Source file src/runtime/mheap.go

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Page heap.
     6  //
     7  // See malloc.go for overview.
     8  
     9  package runtime
    10  
    11  import (
    12  	"internal/cpu"
    13  	"internal/goarch"
    14  	"runtime/internal/atomic"
    15  	"unsafe"
    16  )
    17  
    18  const (
    19  	// minPhysPageSize is a lower-bound on the physical page size. The
    20  	// true physical page size may be larger than this. In contrast,
    21  	// sys.PhysPageSize is an upper-bound on the physical page size.
    22  	minPhysPageSize = 4096
    23  
    24  	// maxPhysPageSize is the maximum page size the runtime supports.
    25  	maxPhysPageSize = 512 << 10
    26  
    27  	// maxPhysHugePageSize sets an upper-bound on the maximum huge page size
    28  	// that the runtime supports.
    29  	maxPhysHugePageSize = pallocChunkBytes
    30  
    31  	// pagesPerReclaimerChunk indicates how many pages to scan from the
    32  	// pageInUse bitmap at a time. Used by the page reclaimer.
    33  	//
    34  	// Higher values reduce contention on scanning indexes (such as
    35  	// h.reclaimIndex), but increase the minimum latency of the
    36  	// operation.
    37  	//
    38  	// The time required to scan this many pages can vary a lot depending
    39  	// on how many spans are actually freed. Experimentally, it can
    40  	// scan for pages at ~300 GB/ms on a 2.6GHz Core i7, but can only
    41  	// free spans at ~32 MB/ms. Using 512 pages bounds this at
    42  	// roughly 100µs.
    43  	//
    44  	// Must be a multiple of the pageInUse bitmap element size and
    45  	// must also evenly divide pagesPerArena.
    46  	pagesPerReclaimerChunk = 512
    47  
    48  	// physPageAlignedStacks indicates whether stack allocations must be
    49  	// physical page aligned. This is a requirement for MAP_STACK on
    50  	// OpenBSD.
    51  	physPageAlignedStacks = GOOS == "openbsd"
    52  )
    53  
    54  // Main malloc heap.
    55  // The heap itself is the "free" and "scav" treaps,
    56  // but all the other global data is here too.
    57  //
    58  // mheap must not be heap-allocated because it contains mSpanLists,
    59  // which must not be heap-allocated.
    60  //
    61  //go:notinheap
    62  type mheap struct {
    63  	// lock must only be acquired on the system stack, otherwise a g
    64  	// could self-deadlock if its stack grows with the lock held.
    65  	lock  mutex
    66  	pages pageAlloc // page allocation data structure
    67  
    68  	sweepgen uint32 // sweep generation, see comment in mspan; written during STW
    69  
    70  	// allspans is a slice of all mspans ever created. Each mspan
    71  	// appears exactly once.
    72  	//
    73  	// The memory for allspans is manually managed and can be
    74  	// reallocated and move as the heap grows.
    75  	//
    76  	// In general, allspans is protected by mheap_.lock, which
    77  	// prevents concurrent access as well as freeing the backing
    78  	// store. Accesses during STW might not hold the lock, but
    79  	// must ensure that allocation cannot happen around the
    80  	// access (since that may free the backing store).
    81  	allspans []*mspan // all spans out there
    82  
    83  	// _ uint32 // align uint64 fields on 32-bit for atomics
    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.Uint64 // 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  	// TODO(austin): pagesInUse should be a uintptr, but the 386
   109  	// compiler can't 8-byte align fields.
   110  
   111  	// scavengeGoal is the amount of total retained heap memory (measured by
   112  	// heapRetained) that the runtime will try to maintain by returning memory
   113  	// to the OS.
   114  	//
   115  	// Accessed atomically.
   116  	scavengeGoal uint64
   117  
   118  	// Page reclaimer state
   119  
   120  	// reclaimIndex is the page index in allArenas of next page to
   121  	// reclaim. Specifically, it refers to page (i %
   122  	// pagesPerArena) of arena allArenas[i / pagesPerArena].
   123  	//
   124  	// If this is >= 1<<63, the page reclaimer is done scanning
   125  	// the page marks.
   126  	reclaimIndex atomic.Uint64
   127  
   128  	// reclaimCredit is spare credit for extra pages swept. Since
   129  	// the page reclaimer works in large chunks, it may reclaim
   130  	// more than requested. Any spare pages released go to this
   131  	// credit pool.
   132  	reclaimCredit atomic.Uintptr
   133  
   134  	// arenas is the heap arena map. It points to the metadata for
   135  	// the heap for every arena frame of the entire usable virtual
   136  	// address space.
   137  	//
   138  	// Use arenaIndex to compute indexes into this array.
   139  	//
   140  	// For regions of the address space that are not backed by the
   141  	// Go heap, the arena map contains nil.
   142  	//
   143  	// Modifications are protected by mheap_.lock. Reads can be
   144  	// performed without locking; however, a given entry can
   145  	// transition from nil to non-nil at any time when the lock
   146  	// isn't held. (Entries never transitions back to nil.)
   147  	//
   148  	// In general, this is a two-level mapping consisting of an L1
   149  	// map and possibly many L2 maps. This saves space when there
   150  	// are a huge number of arena frames. However, on many
   151  	// platforms (even 64-bit), arenaL1Bits is 0, making this
   152  	// effectively a single-level map. In this case, arenas[0]
   153  	// will never be nil.
   154  	arenas [1 << arenaL1Bits]*[1 << arenaL2Bits]*heapArena
   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  	// allArenas is the arenaIndex of every mapped arena. This can
   172  	// be used to iterate through the 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  	allArenas []arenaIdx
   179  
   180  	// sweepArenas is a snapshot of allArenas taken at the
   181  	// beginning of the sweep cycle. This can be read safely by
   182  	// simply blocking GC (by disabling preemption).
   183  	sweepArenas []arenaIdx
   184  
   185  	// markArenas is a snapshot of allArenas taken at the beginning
   186  	// of the mark cycle. Because allArenas is append-only, neither
   187  	// this slice nor its contents will change during the mark, so
   188  	// it can be read safely.
   189  	markArenas []arenaIdx
   190  
   191  	// curArena is the arena that the heap is currently growing
   192  	// into. This should always be physPageSize-aligned.
   193  	curArena struct {
   194  		base, end uintptr
   195  	}
   196  
   197  	_ uint32 // ensure 64-bit alignment of central
   198  
   199  	// central free lists for small size classes.
   200  	// the padding makes sure that the mcentrals are
   201  	// spaced CacheLinePadSize bytes apart, so that each mcentral.lock
   202  	// gets its own cache line.
   203  	// central is indexed by spanClass.
