github.com/mtsmfm/go/src@v0.0.0-20221020090648-44bdcb9f8fde/runtime/mheap.go (about)

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Page heap.
     6  //
     7  // See malloc.go for overview.
     8  
     9  package runtime
    10  
    11  import (
    12  	"internal/cpu"
    13  	"internal/goarch"
    14  	"runtime/internal/atomic"
    15  	"runtime/internal/sys"
    16  	"unsafe"
    17  )
    18  
    19  const (
    20  	// minPhysPageSize is a lower-bound on the physical page size. The
    21  	// true physical page size may be larger than this. In contrast,
    22  	// sys.PhysPageSize is an upper-bound on the physical page size.
    23  	minPhysPageSize = 4096
    24  
    25  	// maxPhysPageSize is the maximum page size the runtime supports.
    26  	maxPhysPageSize = 512 << 10
    27  
    28  	// maxPhysHugePageSize sets an upper-bound on the maximum huge page size
    29  	// that the runtime supports.
    30  	maxPhysHugePageSize = pallocChunkBytes
    31  
    32  	// pagesPerReclaimerChunk indicates how many pages to scan from the
    33  	// pageInUse bitmap at a time. Used by the page reclaimer.
    34  	//
    35  	// Higher values reduce contention on scanning indexes (such as
    36  	// h.reclaimIndex), but increase the minimum latency of the
    37  	// operation.
    38  	//
    39  	// The time required to scan this many pages can vary a lot depending
    40  	// on how many spans are actually freed. Experimentally, it can
    41  	// scan for pages at ~300 GB/ms on a 2.6GHz Core i7, but can only
    42  	// free spans at ~32 MB/ms. Using 512 pages bounds this at
    43  	// roughly 100µs.
    44  	//
    45  	// Must be a multiple of the pageInUse bitmap element size and
    46  	// must also evenly divide pagesPerArena.
    47  	pagesPerReclaimerChunk = 512
    48  
    49  	// physPageAlignedStacks indicates whether stack allocations must be
    50  	// physical page aligned. This is a requirement for MAP_STACK on
    51  	// OpenBSD.
    52  	physPageAlignedStacks = GOOS == "openbsd"
    53  )
    54  
    55  // Main malloc heap.
    56  // The heap itself is the "free" and "scav" treaps,
    57  // but all the other global data is here too.
    58  //
    59  // mheap must not be heap-allocated because it contains mSpanLists,
    60  // which must not be heap-allocated.
    61  type mheap struct {
    62  	_ sys.NotInHeap
    63  
    64  	// lock must only be acquired on the system stack, otherwise a g
    65  	// could self-deadlock if its stack grows with the lock held.
    66  	lock mutex
    67  
    68  	pages pageAlloc // page allocation data structure
    69  
    70  	sweepgen uint32 // sweep generation, see comment in mspan; written during STW
    71  
    72  	// allspans is a slice of all mspans ever created. Each mspan
    73  	// appears exactly once.
    74  	//
    75  	// The memory for allspans is manually managed and can be
    76  	// reallocated and move as the heap grows.
    77  	//
    78  	// In general, allspans is protected by mheap_.lock, which
    79  	// prevents concurrent access as well as freeing the backing
    80  	// store. Accesses during STW might not hold the lock, but
    81  	// must ensure that allocation cannot happen around the
    82  	// access (since that may free the backing store).
    83  	allspans []*mspan // all spans out there
    84  
    85  	// Proportional sweep
    86  	//
    87  	// These parameters represent a linear function from gcController.heapLive
    88  	// to page sweep count. The proportional sweep system works to
    89  	// stay in the black by keeping the current page sweep count
    90  	// above this line at the current gcController.heapLive.
    91  	//
    92  	// The line has slope sweepPagesPerByte and passes through a
    93  	// basis point at (sweepHeapLiveBasis, pagesSweptBasis). At
    94  	// any given time, the system is at (gcController.heapLive,
    95  	// pagesSwept) in this space.
    96  	//
    97  	// It is important that the line pass through a point we
    98  	// control rather than simply starting at a 0,0 origin
    99  	// because that lets us adjust sweep pacing at any time while
   100  	// accounting for current progress. If we could only adjust
   101  	// the slope, it would create a discontinuity in debt if any
   102  	// progress has already been made.
   103  	pagesInUse         atomic.Uintptr // pages of spans in stats mSpanInUse
   104  	pagesSwept         atomic.Uint64  // pages swept this cycle
   105  	pagesSweptBasis    atomic.Uint64  // pagesSwept to use as the origin of the sweep ratio
   106  	sweepHeapLiveBasis uint64         // value of gcController.heapLive to use as the origin of sweep ratio; written with lock, read without
   107  	sweepPagesPerByte  float64        // proportional sweep ratio; written with lock, read without
   108  
   109  	// Page reclaimer state
   110  
   111  	// reclaimIndex is the page index in allArenas of next page to
   112  	// reclaim. Specifically, it refers to page (i %
   113  	// pagesPerArena) of arena allArenas[i / pagesPerArena].
   114  	//
   115  	// If this is >= 1<<63, the page reclaimer is done scanning
   116  	// the page marks.
   117  	reclaimIndex atomic.Uint64
   118  
   119  	// reclaimCredit is spare credit for extra pages swept. Since
   120  	// the page reclaimer works in large chunks, it may reclaim
   121  	// more than requested. Any spare pages released go to this
   122  	// credit pool.
   123  	reclaimCredit atomic.Uintptr
   124  
   125  	// arenas is the heap arena map. It points to the metadata for
   126  	// the heap for every arena frame of the entire usable virtual
   127  	// address space.
   128  	//
   129  	// Use arenaIndex to compute indexes into this array.
   130  	//
   131  	// For regions of the address space that are not backed by the
   132  	// Go heap, the arena map contains nil.
   133  	//
   134  	// Modifications are protected by mheap_.lock. Reads can be
   135  	// performed without locking; however, a given entry can
   136  	// transition from nil to non-nil at any time when the lock
   137  	// isn't held. (Entries never transitions back to nil.)
   138  	//
   139  	// In general, this is a two-level mapping consisting of an L1
   140  	// map and possibly many L2 maps. This saves space when there
   141  	// are a huge number of arena frames. However, on many
   142  	// platforms (even 64-bit), arenaL1Bits is 0, making this
   143  	// effectively a single-level map. In this case, arenas[0]
   144  	// will never be nil.
   145  	arenas [1 << arenaL1Bits]*[1 << arenaL2Bits]*heapArena
   146  
   147  	// heapArenaAlloc is pre-reserved space for allocating heapArena
   148  	// objects. This is only used on 32-bit, where we pre-reserve
   149  	// this space to avoid interleaving it with the heap itself.
   150  	heapArenaAlloc linearAlloc
   151  
   152  	// arenaHints is a list of addresses at which to attempt to
   153  	// add more heap arenas. This is initially populated with a
   154  	// set of general hint addresses, and grown with the bounds of
   155  	// actual heap arena ranges.
   156  	arenaHints *arenaHint
   157  
   158  	// arena is a pre-reserved space for allocating heap arenas
   159  	// (the actual arenas). This is only used on 32-bit.
   160  	arena linearAlloc
   161  
   162  	// allArenas is the arenaIndex of every mapped arena. This can
   163  	// be used to iterate through the address space.
   164  	//
   165  	// Access is protected by mheap_.lock. However, since this is
   166  	// append-only and old backing arrays are never freed, it is
   167  	// safe to acquire mheap_.lock, copy the slice header, and
   168  	// then release mheap_.lock.
   169  	allArenas []arenaIdx
   170  
   171  	// sweepArenas is a snapshot of allArenas taken at the
   172  	// beginning of the sweep cycle. This can be read safely by
   173  	// simply blocking GC (by disabling preemption).
   174  	sweepArenas []arenaIdx
   175  
   176  	// markArenas is a snapshot of allArenas taken at the beginning
   177  	// of the mark cycle. Because allArenas is append-only, neither
   178  	// this slice nor its contents will change during the mark, so
   179  	// it can be read safely.
   180  	markArenas []arenaIdx
   181  
   182  	// curArena is the arena that the heap is currently growing
   183  	// into. This should always be physPageSize-aligned.
   184  	curArena struct {
   185  		base, end uintptr
   186  	}
   187  
   188  	// central free lists for small size classes.
   189  	// the padding makes sure that the mcentrals are
   190  	// spaced CacheLinePadSize bytes apart, so that each mcentral.lock
   191  	// gets its own cache line.
   192  	// central is indexed by spanClass.
   193  	central [numSpanClasses]struct {
   194  		mcentral mcentral
   195  		pad      [(cpu.CacheLinePadSize - unsafe.Sizeof(mcentral{})%cpu.CacheLinePadSize) % cpu.CacheLinePadSize]byte
   196  	}
   197  
   198  	spanalloc             fixalloc // allocator for span*
   199  	cachealloc            fixalloc // allocator for mcache*
   200  	specialfinalizeralloc fixalloc // allocator for specialfinalizer*
   201  	specialprofilealloc   fixalloc // allocator for specialprofile*
   202  	specialReachableAlloc fixalloc // allocator for specialReachable
   203  	speciallock           mutex    // lock for special record allocators.
   204  	arenaHintAlloc        fixalloc // allocator for arenaHints
   205  
   206  	// User arena state.
   207  	//
   208  	// Protected by mheap_.lock.
   209  	userArena struct {
   210  		// arenaHints is a list of addresses at which to attempt to
   211  		// add more heap arenas for user arena chunks. This is initially
   212  		// populated with a set of general hint addresses, and grown with
   213  		// the bounds of actual heap arena ranges.
   214  		arenaHints *arenaHint
   215  
   216  		// quarantineList is a list of user arena spans that have been set to fault, but
   217  		// are waiting for all pointers into them to go away. Sweeping handles
   218  		// identifying when this is true, and moves the span to the ready list.
   219  		quarantineList mSpanList
   220  
   221  		// readyList is a list of empty user arena spans that are ready for reuse.
   222  		readyList mSpanList
   223  	}
   224  
   225  	unused *specialfinalizer // never set, just here to force the specialfinalizer type into DWARF
   226  }
   227  
   228  var mheap_ mheap
   229  
   230  // A heapArena stores metadata for a heap arena. heapArenas are stored
   231  // outside of the Go heap and accessed via the mheap_.arenas index.
   232  type heapArena struct {
   233  	_ sys.NotInHeap
   234  
   235  	// bitmap stores the pointer/scalar bitmap for the words in
   236  	// this arena. See mbitmap.go for a description.
   237  	// This array uses 1 bit per word of heap, or 1.6% of the heap size (for 64-bit).
   238  	bitmap [heapArenaBitmapWords]uintptr
   239  
   240  	// If the ith bit of noMorePtrs is true, then there are no more
   241  	// pointers for the object containing the word described by the
   242  	// high bit of bitmap[i].
   243  	// In that case, bitmap[i+1], ... must be zero until the start
   244  	// of the next object.
   245  	// We never operate on these entries using bit-parallel techniques,
   246  	// so it is ok if they are small. Also, they can't be bigger than
   247  	// uint16 because at that size a single noMorePtrs entry
   248  	// represents 8K of memory, the minimum size of a span. Any larger
   249  	// and we'd have to worry about concurrent updates.
   250  	// This array uses 1 bit per word of bitmap, or .024% of the heap size (for 64-bit).
   251  	noMorePtrs [heapArenaBitmapWords / 8]uint8
   252  
   253  	// spans maps from virtual address page ID within this arena to *mspan.
   254  	// For allocated spans, their pages map to the span itself.
   255  	// For free spans, only the lowest and highest pages map to the span itself.
   256  	// Internal pages map to an arbitrary span.
   257  	// For pages that have never been allocated, spans entries are nil.
   258  	//
   259  	// Modifications are protected by mheap.lock. Reads can be
   260  	// performed without locking, but ONLY from indexes that are
   261  	// known to contain in-use or stack spans. This means there
   262  	// must not be a safe-point between establishing that an
   263  	// address is live and looking it up in the spans array.
