2467  //
  2468  // specialWeakHandle is allocated from non-GC'd memory, so any heap
  2469  // pointers must be specially handled.
  2470  type specialWeakHandle struct {
  2471  	_       sys.NotInHeap
  2472  	special special
  2473  	// handle is a reference to the actual weak pointer.
  2474  	// It is always heap-allocated and must be explicitly kept
  2475  	// live so long as this special exists.
  2476  	handle *atomic.Uintptr
  2477  }
  2478  
  2479  //go:linkname internal_weak_runtime_registerWeakPointer weak.runtime_registerWeakPointer
  2480  func internal_weak_runtime_registerWeakPointer(p unsafe.Pointer) unsafe.Pointer {
  2481  	return unsafe.Pointer(getOrAddWeakHandle(p))
  2482  }
  2483  
  2484  //go:linkname internal_weak_runtime_makeStrongFromWeak weak.runtime_makeStrongFromWeak
  2485  func internal_weak_runtime_makeStrongFromWeak(u unsafe.Pointer) unsafe.Pointer {
  2486  	handle := (*atomic.Uintptr)(u)
  2487  
  2488  	// Prevent preemption. We want to make sure that another GC cycle can't start
  2489  	// and that work.strongFromWeak.block can't change out from under us.
  2490  	mp := acquirem()
  2491  
  2492  	// Yield to the GC if necessary.
  2493  	if work.strongFromWeak.block {
  2494  		releasem(mp)
  2495  
  2496  		// Try to park and wait for mark termination.
  2497  		// N.B. gcParkStrongFromWeak calls acquirem before returning.
  2498  		mp = gcParkStrongFromWeak()
  2499  	}
  2500  
  2501  	p := handle.Load()
  2502  	if p == 0 {
  2503  		releasem(mp)
  2504  		return nil
  2505  	}
  2506  	// Be careful. p may or may not refer to valid memory anymore, as it could've been
  2507  	// swept and released already. It's always safe to ensure a span is swept, though,
  2508  	// even if it's just some random span.
  2509  	span := spanOfHeap(p)
  2510  	if span == nil {
  2511  		// If it's immortal, then just return the pointer.
  2512  		//
  2513  		// Stay non-preemptible so the GC can't see us convert this potentially
  2514  		// completely bogus value to an unsafe.Pointer.
  2515  		if isGoPointerWithoutSpan(unsafe.Pointer(p)) {
  2516  			releasem(mp)
  2517  			return unsafe.Pointer(p)
  2518  		}
  2519  		// It's heap-allocated, so the span probably just got swept and released.
  2520  		releasem(mp)
  2521  		return nil
  2522  	}
  2523  	// Ensure the span is swept.
  2524  	span.ensureSwept()
  2525  
  2526  	// Now we can trust whatever we get from handle, so make a strong pointer.
  2527  	//
  2528  	// Even if we just swept some random span that doesn't contain this object, because
  2529  	// this object is long dead and its memory has since been reused, we'll just observe nil.
  2530  	ptr := unsafe.Pointer(handle.Load())
  2531  
  2532  	// This is responsible for maintaining the same GC-related
  2533  	// invariants as the Yuasa part of the write barrier. During
  2534  	// the mark phase, it's possible that we just created the only
  2535  	// valid pointer to the object pointed to by ptr. If it's only
  2536  	// ever referenced from our stack, and our stack is blackened
  2537  	// already, we could fail to mark it. So, mark it now.
  2538  	if gcphase != _GCoff {
  2539  		shade(uintptr(ptr))
  2540  	}
  2541  	releasem(mp)
  2542  
  2543  	// Explicitly keep ptr alive. This seems unnecessary since we return ptr,
  2544  	// but let's be explicit since it's important we keep ptr alive across the
  2545  	// call to shade.
  2546  	KeepAlive(ptr)
  2547  	return ptr
  2548  }
  2549  
  2550  // gcParkStrongFromWeak puts the current goroutine on the weak->strong queue and parks.
  2551  func gcParkStrongFromWeak() *m {
  2552  	// Prevent preemption as we check strongFromWeak, so it can't change out from under us.