   204  	central [numSpanClasses]struct {
   205  		mcentral mcentral
   206  		pad      [cpu.CacheLinePadSize - unsafe.Sizeof(mcentral{})%cpu.CacheLinePadSize]byte
   207  	}
   208  
   209  	spanalloc             fixalloc // allocator for span*
   210  	cachealloc            fixalloc // allocator for mcache*
   211  	specialfinalizeralloc fixalloc // allocator for specialfinalizer*
   212  	specialprofilealloc   fixalloc // allocator for specialprofile*
   213  	specialReachableAlloc fixalloc // allocator for specialReachable
   214  	speciallock           mutex    // lock for special record allocators.
   215  	arenaHintAlloc        fixalloc // allocator for arenaHints
   216  
   217  	unused *specialfinalizer // never set, just here to force the specialfinalizer type into DWARF
   218  }
   219  
   220  var mheap_ mheap
   221  
   222  // A heapArena stores metadata for a heap arena. heapArenas are stored
   223  // outside of the Go heap and accessed via the mheap_.arenas index.
   224  //
   225  //go:notinheap
   226  type heapArena struct {
   227  	// bitmap stores the pointer/scalar bitmap for the words in
   228  	// this arena. See mbitmap.go for a description. Use the
   229  	// heapBits type to access this.
   230  	bitmap [heapArenaBitmapBytes]byte
   231  
   232  	// spans maps from virtual address page ID within this arena to *mspan.
   233  	// For allocated spans, their pages map to the span itself.
   234  	// For free spans, only the lowest and highest pages map to the span itself.
   235  	// Internal pages map to an arbitrary span.
   236  	// For pages that have never been allocated, spans entries are nil.
   237  	//
   238  	// Modifications are protected by mheap.lock. Reads can be
   239  	// performed without locking, but ONLY from indexes that are
   240  	// known to contain in-use or stack spans. This means there
   241  	// must not be a safe-point between establishing that an
   242  	// address is live and looking it up in the spans array.
   243  	spans [pagesPerArena]*mspan
   244  
   245  	// pageInUse is a bitmap that indicates which spans are in
   246  	// state mSpanInUse. This bitmap is indexed by page number,
   247  	// but only the bit corresponding to the first page in each
   248  	// span is used.
   249  	//
   250  	// Reads and writes are atomic.
   251  	pageInUse [pagesPerArena / 8]uint8
   252  
   253  	// pageMarks is a bitmap that indicates which spans have any
   254  	// marked objects on them. Like pageInUse, only the bit
   255  	// corresponding to the first page in each span is used.
   256  	//
   257  	// Writes are done atomically during marking. Reads are
   258  	// non-atomic and lock-free since they only occur during
   259  	// sweeping (and hence never race with writes).
   260  	//
   261  	// This is used to quickly find whole spans that can be freed.
   262  	//
   263  	// TODO(austin): It would be nice if this was uint64 for
   264  	// faster scanning, but we don't have 64-bit atomic bit
   265  	// operations.
   266  	pageMarks [pagesPerArena / 8]uint8
   267  
   268  	// pageSpecials is a bitmap that indicates which spans have
   269  	// specials (finalizers or other). Like pageInUse, only the bit
   270  	// corresponding to the first page in each span is used.
   271  	//
   272  	// Writes are done atomically whenever a special is added to
   273  	// a span and whenever the last special is removed from a span.
   274  	// Reads are done atomically to find spans containing specials
   275  	// during marking.
   276  	pageSpecials [pagesPerArena / 8]uint8
   277  
   278  	// checkmarks stores the debug.gccheckmark state. It is only
   279  	// used if debug.gccheckmark > 0.
   280  	checkmarks *checkmarksMap
   281  
   282  	// zeroedBase marks the first byte of the first page in this
   283  	// arena which hasn't been used yet and is therefore already
   284  	// zero. zeroedBase is relative to the arena base.
   285  	// Increases monotonically until it hits heapArenaBytes.
   286  	//
   287  	// This field is sufficient to determine if an allocation
   288  	// needs to be zeroed because the page allocator follows an
   289  	// address-ordered first-fit policy.
   290  	//
   291  	// Read atomically and written with an atomic CAS.
   292  	zeroedBase uintptr
   293  }
   294  
   295  // arenaHint is a hint for where to grow the heap arenas. See
   296  // mheap_.arenaHints.
   297  //
   298  //go:notinheap
   299  type arenaHint struct {
   300  	addr uintptr
   301  	down bool
   302  	next *arenaHint
   303  }
   304  
   305  // An mspan is a run of pages.
   306  //
   307  // When a mspan is in the heap free treap, state == mSpanFree
   308  // and heapmap(s->start) == span, heapmap(s->start+s->npages-1) == span.
   309  // If the mspan is in the heap scav treap, then in addition to the
   310  // above scavenged == true. scavenged == false in all other cases.
   311  //
   312  // When a mspan is allocated, state == mSpanInUse or mSpanManual
   313  // and heapmap(i) == span for all s->start <= i < s->start+s->npages.
   314  
   315  // Every mspan is in one doubly-linked list, either in the mheap's
   316  // busy list or one of the mcentral's span lists.
   317  
   318  // An mspan representing actual memory has state mSpanInUse,
   319  // mSpanManual, or mSpanFree. Transitions between these states are
   320  // constrained as follows:
   321  //
   322  // * A span may transition from free to in-use or manual during any GC
   323  //   phase.
   324  //
   325  // * During sweeping (gcphase == _GCoff), a span may transition from
   326  //   in-use to free (as a result of sweeping) or manual to free (as a
   327  //   result of stacks being freed).
   328  //
   329  // * During GC (gcphase != _GCoff), a span *must not* transition from
   330  //   manual or in-use to free. Because concurrent GC may read a pointer
   331  //   and then look up its span, the span state must be monotonic.
   332  //
   333  // Setting mspan.state to mSpanInUse or mSpanManual must be done
   334  // atomically and only after all other span fields are valid.
   335  // Likewise, if inspecting a span is contingent on it being
   336  // mSpanInUse, the state should be loaded atomically and checked
   337  // before depending on other fields. This allows the garbage collector
   338  // to safely deal with potentially invalid pointers, since resolving
   339  // such pointers may race with a span being allocated.
   340  type mSpanState uint8
   341  
   342  const (
   343  	mSpanDead   mSpanState = iota
   344  	mSpanInUse             // allocated for garbage collected heap
   345  	mSpanManual            // allocated for manual management (e.g., stack allocator)
   346  )
   347  
   348  // mSpanStateNames are the names of the span states, indexed by
   349  // mSpanState.
   350  var mSpanStateNames = []string{
   351  	"mSpanDead",
   352  	"mSpanInUse",
   353  	"mSpanManual",
   354  	"mSpanFree",
   355  }
   356  
   357  // mSpanStateBox holds an mSpanState and provides atomic operations on
   358  // it. This is a separate type to disallow accidental comparison or
   359  // assignment with mSpanState.