   264  	spans [pagesPerArena]*mspan
   265  
   266  	// pageInUse is a bitmap that indicates which spans are in
   267  	// state mSpanInUse. This bitmap is indexed by page number,
   268  	// but only the bit corresponding to the first page in each
   269  	// span is used.
   270  	//
   271  	// Reads and writes are atomic.
   272  	pageInUse [pagesPerArena / 8]uint8
   273  
   274  	// pageMarks is a bitmap that indicates which spans have any
   275  	// marked objects on them. Like pageInUse, only the bit
   276  	// corresponding to the first page in each span is used.
   277  	//
   278  	// Writes are done atomically during marking. Reads are
   279  	// non-atomic and lock-free since they only occur during
   280  	// sweeping (and hence never race with writes).
   281  	//
   282  	// This is used to quickly find whole spans that can be freed.
   283  	//
   284  	// TODO(austin): It would be nice if this was uint64 for
   285  	// faster scanning, but we don't have 64-bit atomic bit
   286  	// operations.
   287  	pageMarks [pagesPerArena / 8]uint8
   288  
   289  	// pageSpecials is a bitmap that indicates which spans have
   290  	// specials (finalizers or other). Like pageInUse, only the bit
   291  	// corresponding to the first page in each span is used.
   292  	//
   293  	// Writes are done atomically whenever a special is added to
   294  	// a span and whenever the last special is removed from a span.
   295  	// Reads are done atomically to find spans containing specials
   296  	// during marking.
   297  	pageSpecials [pagesPerArena / 8]uint8
   298  
   299  	// checkmarks stores the debug.gccheckmark state. It is only
   300  	// used if debug.gccheckmark > 0.
   301  	checkmarks *checkmarksMap
   302  
   303  	// zeroedBase marks the first byte of the first page in this
   304  	// arena which hasn't been used yet and is therefore already
   305  	// zero. zeroedBase is relative to the arena base.
   306  	// Increases monotonically until it hits heapArenaBytes.
   307  	//
   308  	// This field is sufficient to determine if an allocation
   309  	// needs to be zeroed because the page allocator follows an
   310  	// address-ordered first-fit policy.
   311  	//
   312  	// Read atomically and written with an atomic CAS.
   313  	zeroedBase uintptr
   314  }
   315  
   316  // arenaHint is a hint for where to grow the heap arenas. See
   317  // mheap_.arenaHints.
   318  type arenaHint struct {
   319  	_    sys.NotInHeap
   320  	addr uintptr
   321  	down bool
   322  	next *arenaHint
   323  }
   324  
   325  // An mspan is a run of pages.
   326  //
   327  // When a mspan is in the heap free treap, state == mSpanFree
   328  // and heapmap(s->start) == span, heapmap(s->start+s->npages-1) == span.
   329  // If the mspan is in the heap scav treap, then in addition to the
   330  // above scavenged == true. scavenged == false in all other cases.
   331  //
   332  // When a mspan is allocated, state == mSpanInUse or mSpanManual
   333  // and heapmap(i) == span for all s->start <= i < s->start+s->npages.
   334  
   335  // Every mspan is in one doubly-linked list, either in the mheap's
   336  // busy list or one of the mcentral's span lists.
   337  
   338  // An mspan representing actual memory has state mSpanInUse,
   339  // mSpanManual, or mSpanFree. Transitions between these states are
   340  // constrained as follows:
   341  //
   342  //   - A span may transition from free to in-use or manual during any GC
   343  //     phase.
   344  //
   345  //   - During sweeping (gcphase == _GCoff), a span may transition from
   346  //     in-use to free (as a result of sweeping) or manual to free (as a
   347  //     result of stacks being freed).
   348  //
   349  //   - During GC (gcphase != _GCoff), a span *must not* transition from
   350  //     manual or in-use to free. Because concurrent GC may read a pointer
   351  //     and then look up its span, the span state must be monotonic.
   352  //
   353  // Setting mspan.state to mSpanInUse or mSpanManual must be done
   354  // atomically and only after all other span fields are valid.
   355  // Likewise, if inspecting a span is contingent on it being
   356  // mSpanInUse, the state should be loaded atomically and checked
   357  // before depending on other fields. This allows the garbage collector
   358  // to safely deal with potentially invalid pointers, since resolving
   359  // such pointers may race with a span being allocated.
   360  type mSpanState uint8
   361  
   362  const (
   363  	mSpanDead   mSpanState = iota
   364  	mSpanInUse             // allocated for garbage collected heap
   365  	mSpanManual            // allocated for manual management (e.g., stack allocator)
   366  )
   367  
   368  // mSpanStateNames are the names of the span states, indexed by
   369  // mSpanState.
   370  var mSpanStateNames = []string{
   371  	"mSpanDead",
   372  	"mSpanInUse",
   373  	"mSpanManual",
   374  }
   375  
   376  // mSpanStateBox holds an atomic.Uint8 to provide atomic operations on
   377  // an mSpanState. This is a separate type to disallow accidental comparison
   378  // or assignment with mSpanState.
   379  type mSpanStateBox struct {
   380  	s atomic.Uint8
   381  }
   382  
   383  // It is nosplit to match get, below.
   384  
   385  //go:nosplit
   386  func (b *mSpanStateBox) set(s mSpanState) {
   387  	b.s.Store(uint8(s))
   388  }
   389  
   390  // It is nosplit because it's called indirectly by typedmemclr,
   391  // which must not be preempted.
   392  
   393  //go:nosplit
   394  func (b *mSpanStateBox) get() mSpanState {
   395  	return mSpanState(b.s.Load())
   396  }
   397  
   398  // mSpanList heads a linked list of spans.
   399  type mSpanList struct {
   400  	_     sys.NotInHeap
   401  	first *mspan // first span in list, or nil if none
   402  	last  *mspan // last span in list, or nil if none
   403  }
   404  
   405  type mspan struct {
   406  	_    sys.NotInHeap
   407  	next *mspan     // next span in list, or nil if none
   408  	prev *mspan     // previous span in list, or nil if none
   409  	list *mSpanList // For debugging. TODO: Remove.
   410  
   411  	startAddr uintptr // address of first byte of span aka s.base()
   412  	npages    uintptr // number of pages in span
   413  
   414  	manualFreeList gclinkptr // list of free objects in mSpanManual spans
   415  
   416  	// freeindex is the slot index between 0 and nelems at which to begin scanning
   417  	// for the next free object in this span.
   418  	// Each allocation scans allocBits starting at freeindex until it encounters a 0
   419  	// indicating a free object. freeindex is then adjusted so that subsequent scans begin
   420  	// just past the newly discovered free object.
   421  	//
   422  	// If freeindex == nelem, this span has no free objects.
   423  	//
   424  	// allocBits is a bitmap of objects in this span.
   425  	// If n >= freeindex and allocBits[n/8] & (1<<(n%8)) is 0
   426  	// then object n is free;
   427  	// otherwise, object n is allocated. Bits starting at nelem are
   428  	// undefined and should never be referenced.
   429  	//
   430  	// Object n starts at address n*elemsize + (start << pageShift).
   431  	freeindex uintptr
   432  	// TODO: Look up nelems from sizeclass and remove this field if it
   433  	// helps performance.
   434  	nelems uintptr // number of object in the span.
   435  
   436  	// Cache of the allocBits at freeindex. allocCache is shifted
   437  	// such that the lowest bit corresponds to the bit freeindex.
   438  	// allocCache holds the complement of allocBits, thus allowing
   439  	// ctz (count trailing zero) to use it directly.
   440  	// allocCache may contain bits beyond s.nelems; the caller must ignore
   441  	// these.
   442  	allocCache uint64
   443  
   444  	// allocBits and gcmarkBits hold pointers to a span's mark and
   445  	// allocation bits. The pointers are 8 byte aligned.
   446  	// There are three arenas where this data is held.
   447  	// free: Dirty arenas that are no longer accessed
   448  	//       and can be reused.
   449  	// next: Holds information to be used in the next GC cycle.
   450  	// current: Information being used during this GC cycle.
   451  	// previous: Information being used during the last GC cycle.
   452  	// A new GC cycle starts with the call to finishsweep_m.
   453  	// finishsweep_m moves the previous arena to the free arena,
   454  	// the current arena to the previous arena, and
   455  	// the next arena to the current arena.
   456  	// The next arena is populated as the spans request
   457  	// memory to hold gcmarkBits for the next GC cycle as well
   458  	// as allocBits for newly allocated spans.
   459  	//
   460  	// The pointer arithmetic is done "by hand" instead of using
   461  	// arrays to avoid bounds checks along critical performance
   462  	// paths.
   463  	// The sweep will free the old allocBits and set allocBits to the
   464  	// gcmarkBits. The gcmarkBits are replaced with a fresh zeroed
   465  	// out memory.
   466  	allocBits  *gcBits
   467  	gcmarkBits *gcBits
   468  
   469  	// sweep generation:
   470  	// if sweepgen == h->sweepgen - 2, the span needs sweeping
   471  	// if sweepgen == h->sweepgen - 1, the span is currently being swept
   472  	// if sweepgen == h->sweepgen, the span is swept and ready to use
   473  	// if sweepgen == h->sweepgen + 1, the span was cached before sweep began and is still cached, and needs sweeping
   474  	// if sweepgen == h->sweepgen + 3, the span was swept and then cached and is still cached
   475  	// h->sweepgen is incremented by 2 after every GC
   476  
   477  	sweepgen              uint32
   478  	divMul                uint32        // for divide by elemsize
   479  	allocCount            uint16        // number of allocated objects
   480  	spanclass             spanClass     // size class and noscan (uint8)
   481  	state                 mSpanStateBox // mSpanInUse etc; accessed atomically (get/set methods)
   482  	needzero              uint8         // needs to be zeroed before allocation
   483  	isUserArenaChunk      bool          // whether or not this span represents a user arena
   484  	allocCountBeforeCache uint16        // a copy of allocCount that is stored just before this span is cached
   485  	elemsize              uintptr       // computed from sizeclass or from npages
   486  	limit                 uintptr       // end of data in span
   487  	speciallock           mutex         // guards specials list
   488  	specials              *special      // linked list of special records sorted by offset.
   489  	userArenaChunkFree    addrRange     // interval for managing chunk allocation
   490  }
   491  
   492  func (s *mspan) base() uintptr {
   493  	return s.startAddr
   494  }
   495  
   496  func (s *mspan) layout() (size, n, total uintptr) {
   497  	total = s.npages << _PageShift
   498  	size = s.elemsize
   499  	if size > 0 {
   500  		n = total / size
   501  	}
   502  	return
   503  }
   504  
   505  // recordspan adds a newly allocated span to h.allspans.
   506  //
   507  // This only happens the first time a span is allocated from
   508  // mheap.spanalloc (it is not called when a span is reused).
   509  //
   510  // Write barriers are disallowed here because it can be called from
   511  // gcWork when allocating new workbufs. However, because it's an
   512  // indirect call from the fixalloc initializer, the compiler can't see
   513  // this.
   514  //
   515  // The heap lock must be held.