  2553  	mp := acquirem()
  2554  
  2555  	for work.strongFromWeak.block {
  2556  		lock(&work.strongFromWeak.lock)
  2557  		releasem(mp) // N.B. Holding the lock prevents preemption.
  2558  
  2559  		// Queue ourselves up.
  2560  		work.strongFromWeak.q.pushBack(getg())
  2561  
  2562  		// Park.
  2563  		goparkunlock(&work.strongFromWeak.lock, waitReasonGCWeakToStrongWait, traceBlockGCWeakToStrongWait, 2)
  2564  
  2565  		// Re-acquire the current M since we're going to check the condition again.
  2566  		mp = acquirem()
  2567  
  2568  		// Re-check condition. We may have awoken in the next GC's mark termination phase.
  2569  	}
  2570  	return mp
  2571  }
  2572  
  2573  // gcWakeAllStrongFromWeak wakes all currently blocked weak->strong
  2574  // conversions. This is used at the end of a GC cycle.
  2575  //
  2576  // work.strongFromWeak.block must be false to prevent woken goroutines
  2577  // from immediately going back to sleep.
  2578  func gcWakeAllStrongFromWeak() {
  2579  	lock(&work.strongFromWeak.lock)
  2580  	list := work.strongFromWeak.q.popList()
  2581  	injectglist(&list)
  2582  	unlock(&work.strongFromWeak.lock)
  2583  }
  2584  
  2585  // Retrieves or creates a weak pointer handle for the object p.
  2586  func getOrAddWeakHandle(p unsafe.Pointer) *atomic.Uintptr {
  2587  	if debug.sbrk != 0 {
  2588  		// debug.sbrk never frees memory, so it'll never go nil. However, we do still
  2589  		// need a weak handle that's specific to p. Use the immortal weak handle map.
  2590  		// Keep p alive across the call to getOrAdd defensively, though it doesn't
  2591  		// really matter in this particular case.
  2592  		handle := mheap_.immortalWeakHandles.getOrAdd(uintptr(p))
  2593  		KeepAlive(p)
  2594  		return handle
  2595  	}
  2596  
  2597  	// First try to retrieve without allocating.
  2598  	if handle := getWeakHandle(p); handle != nil {
  2599  		// Keep p alive for the duration of the function to ensure
  2600  		// that it cannot die while we're trying to do this.
  2601  		KeepAlive(p)
  2602  		return handle
  2603  	}
  2604  
  2605  	lock(&mheap_.speciallock)
  2606  	s := (*specialWeakHandle)(mheap_.specialWeakHandleAlloc.alloc())
  2607  	unlock(&mheap_.speciallock)
  2608  
  2609  	handle := new(atomic.Uintptr)
  2610  	s.special.kind = _KindSpecialWeakHandle
  2611  	s.handle = handle
  2612  	handle.Store(uintptr(p))
  2613  	if addspecial(p, &s.special, false) {
  2614  		// This is responsible for maintaining the same
  2615  		// GC-related invariants as markrootSpans in any
  2616  		// situation where it's possible that markrootSpans
  2617  		// has already run but mark termination hasn't yet.
  2618  		if gcphase != _GCoff {
  2619  			mp := acquirem()
  2620  			gcw := &mp.p.ptr().gcw
  2621  			// Mark the weak handle itself, since the
  2622  			// special isn't part of the GC'd heap.
  2623  			scanblock(uintptr(unsafe.Pointer(&s.handle)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  2624  			releasem(mp)
  2625  		}
  2626  
  2627  		// Keep p alive for the duration of the function to ensure
  2628  		// that it cannot die while we're trying to do this.
  2629  		//
  2630  		// Same for handle, which is only stored in the special.
  2631  		// There's a window where it might die if we don't keep it
  2632  		// alive explicitly. Returning it here is probably good enough,
  2633  		// but let's be defensive and explicit. See #70455.
  2634  		KeepAlive(p)
  2635  		KeepAlive(handle)
  2636  		return handle
  2637  	}
  2638  
  2639  	// There was an existing handle. Free the special
  2640  	// and try again. We must succeed because we're explicitly
  2641  	// keeping p live until the end of this function. Either
  2642  	// we, or someone else, must have succeeded, because we can
  2643  	// only fail in the event of a race, and p will still be
  2644  	// be valid no matter how much time we spend here.