   360  type mSpanStateBox struct {
   361  	s mSpanState
   362  }
   363  
   364  func (b *mSpanStateBox) set(s mSpanState) {
   365  	atomic.Store8((*uint8)(&b.s), uint8(s))
   366  }
   367  
   368  func (b *mSpanStateBox) get() mSpanState {
   369  	return mSpanState(atomic.Load8((*uint8)(&b.s)))
   370  }
   371  
   372  // mSpanList heads a linked list of spans.
   373  //
   374  //go:notinheap
   375  type mSpanList struct {
   376  	first *mspan // first span in list, or nil if none
   377  	last  *mspan // last span in list, or nil if none
   378  }
   379  
   380  //go:notinheap
   381  type mspan struct {
   382  	next *mspan     // next span in list, or nil if none
   383  	prev *mspan     // previous span in list, or nil if none
   384  	list *mSpanList // For debugging. TODO: Remove.
   385  
   386  	startAddr uintptr // address of first byte of span aka s.base()
   387  	npages    uintptr // number of pages in span
   388  
   389  	manualFreeList gclinkptr // list of free objects in mSpanManual spans
   390  
   391  	// freeindex is the slot index between 0 and nelems at which to begin scanning
   392  	// for the next free object in this span.
   393  	// Each allocation scans allocBits starting at freeindex until it encounters a 0
   394  	// indicating a free object. freeindex is then adjusted so that subsequent scans begin
   395  	// just past the newly discovered free object.
   396  	//
   397  	// If freeindex == nelem, this span has no free objects.
   398  	//
   399  	// allocBits is a bitmap of objects in this span.
   400  	// If n >= freeindex and allocBits[n/8] & (1<<(n%8)) is 0
   401  	// then object n is free;
   402  	// otherwise, object n is allocated. Bits starting at nelem are
   403  	// undefined and should never be referenced.
   404  	//
   405  	// Object n starts at address n*elemsize + (start << pageShift).
   406  	freeindex uintptr
   407  	// TODO: Look up nelems from sizeclass and remove this field if it
   408  	// helps performance.
   409  	nelems uintptr // number of object in the span.
   410  
   411  	// Cache of the allocBits at freeindex. allocCache is shifted
   412  	// such that the lowest bit corresponds to the bit freeindex.
   413  	// allocCache holds the complement of allocBits, thus allowing
   414  	// ctz (count trailing zero) to use it directly.
   415  	// allocCache may contain bits beyond s.nelems; the caller must ignore
   416  	// these.
   417  	allocCache uint64
   418  
   419  	// allocBits and gcmarkBits hold pointers to a span's mark and
   420  	// allocation bits. The pointers are 8 byte aligned.
   421  	// There are three arenas where this data is held.
   422  	// free: Dirty arenas that are no longer accessed
   423  	//       and can be reused.
   424  	// next: Holds information to be used in the next GC cycle.
   425  	// current: Information being used during this GC cycle.
   426  	// previous: Information being used during the last GC cycle.
   427  	// A new GC cycle starts with the call to finishsweep_m.
   428  	// finishsweep_m moves the previous arena to the free arena,
   429  	// the current arena to the previous arena, and
   430  	// the next arena to the current arena.
   431  	// The next arena is populated as the spans request
   432  	// memory to hold gcmarkBits for the next GC cycle as well
   433  	// as allocBits for newly allocated spans.
   434  	//
   435  	// The pointer arithmetic is done "by hand" instead of using
   436  	// arrays to avoid bounds checks along critical performance
   437  	// paths.
   438  	// The sweep will free the old allocBits and set allocBits to the
   439  	// gcmarkBits. The gcmarkBits are replaced with a fresh zeroed
   440  	// out memory.
   441  	allocBits  *gcBits
   442  	gcmarkBits *gcBits
   443  
   444  	// sweep generation:
   445  	// if sweepgen == h->sweepgen - 2, the span needs sweeping
   446  	// if sweepgen == h->sweepgen - 1, the span is currently being swept
   447  	// if sweepgen == h->sweepgen, the span is swept and ready to use
   448  	// if sweepgen == h->sweepgen + 1, the span was cached before sweep began and is still cached, and needs sweeping
   449  	// if sweepgen == h->sweepgen + 3, the span was swept and then cached and is still cached
   450  	// h->sweepgen is incremented by 2 after every GC
   451  
   452  	sweepgen    uint32
   453  	divMul      uint32        // for divide by elemsize
   454  	allocCount  uint16        // number of allocated objects
   455  	spanclass   spanClass     // size class and noscan (uint8)
   456  	state       mSpanStateBox // mSpanInUse etc; accessed atomically (get/set methods)
   457  	needzero    uint8         // needs to be zeroed before allocation
   458  	elemsize    uintptr       // computed from sizeclass or from npages
   459  	limit       uintptr       // end of data in span
   460  	speciallock mutex         // guards specials list
   461  	specials    *special      // linked list of special records sorted by offset.
   462  }
   463  
   464  func (s *mspan) base() uintptr {
   465  	return s.startAddr
   466  }
   467  
   468  func (s *mspan) layout() (size, n, total uintptr) {
   469  	total = s.npages << _PageShift
   470  	size = s.elemsize
   471  	if size > 0 {
   472  		n = total / size
   473  	}
   474  	return
   475  }
   476  
   477  // recordspan adds a newly allocated span to h.allspans.
   478  //
   479  // This only happens the first time a span is allocated from
   480  // mheap.spanalloc (it is not called when a span is reused).
   481  //
   482  // Write barriers are disallowed here because it can be called from
   483  // gcWork when allocating new workbufs. However, because it's an
   484  // indirect call from the fixalloc initializer, the compiler can't see
   485  // this.
   486  //
   487  // The heap lock must be held.
   488  //
   489  //go:nowritebarrierrec
   490  func recordspan(vh unsafe.Pointer, p unsafe.Pointer) {
   491  	h := (*mheap)(vh)
   492  	s := (*mspan)(p)
   493  
   494  	assertLockHeld(&h.lock)
   495  
   496  	if len(h.allspans) >= cap(h.allspans) {
   497  		n := 64 * 1024 / goarch.PtrSize
   498  		if n < cap(h.allspans)*3/2 {
   499  			n = cap(h.allspans) * 3 / 2
   500  		}
   501  		var new []*mspan
   502  		sp := (*slice)(unsafe.Pointer(&new))
   503  		sp.array = sysAlloc(uintptr(n)*goarch.PtrSize, &memstats.other_sys)
   504  		if sp.array == nil {
   505  			throw("runtime: cannot allocate memory")
   506  		}
   507  		sp.len = len(h.allspans)
   508  		sp.cap = n
   509  		if len(h.allspans) > 0 {
   510  			copy(new, h.allspans)
   511  		}
   512  		oldAllspans := h.allspans
   513  		*(*notInHeapSlice)(unsafe.Pointer(&h.allspans)) = *(*notInHeapSlice)(unsafe.Pointer(&new))
   514  		if len(oldAllspans) != 0 {
   515  			sysFree(unsafe.Pointer(&oldAllspans[0]), uintptr(cap(oldAllspans))*unsafe.Sizeof(oldAllspans[0]), &memstats.other_sys)
   516  		}
   517  	}
   518  	h.allspans = h.allspans[:len(h.allspans)+1]
   519  	h.allspans[len(h.allspans)-1] = s
   520  }
   521  
   522  // A spanClass represents the size class and noscan-ness of a span.