   516  //
   517  //go:nowritebarrierrec
   518  func recordspan(vh unsafe.Pointer, p unsafe.Pointer) {
   519  	h := (*mheap)(vh)
   520  	s := (*mspan)(p)
   521  
   522  	assertLockHeld(&h.lock)
   523  
   524  	if len(h.allspans) >= cap(h.allspans) {
   525  		n := 64 * 1024 / goarch.PtrSize
   526  		if n < cap(h.allspans)*3/2 {
   527  			n = cap(h.allspans) * 3 / 2
   528  		}
   529  		var new []*mspan
   530  		sp := (*slice)(unsafe.Pointer(&new))
   531  		sp.array = sysAlloc(uintptr(n)*goarch.PtrSize, &memstats.other_sys)
   532  		if sp.array == nil {
   533  			throw("runtime: cannot allocate memory")
   534  		}
   535  		sp.len = len(h.allspans)
   536  		sp.cap = n
   537  		if len(h.allspans) > 0 {
   538  			copy(new, h.allspans)
   539  		}
   540  		oldAllspans := h.allspans
   541  		*(*notInHeapSlice)(unsafe.Pointer(&h.allspans)) = *(*notInHeapSlice)(unsafe.Pointer(&new))
   542  		if len(oldAllspans) != 0 {
   543  			sysFree(unsafe.Pointer(&oldAllspans[0]), uintptr(cap(oldAllspans))*unsafe.Sizeof(oldAllspans[0]), &memstats.other_sys)
   544  		}
   545  	}
   546  	h.allspans = h.allspans[:len(h.allspans)+1]
   547  	h.allspans[len(h.allspans)-1] = s
   548  }
   549  
   550  // A spanClass represents the size class and noscan-ness of a span.
   551  //
   552  // Each size class has a noscan spanClass and a scan spanClass. The
   553  // noscan spanClass contains only noscan objects, which do not contain
   554  // pointers and thus do not need to be scanned by the garbage
   555  // collector.
   556  type spanClass uint8
   557  
   558  const (
   559  	numSpanClasses = _NumSizeClasses << 1
   560  	tinySpanClass  = spanClass(tinySizeClass<<1 | 1)
   561  )
   562  
   563  func makeSpanClass(sizeclass uint8, noscan bool) spanClass {
   564  	return spanClass(sizeclass<<1) | spanClass(bool2int(noscan))
   565  }
   566  
   567  func (sc spanClass) sizeclass() int8 {
   568  	return int8(sc >> 1)
   569  }
   570  
   571  func (sc spanClass) noscan() bool {
   572  	return sc&1 != 0
   573  }
   574  
   575  // arenaIndex returns the index into mheap_.arenas of the arena
   576  // containing metadata for p. This index combines of an index into the
   577  // L1 map and an index into the L2 map and should be used as
   578  // mheap_.arenas[ai.l1()][ai.l2()].
   579  //
   580  // If p is outside the range of valid heap addresses, either l1() or
   581  // l2() will be out of bounds.
   582  //
   583  // It is nosplit because it's called by spanOf and several other
   584  // nosplit functions.
   585  //
   586  //go:nosplit
   587  func arenaIndex(p uintptr) arenaIdx {
   588  	return arenaIdx((p - arenaBaseOffset) / heapArenaBytes)
   589  }
   590  
   591  // arenaBase returns the low address of the region covered by heap
   592  // arena i.
   593  func arenaBase(i arenaIdx) uintptr {
   594  	return uintptr(i)*heapArenaBytes + arenaBaseOffset
   595  }
   596  
   597  type arenaIdx uint
   598  
   599  // l1 returns the "l1" portion of an arenaIdx.
   600  //
   601  // Marked nosplit because it's called by spanOf and other nosplit
   602  // functions.
   603  //
   604  //go:nosplit
   605  func (i arenaIdx) l1() uint {
   606  	if arenaL1Bits == 0 {
   607  		// Let the compiler optimize this away if there's no
   608  		// L1 map.
   609  		return 0
   610  	} else {
   611  		return uint(i) >> arenaL1Shift
   612  	}
   613  }
   614  
   615  // l2 returns the "l2" portion of an arenaIdx.
   616  //
   617  // Marked nosplit because it's called by spanOf and other nosplit funcs.
   618  // functions.
   619  //
   620  //go:nosplit
   621  func (i arenaIdx) l2() uint {
   622  	if arenaL1Bits == 0 {
   623  		return uint(i)
   624  	} else {
   625  		return uint(i) & (1<<arenaL2Bits - 1)
   626  	}
   627  }
   628  
   629  // inheap reports whether b is a pointer into a (potentially dead) heap object.
   630  // It returns false for pointers into mSpanManual spans.
   631  // Non-preemptible because it is used by write barriers.
   632  //
   633  //go:nowritebarrier
   634  //go:nosplit
   635  func inheap(b uintptr) bool {
   636  	return spanOfHeap(b) != nil
   637  }
   638  
   639  // inHeapOrStack is a variant of inheap that returns true for pointers
   640  // into any allocated heap span.
   641  //
   642  //go:nowritebarrier
   643  //go:nosplit
   644  func inHeapOrStack(b uintptr) bool {
   645  	s := spanOf(b)
   646  	if s == nil || b < s.base() {
   647  		return false
   648  	}
   649  	switch s.state.get() {
   650  	case mSpanInUse, mSpanManual:
   651  		return b < s.limit
   652  	default:
   653  		return false
   654  	}
   655  }
   656  
   657  // spanOf returns the span of p. If p does not point into the heap
   658  // arena or no span has ever contained p, spanOf returns nil.
   659  //
   660  // If p does not point to allocated memory, this may return a non-nil
   661  // span that does *not* contain p. If this is a possibility, the
   662  // caller should either call spanOfHeap or check the span bounds
   663  // explicitly.
   664  //
   665  // Must be nosplit because it has callers that are nosplit.
   666  //
   667  //go:nosplit
   668  func spanOf(p uintptr) *mspan {
   669  	// This function looks big, but we use a lot of constant
   670  	// folding around arenaL1Bits to get it under the inlining
   671  	// budget. Also, many of the checks here are safety checks
   672  	// that Go needs to do anyway, so the generated code is quite
   673  	// short.
   674  	ri := arenaIndex(p)
   675  	if arenaL1Bits == 0 {
   676  		// If there's no L1, then ri.l1() can't be out of bounds but ri.l2() can.
   677  		if ri.l2() >= uint(len(mheap_.arenas[0])) {
   678  			return nil
   679  		}
   680  	} else {
   681  		// If there's an L1, then ri.l1() can be out of bounds but ri.l2() can't.
   682  		if ri.l1() >= uint(len(mheap_.arenas)) {
   683  			return nil
   684  		}
   685  	}
   686  	l2 := mheap_.arenas[ri.l1()]
   687  	if arenaL1Bits != 0 && l2 == nil { // Should never happen if there's no L1.
   688  		return nil
   689  	}
   690  	ha := l2[ri.l2()]
   691  	if ha == nil {
   692  		return nil
   693  	}
   694  	return ha.spans[(p/pageSize)%pagesPerArena]
   695  }
   696  
   697  // spanOfUnchecked is equivalent to spanOf, but the caller must ensure
   698  // that p points into an allocated heap arena.
   699  //
   700  // Must be nosplit because it has callers that are nosplit.
   701  //
   702  //go:nosplit
   703  func spanOfUnchecked(p uintptr) *mspan {
   704  	ai := arenaIndex(p)
   705  	return mheap_.arenas[ai.l1()][ai.l2()].spans[(p/pageSize)%pagesPerArena]
   706  }
   707  
   708  // spanOfHeap is like spanOf, but returns nil if p does not point to a
   709  // heap object.
   710  //
   711  // Must be nosplit because it has callers that are nosplit.
   712  //
   713  //go:nosplit
   714  func spanOfHeap(p uintptr) *mspan {
   715  	s := spanOf(p)
   716  	// s is nil if it's never been allocated. Otherwise, we check
   717  	// its state first because we don't trust this pointer, so we
   718  	// have to synchronize with span initialization. Then, it's
   719  	// still possible we picked up a stale span pointer, so we
   720  	// have to check the span's bounds.
   721  	if s == nil || s.state.get() != mSpanInUse || p < s.base() || p >= s.limit {
   722  		return nil
   723  	}
   724  	return s
   725  }
   726  
   727  // pageIndexOf returns the arena, page index, and page mask for pointer p.
   728  // The caller must ensure p is in the heap.
   729  func pageIndexOf(p uintptr) (arena *heapArena, pageIdx uintptr, pageMask uint8) {
   730  	ai := arenaIndex(p)
   731  	arena = mheap_.arenas[ai.l1()][ai.l2()]
   732  	pageIdx = ((p / pageSize) / 8) % uintptr(len(arena.pageInUse))
   733  	pageMask = byte(1 << ((p / pageSize) % 8))
   734  	return
   735  }
   736  
   737  // Initialize the heap.
   738  func (h *mheap) init() {
   739  	lockInit(&h.lock, lockRankMheap)
   740  	lockInit(&h.speciallock, lockRankMheapSpecial)
   741  
   742  	h.spanalloc.init(unsafe.Sizeof(mspan{}), recordspan, unsafe.Pointer(h), &memstats.mspan_sys)
   743  	h.cachealloc.init(unsafe.Sizeof(mcache{}), nil, nil, &memstats.mcache_sys)
   744  	h.specialfinalizeralloc.init(unsafe.Sizeof(specialfinalizer{}), nil, nil, &memstats.other_sys)
   745  	h.specialprofilealloc.init(unsafe.Sizeof(specialprofile{}), nil, nil, &memstats.other_sys)
   746  	h.specialReachableAlloc.init(unsafe.Sizeof(specialReachable{}), nil, nil, &memstats.other_sys)
   747  	h.arenaHintAlloc.init(unsafe.Sizeof(arenaHint{}), nil, nil, &memstats.other_sys)
   748  
   749  	// Don't zero mspan allocations. Background sweeping can
   750  	// inspect a span concurrently with allocating it, so it's
   751  	// important that the span's sweepgen survive across freeing
   752  	// and re-allocating a span to prevent background sweeping
   753  	// from improperly cas'ing it from 0.
   754  	//
   755  	// This is safe because mspan contains no heap pointers.
   756  	h.spanalloc.zero = false
   757  
   758  	// h->mapcache needs no init
   759  
   760  	for i := range h.central {
   761  		h.central[i].mcentral.init(spanClass(i))
   762  	}
   763  
   764  	h.pages.init(&h.lock, &memstats.gcMiscSys)
   765  }
   766  
   767  // reclaim sweeps and reclaims at least npage pages into the heap.
   768  // It is called before allocating npage pages to keep growth in check.
   769  //
   770  // reclaim implements the page-reclaimer half of the sweeper.
   771  //
   772  // h.lock must NOT be held.
   773  func (h *mheap) reclaim(npage uintptr) {
   774  	// TODO(austin): Half of the time spent freeing spans is in
   775  	// locking/unlocking the heap (even with low contention). We
   776  	// could make the slow path here several times faster by
   777  	// batching heap frees.
   778  
   779  	// Bail early if there's no more reclaim work.
   780  	if h.reclaimIndex.Load() >= 1<<63 {
   781  		return
   782  	}
   783  
   784  	// Disable preemption so the GC can't start while we're
   785  	// sweeping, so we can read h.sweepArenas, and so
   786  	// traceGCSweepStart/Done pair on the P.
   787  	mp := acquirem()
   788  
   789  	if trace.enabled {
   790  		traceGCSweepStart()
   791  	}
   792  
   793  	arenas := h.sweepArenas
   794  	locked := false
   795  	for npage > 0 {
   796  		// Pull from accumulated credit first.
   797  		if credit := h.reclaimCredit.Load(); credit > 0 {
   798  			take := credit
   799  			if take > npage {
   800  				// Take only what we need.
   801  				take = npage
   802  			}
   803  			if h.reclaimCredit.CompareAndSwap(credit, credit-take) {
   804  				npage -= take
   805  			}
   806  			continue
   807  		}
   808  
   809  		// Claim a chunk of work.
   810  		idx := uintptr(h.reclaimIndex.Add(pagesPerReclaimerChunk) - pagesPerReclaimerChunk)
   811  		if idx/pagesPerArena >= uintptr(len(arenas)) {
   812  			// Page reclaiming is done.
   813  			h.reclaimIndex.Store(1 << 63)
   814  			break
   815  		}
   816  
   817  		if !locked {
   818  			// Lock the heap for reclaimChunk.
   819  			lock(&h.lock)
   820  			locked = true
   821  		}
   822  
   823  		// Scan this chunk.