  2645  	lock(&mheap_.speciallock)
  2646  	mheap_.specialWeakHandleAlloc.free(unsafe.Pointer(s))
  2647  	unlock(&mheap_.speciallock)
  2648  
  2649  	handle = getWeakHandle(p)
  2650  	if handle == nil {
  2651  		throw("failed to get or create weak handle")
  2652  	}
  2653  
  2654  	// Keep p alive for the duration of the function to ensure
  2655  	// that it cannot die while we're trying to do this.
  2656  	//
  2657  	// Same for handle, just to be defensive.
  2658  	KeepAlive(p)
  2659  	KeepAlive(handle)
  2660  	return handle
  2661  }
  2662  
  2663  func getWeakHandle(p unsafe.Pointer) *atomic.Uintptr {
  2664  	span := spanOfHeap(uintptr(p))
  2665  	if span == nil {
  2666  		if isGoPointerWithoutSpan(p) {
  2667  			return mheap_.immortalWeakHandles.getOrAdd(uintptr(p))
  2668  		}
  2669  		throw("getWeakHandle on invalid pointer")
  2670  	}
  2671  
  2672  	// Ensure that the span is swept.
  2673  	// Sweeping accesses the specials list w/o locks, so we have
  2674  	// to synchronize with it. And it's just much safer.
  2675  	mp := acquirem()
  2676  	span.ensureSwept()
  2677  
  2678  	offset := uintptr(p) - span.base()
  2679  
  2680  	lock(&span.speciallock)
  2681  
  2682  	// Find the existing record and return the handle if one exists.
  2683  	var handle *atomic.Uintptr
  2684  	iter, exists := span.specialFindSplicePoint(offset, _KindSpecialWeakHandle)
  2685  	if exists {
  2686  		handle = ((*specialWeakHandle)(unsafe.Pointer(*iter))).handle
  2687  	}
  2688  	unlock(&span.speciallock)
  2689  	releasem(mp)
  2690  
  2691  	// Keep p alive for the duration of the function to ensure
  2692  	// that it cannot die while we're trying to do this.
  2693  	KeepAlive(p)
  2694  	return handle
  2695  }
  2696  
  2697  type immortalWeakHandleMap struct {
  2698  	root atomic.UnsafePointer // *immortalWeakHandle (can't use generics because it's notinheap)
  2699  }
  2700  
  2701  // immortalWeakHandle is a lock-free append-only hash-trie.
  2702  //
  2703  // Key features:
  2704  //   - 2-ary trie. Child nodes are indexed by the highest bit (remaining) of the hash of the address.
  2705  //   - New nodes are placed at the first empty level encountered.
  2706  //   - When the first child is added to a node, the existing value is not moved into a child.
  2707  //     This means that we must check the value at each level, not just at the leaf.
  2708  //   - No deletion or rebalancing.
  2709  //   - Intentionally devolves into a linked list on hash collisions (the hash bits will all
  2710  //     get shifted out during iteration, and new nodes will just be appended to the 0th child).
  2711  type immortalWeakHandle struct {
  2712  	_ sys.NotInHeap
  2713  
  2714  	children [2]atomic.UnsafePointer // *immortalObjectMapNode (can't use generics because it's notinheap)
  2715  	ptr      uintptr                 // &ptr is the weak handle
  2716  }
  2717  
  2718  // handle returns a canonical weak handle.
  2719  func (h *immortalWeakHandle) handle() *atomic.Uintptr {
  2720  	// N.B. Since we just need an *atomic.Uintptr that never changes, we can trivially
  2721  	// reference ptr to save on some memory in immortalWeakHandle and avoid extra atomics
  2722  	// in getOrAdd.
  2723  	return (*atomic.Uintptr)(unsafe.Pointer(&h.ptr))
  2724  }
  2725  
  2726  // getOrAdd introduces p, which must be a pointer to immortal memory (for example, a linker-allocated
  2727  // object) and returns a weak handle. The weak handle will never become nil.