   523  //
   524  // Each size class has a noscan spanClass and a scan spanClass. The
   525  // noscan spanClass contains only noscan objects, which do not contain
   526  // pointers and thus do not need to be scanned by the garbage
   527  // collector.
   528  type spanClass uint8
   529  
   530  const (
   531  	numSpanClasses = _NumSizeClasses << 1
   532  	tinySpanClass  = spanClass(tinySizeClass<<1 | 1)
   533  )
   534  
   535  func makeSpanClass(sizeclass uint8, noscan bool) spanClass {
   536  	return spanClass(sizeclass<<1) | spanClass(bool2int(noscan))
   537  }
   538  
   539  func (sc spanClass) sizeclass() int8 {
   540  	return int8(sc >> 1)
   541  }
   542  
   543  func (sc spanClass) noscan() bool {
   544  	return sc&1 != 0
   545  }
   546  
   547  // arenaIndex returns the index into mheap_.arenas of the arena
   548  // containing metadata for p. This index combines of an index into the
   549  // L1 map and an index into the L2 map and should be used as
   550  // mheap_.arenas[ai.l1()][ai.l2()].
   551  //
   552  // If p is outside the range of valid heap addresses, either l1() or
   553  // l2() will be out of bounds.
   554  //
   555  // It is nosplit because it's called by spanOf and several other
   556  // nosplit functions.
   557  //
   558  //go:nosplit
   559  func arenaIndex(p uintptr) arenaIdx {
   560  	return arenaIdx((p - arenaBaseOffset) / heapArenaBytes)
   561  }
   562  
   563  // arenaBase returns the low address of the region covered by heap
   564  // arena i.
   565  func arenaBase(i arenaIdx) uintptr {
   566  	return uintptr(i)*heapArenaBytes + arenaBaseOffset
   567  }
   568  
   569  type arenaIdx uint
   570  
   571  func (i arenaIdx) l1() uint {
   572  	if arenaL1Bits == 0 {
   573  		// Let the compiler optimize this away if there's no
   574  		// L1 map.
   575  		return 0
   576  	} else {
   577  		return uint(i) >> arenaL1Shift
   578  	}
   579  }
   580  
   581  func (i arenaIdx) l2() uint {
   582  	if arenaL1Bits == 0 {
   583  		return uint(i)
   584  	} else {
   585  		return uint(i) & (1<<arenaL2Bits - 1)
   586  	}
   587  }
   588  
   589  // inheap reports whether b is a pointer into a (potentially dead) heap object.
   590  // It returns false for pointers into mSpanManual spans.
   591  // Non-preemptible because it is used by write barriers.
   592  //go:nowritebarrier
   593  //go:nosplit
   594  func inheap(b uintptr) bool {
   595  	return spanOfHeap(b) != nil
   596  }
   597  
   598  // inHeapOrStack is a variant of inheap that returns true for pointers
   599  // into any allocated heap span.
   600  //
   601  //go:nowritebarrier
   602  //go:nosplit
   603  func inHeapOrStack(b uintptr) bool {
   604  	s := spanOf(b)
   605  	if s == nil || b < s.base() {
   606  		return false
   607  	}
   608  	switch s.state.get() {
   609  	case mSpanInUse, mSpanManual:
   610  		return b < s.limit
   611  	default:
   612  		return false
   613  	}
   614  }
   615  
   616  // spanOf returns the span of p. If p does not point into the heap
   617  // arena or no span has ever contained p, spanOf returns nil.
   618  //
   619  // If p does not point to allocated memory, this may return a non-nil
   620  // span that does *not* contain p. If this is a possibility, the
   621  // caller should either call spanOfHeap or check the span bounds
   622  // explicitly.
   623  //
   624  // Must be nosplit because it has callers that are nosplit.
   625  //
   626  //go:nosplit
   627  func spanOf(p uintptr) *mspan {
   628  	// This function looks big, but we use a lot of constant
   629  	// folding around arenaL1Bits to get it under the inlining
   630  	// budget. Also, many of the checks here are safety checks
   631  	// that Go needs to do anyway, so the generated code is quite
   632  	// short.
   633  	ri := arenaIndex(p)
   634  	if arenaL1Bits == 0 {
   635  		// If there's no L1, then ri.l1() can't be out of bounds but ri.l2() can.
   636  		if ri.l2() >= uint(len(mheap_.arenas[0])) {
   637  			return nil
   638  		}
   639  	} else {
   640  		// If there's an L1, then ri.l1() can be out of bounds but ri.l2() can't.
   641  		if ri.l1() >= uint(len(mheap_.arenas)) {
   642  			return nil
   643  		}
   644  	}
   645  	l2 := mheap_.arenas[ri.l1()]
   646  	if arenaL1Bits != 0 && l2 == nil { // Should never happen if there's no L1.
   647  		return nil
   648  	}
   649  	ha := l2[ri.l2()]
   650  	if ha == nil {
   651  		return nil
   652  	}
   653  	return ha.spans[(p/pageSize)%pagesPerArena]
   654  }
   655  
   656  // spanOfUnchecked is equivalent to spanOf, but the caller must ensure
   657  // that p points into an allocated heap arena.
   658  //
   659  // Must be nosplit because it has callers that are nosplit.
   660  //
   661  //go:nosplit
   662  func spanOfUnchecked(p uintptr) *mspan {
   663  	ai := arenaIndex(p)
   664  	return mheap_.arenas[ai.l1()][ai.l2()].spans[(p/pageSize)%pagesPerArena]
   665  }
   666  
   667  // spanOfHeap is like spanOf, but returns nil if p does not point to a
   668  // heap object.
   669  //
   670  // Must be nosplit because it has callers that are nosplit.
   671  //
   672  //go:nosplit
   673  func spanOfHeap(p uintptr) *mspan {
   674  	s := spanOf(p)
   675  	// s is nil if it's never been allocated. Otherwise, we check
   676  	// its state first because we don't trust this pointer, so we
   677  	// have to synchronize with span initialization. Then, it's
   678  	// still possible we picked up a stale span pointer, so we
   679  	// have to check the span's bounds.
   680  	if s == nil || s.state.get() != mSpanInUse || p < s.base() || p >= s.limit {
   681  		return nil
   682  	}
   683  	return s
   684  }
   685  
   686  // pageIndexOf returns the arena, page index, and page mask for pointer p.
   687  // The caller must ensure p is in the heap.