   824  		nfound := h.reclaimChunk(arenas, idx, pagesPerReclaimerChunk)
   825  		if nfound <= npage {
   826  			npage -= nfound
   827  		} else {
   828  			// Put spare pages toward global credit.
   829  			h.reclaimCredit.Add(nfound - npage)
   830  			npage = 0
   831  		}
   832  	}
   833  	if locked {
   834  		unlock(&h.lock)
   835  	}
   836  
   837  	if trace.enabled {
   838  		traceGCSweepDone()
   839  	}
   840  	releasem(mp)
   841  }
   842  
   843  // reclaimChunk sweeps unmarked spans that start at page indexes [pageIdx, pageIdx+n).
   844  // It returns the number of pages returned to the heap.
   845  //
   846  // h.lock must be held and the caller must be non-preemptible. Note: h.lock may be
   847  // temporarily unlocked and re-locked in order to do sweeping or if tracing is
   848  // enabled.
   849  func (h *mheap) reclaimChunk(arenas []arenaIdx, pageIdx, n uintptr) uintptr {
   850  	// The heap lock must be held because this accesses the
   851  	// heapArena.spans arrays using potentially non-live pointers.
   852  	// In particular, if a span were freed and merged concurrently
   853  	// with this probing heapArena.spans, it would be possible to
   854  	// observe arbitrary, stale span pointers.
   855  	assertLockHeld(&h.lock)
   856  
   857  	n0 := n
   858  	var nFreed uintptr
   859  	sl := sweep.active.begin()
   860  	if !sl.valid {
   861  		return 0
   862  	}
   863  	for n > 0 {
   864  		ai := arenas[pageIdx/pagesPerArena]
   865  		ha := h.arenas[ai.l1()][ai.l2()]
   866  
   867  		// Get a chunk of the bitmap to work on.
   868  		arenaPage := uint(pageIdx % pagesPerArena)
   869  		inUse := ha.pageInUse[arenaPage/8:]
   870  		marked := ha.pageMarks[arenaPage/8:]
   871  		if uintptr(len(inUse)) > n/8 {
   872  			inUse = inUse[:n/8]
   873  			marked = marked[:n/8]
   874  		}
   875  
   876  		// Scan this bitmap chunk for spans that are in-use
   877  		// but have no marked objects on them.
   878  		for i := range inUse {
   879  			inUseUnmarked := atomic.Load8(&inUse[i]) &^ marked[i]
   880  			if inUseUnmarked == 0 {
   881  				continue
   882  			}
   883  
   884  			for j := uint(0); j < 8; j++ {
   885  				if inUseUnmarked&(1<<j) != 0 {
   886  					s := ha.spans[arenaPage+uint(i)*8+j]
   887  					if s, ok := sl.tryAcquire(s); ok {
   888  						npages := s.npages
   889  						unlock(&h.lock)
   890  						if s.sweep(false) {
   891  							nFreed += npages
   892  						}
   893  						lock(&h.lock)
   894  						// Reload inUse. It's possible nearby
   895  						// spans were freed when we dropped the
   896  						// lock and we don't want to get stale
   897  						// pointers from the spans array.
   898  						inUseUnmarked = atomic.Load8(&inUse[i]) &^ marked[i]
   899  					}
   900  				}
   901  			}
   902  		}
   903  
   904  		// Advance.
   905  		pageIdx += uintptr(len(inUse) * 8)
   906  		n -= uintptr(len(inUse) * 8)
   907  	}
   908  	sweep.active.end(sl)
   909  	if trace.enabled {
   910  		unlock(&h.lock)
   911  		// Account for pages scanned but not reclaimed.
   912  		traceGCSweepSpan((n0 - nFreed) * pageSize)
   913  		lock(&h.lock)
   914  	}
   915  
   916  	assertLockHeld(&h.lock) // Must be locked on return.
   917  	return nFreed
   918  }
   919  
   920  // spanAllocType represents the type of allocation to make, or
   921  // the type of allocation to be freed.
   922  type spanAllocType uint8
   923  
   924  const (
   925  	spanAllocHeap          spanAllocType = iota // heap span
   926  	spanAllocStack                              // stack span
   927  	spanAllocPtrScalarBits                      // unrolled GC prog bitmap span
   928  	spanAllocWorkBuf                            // work buf span
   929  )
   930  
   931  // manual returns true if the span allocation is manually managed.
   932  func (s spanAllocType) manual() bool {
   933  	return s != spanAllocHeap
   934  }
   935  
   936  // alloc allocates a new span of npage pages from the GC'd heap.
   937  //
   938  // spanclass indicates the span's size class and scannability.
   939  //
   940  // Returns a span that has been fully initialized. span.needzero indicates
   941  // whether the span has been zeroed. Note that it may not be.
   942  func (h *mheap) alloc(npages uintptr, spanclass spanClass) *mspan {
   943  	// Don't do any operations that lock the heap on the G stack.
   944  	// It might trigger stack growth, and the stack growth code needs
   945  	// to be able to allocate heap.
   946  	var s *mspan
   947  	systemstack(func() {
   948  		// To prevent excessive heap growth, before allocating n pages
   949  		// we need to sweep and reclaim at least n pages.
   950  		if !isSweepDone() {
   951  			h.reclaim(npages)
   952  		}
   953  		s = h.allocSpan(npages, spanAllocHeap, spanclass)
   954  	})
   955  	return s
   956  }
   957  
   958  // allocManual allocates a manually-managed span of npage pages.
   959  // allocManual returns nil if allocation fails.
   960  //
   961  // allocManual adds the bytes used to *stat, which should be a
   962  // memstats in-use field. Unlike allocations in the GC'd heap, the
   963  // allocation does *not* count toward heapInUse.
   964  //
   965  // The memory backing the returned span may not be zeroed if
   966  // span.needzero is set.
   967  //
   968  // allocManual must be called on the system stack because it may
   969  // acquire the heap lock via allocSpan. See mheap for details.
   970  //
   971  // If new code is written to call allocManual, do NOT use an
   972  // existing spanAllocType value and instead declare a new one.
   973  //
   974  //go:systemstack
   975  func (h *mheap) allocManual(npages uintptr, typ spanAllocType) *mspan {
   976  	if !typ.manual() {
   977  		throw("manual span allocation called with non-manually-managed type")
   978  	}
   979  	return h.allocSpan(npages, typ, 0)
   980  }
   981  
   982  // setSpans modifies the span map so [spanOf(base), spanOf(base+npage*pageSize))
   983  // is s.
   984  func (h *mheap) setSpans(base, npage uintptr, s *mspan) {
   985  	p := base / pageSize
   986  	ai := arenaIndex(base)
   987  	ha := h.arenas[ai.l1()][ai.l2()]
   988  	for n := uintptr(0); n < npage; n++ {
   989  		i := (p + n) % pagesPerArena
   990  		if i == 0 {
   991  			ai = arenaIndex(base + n*pageSize)
   992  			ha = h.arenas[ai.l1()][ai.l2()]
   993  		}
   994  		ha.spans[i] = s
   995  	}
   996  }
   997  
   998  // allocNeedsZero checks if the region of address space [base, base+npage*pageSize),
   999  // assumed to be allocated, needs to be zeroed, updating heap arena metadata for
  1000  // future allocations.
  1001  //
  1002  // This must be called each time pages are allocated from the heap, even if the page
  1003  // allocator can otherwise prove the memory it's allocating is already zero because
  1004  // they're fresh from the operating system. It updates heapArena metadata that is
  1005  // critical for future page allocations.
  1006  //
  1007  // There are no locking constraints on this method.
  1008  func (h *mheap) allocNeedsZero(base, npage uintptr) (needZero bool) {
  1009  	for npage > 0 {
  1010  		ai := arenaIndex(base)
  1011  		ha := h.arenas[ai.l1()][ai.l2()]
  1012  
  1013  		zeroedBase := atomic.Loaduintptr(&ha.zeroedBase)
  1014  		arenaBase := base % heapArenaBytes
  1015  		if arenaBase < zeroedBase {
  1016  			// We extended into the non-zeroed part of the
  1017  			// arena, so this region needs to be zeroed before use.
  1018  			//
  1019  			// zeroedBase is monotonically increasing, so if we see this now then
  1020  			// we can be sure we need to zero this memory region.
  1021  			//
  1022  			// We still need to update zeroedBase for this arena, and
  1023  			// potentially more arenas.
  1024  			needZero = true
  1025  		}
  1026  		// We may observe arenaBase > zeroedBase if we're racing with one or more
  1027  		// allocations which are acquiring memory directly before us in the address
  1028  		// space. But, because we know no one else is acquiring *this* memory, it's
  1029  		// still safe to not zero.
  1030  
  1031  		// Compute how far into the arena we extend into, capped
  1032  		// at heapArenaBytes.
  1033  		arenaLimit := arenaBase + npage*pageSize
  1034  		if arenaLimit > heapArenaBytes {
  1035  			arenaLimit = heapArenaBytes
  1036  		}
  1037  		// Increase ha.zeroedBase so it's >= arenaLimit.
  1038  		// We may be racing with other updates.
  1039  		for arenaLimit > zeroedBase {
  1040  			if atomic.Casuintptr(&ha.zeroedBase, zeroedBase, arenaLimit) {
  1041  				break
  1042  			}
  1043  			zeroedBase = atomic.Loaduintptr(&ha.zeroedBase)
  1044  			// Double check basic conditions of zeroedBase.
  1045  			if zeroedBase <= arenaLimit && zeroedBase > arenaBase {
  1046  				// The zeroedBase moved into the space we were trying to
  1047  				// claim. That's very bad, and indicates someone allocated
  1048  				// the same region we did.
  1049  				throw("potentially overlapping in-use allocations detected")
  1050  			}
  1051  		}
  1052  
  1053  		// Move base forward and subtract from npage to move into
  1054  		// the next arena, or finish.
  1055  		base += arenaLimit - arenaBase
  1056  		npage -= (arenaLimit - arenaBase) / pageSize
  1057  	}
  1058  	return
  1059  }
  1060  
  1061  // tryAllocMSpan attempts to allocate an mspan object from
  1062  // the P-local cache, but may fail.
  1063  //
  1064  // h.lock need not be held.
  1065  //
  1066  // This caller must ensure that its P won't change underneath
  1067  // it during this function. Currently to ensure that we enforce
  1068  // that the function is run on the system stack, because that's
  1069  // the only place it is used now. In the future, this requirement
  1070  // may be relaxed if its use is necessary elsewhere.
  1071  //
  1072  //go:systemstack
  1073  func (h *mheap) tryAllocMSpan() *mspan {
  1074  	pp := getg().m.p.ptr()
  1075  	// If we don't have a p or the cache is empty, we can't do
  1076  	// anything here.
  1077  	if pp == nil || pp.mspancache.len == 0 {
  1078  		return nil
  1079  	}
  1080  	// Pull off the last entry in the cache.
  1081  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1082  	pp.mspancache.len--
  1083  	return s
  1084  }
  1085  
  1086  // allocMSpanLocked allocates an mspan object.
  1087  //
  1088  // h.lock must be held.
  1089  //
  1090  // allocMSpanLocked must be called on the system stack because
  1091  // its caller holds the heap lock. See mheap for details.
  1092  // Running on the system stack also ensures that we won't
  1093  // switch Ps during this function. See tryAllocMSpan for details.
  1094  //
  1095  //go:systemstack
  1096  func (h *mheap) allocMSpanLocked() *mspan {
  1097  	assertLockHeld(&h.lock)
  1098  
  1099  	pp := getg().m.p.ptr()
  1100  	if pp == nil {
  1101  		// We don't have a p so just do the normal thing.
  1102  		return (*mspan)(h.spanalloc.alloc())
  1103  	}
  1104  	// Refill the cache if necessary.