  2728  func (tab *immortalWeakHandleMap) getOrAdd(p uintptr) *atomic.Uintptr {
  2729  	var newNode *immortalWeakHandle
  2730  	m := &tab.root
  2731  	hash := memhash(abi.NoEscape(unsafe.Pointer(&p)), 0, goarch.PtrSize)
  2732  	hashIter := hash
  2733  	for {
  2734  		n := (*immortalWeakHandle)(m.Load())
  2735  		if n == nil {
  2736  			// Try to insert a new map node. We may end up discarding
  2737  			// this node if we fail to insert because it turns out the
  2738  			// value is already in the map.
  2739  			//
  2740  			// The discard will only happen if two threads race on inserting
  2741  			// the same value. Both might create nodes, but only one will
  2742  			// succeed on insertion. If two threads race to insert two
  2743  			// different values, then both nodes will *always* get inserted,
  2744  			// because the equality checking below will always fail.
  2745  			//
  2746  			// Performance note: contention on insertion is likely to be
  2747  			// higher for small maps, but since this data structure is
  2748  			// append-only, either the map stays small because there isn't
  2749  			// much activity, or the map gets big and races to insert on
  2750  			// the same node are much less likely.
  2751  			if newNode == nil {
  2752  				newNode = (*immortalWeakHandle)(persistentalloc(unsafe.Sizeof(immortalWeakHandle{}), goarch.PtrSize, &memstats.gcMiscSys))
  2753  				newNode.ptr = p
  2754  			}
  2755  			if m.CompareAndSwapNoWB(nil, unsafe.Pointer(newNode)) {
  2756  				return newNode.handle()
  2757  			}
  2758  			// Reload n. Because pointers are only stored once,
  2759  			// we must have lost the race, and therefore n is not nil
  2760  			// anymore.
  2761  			n = (*immortalWeakHandle)(m.Load())
  2762  		}
  2763  		if n.ptr == p {
  2764  			return n.handle()
  2765  		}
  2766  		m = &n.children[hashIter>>(8*goarch.PtrSize-1)]
  2767  		hashIter <<= 1
  2768  	}
  2769  }
  2770  
  2771  // The described object is being heap profiled.
  2772  type specialprofile struct {
  2773  	_       sys.NotInHeap
  2774  	special special
  2775  	b       *bucket
  2776  }
  2777  
  2778  // Set the heap profile bucket associated with addr to b.
  2779  func setprofilebucket(p unsafe.Pointer, b *bucket) {
  2780  	lock(&mheap_.speciallock)
  2781  	s := (*specialprofile)(mheap_.specialprofilealloc.alloc())
  2782  	unlock(&mheap_.speciallock)
  2783  	s.special.kind = _KindSpecialProfile
  2784  	s.b = b
  2785  	if !addspecial(p, &s.special, false) {
  2786  		throw("setprofilebucket: profile already set")
  2787  	}
  2788  }
  2789  
  2790  // specialReachable tracks whether an object is reachable on the next
  2791  // GC cycle. This is used by testing.
  2792  type specialReachable struct {
  2793  	special   special
  2794  	done      bool
  2795  	reachable bool
  2796  }
  2797  
  2798  // specialPinCounter tracks whether an object is pinned multiple times.
  2799  type specialPinCounter struct {
  2800  	special special
  2801  	counter uintptr
  2802  }
  2803  
  2804  // specialsIter helps iterate over specials lists.
  2805  type specialsIter struct {
  2806  	pprev **special
  2807  	s     *special
  2808  }
  2809  
  2810  func newSpecialsIter(span *mspan) specialsIter {
  2811  	return specialsIter{&span.specials, span.specials}
  2812  }
  2813  
  2814  func (i *specialsIter) valid() bool {
  2815  	return i.s != nil
  2816  }
  2817  
  2818  func (i *specialsIter) next() {
  2819  	i.pprev = &i.s.next
  2820  	i.s = *i.pprev
  2821  }
  2822  
  2823  // unlinkAndNext removes the current special from the list and moves
  2824  // the iterator to the next special. It returns the unlinked special.