   688  func pageIndexOf(p uintptr) (arena *heapArena, pageIdx uintptr, pageMask uint8) {
   689  	ai := arenaIndex(p)
   690  	arena = mheap_.arenas[ai.l1()][ai.l2()]
   691  	pageIdx = ((p / pageSize) / 8) % uintptr(len(arena.pageInUse))
   692  	pageMask = byte(1 << ((p / pageSize) % 8))
   693  	return
   694  }
   695  
   696  // Initialize the heap.
   697  func (h *mheap) init() {
   698  	lockInit(&h.lock, lockRankMheap)
   699  	lockInit(&h.speciallock, lockRankMheapSpecial)
   700  
   701  	h.spanalloc.init(unsafe.Sizeof(mspan{}), recordspan, unsafe.Pointer(h), &memstats.mspan_sys)
   702  	h.cachealloc.init(unsafe.Sizeof(mcache{}), nil, nil, &memstats.mcache_sys)
   703  	h.specialfinalizeralloc.init(unsafe.Sizeof(specialfinalizer{}), nil, nil, &memstats.other_sys)
   704  	h.specialprofilealloc.init(unsafe.Sizeof(specialprofile{}), nil, nil, &memstats.other_sys)
   705  	h.specialReachableAlloc.init(unsafe.Sizeof(specialReachable{}), nil, nil, &memstats.other_sys)
   706  	h.arenaHintAlloc.init(unsafe.Sizeof(arenaHint{}), nil, nil, &memstats.other_sys)
   707  
   708  	// Don't zero mspan allocations. Background sweeping can
   709  	// inspect a span concurrently with allocating it, so it's
   710  	// important that the span's sweepgen survive across freeing
   711  	// and re-allocating a span to prevent background sweeping
   712  	// from improperly cas'ing it from 0.
   713  	//
   714  	// This is safe because mspan contains no heap pointers.
   715  	h.spanalloc.zero = false
   716  
   717  	// h->mapcache needs no init
   718  
   719  	for i := range h.central {
   720  		h.central[i].mcentral.init(spanClass(i))
   721  	}
   722  
   723  	h.pages.init(&h.lock, &memstats.gcMiscSys)
   724  }
   725  
   726  // reclaim sweeps and reclaims at least npage pages into the heap.
   727  // It is called before allocating npage pages to keep growth in check.
   728  //
   729  // reclaim implements the page-reclaimer half of the sweeper.
   730  //
   731  // h.lock must NOT be held.
   732  func (h *mheap) reclaim(npage uintptr) {
   733  	// TODO(austin): Half of the time spent freeing spans is in
   734  	// locking/unlocking the heap (even with low contention). We
   735  	// could make the slow path here several times faster by
   736  	// batching heap frees.
   737  
   738  	// Bail early if there's no more reclaim work.
   739  	if h.reclaimIndex.Load() >= 1<<63 {
   740  		return
   741  	}
   742  
   743  	// Disable preemption so the GC can't start while we're
   744  	// sweeping, so we can read h.sweepArenas, and so
   745  	// traceGCSweepStart/Done pair on the P.
   746  	mp := acquirem()
   747  
   748  	if trace.enabled {
   749  		traceGCSweepStart()
   750  	}
   751  
   752  	arenas := h.sweepArenas
   753  	locked := false
   754  	for npage > 0 {
   755  		// Pull from accumulated credit first.
   756  		if credit := h.reclaimCredit.Load(); credit > 0 {
   757  			take := credit
   758  			if take > npage {
   759  				// Take only what we need.
   760  				take = npage
   761  			}
   762  			if h.reclaimCredit.CompareAndSwap(credit, credit-take) {
   763  				npage -= take
   764  			}
   765  			continue
   766  		}
   767  
   768  		// Claim a chunk of work.
   769  		idx := uintptr(h.reclaimIndex.Add(pagesPerReclaimerChunk) - pagesPerReclaimerChunk)
   770  		if idx/pagesPerArena >= uintptr(len(arenas)) {
   771  			// Page reclaiming is done.
   772  			h.reclaimIndex.Store(1 << 63)
   773  			break
   774  		}
   775  
   776  		if !locked {
   777  			// Lock the heap for reclaimChunk.
   778  			lock(&h.lock)
   779  			locked = true
   780  		}
   781  
   782  		// Scan this chunk.
   783  		nfound := h.reclaimChunk(arenas, idx, pagesPerReclaimerChunk)
   784  		if nfound <= npage {
   785  			npage -= nfound
   786  		} else {
   787  			// Put spare pages toward global credit.
   788  			h.reclaimCredit.Add(nfound - npage)
   789  			npage = 0
   790  		}
   791  	}
   792  	if locked {
   793  		unlock(&h.lock)
   794  	}
   795  
   796  	if trace.enabled {
   797  		traceGCSweepDone()
   798  	}
   799  	releasem(mp)
   800  }
   801  
   802  // reclaimChunk sweeps unmarked spans that start at page indexes [pageIdx, pageIdx+n).
   803  // It returns the number of pages returned to the heap.
   804  //
   805  // h.lock must be held and the caller must be non-preemptible. Note: h.lock may be
   806  // temporarily unlocked and re-locked in order to do sweeping or if tracing is
   807  // enabled.
   808  func (h *mheap) reclaimChunk(arenas []arenaIdx, pageIdx, n uintptr) uintptr {
   809  	// The heap lock must be held because this accesses the
   810  	// heapArena.spans arrays using potentially non-live pointers.
   811  	// In particular, if a span were freed and merged concurrently
   812  	// with this probing heapArena.spans, it would be possible to
   813  	// observe arbitrary, stale span pointers.
   814  	assertLockHeld(&h.lock)
   815  
   816  	n0 := n
   817  	var nFreed uintptr
   818  	sl := sweep.active.begin()
   819  	if !sl.valid {
   820  		return 0
   821  	}
   822  	for n > 0 {
   823  		ai := arenas[pageIdx/pagesPerArena]
   824  		ha := h.arenas[ai.l1()][ai.l2()]
   825  
   826  		// Get a chunk of the bitmap to work on.
   827  		arenaPage := uint(pageIdx % pagesPerArena)
   828  		inUse := ha.pageInUse[arenaPage/8:]
   829  		marked := ha.pageMarks[arenaPage/8:]
   830  		if uintptr(len(inUse)) > n/8 {
   831  			inUse = inUse[:n/8]
   832  			marked = marked[:n/8]
   833  		}
   834  
   835  		// Scan this bitmap chunk for spans that are in-use
   836  		// but have no marked objects on them.
   837  		for i := range inUse {
   838  			inUseUnmarked := atomic.Load8(&inUse[i]) &^ marked[i]
   839  			if inUseUnmarked == 0 {
   840  				continue
   841  			}
   842  
   843  			for j := uint(0); j < 8; j++ {
   844  				if inUseUnmarked&(1<<j) != 0 {
   845  					s := ha.spans[arenaPage+uint(i)*8+j]
   846  					if s, ok := sl.tryAcquire(s); ok {
   847  						npages := s.npages
   848  						unlock(&h.lock)
   849  						if s.sweep(false) {
   850  							nFreed += npages
   851  						}
   852  						lock(&h.lock)
   853  						// Reload inUse. It's possible nearby
   854  						// spans were freed when we dropped the
   855  						// lock and we don't want to get stale
   856  						// pointers from the spans array.