  1105  	if pp.mspancache.len == 0 {
  1106  		const refillCount = len(pp.mspancache.buf) / 2
  1107  		for i := 0; i < refillCount; i++ {
  1108  			pp.mspancache.buf[i] = (*mspan)(h.spanalloc.alloc())
  1109  		}
  1110  		pp.mspancache.len = refillCount
  1111  	}
  1112  	// Pull off the last entry in the cache.
  1113  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1114  	pp.mspancache.len--
  1115  	return s
  1116  }
  1117  
  1118  // freeMSpanLocked free an mspan object.
  1119  //
  1120  // h.lock must be held.
  1121  //
  1122  // freeMSpanLocked must be called on the system stack because
  1123  // its caller holds the heap lock. See mheap for details.
  1124  // Running on the system stack also ensures that we won't
  1125  // switch Ps during this function. See tryAllocMSpan for details.
  1126  //
  1127  //go:systemstack
  1128  func (h *mheap) freeMSpanLocked(s *mspan) {
  1129  	assertLockHeld(&h.lock)
  1130  
  1131  	pp := getg().m.p.ptr()
  1132  	// First try to free the mspan directly to the cache.
  1133  	if pp != nil && pp.mspancache.len < len(pp.mspancache.buf) {
  1134  		pp.mspancache.buf[pp.mspancache.len] = s
  1135  		pp.mspancache.len++
  1136  		return
  1137  	}
  1138  	// Failing that (or if we don't have a p), just free it to
  1139  	// the heap.
  1140  	h.spanalloc.free(unsafe.Pointer(s))
  1141  }
  1142  
  1143  // allocSpan allocates an mspan which owns npages worth of memory.
  1144  //
  1145  // If typ.manual() == false, allocSpan allocates a heap span of class spanclass
  1146  // and updates heap accounting. If manual == true, allocSpan allocates a
  1147  // manually-managed span (spanclass is ignored), and the caller is
  1148  // responsible for any accounting related to its use of the span. Either
  1149  // way, allocSpan will atomically add the bytes in the newly allocated
  1150  // span to *sysStat.
  1151  //
  1152  // The returned span is fully initialized.
  1153  //
  1154  // h.lock must not be held.
  1155  //
  1156  // allocSpan must be called on the system stack both because it acquires
  1157  // the heap lock and because it must block GC transitions.
  1158  //
  1159  //go:systemstack
  1160  func (h *mheap) allocSpan(npages uintptr, typ spanAllocType, spanclass spanClass) (s *mspan) {
  1161  	// Function-global state.
  1162  	gp := getg()
  1163  	base, scav := uintptr(0), uintptr(0)
  1164  	growth := uintptr(0)
  1165  
  1166  	// On some platforms we need to provide physical page aligned stack
  1167  	// allocations. Where the page size is less than the physical page
  1168  	// size, we already manage to do this by default.
  1169  	needPhysPageAlign := physPageAlignedStacks && typ == spanAllocStack && pageSize < physPageSize
  1170  
  1171  	// If the allocation is small enough, try the page cache!
  1172  	// The page cache does not support aligned allocations, so we cannot use
  1173  	// it if we need to provide a physical page aligned stack allocation.
  1174  	pp := gp.m.p.ptr()
  1175  	if !needPhysPageAlign && pp != nil && npages < pageCachePages/4 {
  1176  		c := &pp.pcache
  1177  
  1178  		// If the cache is empty, refill it.
  1179  		if c.empty() {
  1180  			lock(&h.lock)
  1181  			*c = h.pages.allocToCache()
  1182  			unlock(&h.lock)
  1183  		}
  1184  
  1185  		// Try to allocate from the cache.
  1186  		base, scav = c.alloc(npages)
  1187  		if base != 0 {
  1188  			s = h.tryAllocMSpan()
  1189  			if s != nil {
  1190  				goto HaveSpan
  1191  			}
  1192  			// We have a base but no mspan, so we need
  1193  			// to lock the heap.
  1194  		}
  1195  	}
  1196  
  1197  	// For one reason or another, we couldn't get the
  1198  	// whole job done without the heap lock.
  1199  	lock(&h.lock)
  1200  
  1201  	if needPhysPageAlign {
  1202  		// Overallocate by a physical page to allow for later alignment.
  1203  		extraPages := physPageSize / pageSize
  1204  
  1205  		// Find a big enough region first, but then only allocate the
  1206  		// aligned portion. We can't just allocate and then free the
  1207  		// edges because we need to account for scavenged memory, and
  1208  		// that's difficult with alloc.
  1209  		//
  1210  		// Note that we skip updates to searchAddr here. It's OK if
  1211  		// it's stale and higher than normal; it'll operate correctly,
  1212  		// just come with a performance cost.
  1213  		base, _ = h.pages.find(npages + extraPages)
  1214  		if base == 0 {
  1215  			var ok bool
  1216  			growth, ok = h.grow(npages + extraPages)
  1217  			if !ok {
  1218  				unlock(&h.lock)
  1219  				return nil
  1220  			}
  1221  			base, _ = h.pages.find(npages + extraPages)
  1222  			if base == 0 {
  1223  				throw("grew heap, but no adequate free space found")
  1224  			}
  1225  		}
  1226  		base = alignUp(base, physPageSize)
  1227  		scav = h.pages.allocRange(base, npages)
  1228  	}
  1229  
  1230  	if base == 0 {
  1231  		// Try to acquire a base address.
  1232  		base, scav = h.pages.alloc(npages)
  1233  		if base == 0 {
  1234  			var ok bool
  1235  			growth, ok = h.grow(npages)
  1236  			if !ok {
  1237  				unlock(&h.lock)
  1238  				return nil
  1239  			}
  1240  			base, scav = h.pages.alloc(npages)
  1241  			if base == 0 {
  1242  				throw("grew heap, but no adequate free space found")
  1243  			}
  1244  		}
  1245  	}
  1246  	if s == nil {
  1247  		// We failed to get an mspan earlier, so grab
  1248  		// one now that we have the heap lock.
  1249  		s = h.allocMSpanLocked()
  1250  	}
  1251  	unlock(&h.lock)
  1252  
  1253  HaveSpan:
  1254  	// Decide if we need to scavenge in response to what we just allocated.
  1255  	// Specifically, we track the maximum amount of memory to scavenge of all
  1256  	// the alternatives below, assuming that the maximum satisfies *all*
  1257  	// conditions we check (e.g. if we need to scavenge X to satisfy the
  1258  	// memory limit and Y to satisfy heap-growth scavenging, and Y > X, then
  1259  	// it's fine to pick Y, because the memory limit is still satisfied).
  1260  	//
  1261  	// It's fine to do this after allocating because we expect any scavenged
  1262  	// pages not to get touched until we return. Simultaneously, it's important
  1263  	// to do this before calling sysUsed because that may commit address space.
  1264  	bytesToScavenge := uintptr(0)
  1265  	if limit := gcController.memoryLimit.Load(); go119MemoryLimitSupport && !gcCPULimiter.limiting() {
  1266  		// Assist with scavenging to maintain the memory limit by the amount
  1267  		// that we expect to page in.
  1268  		inuse := gcController.mappedReady.Load()
  1269  		// Be careful about overflow, especially with uintptrs. Even on 32-bit platforms
  1270  		// someone can set a really big memory limit that isn't maxInt64.
  1271  		if uint64(scav)+inuse > uint64(limit) {
  1272  			bytesToScavenge = uintptr(uint64(scav) + inuse - uint64(limit))
  1273  		}
  1274  	}
  1275  	if goal := scavenge.gcPercentGoal.Load(); goal != ^uint64(0) && growth > 0 {
  1276  		// We just caused a heap growth, so scavenge down what will soon be used.
  1277  		// By scavenging inline we deal with the failure to allocate out of
  1278  		// memory fragments by scavenging the memory fragments that are least
  1279  		// likely to be re-used.
  1280  		//
  1281  		// Only bother with this because we're not using a memory limit. We don't
  1282  		// care about heap growths as long as we're under the memory limit, and the
  1283  		// previous check for scaving already handles that.
  1284  		if retained := heapRetained(); retained+uint64(growth) > goal {
  1285  			// The scavenging algorithm requires the heap lock to be dropped so it
  1286  			// can acquire it only sparingly. This is a potentially expensive operation
  1287  			// so it frees up other goroutines to allocate in the meanwhile. In fact,
  1288  			// they can make use of the growth we just created.
  1289  			todo := growth
  1290  			if overage := uintptr(retained + uint64(growth) - goal); todo > overage {
  1291  				todo = overage
  1292  			}
  1293  			if todo > bytesToScavenge {
  1294  				bytesToScavenge = todo
  1295  			}
  1296  		}
  1297  	}
  1298  	// There are a few very limited cirumstances where we won't have a P here.
  1299  	// It's OK to simply skip scavenging in these cases. Something else will notice
  1300  	// and pick up the tab.
  1301  	if pp != nil && bytesToScavenge > 0 {
  1302  		// Measure how long we spent scavenging and add that measurement to the assist
  1303  		// time so we can track it for the GC CPU limiter.
  1304  		//
  1305  		// Limiter event tracking might be disabled if we end up here
  1306  		// while on a mark worker.
  1307  		start := nanotime()
  1308  		track := pp.limiterEvent.start(limiterEventScavengeAssist, start)
  1309  
  1310  		// Scavenge, but back out if the limiter turns on.
  1311  		h.pages.scavenge(bytesToScavenge, func() bool {
  1312  			return gcCPULimiter.limiting()
  1313  		})
  1314  
  1315  		// Finish up accounting.
  1316  		now := nanotime()
  1317  		if track {
  1318  			pp.limiterEvent.stop(limiterEventScavengeAssist, now)
  1319  		}
  1320  		scavenge.assistTime.Add(now - start)
  1321  	}
  1322  
  1323  	// Initialize the span.
  1324  	h.initSpan(s, typ, spanclass, base, npages)
  1325  
  1326  	// Commit and account for any scavenged memory that the span now owns.
  1327  	nbytes := npages * pageSize
  1328  	if scav != 0 {
  1329  		// sysUsed all the pages that are actually available
  1330  		// in the span since some of them might be scavenged.
  1331  		sysUsed(unsafe.Pointer(base), nbytes, scav)
  1332  		gcController.heapReleased.add(-int64(scav))
  1333  	}
  1334  	// Update stats.
  1335  	gcController.heapFree.add(-int64(nbytes - scav))
  1336  	if typ == spanAllocHeap {
  1337  		gcController.heapInUse.add(int64(nbytes))
  1338  	}
  1339  	// Update consistent stats.
  1340  	stats := memstats.heapStats.acquire()
  1341  	atomic.Xaddint64(&stats.committed, int64(scav))
  1342  	atomic.Xaddint64(&stats.released, -int64(scav))
  1343  	switch typ {
  1344  	case spanAllocHeap:
  1345  		atomic.Xaddint64(&stats.inHeap, int64(nbytes))
  1346  	case spanAllocStack:
  1347  		atomic.Xaddint64(&stats.inStacks, int64(nbytes))
  1348  	case spanAllocPtrScalarBits:
  1349  		atomic.Xaddint64(&stats.inPtrScalarBits, int64(nbytes))
  1350  	case spanAllocWorkBuf:
  1351  		atomic.Xaddint64(&stats.inWorkBufs, int64(nbytes))
  1352  	}
  1353  	memstats.heapStats.release()
  1354  
  1355  	return s
  1356  }
  1357  
  1358  // initSpan initializes a blank span s which will represent the range
  1359  // [base, base+npages*pageSize). typ is the type of span being allocated.
  1360  func (h *mheap) initSpan(s *mspan, typ spanAllocType, spanclass spanClass, base, npages uintptr) {
  1361  	// At this point, both s != nil and base != 0, and the heap
  1362  	// lock is no longer held. Initialize the span.
  1363  	s.init(base, npages)
  1364  	if h.allocNeedsZero(base, npages) {
  1365  		s.needzero = 1
  1366  	}
  1367  	nbytes := npages * pageSize
  1368  	if typ.manual() {
  1369  		s.manualFreeList = 0
  1370  		s.nelems = 0
  1371  		s.limit = s.base() + s.npages*pageSize
  1372  		s.state.set(mSpanManual)
  1373  	} else {
  1374  		// We must set span properties before the span is published anywhere
  1375  		// since we're not holding the heap lock.