  2825  func (i *specialsIter) unlinkAndNext() *special {
  2826  	cur := i.s
  2827  	i.s = cur.next
  2828  	*i.pprev = i.s
  2829  	return cur
  2830  }
  2831  
  2832  // freeSpecial performs any cleanup on special s and deallocates it.
  2833  // s must already be unlinked from the specials list.
  2834  func freeSpecial(s *special, p unsafe.Pointer, size uintptr) {
  2835  	switch s.kind {
  2836  	case _KindSpecialFinalizer:
  2837  		sf := (*specialfinalizer)(unsafe.Pointer(s))
  2838  		queuefinalizer(p, sf.fn, sf.nret, sf.fint, sf.ot)
  2839  		lock(&mheap_.speciallock)
  2840  		mheap_.specialfinalizeralloc.free(unsafe.Pointer(sf))
  2841  		unlock(&mheap_.speciallock)
  2842  	case _KindSpecialWeakHandle:
  2843  		sw := (*specialWeakHandle)(unsafe.Pointer(s))
  2844  		sw.handle.Store(0)
  2845  		lock(&mheap_.speciallock)
  2846  		mheap_.specialWeakHandleAlloc.free(unsafe.Pointer(s))
  2847  		unlock(&mheap_.speciallock)
  2848  	case _KindSpecialProfile:
  2849  		sp := (*specialprofile)(unsafe.Pointer(s))
  2850  		mProf_Free(sp.b, size)
  2851  		lock(&mheap_.speciallock)
  2852  		mheap_.specialprofilealloc.free(unsafe.Pointer(sp))
  2853  		unlock(&mheap_.speciallock)
  2854  	case _KindSpecialReachable:
  2855  		sp := (*specialReachable)(unsafe.Pointer(s))
  2856  		sp.done = true
  2857  		// The creator frees these.
  2858  	case _KindSpecialPinCounter:
  2859  		lock(&mheap_.speciallock)
  2860  		mheap_.specialPinCounterAlloc.free(unsafe.Pointer(s))
  2861  		unlock(&mheap_.speciallock)
  2862  	case _KindSpecialCleanup:
  2863  		sc := (*specialCleanup)(unsafe.Pointer(s))
  2864  		// Cleanups, unlike finalizers, do not resurrect the objects
  2865  		// they're attached to, so we only need to pass the cleanup
  2866  		// function, not the object.
  2867  		gcCleanups.enqueue(sc.fn)
  2868  		lock(&mheap_.speciallock)
  2869  		mheap_.specialCleanupAlloc.free(unsafe.Pointer(sc))
  2870  		unlock(&mheap_.speciallock)
  2871  	case _KindSpecialCheckFinalizer:
  2872  		sc := (*specialCheckFinalizer)(unsafe.Pointer(s))
  2873  		lock(&mheap_.speciallock)
  2874  		mheap_.specialCheckFinalizerAlloc.free(unsafe.Pointer(sc))
  2875  		unlock(&mheap_.speciallock)
  2876  	case _KindSpecialTinyBlock:
  2877  		st := (*specialTinyBlock)(unsafe.Pointer(s))
  2878  		lock(&mheap_.speciallock)
  2879  		mheap_.specialTinyBlockAlloc.free(unsafe.Pointer(st))
  2880  		unlock(&mheap_.speciallock)
  2881  	case _KindSpecialBubble:
  2882  		st := (*specialBubble)(unsafe.Pointer(s))
  2883  		lock(&mheap_.speciallock)
  2884  		mheap_.specialBubbleAlloc.free(unsafe.Pointer(st))
  2885  		unlock(&mheap_.speciallock)
  2886  	default:
  2887  		throw("bad special kind")
  2888  		panic("not reached")
  2889  	}
  2890  }
  2891  
  2892  // gcBits is an alloc/mark bitmap. This is always used as gcBits.x.
  2893  type gcBits struct {
  2894  	_ sys.NotInHeap
  2895  	x uint8
  2896  }
  2897  
  2898  // bytep returns a pointer to the n'th byte of b.
  2899  func (b *gcBits) bytep(n uintptr) *uint8 {
  2900  	return addb(&b.x, n)
  2901  }
  2902  
  2903  // bitp returns a pointer to the byte containing bit n and a mask for
  2904  // selecting that bit from *bytep.