   857  						inUseUnmarked = atomic.Load8(&inUse[i]) &^ marked[i]
   858  					}
   859  				}
   860  			}
   861  		}
   862  
   863  		// Advance.
   864  		pageIdx += uintptr(len(inUse) * 8)
   865  		n -= uintptr(len(inUse) * 8)
   866  	}
   867  	sweep.active.end(sl)
   868  	if trace.enabled {
   869  		unlock(&h.lock)
   870  		// Account for pages scanned but not reclaimed.
   871  		traceGCSweepSpan((n0 - nFreed) * pageSize)
   872  		lock(&h.lock)
   873  	}
   874  
   875  	assertLockHeld(&h.lock) // Must be locked on return.
   876  	return nFreed
   877  }
   878  
   879  // spanAllocType represents the type of allocation to make, or
   880  // the type of allocation to be freed.
   881  type spanAllocType uint8
   882  
   883  const (
   884  	spanAllocHeap          spanAllocType = iota // heap span
   885  	spanAllocStack                              // stack span
   886  	spanAllocPtrScalarBits                      // unrolled GC prog bitmap span
   887  	spanAllocWorkBuf                            // work buf span
   888  )
   889  
   890  // manual returns true if the span allocation is manually managed.
   891  func (s spanAllocType) manual() bool {
   892  	return s != spanAllocHeap
   893  }
   894  
   895  // alloc allocates a new span of npage pages from the GC'd heap.
   896  //
   897  // spanclass indicates the span's size class and scannability.
   898  //
   899  // Returns a span that has been fully initialized. span.needzero indicates
   900  // whether the span has been zeroed. Note that it may not be.
   901  func (h *mheap) alloc(npages uintptr, spanclass spanClass) *mspan {
   902  	// Don't do any operations that lock the heap on the G stack.
   903  	// It might trigger stack growth, and the stack growth code needs
   904  	// to be able to allocate heap.
   905  	var s *mspan
   906  	systemstack(func() {
   907  		// To prevent excessive heap growth, before allocating n pages
   908  		// we need to sweep and reclaim at least n pages.
   909  		if !isSweepDone() {
   910  			h.reclaim(npages)
   911  		}
   912  		s = h.allocSpan(npages, spanAllocHeap, spanclass)
   913  	})
   914  	return s
   915  }
   916  
   917  // allocManual allocates a manually-managed span of npage pages.
   918  // allocManual returns nil if allocation fails.
   919  //
   920  // allocManual adds the bytes used to *stat, which should be a
   921  // memstats in-use field. Unlike allocations in the GC'd heap, the
   922  // allocation does *not* count toward heap_inuse or heap_sys.
   923  //
   924  // The memory backing the returned span may not be zeroed if
   925  // span.needzero is set.
   926  //
   927  // allocManual must be called on the system stack because it may
   928  // acquire the heap lock via allocSpan. See mheap for details.
   929  //
   930  // If new code is written to call allocManual, do NOT use an
   931  // existing spanAllocType value and instead declare a new one.
   932  //
   933  //go:systemstack
   934  func (h *mheap) allocManual(npages uintptr, typ spanAllocType) *mspan {
   935  	if !typ.manual() {
   936  		throw("manual span allocation called with non-manually-managed type")
   937  	}
   938  	return h.allocSpan(npages, typ, 0)
   939  }
   940  
   941  // setSpans modifies the span map so [spanOf(base), spanOf(base+npage*pageSize))
   942  // is s.
   943  func (h *mheap) setSpans(base, npage uintptr, s *mspan) {
   944  	p := base / pageSize
   945  	ai := arenaIndex(base)
   946  	ha := h.arenas[ai.l1()][ai.l2()]
   947  	for n := uintptr(0); n < npage; n++ {
   948  		i := (p + n) % pagesPerArena
   949  		if i == 0 {
   950  			ai = arenaIndex(base + n*pageSize)
   951  			ha = h.arenas[ai.l1()][ai.l2()]
   952  		}
   953  		ha.spans[i] = s
   954  	}
   955  }
   956  
   957  // allocNeedsZero checks if the region of address space [base, base+npage*pageSize),
   958  // assumed to be allocated, needs to be zeroed, updating heap arena metadata for
   959  // future allocations.
   960  //
   961  // This must be called each time pages are allocated from the heap, even if the page
   962  // allocator can otherwise prove the memory it's allocating is already zero because
   963  // they're fresh from the operating system. It updates heapArena metadata that is
   964  // critical for future page allocations.
   965  //
   966  // There are no locking constraints on this method.
   967  func (h *mheap) allocNeedsZero(base, npage uintptr) (needZero bool) {
   968  	for npage > 0 {
   969  		ai := arenaIndex(base)
   970  		ha := h.arenas[ai.l1()][ai.l2()]
   971  
   972  		zeroedBase := atomic.Loaduintptr(&ha.zeroedBase)
   973  		arenaBase := base % heapArenaBytes
   974  		if arenaBase < zeroedBase {
   975  			// We extended into the non-zeroed part of the
   976  			// arena, so this region needs to be zeroed before use.
   977  			//
   978  			// zeroedBase is monotonically increasing, so if we see this now then
   979  			// we can be sure we need to zero this memory region.
   980  			//
   981  			// We still need to update zeroedBase for this arena, and
   982  			// potentially more arenas.
   983  			needZero = true
   984  		}
   985  		// We may observe arenaBase > zeroedBase if we're racing with one or more
   986  		// allocations which are acquiring memory directly before us in the address
   987  		// space. But, because we know no one else is acquiring *this* memory, it's
   988  		// still safe to not zero.
   989  
   990  		// Compute how far into the arena we extend into, capped
   991  		// at heapArenaBytes.
   992  		arenaLimit := arenaBase + npage*pageSize
   993  		if arenaLimit > heapArenaBytes {
   994  			arenaLimit = heapArenaBytes
   995  		}
   996  		// Increase ha.zeroedBase so it's >= arenaLimit.
   997  		// We may be racing with other updates.
   998  		for arenaLimit > zeroedBase {
   999  			if atomic.Casuintptr(&ha.zeroedBase, zeroedBase, arenaLimit) {
  1000  				break
  1001  			}
  1002  			zeroedBase = atomic.Loaduintptr(&ha.zeroedBase)
  1003  			// Double check basic conditions of zeroedBase.
  1004  			if zeroedBase <= arenaLimit && zeroedBase > arenaBase {
  1005  				// The zeroedBase moved into the space we were trying to
  1006  				// claim. That's very bad, and indicates someone allocated
  1007  				// the same region we did.
  1008  				throw("potentially overlapping in-use allocations detected")
  1009  			}
  1010  		}
  1011  
  1012  		// Move base forward and subtract from npage to move into
  1013  		// the next arena, or finish.