  1376  		s.spanclass = spanclass
  1377  		if sizeclass := spanclass.sizeclass(); sizeclass == 0 {
  1378  			s.elemsize = nbytes
  1379  			s.nelems = 1
  1380  			s.divMul = 0
  1381  		} else {
  1382  			s.elemsize = uintptr(class_to_size[sizeclass])
  1383  			s.nelems = nbytes / s.elemsize
  1384  			s.divMul = class_to_divmagic[sizeclass]
  1385  		}
  1386  
  1387  		// Initialize mark and allocation structures.
  1388  		s.freeindex = 0
  1389  		s.allocCache = ^uint64(0) // all 1s indicating all free.
  1390  		s.gcmarkBits = newMarkBits(s.nelems)
  1391  		s.allocBits = newAllocBits(s.nelems)
  1392  
  1393  		// It's safe to access h.sweepgen without the heap lock because it's
  1394  		// only ever updated with the world stopped and we run on the
  1395  		// systemstack which blocks a STW transition.
  1396  		atomic.Store(&s.sweepgen, h.sweepgen)
  1397  
  1398  		// Now that the span is filled in, set its state. This
  1399  		// is a publication barrier for the other fields in
  1400  		// the span. While valid pointers into this span
  1401  		// should never be visible until the span is returned,
  1402  		// if the garbage collector finds an invalid pointer,
  1403  		// access to the span may race with initialization of
  1404  		// the span. We resolve this race by atomically
  1405  		// setting the state after the span is fully
  1406  		// initialized, and atomically checking the state in
  1407  		// any situation where a pointer is suspect.
  1408  		s.state.set(mSpanInUse)
  1409  	}
  1410  
  1411  	// Publish the span in various locations.
  1412  
  1413  	// This is safe to call without the lock held because the slots
  1414  	// related to this span will only ever be read or modified by
  1415  	// this thread until pointers into the span are published (and
  1416  	// we execute a publication barrier at the end of this function
  1417  	// before that happens) or pageInUse is updated.
  1418  	h.setSpans(s.base(), npages, s)
  1419  
  1420  	if !typ.manual() {
  1421  		// Mark in-use span in arena page bitmap.
  1422  		//
  1423  		// This publishes the span to the page sweeper, so
  1424  		// it's imperative that the span be completely initialized
  1425  		// prior to this line.
  1426  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1427  		atomic.Or8(&arena.pageInUse[pageIdx], pageMask)
  1428  
  1429  		// Update related page sweeper stats.
  1430  		h.pagesInUse.Add(npages)
  1431  	}
  1432  
  1433  	// Make sure the newly allocated span will be observed
  1434  	// by the GC before pointers into the span are published.
  1435  	publicationBarrier()
  1436  }
  1437  
  1438  // Try to add at least npage pages of memory to the heap,
  1439  // returning how much the heap grew by and whether it worked.
  1440  //
  1441  // h.lock must be held.
  1442  func (h *mheap) grow(npage uintptr) (uintptr, bool) {
  1443  	assertLockHeld(&h.lock)
  1444  
  1445  	// We must grow the heap in whole palloc chunks.
  1446  	// We call sysMap below but note that because we
  1447  	// round up to pallocChunkPages which is on the order
  1448  	// of MiB (generally >= to the huge page size) we
  1449  	// won't be calling it too much.
  1450  	ask := alignUp(npage, pallocChunkPages) * pageSize
  1451  
  1452  	totalGrowth := uintptr(0)
  1453  	// This may overflow because ask could be very large
  1454  	// and is otherwise unrelated to h.curArena.base.
  1455  	end := h.curArena.base + ask
  1456  	nBase := alignUp(end, physPageSize)
  1457  	if nBase > h.curArena.end || /* overflow */ end < h.curArena.base {
  1458  		// Not enough room in the current arena. Allocate more
  1459  		// arena space. This may not be contiguous with the
  1460  		// current arena, so we have to request the full ask.
  1461  		av, asize := h.sysAlloc(ask, &h.arenaHints, true)
  1462  		if av == nil {
  1463  			inUse := gcController.heapFree.load() + gcController.heapReleased.load() + gcController.heapInUse.load()
  1464  			print("runtime: out of memory: cannot allocate ", ask, "-byte block (", inUse, " in use)\n")
  1465  			return 0, false
  1466  		}
  1467  
  1468  		if uintptr(av) == h.curArena.end {
  1469  			// The new space is contiguous with the old
  1470  			// space, so just extend the current space.
  1471  			h.curArena.end = uintptr(av) + asize
  1472  		} else {
  1473  			// The new space is discontiguous. Track what
  1474  			// remains of the current space and switch to
  1475  			// the new space. This should be rare.
  1476  			if size := h.curArena.end - h.curArena.base; size != 0 {
  1477  				// Transition this space from Reserved to Prepared and mark it
  1478  				// as released since we'll be able to start using it after updating
  1479  				// the page allocator and releasing the lock at any time.
  1480  				sysMap(unsafe.Pointer(h.curArena.base), size, &gcController.heapReleased)
  1481  				// Update stats.
  1482  				stats := memstats.heapStats.acquire()
  1483  				atomic.Xaddint64(&stats.released, int64(size))
  1484  				memstats.heapStats.release()
  1485  				// Update the page allocator's structures to make this
  1486  				// space ready for allocation.
  1487  				h.pages.grow(h.curArena.base, size)
  1488  				totalGrowth += size
  1489  			}
  1490  			// Switch to the new space.
  1491  			h.curArena.base = uintptr(av)
  1492  			h.curArena.end = uintptr(av) + asize
  1493  		}
  1494  
  1495  		// Recalculate nBase.
  1496  		// We know this won't overflow, because sysAlloc returned
  1497  		// a valid region starting at h.curArena.base which is at
  1498  		// least ask bytes in size.
  1499  		nBase = alignUp(h.curArena.base+ask, physPageSize)
  1500  	}
  1501  
  1502  	// Grow into the current arena.
  1503  	v := h.curArena.base
  1504  	h.curArena.base = nBase
  1505  
  1506  	// Transition the space we're going to use from Reserved to Prepared.
  1507  	//
  1508  	// The allocation is always aligned to the heap arena
  1509  	// size which is always > physPageSize, so its safe to
  1510  	// just add directly to heapReleased.
  1511  	sysMap(unsafe.Pointer(v), nBase-v, &gcController.heapReleased)
  1512  
  1513  	// The memory just allocated counts as both released
  1514  	// and idle, even though it's not yet backed by spans.
  1515  	stats := memstats.heapStats.acquire()
  1516  	atomic.Xaddint64(&stats.released, int64(nBase-v))
  1517  	memstats.heapStats.release()
  1518  
  1519  	// Update the page allocator's structures to make this
  1520  	// space ready for allocation.
  1521  	h.pages.grow(v, nBase-v)
  1522  	totalGrowth += nBase - v
  1523  	return totalGrowth, true
  1524  }
  1525  
  1526  // Free the span back into the heap.
  1527  func (h *mheap) freeSpan(s *mspan) {
  1528  	systemstack(func() {
  1529  		lock(&h.lock)
  1530  		if msanenabled {
  1531  			// Tell msan that this entire span is no longer in use.
  1532  			base := unsafe.Pointer(s.base())
  1533  			bytes := s.npages << _PageShift
  1534  			msanfree(base, bytes)
  1535  		}
  1536  		if asanenabled {
  1537  			// Tell asan that this entire span is no longer in use.
  1538  			base := unsafe.Pointer(s.base())
  1539  			bytes := s.npages << _PageShift
  1540  			asanpoison(base, bytes)
  1541  		}
  1542  		h.freeSpanLocked(s, spanAllocHeap)
  1543  		unlock(&h.lock)
  1544  	})
  1545  }
  1546  
  1547  // freeManual frees a manually-managed span returned by allocManual.
  1548  // typ must be the same as the spanAllocType passed to the allocManual that
  1549  // allocated s.
  1550  //
  1551  // This must only be called when gcphase == _GCoff. See mSpanState for
  1552  // an explanation.
  1553  //
  1554  // freeManual must be called on the system stack because it acquires
  1555  // the heap lock. See mheap for details.
  1556  //
  1557  //go:systemstack
  1558  func (h *mheap) freeManual(s *mspan, typ spanAllocType) {
  1559  	s.needzero = 1
  1560  	lock(&h.lock)
  1561  	h.freeSpanLocked(s, typ)
  1562  	unlock(&h.lock)
  1563  }
  1564  
  1565  func (h *mheap) freeSpanLocked(s *mspan, typ spanAllocType) {
  1566  	assertLockHeld(&h.lock)
  1567  
  1568  	switch s.state.get() {
  1569  	case mSpanManual:
  1570  		if s.allocCount != 0 {
  1571  			throw("mheap.freeSpanLocked - invalid stack free")
  1572  		}
  1573  	case mSpanInUse:
  1574  		if s.isUserArenaChunk {
  1575  			throw("mheap.freeSpanLocked - invalid free of user arena chunk")
  1576  		}
  1577  		if s.allocCount != 0 || s.sweepgen != h.sweepgen {
  1578  			print("mheap.freeSpanLocked - span ", s, " ptr ", hex(s.base()), " allocCount ", s.allocCount, " sweepgen ", s.sweepgen, "/", h.sweepgen, "\n")
  1579  			throw("mheap.freeSpanLocked - invalid free")
  1580  		}
  1581  		h.pagesInUse.Add(-s.npages)
  1582  
  1583  		// Clear in-use bit in arena page bitmap.
  1584  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1585  		atomic.And8(&arena.pageInUse[pageIdx], ^pageMask)
  1586  	default:
  1587  		throw("mheap.freeSpanLocked - invalid span state")
  1588  	}
  1589  
  1590  	// Update stats.
  1591  	//
  1592  	// Mirrors the code in allocSpan.
  1593  	nbytes := s.npages * pageSize
  1594  	gcController.heapFree.add(int64(nbytes))
  1595  	if typ == spanAllocHeap {
  1596  		gcController.heapInUse.add(-int64(nbytes))
  1597  	}
  1598  	// Update consistent stats.
  1599  	stats := memstats.heapStats.acquire()
  1600  	switch typ {
  1601  	case spanAllocHeap:
  1602  		atomic.Xaddint64(&stats.inHeap, -int64(nbytes))
  1603  	case spanAllocStack:
  1604  		atomic.Xaddint64(&stats.inStacks, -int64(nbytes))
  1605  	case spanAllocPtrScalarBits:
  1606  		atomic.Xaddint64(&stats.inPtrScalarBits, -int64(nbytes))
  1607  	case spanAllocWorkBuf:
  1608  		atomic.Xaddint64(&stats.inWorkBufs, -int64(nbytes))
  1609  	}
  1610  	memstats.heapStats.release()
  1611  
  1612  	// Mark the space as free.
  1613  	h.pages.free(s.base(), s.npages, false)
  1614  
  1615  	// Free the span structure. We no longer have a use for it.
  1616  	s.state.set(mSpanDead)
  1617  	h.freeMSpanLocked(s)
  1618  }
  1619  
  1620  // scavengeAll acquires the heap lock (blocking any additional
  1621  // manipulation of the page allocator) and iterates over the whole
  1622  // heap, scavenging every free page available.
  1623  func (h *mheap) scavengeAll() {
  1624  	// Disallow malloc or panic while holding the heap lock. We do
  1625  	// this here because this is a non-mallocgc entry-point to
  1626  	// the mheap API.