  2905  func (b *gcBits) bitp(n uintptr) (bytep *uint8, mask uint8) {
  2906  	return b.bytep(n / 8), 1 << (n % 8)
  2907  }
  2908  
  2909  const gcBitsChunkBytes = uintptr(64 << 10)
  2910  const gcBitsHeaderBytes = unsafe.Sizeof(gcBitsHeader{})
  2911  
  2912  type gcBitsHeader struct {
  2913  	free uintptr // free is the index into bits of the next free byte.
  2914  	next uintptr // *gcBits triggers recursive type bug. (issue 14620)
  2915  }
  2916  
  2917  type gcBitsArena struct {
  2918  	_ sys.NotInHeap
  2919  	// gcBitsHeader // side step recursive type bug (issue 14620) by including fields by hand.
  2920  	free uintptr // free is the index into bits of the next free byte; read/write atomically
  2921  	next *gcBitsArena
  2922  	bits [gcBitsChunkBytes - gcBitsHeaderBytes]gcBits
  2923  }
  2924  
  2925  var gcBitsArenas struct {
  2926  	lock     mutex
  2927  	free     *gcBitsArena
  2928  	next     *gcBitsArena // Read atomically. Write atomically under lock.
  2929  	current  *gcBitsArena
  2930  	previous *gcBitsArena
  2931  }
  2932  
  2933  // tryAlloc allocates from b or returns nil if b does not have enough room.
  2934  // This is safe to call concurrently.
  2935  func (b *gcBitsArena) tryAlloc(bytes uintptr) *gcBits {
  2936  	if b == nil || atomic.Loaduintptr(&b.free)+bytes > uintptr(len(b.bits)) {
  2937  		return nil
  2938  	}
  2939  	// Try to allocate from this block.
  2940  	end := atomic.Xadduintptr(&b.free, bytes)
  2941  	if end > uintptr(len(b.bits)) {
  2942  		return nil
  2943  	}
  2944  	// There was enough room.
  2945  	start := end - bytes
  2946  	return &b.bits[start]
  2947  }
  2948  
  2949  // newMarkBits returns a pointer to 8 byte aligned bytes
  2950  // to be used for a span's mark bits.
  2951  func newMarkBits(nelems uintptr) *gcBits {
  2952  	blocksNeeded := (nelems + 63) / 64
  2953  	bytesNeeded := blocksNeeded * 8
  2954  
  2955  	// Try directly allocating from the current head arena.
  2956  	head := (*gcBitsArena)(atomic.Loadp(unsafe.Pointer(&gcBitsArenas.next)))
  2957  	if p := head.tryAlloc(bytesNeeded); p != nil {
  2958  		return p
  2959  	}
  2960  
  2961  	// There's not enough room in the head arena. We may need to
  2962  	// allocate a new arena.
  2963  	lock(&gcBitsArenas.lock)
  2964  	// Try the head arena again, since it may have changed. Now
  2965  	// that we hold the lock, the list head can't change, but its
  2966  	// free position still can.
  2967  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2968  		unlock(&gcBitsArenas.lock)
  2969  		return p
  2970  	}
  2971  
  2972  	// Allocate a new arena. This may temporarily drop the lock.
  2973  	fresh := newArenaMayUnlock()
  2974  	// If newArenaMayUnlock dropped the lock, another thread may
  2975  	// have put a fresh arena on the "next" list. Try allocating
  2976  	// from next again.
  2977  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2978  		// Put fresh back on the free list.
  2979  		// TODO: Mark it "already zeroed"
  2980  		fresh.next = gcBitsArenas.free
  2981  		gcBitsArenas.free = fresh
  2982  		unlock(&gcBitsArenas.lock)
  2983  		return p
  2984  	}
  2985  
  2986  	// Allocate from the fresh arena. We haven't linked it in yet, so
  2987  	// this cannot race and is guaranteed to succeed.
  2988  	p := fresh.tryAlloc(bytesNeeded)
  2989  	if p == nil {
  2990  		throw("markBits overflow")
  2991  	}
  2992  
  2993  	// Add the fresh arena to the "next" list.