  1014  		base += arenaLimit - arenaBase
  1015  		npage -= (arenaLimit - arenaBase) / pageSize
  1016  	}
  1017  	return
  1018  }
  1019  
  1020  // tryAllocMSpan attempts to allocate an mspan object from
  1021  // the P-local cache, but may fail.
  1022  //
  1023  // h.lock need not be held.
  1024  //
  1025  // This caller must ensure that its P won't change underneath
  1026  // it during this function. Currently to ensure that we enforce
  1027  // that the function is run on the system stack, because that's
  1028  // the only place it is used now. In the future, this requirement
  1029  // may be relaxed if its use is necessary elsewhere.
  1030  //
  1031  //go:systemstack
  1032  func (h *mheap) tryAllocMSpan() *mspan {
  1033  	pp := getg().m.p.ptr()
  1034  	// If we don't have a p or the cache is empty, we can't do
  1035  	// anything here.
  1036  	if pp == nil || pp.mspancache.len == 0 {
  1037  		return nil
  1038  	}
  1039  	// Pull off the last entry in the cache.
  1040  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1041  	pp.mspancache.len--
  1042  	return s
  1043  }
  1044  
  1045  // allocMSpanLocked allocates an mspan object.
  1046  //
  1047  // h.lock must be held.
  1048  //
  1049  // allocMSpanLocked must be called on the system stack because
  1050  // its caller holds the heap lock. See mheap for details.
  1051  // Running on the system stack also ensures that we won't
  1052  // switch Ps during this function. See tryAllocMSpan for details.
  1053  //
  1054  //go:systemstack
  1055  func (h *mheap) allocMSpanLocked() *mspan {
  1056  	assertLockHeld(&h.lock)
  1057  
  1058  	pp := getg().m.p.ptr()
  1059  	if pp == nil {
  1060  		// We don't have a p so just do the normal thing.
  1061  		return (*mspan)(h.spanalloc.alloc())
  1062  	}
  1063  	// Refill the cache if necessary.
  1064  	if pp.mspancache.len == 0 {
  1065  		const refillCount = len(pp.mspancache.buf) / 2
  1066  		for i := 0; i < refillCount; i++ {
  1067  			pp.mspancache.buf[i] = (*mspan)(h.spanalloc.alloc())
  1068  		}
  1069  		pp.mspancache.len = refillCount
  1070  	}
  1071  	// Pull off the last entry in the cache.
  1072  	s := pp.mspancache.buf[pp.mspancache.len-1]
  1073  	pp.mspancache.len--
  1074  	return s
  1075  }
  1076  
  1077  // freeMSpanLocked free an mspan object.
  1078  //
  1079  // h.lock must be held.
  1080  //
  1081  // freeMSpanLocked must be called on the system stack because
  1082  // its caller holds the heap lock. See mheap for details.
  1083  // Running on the system stack also ensures that we won't
  1084  // switch Ps during this function. See tryAllocMSpan for details.
  1085  //
  1086  //go:systemstack
  1087  func (h *mheap) freeMSpanLocked(s *mspan) {
  1088  	assertLockHeld(&h.lock)
  1089  
  1090  	pp := getg().m.p.ptr()
  1091  	// First try to free the mspan directly to the cache.
  1092  	if pp != nil && pp.mspancache.len < len(pp.mspancache.buf) {
  1093  		pp.mspancache.buf[pp.mspancache.len] = s
  1094  		pp.mspancache.len++
  1095  		return
  1096  	}
  1097  	// Failing that (or if we don't have a p), just free it to
  1098  	// the heap.
  1099  	h.spanalloc.free(unsafe.Pointer(s))
  1100  }
  1101  
  1102  // allocSpan allocates an mspan which owns npages worth of memory.
  1103  //
  1104  // If typ.manual() == false, allocSpan allocates a heap span of class spanclass
  1105  // and updates heap accounting. If manual == true, allocSpan allocates a
  1106  // manually-managed span (spanclass is ignored), and the caller is
  1107  // responsible for any accounting related to its use of the span. Either
  1108  // way, allocSpan will atomically add the bytes in the newly allocated
  1109  // span to *sysStat.
  1110  //
  1111  // The returned span is fully initialized.
  1112  //
  1113  // h.lock must not be held.
  1114  //
  1115  // allocSpan must be called on the system stack both because it acquires
  1116  // the heap lock and because it must block GC transitions.
  1117  //
  1118  //go:systemstack
  1119  func (h *mheap) allocSpan(npages uintptr, typ spanAllocType, spanclass spanClass) (s *mspan) {
  1120  	// Function-global state.
  1121  	gp := getg()
  1122  	base, scav := uintptr(0), uintptr(0)
  1123  	growth := uintptr(0)
  1124  
  1125  	// On some platforms we need to provide physical page aligned stack
  1126  	// allocations. Where the page size is less than the physical page
  1127  	// size, we already manage to do this by default.
  1128  	needPhysPageAlign := physPageAlignedStacks && typ == spanAllocStack && pageSize < physPageSize
  1129  
  1130  	// If the allocation is small enough, try the page cache!
  1131  	// The page cache does not support aligned allocations, so we cannot use
  1132  	// it if we need to provide a physical page aligned stack allocation.
  1133  	pp := gp.m.p.ptr()
  1134  	if !needPhysPageAlign && pp != nil && npages < pageCachePages/4 {
  1135  		c := &pp.pcache
  1136  
  1137  		// If the cache is empty, refill it.
  1138  		if c.empty() {
  1139  			lock(&h.lock)
  1140  			*c = h.pages.allocToCache()
  1141  			unlock(&h.lock)
  1142  		}
  1143  
  1144  		// Try to allocate from the cache.
  1145  		base, scav = c.alloc(npages)
  1146  		if base != 0 {
  1147  			s = h.tryAllocMSpan()
  1148  			if s != nil {
  1149  				goto HaveSpan
  1150  			}
  1151  			// We have a base but no mspan, so we need
  1152  			// to lock the heap.
  1153  		}
  1154  	}
  1155  
  1156  	// For one reason or another, we couldn't get the
  1157  	// whole job done without the heap lock.
  1158  	lock(&h.lock)
  1159  
  1160  	if needPhysPageAlign {
  1161  		// Overallocate by a physical page to allow for later alignment.
  1162  		npages += physPageSize / pageSize
  1163  	}
  1164  
  1165  	if base == 0 {
  1166  		// Try to acquire a base address.
  1167  		base, scav = h.pages.alloc(npages)
  1168  		if base == 0 {
  1169  			var ok bool
  1170  			growth, ok = h.grow(npages)
  1171  			if !ok {
  1172  				unlock(&h.lock)
  1173  				return nil
  1174  			}
  1175  			base, scav = h.pages.alloc(npages)
  1176  			if base == 0 {
  1177  				throw("grew heap, but no adequate free space found")
  1178  			}
  1179  		}
  1180  	}
  1181  	if s == nil {
  1182  		// We failed to get an mspan earlier, so grab
  1183  		// one now that we have the heap lock.