  1627  	gp := getg()
  1628  	gp.m.mallocing++
  1629  
  1630  	released := h.pages.scavenge(^uintptr(0), nil)
  1631  
  1632  	gp.m.mallocing--
  1633  
  1634  	if debug.scavtrace > 0 {
  1635  		printScavTrace(released, true)
  1636  	}
  1637  }
  1638  
  1639  //go:linkname runtime_debug_freeOSMemory runtime/debug.freeOSMemory
  1640  func runtime_debug_freeOSMemory() {
  1641  	GC()
  1642  	systemstack(func() { mheap_.scavengeAll() })
  1643  }
  1644  
  1645  // Initialize a new span with the given start and npages.
  1646  func (span *mspan) init(base uintptr, npages uintptr) {
  1647  	// span is *not* zeroed.
  1648  	span.next = nil
  1649  	span.prev = nil
  1650  	span.list = nil
  1651  	span.startAddr = base
  1652  	span.npages = npages
  1653  	span.allocCount = 0
  1654  	span.spanclass = 0
  1655  	span.elemsize = 0
  1656  	span.speciallock.key = 0
  1657  	span.specials = nil
  1658  	span.needzero = 0
  1659  	span.freeindex = 0
  1660  	span.allocBits = nil
  1661  	span.gcmarkBits = nil
  1662  	span.state.set(mSpanDead)
  1663  	lockInit(&span.speciallock, lockRankMspanSpecial)
  1664  }
  1665  
  1666  func (span *mspan) inList() bool {
  1667  	return span.list != nil
  1668  }
  1669  
  1670  // Initialize an empty doubly-linked list.
  1671  func (list *mSpanList) init() {
  1672  	list.first = nil
  1673  	list.last = nil
  1674  }
  1675  
  1676  func (list *mSpanList) remove(span *mspan) {
  1677  	if span.list != list {
  1678  		print("runtime: failed mSpanList.remove span.npages=", span.npages,
  1679  			" span=", span, " prev=", span.prev, " span.list=", span.list, " list=", list, "\n")
  1680  		throw("mSpanList.remove")
  1681  	}
  1682  	if list.first == span {
  1683  		list.first = span.next
  1684  	} else {
  1685  		span.prev.next = span.next
  1686  	}
  1687  	if list.last == span {
  1688  		list.last = span.prev
  1689  	} else {
  1690  		span.next.prev = span.prev
  1691  	}
  1692  	span.next = nil
  1693  	span.prev = nil
  1694  	span.list = nil
  1695  }
  1696  
  1697  func (list *mSpanList) isEmpty() bool {
  1698  	return list.first == nil
  1699  }
  1700  
  1701  func (list *mSpanList) insert(span *mspan) {
  1702  	if span.next != nil || span.prev != nil || span.list != nil {
  1703  		println("runtime: failed mSpanList.insert", span, span.next, span.prev, span.list)
  1704  		throw("mSpanList.insert")
  1705  	}
  1706  	span.next = list.first
  1707  	if list.first != nil {
  1708  		// The list contains at least one span; link it in.
  1709  		// The last span in the list doesn't change.
  1710  		list.first.prev = span
  1711  	} else {
  1712  		// The list contains no spans, so this is also the last span.
  1713  		list.last = span
  1714  	}
  1715  	list.first = span
  1716  	span.list = list
  1717  }
  1718  
  1719  func (list *mSpanList) insertBack(span *mspan) {
  1720  	if span.next != nil || span.prev != nil || span.list != nil {
  1721  		println("runtime: failed mSpanList.insertBack", span, span.next, span.prev, span.list)
  1722  		throw("mSpanList.insertBack")
  1723  	}
  1724  	span.prev = list.last
  1725  	if list.last != nil {
  1726  		// The list contains at least one span.
  1727  		list.last.next = span
  1728  	} else {
  1729  		// The list contains no spans, so this is also the first span.
  1730  		list.first = span
  1731  	}
  1732  	list.last = span
  1733  	span.list = list
  1734  }
  1735  
  1736  // takeAll removes all spans from other and inserts them at the front
  1737  // of list.
  1738  func (list *mSpanList) takeAll(other *mSpanList) {
  1739  	if other.isEmpty() {
  1740  		return
  1741  	}
  1742  
  1743  	// Reparent everything in other to list.
  1744  	for s := other.first; s != nil; s = s.next {
  1745  		s.list = list
  1746  	}
  1747  
  1748  	// Concatenate the lists.
  1749  	if list.isEmpty() {
  1750  		*list = *other
  1751  	} else {
  1752  		// Neither list is empty. Put other before list.
  1753  		other.last.next = list.first
  1754  		list.first.prev = other.last
  1755  		list.first = other.first
  1756  	}
  1757  
  1758  	other.first, other.last = nil, nil
  1759  }
  1760  
  1761  const (
  1762  	_KindSpecialFinalizer = 1
  1763  	_KindSpecialProfile   = 2
  1764  	// _KindSpecialReachable is a special used for tracking
  1765  	// reachability during testing.
  1766  	_KindSpecialReachable = 3
  1767  	// Note: The finalizer special must be first because if we're freeing
  1768  	// an object, a finalizer special will cause the freeing operation
  1769  	// to abort, and we want to keep the other special records around
  1770  	// if that happens.
  1771  )
  1772  
  1773  type special struct {
  1774  	_      sys.NotInHeap
  1775  	next   *special // linked list in span
  1776  	offset uint16   // span offset of object
  1777  	kind   byte     // kind of special
  1778  }
  1779  
  1780  // spanHasSpecials marks a span as having specials in the arena bitmap.
  1781  func spanHasSpecials(s *mspan) {
  1782  	arenaPage := (s.base() / pageSize) % pagesPerArena
  1783  	ai := arenaIndex(s.base())
  1784  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  1785  	atomic.Or8(&ha.pageSpecials[arenaPage/8], uint8(1)<<(arenaPage%8))
  1786  }
  1787  
  1788  // spanHasNoSpecials marks a span as having no specials in the arena bitmap.
  1789  func spanHasNoSpecials(s *mspan) {
  1790  	arenaPage := (s.base() / pageSize) % pagesPerArena
  1791  	ai := arenaIndex(s.base())
  1792  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  1793  	atomic.And8(&ha.pageSpecials[arenaPage/8], ^(uint8(1) << (arenaPage % 8)))
  1794  }
  1795  
  1796  // Adds the special record s to the list of special records for
  1797  // the object p. All fields of s should be filled in except for
  1798  // offset & next, which this routine will fill in.
  1799  // Returns true if the special was successfully added, false otherwise.
  1800  // (The add will fail only if a record with the same p and s->kind
  1801  // already exists.)
  1802  func addspecial(p unsafe.Pointer, s *special) bool {
  1803  	span := spanOfHeap(uintptr(p))
  1804  	if span == nil {
  1805  		throw("addspecial on invalid pointer")
  1806  	}
  1807  
  1808  	// Ensure that the span is swept.
  1809  	// Sweeping accesses the specials list w/o locks, so we have
  1810  	// to synchronize with it. And it's just much safer.
  1811  	mp := acquirem()
  1812  	span.ensureSwept()
  1813  
  1814  	offset := uintptr(p) - span.base()
  1815  	kind := s.kind
  1816  
  1817  	lock(&span.speciallock)
  1818  
  1819  	// Find splice point, check for existing record.
  1820  	t := &span.specials
  1821  	for {
  1822  		x := *t
  1823  		if x == nil {
  1824  			break
  1825  		}
  1826  		if offset == uintptr(x.offset) && kind == x.kind {
  1827  			unlock(&span.speciallock)
  1828  			releasem(mp)
  1829  			return false // already exists
  1830  		}
  1831  		if offset < uintptr(x.offset) || (offset == uintptr(x.offset) && kind < x.kind) {
  1832  			break
  1833  		}
  1834  		t = &x.next
  1835  	}
  1836  
  1837  	// Splice in record, fill in offset.
  1838  	s.offset = uint16(offset)
  1839  	s.next = *t
  1840  	*t = s
  1841  	spanHasSpecials(span)
  1842  	unlock(&span.speciallock)
  1843  	releasem(mp)
  1844  
  1845  	return true
  1846  }
  1847  
  1848  // Removes the Special record of the given kind for the object p.
  1849  // Returns the record if the record existed, nil otherwise.
  1850  // The caller must FixAlloc_Free the result.
  1851  func removespecial(p unsafe.Pointer, kind uint8) *special {
  1852  	span := spanOfHeap(uintptr(p))
  1853  	if span == nil {
  1854  		throw("removespecial on invalid pointer")
  1855  	}
  1856  
  1857  	// Ensure that the span is swept.
  1858  	// Sweeping accesses the specials list w/o locks, so we have
  1859  	// to synchronize with it. And it's just much safer.
  1860  	mp := acquirem()
  1861  	span.ensureSwept()
  1862  
  1863  	offset := uintptr(p) - span.base()
  1864  
  1865  	var result *special
  1866  	lock(&span.speciallock)
  1867  	t := &span.specials
  1868  	for {
  1869  		s := *t
  1870  		if s == nil {
  1871  			break
  1872  		}
  1873  		// This function is used for finalizers only, so we don't check for
  1874  		// "interior" specials (p must be exactly equal to s->offset).
  1875  		if offset == uintptr(s.offset) && kind == s.kind {
  1876  			*t = s.next
  1877  			result = s
  1878  			break
  1879  		}
  1880  		t = &s.next
  1881  	}
  1882  	if span.specials == nil {
  1883  		spanHasNoSpecials(span)
  1884  	}
  1885  	unlock(&span.speciallock)
  1886  	releasem(mp)
  1887  	return result
  1888  }
  1889  
  1890  // The described object has a finalizer set for it.
  1891  //
  1892  // specialfinalizer is allocated from non-GC'd memory, so any heap
  1893  // pointers must be specially handled.
  1894  type specialfinalizer struct {
  1895  	_       sys.NotInHeap
  1896  	special special
  1897  	fn      *funcval // May be a heap pointer.
  1898  	nret    uintptr
  1899  	fint    *_type   // May be a heap pointer, but always live.
  1900  	ot      *ptrtype // May be a heap pointer, but always live.
  1901  }
  1902  
  1903  // Adds a finalizer to the object p. Returns true if it succeeded.
  1904  func addfinalizer(p unsafe.Pointer, f *funcval, nret uintptr, fint *_type, ot *ptrtype) bool {
  1905  	lock(&mheap_.speciallock)
  1906  	s := (*specialfinalizer)(mheap_.specialfinalizeralloc.alloc())
  1907  	unlock(&mheap_.speciallock)
  1908  	s.special.kind = _KindSpecialFinalizer
  1909  	s.fn = f
  1910  	s.nret = nret
  1911  	s.fint = fint
  1912  	s.ot = ot
  1913  	if addspecial(p, &s.special) {
  1914  		// This is responsible for maintaining the same
  1915  		// GC-related invariants as markrootSpans in any
  1916  		// situation where it's possible that markrootSpans
  1917  		// has already run but mark termination hasn't yet.
  1918  		if gcphase != _GCoff {
  1919  			base, span, _ := findObject(uintptr(p), 0, 0)
  1920  			mp := acquirem()
  1921  			gcw := &mp.p.ptr().gcw
  1922  			// Mark everything reachable from the object
  1923  			// so it's retained for the finalizer.
  1924  			if !span.spanclass.noscan() {
  1925  				scanobject(base, gcw)
  1926  			}
  1927  			// Mark the finalizer itself, since the
  1928  			// special isn't part of the GC'd heap.
  1929  			scanblock(uintptr(unsafe.Pointer(&s.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  1930  			releasem(mp)
  1931  		}
  1932  		return true
  1933  	}
  1934  
  1935  	// There was an old finalizer
  1936  	lock(&mheap_.speciallock)
  1937  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  1938  	unlock(&mheap_.speciallock)
  1939  	return false
  1940  }
  1941  
  1942  // Removes the finalizer (if any) from the object p.