  2994  	fresh.next = gcBitsArenas.next
  2995  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), unsafe.Pointer(fresh))
  2996  
  2997  	unlock(&gcBitsArenas.lock)
  2998  	return p
  2999  }
  3000  
  3001  // newAllocBits returns a pointer to 8 byte aligned bytes
  3002  // to be used for this span's alloc bits.
  3003  // newAllocBits is used to provide newly initialized spans
  3004  // allocation bits. For spans not being initialized the
  3005  // mark bits are repurposed as allocation bits when
  3006  // the span is swept.
  3007  func newAllocBits(nelems uintptr) *gcBits {
  3008  	return newMarkBits(nelems)
  3009  }
  3010  
  3011  // nextMarkBitArenaEpoch establishes a new epoch for the arenas
  3012  // holding the mark bits. The arenas are named relative to the
  3013  // current GC cycle which is demarcated by the call to finishweep_m.
  3014  //
  3015  // All current spans have been swept.
  3016  // During that sweep each span allocated room for its gcmarkBits in
  3017  // gcBitsArenas.next block. gcBitsArenas.next becomes the gcBitsArenas.current
  3018  // where the GC will mark objects and after each span is swept these bits
  3019  // will be used to allocate objects.
  3020  // gcBitsArenas.current becomes gcBitsArenas.previous where the span's
  3021  // gcAllocBits live until all the spans have been swept during this GC cycle.
  3022  // The span's sweep extinguishes all the references to gcBitsArenas.previous
  3023  // by pointing gcAllocBits into the gcBitsArenas.current.
  3024  // The gcBitsArenas.previous is released to the gcBitsArenas.free list.
  3025  func nextMarkBitArenaEpoch() {
  3026  	lock(&gcBitsArenas.lock)
  3027  	if gcBitsArenas.previous != nil {
  3028  		if gcBitsArenas.free == nil {
  3029  			gcBitsArenas.free = gcBitsArenas.previous
  3030  		} else {
  3031  			// Find end of previous arenas.
  3032  			last := gcBitsArenas.previous
  3033  			for last = gcBitsArenas.previous; last.next != nil; last = last.next {
  3034  			}
  3035  			last.next = gcBitsArenas.free
  3036  			gcBitsArenas.free = gcBitsArenas.previous
  3037  		}
  3038  	}
  3039  	gcBitsArenas.previous = gcBitsArenas.current
  3040  	gcBitsArenas.current = gcBitsArenas.next
  3041  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), nil) // newMarkBits calls newArena when needed
  3042  	unlock(&gcBitsArenas.lock)
  3043  }
  3044  
  3045  // newArenaMayUnlock allocates and zeroes a gcBits arena.
  3046  // The caller must hold gcBitsArena.lock. This may temporarily release it.
  3047  func newArenaMayUnlock() *gcBitsArena {
  3048  	var result *gcBitsArena
  3049  	if gcBitsArenas.free == nil {
  3050  		unlock(&gcBitsArenas.lock)
  3051  		result = (*gcBitsArena)(sysAlloc(gcBitsChunkBytes, &memstats.gcMiscSys, "gc bits"))
  3052  		if result == nil {
  3053  			throw("runtime: cannot allocate memory")
  3054  		}
  3055  		lock(&gcBitsArenas.lock)
  3056  	} else {
  3057  		result = gcBitsArenas.free
  3058  		gcBitsArenas.free = gcBitsArenas.free.next
  3059  		memclrNoHeapPointers(unsafe.Pointer(result), gcBitsChunkBytes)
  3060  	}
  3061  	result.next = nil
  3062  	// If result.bits is not 8 byte aligned adjust index so
  3063  	// that &result.bits[result.free] is 8 byte aligned.
  3064  	if unsafe.Offsetof(gcBitsArena{}.bits)&7 == 0 {
  3065  		result.free = 0
  3066  	} else {
  3067  		result.free = 8 - (uintptr(unsafe.Pointer(&result.bits[0])) & 7)
  3068  	}
  3069  	return result
  3070  }
  3071  

View as plain text