  1943  func removefinalizer(p unsafe.Pointer) {
  1944  	s := (*specialfinalizer)(unsafe.Pointer(removespecial(p, _KindSpecialFinalizer)))
  1945  	if s == nil {
  1946  		return // there wasn't a finalizer to remove
  1947  	}
  1948  	lock(&mheap_.speciallock)
  1949  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  1950  	unlock(&mheap_.speciallock)
  1951  }
  1952  
  1953  // The described object is being heap profiled.
  1954  type specialprofile struct {
  1955  	_       sys.NotInHeap
  1956  	special special
  1957  	b       *bucket
  1958  }
  1959  
  1960  // Set the heap profile bucket associated with addr to b.
  1961  func setprofilebucket(p unsafe.Pointer, b *bucket) {
  1962  	lock(&mheap_.speciallock)
  1963  	s := (*specialprofile)(mheap_.specialprofilealloc.alloc())
  1964  	unlock(&mheap_.speciallock)
  1965  	s.special.kind = _KindSpecialProfile
  1966  	s.b = b
  1967  	if !addspecial(p, &s.special) {
  1968  		throw("setprofilebucket: profile already set")
  1969  	}
  1970  }
  1971  
  1972  // specialReachable tracks whether an object is reachable on the next
  1973  // GC cycle. This is used by testing.
  1974  type specialReachable struct {
  1975  	special   special
  1976  	done      bool
  1977  	reachable bool
  1978  }
  1979  
  1980  // specialsIter helps iterate over specials lists.
  1981  type specialsIter struct {
  1982  	pprev **special
  1983  	s     *special
  1984  }
  1985  
  1986  func newSpecialsIter(span *mspan) specialsIter {
  1987  	return specialsIter{&span.specials, span.specials}
  1988  }
  1989  
  1990  func (i *specialsIter) valid() bool {
  1991  	return i.s != nil
  1992  }
  1993  
  1994  func (i *specialsIter) next() {
  1995  	i.pprev = &i.s.next
  1996  	i.s = *i.pprev
  1997  }
  1998  
  1999  // unlinkAndNext removes the current special from the list and moves
  2000  // the iterator to the next special. It returns the unlinked special.
  2001  func (i *specialsIter) unlinkAndNext() *special {
  2002  	cur := i.s
  2003  	i.s = cur.next
  2004  	*i.pprev = i.s
  2005  	return cur
  2006  }
  2007  
  2008  // freeSpecial performs any cleanup on special s and deallocates it.
  2009  // s must already be unlinked from the specials list.
  2010  func freeSpecial(s *special, p unsafe.Pointer, size uintptr) {
  2011  	switch s.kind {
  2012  	case _KindSpecialFinalizer:
  2013  		sf := (*specialfinalizer)(unsafe.Pointer(s))
  2014  		queuefinalizer(p, sf.fn, sf.nret, sf.fint, sf.ot)
  2015  		lock(&mheap_.speciallock)
  2016  		mheap_.specialfinalizeralloc.free(unsafe.Pointer(sf))
  2017  		unlock(&mheap_.speciallock)
  2018  	case _KindSpecialProfile:
  2019  		sp := (*specialprofile)(unsafe.Pointer(s))
  2020  		mProf_Free(sp.b, size)
  2021  		lock(&mheap_.speciallock)
  2022  		mheap_.specialprofilealloc.free(unsafe.Pointer(sp))
  2023  		unlock(&mheap_.speciallock)
  2024  	case _KindSpecialReachable:
  2025  		sp := (*specialReachable)(unsafe.Pointer(s))
  2026  		sp.done = true
  2027  		// The creator frees these.
  2028  	default:
  2029  		throw("bad special kind")
  2030  		panic("not reached")
  2031  	}
  2032  }
  2033  
  2034  // gcBits is an alloc/mark bitmap. This is always used as gcBits.x.
  2035  type gcBits struct {
  2036  	_ sys.NotInHeap
  2037  	x uint8
  2038  }
  2039  
  2040  // bytep returns a pointer to the n'th byte of b.
  2041  func (b *gcBits) bytep(n uintptr) *uint8 {
  2042  	return addb(&b.x, n)
  2043  }
  2044  
  2045  // bitp returns a pointer to the byte containing bit n and a mask for
  2046  // selecting that bit from *bytep.
  2047  func (b *gcBits) bitp(n uintptr) (bytep *uint8, mask uint8) {
  2048  	return b.bytep(n / 8), 1 << (n % 8)
  2049  }
  2050  
  2051  const gcBitsChunkBytes = uintptr(64 << 10)
  2052  const gcBitsHeaderBytes = unsafe.Sizeof(gcBitsHeader{})
  2053  
  2054  type gcBitsHeader struct {
  2055  	free uintptr // free is the index into bits of the next free byte.
  2056  	next uintptr // *gcBits triggers recursive type bug. (issue 14620)
  2057  }
  2058  
  2059  type gcBitsArena struct {
  2060  	_ sys.NotInHeap
  2061  	// gcBitsHeader // side step recursive type bug (issue 14620) by including fields by hand.
  2062  	free uintptr // free is the index into bits of the next free byte; read/write atomically
  2063  	next *gcBitsArena
  2064  	bits [gcBitsChunkBytes - gcBitsHeaderBytes]gcBits
  2065  }
  2066  
  2067  var gcBitsArenas struct {
  2068  	lock     mutex
  2069  	free     *gcBitsArena
  2070  	next     *gcBitsArena // Read atomically. Write atomically under lock.
  2071  	current  *gcBitsArena
  2072  	previous *gcBitsArena
  2073  }
  2074  
  2075  // tryAlloc allocates from b or returns nil if b does not have enough room.
  2076  // This is safe to call concurrently.
  2077  func (b *gcBitsArena) tryAlloc(bytes uintptr) *gcBits {
  2078  	if b == nil || atomic.Loaduintptr(&b.free)+bytes > uintptr(len(b.bits)) {
  2079  		return nil
  2080  	}
  2081  	// Try to allocate from this block.
  2082  	end := atomic.Xadduintptr(&b.free, bytes)
  2083  	if end > uintptr(len(b.bits)) {
  2084  		return nil
  2085  	}
  2086  	// There was enough room.
  2087  	start := end - bytes
  2088  	return &b.bits[start]
  2089  }
  2090  
  2091  // newMarkBits returns a pointer to 8 byte aligned bytes
  2092  // to be used for a span's mark bits.
  2093  func newMarkBits(nelems uintptr) *gcBits {
  2094  	blocksNeeded := uintptr((nelems + 63) / 64)
  2095  	bytesNeeded := blocksNeeded * 8
  2096  
  2097  	// Try directly allocating from the current head arena.
  2098  	head := (*gcBitsArena)(atomic.Loadp(unsafe.Pointer(&gcBitsArenas.next)))
  2099  	if p := head.tryAlloc(bytesNeeded); p != nil {
  2100  		return p
  2101  	}
  2102  
  2103  	// There's not enough room in the head arena. We may need to
  2104  	// allocate a new arena.
  2105  	lock(&gcBitsArenas.lock)
  2106  	// Try the head arena again, since it may have changed. Now
  2107  	// that we hold the lock, the list head can't change, but its
  2108  	// free position still can.
  2109  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2110  		unlock(&gcBitsArenas.lock)
  2111  		return p
  2112  	}
  2113  
  2114  	// Allocate a new arena. This may temporarily drop the lock.
  2115  	fresh := newArenaMayUnlock()
  2116  	// If newArenaMayUnlock dropped the lock, another thread may
  2117  	// have put a fresh arena on the "next" list. Try allocating
  2118  	// from next again.
  2119  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2120  		// Put fresh back on the free list.
  2121  		// TODO: Mark it "already zeroed"
  2122  		fresh.next = gcBitsArenas.free
  2123  		gcBitsArenas.free = fresh
  2124  		unlock(&gcBitsArenas.lock)
  2125  		return p
  2126  	}
  2127  
  2128  	// Allocate from the fresh arena. We haven't linked it in yet, so
  2129  	// this cannot race and is guaranteed to succeed.
  2130  	p := fresh.tryAlloc(bytesNeeded)
  2131  	if p == nil {
  2132  		throw("markBits overflow")
  2133  	}
  2134  
  2135  	// Add the fresh arena to the "next" list.
  2136  	fresh.next = gcBitsArenas.next
  2137  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), unsafe.Pointer(fresh))
  2138  
  2139  	unlock(&gcBitsArenas.lock)
  2140  	return p
  2141  }
  2142  
  2143  // newAllocBits returns a pointer to 8 byte aligned bytes
  2144  // to be used for this span's alloc bits.
  2145  // newAllocBits is used to provide newly initialized spans
  2146  // allocation bits. For spans not being initialized the
  2147  // mark bits are repurposed as allocation bits when
  2148  // the span is swept.
  2149  func newAllocBits(nelems uintptr) *gcBits {
  2150  	return newMarkBits(nelems)
  2151  }
  2152  
  2153  // nextMarkBitArenaEpoch establishes a new epoch for the arenas
  2154  // holding the mark bits. The arenas are named relative to the
  2155  // current GC cycle which is demarcated by the call to finishweep_m.
  2156  //
  2157  // All current spans have been swept.
  2158  // During that sweep each span allocated room for its gcmarkBits in
  2159  // gcBitsArenas.next block. gcBitsArenas.next becomes the gcBitsArenas.current
  2160  // where the GC will mark objects and after each span is swept these bits
  2161  // will be used to allocate objects.
  2162  // gcBitsArenas.current becomes gcBitsArenas.previous where the span's
  2163  // gcAllocBits live until all the spans have been swept during this GC cycle.
  2164  // The span's sweep extinguishes all the references to gcBitsArenas.previous
  2165  // by pointing gcAllocBits into the gcBitsArenas.current.
  2166  // The gcBitsArenas.previous is released to the gcBitsArenas.free list.
  2167  func nextMarkBitArenaEpoch() {
  2168  	lock(&gcBitsArenas.lock)
  2169  	if gcBitsArenas.previous != nil {
  2170  		if gcBitsArenas.free == nil {
  2171  			gcBitsArenas.free = gcBitsArenas.previous
  2172  		} else {
  2173  			// Find end of previous arenas.
  2174  			last := gcBitsArenas.previous
  2175  			for last = gcBitsArenas.previous; last.next != nil; last = last.next {
  2176  			}
  2177  			last.next = gcBitsArenas.free
  2178  			gcBitsArenas.free = gcBitsArenas.previous
  2179  		}
  2180  	}
  2181  	gcBitsArenas.previous = gcBitsArenas.current
  2182  	gcBitsArenas.current = gcBitsArenas.next
  2183  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), nil) // newMarkBits calls newArena when needed
  2184  	unlock(&gcBitsArenas.lock)
  2185  }
  2186  
  2187  // newArenaMayUnlock allocates and zeroes a gcBits arena.
  2188  // The caller must hold gcBitsArena.lock. This may temporarily release it.
  2189  func newArenaMayUnlock() *gcBitsArena {
  2190  	var result *gcBitsArena
  2191  	if gcBitsArenas.free == nil {
  2192  		unlock(&gcBitsArenas.lock)
  2193  		result = (*gcBitsArena)(sysAlloc(gcBitsChunkBytes, &memstats.gcMiscSys))
  2194  		if result == nil {
  2195  			throw("runtime: cannot allocate memory")
  2196  		}
  2197  		lock(&gcBitsArenas.lock)
  2198  	} else {
  2199  		result = gcBitsArenas.free
  2200  		gcBitsArenas.free = gcBitsArenas.free.next
  2201  		memclrNoHeapPointers(unsafe.Pointer(result), gcBitsChunkBytes)
  2202  	}
  2203  	result.next = nil
  2204  	// If result.bits is not 8 byte aligned adjust index so
  2205  	// that &result.bits[result.free] is 8 byte aligned.
  2206  	if uintptr(unsafe.Offsetof(gcBitsArena{}.bits))&7 == 0 {
  2207  		result.free = 0
  2208  	} else {
  2209  		result.free = 8 - (uintptr(unsafe.Pointer(&result.bits[0])) & 7)
  2210  	}
  2211  	return result
  2212  }