  1184  		s = h.allocMSpanLocked()
  1185  	}
  1186  
  1187  	if needPhysPageAlign {
  1188  		allocBase, allocPages := base, npages
  1189  		base = alignUp(allocBase, physPageSize)
  1190  		npages -= physPageSize / pageSize
  1191  
  1192  		// Return memory around the aligned allocation.
  1193  		spaceBefore := base - allocBase
  1194  		if spaceBefore > 0 {
  1195  			h.pages.free(allocBase, spaceBefore/pageSize, false)
  1196  		}
  1197  		spaceAfter := (allocPages-npages)*pageSize - spaceBefore
  1198  		if spaceAfter > 0 {
  1199  			h.pages.free(base+npages*pageSize, spaceAfter/pageSize, false)
  1200  		}
  1201  	}
  1202  
  1203  	unlock(&h.lock)
  1204  
  1205  	if growth > 0 {
  1206  		// We just caused a heap growth, so scavenge down what will soon be used.
  1207  		// By scavenging inline we deal with the failure to allocate out of
  1208  		// memory fragments by scavenging the memory fragments that are least
  1209  		// likely to be re-used.
  1210  		scavengeGoal := atomic.Load64(&h.scavengeGoal)
  1211  		if retained := heapRetained(); retained+uint64(growth) > scavengeGoal {
  1212  			// The scavenging algorithm requires the heap lock to be dropped so it
  1213  			// can acquire it only sparingly. This is a potentially expensive operation
  1214  			// so it frees up other goroutines to allocate in the meanwhile. In fact,
  1215  			// they can make use of the growth we just created.
  1216  			todo := growth
  1217  			if overage := uintptr(retained + uint64(growth) - scavengeGoal); todo > overage {
  1218  				todo = overage
  1219  			}
  1220  			h.pages.scavenge(todo)
  1221  		}
  1222  	}
  1223  
  1224  HaveSpan:
  1225  	// At this point, both s != nil and base != 0, and the heap
  1226  	// lock is no longer held. Initialize the span.
  1227  	s.init(base, npages)
  1228  	if h.allocNeedsZero(base, npages) {
  1229  		s.needzero = 1
  1230  	}
  1231  	nbytes := npages * pageSize
  1232  	if typ.manual() {
  1233  		s.manualFreeList = 0
  1234  		s.nelems = 0
  1235  		s.limit = s.base() + s.npages*pageSize
  1236  		s.state.set(mSpanManual)
  1237  	} else {
  1238  		// We must set span properties before the span is published anywhere
  1239  		// since we're not holding the heap lock.
  1240  		s.spanclass = spanclass
  1241  		if sizeclass := spanclass.sizeclass(); sizeclass == 0 {
  1242  			s.elemsize = nbytes
  1243  			s.nelems = 1
  1244  			s.divMul = 0
  1245  		} else {
  1246  			s.elemsize = uintptr(class_to_size[sizeclass])
  1247  			s.nelems = nbytes / s.elemsize
  1248  			s.divMul = class_to_divmagic[sizeclass]
  1249  		}
  1250  
  1251  		// Initialize mark and allocation structures.
  1252  		s.freeindex = 0
  1253  		s.allocCache = ^uint64(0) // all 1s indicating all free.
  1254  		s.gcmarkBits = newMarkBits(s.nelems)
  1255  		s.allocBits = newAllocBits(s.nelems)
  1256  
  1257  		// It's safe to access h.sweepgen without the heap lock because it's
  1258  		// only ever updated with the world stopped and we run on the
  1259  		// systemstack which blocks a STW transition.
  1260  		atomic.Store(&s.sweepgen, h.sweepgen)
  1261  
  1262  		// Now that the span is filled in, set its state. This
  1263  		// is a publication barrier for the other fields in
  1264  		// the span. While valid pointers into this span
  1265  		// should never be visible until the span is returned,
  1266  		// if the garbage collector finds an invalid pointer,
  1267  		// access to the span may race with initialization of
  1268  		// the span. We resolve this race by atomically
  1269  		// setting the state after the span is fully
  1270  		// initialized, and atomically checking the state in
  1271  		// any situation where a pointer is suspect.
  1272  		s.state.set(mSpanInUse)
  1273  	}
  1274  
  1275  	// Commit and account for any scavenged memory that the span now owns.
  1276  	if scav != 0 {
  1277  		// sysUsed all the pages that are actually available
  1278  		// in the span since some of them might be scavenged.
  1279  		sysUsed(unsafe.Pointer(base), nbytes)
  1280  		atomic.Xadd64(&memstats.heap_released, -int64(scav))
  1281  	}
  1282  	// Update stats.
  1283  	if typ == spanAllocHeap {
  1284  		atomic.Xadd64(&memstats.heap_inuse, int64(nbytes))
  1285  	}
  1286  	if typ.manual() {
  1287  		// Manually managed memory doesn't count toward heap_sys.
  1288  		memstats.heap_sys.add(-int64(nbytes))
  1289  	}
  1290  	// Update consistent stats.
  1291  	stats := memstats.heapStats.acquire()
  1292  	atomic.Xaddint64(&stats.committed, int64(scav))
  1293  	atomic.Xaddint64(&stats.released, -int64(scav))
  1294  	switch typ {
  1295  	case spanAllocHeap:
  1296  		atomic.Xaddint64(&stats.inHeap, int64(nbytes))
  1297  	case spanAllocStack:
  1298  		atomic.Xaddint64(&stats.inStacks, int64(nbytes))
  1299  	case spanAllocPtrScalarBits:
  1300  		atomic.Xaddint64(&stats.inPtrScalarBits, int64(nbytes))
  1301  	case spanAllocWorkBuf:
  1302  		atomic.Xaddint64(&stats.inWorkBufs, int64(nbytes))
  1303  	}
  1304  	memstats.heapStats.release()
  1305  
  1306  	// Publish the span in various locations.
  1307  
  1308  	// This is safe to call without the lock held because the slots
  1309  	// related to this span will only ever be read or modified by
  1310  	// this thread until pointers into the span are published (and
  1311  	// we execute a publication barrier at the end of this function
  1312  	// before that happens) or pageInUse is updated.
  1313  	h.setSpans(s.base(), npages, s)
  1314  
  1315  	if !typ.manual() {
  1316  		// Mark in-use span in arena page bitmap.
  1317  		//
  1318  		// This publishes the span to the page sweeper, so
  1319  		// it's imperative that the span be completely initialized
  1320  		// prior to this line.
  1321  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1322  		atomic.Or8(&arena.pageInUse[pageIdx], pageMask)
  1323  
  1324  		// Update related page sweeper stats.
  1325  		h.pagesInUse.Add(int64(npages))
  1326  	}
  1327  
  1328  	// Make sure the newly allocated span will be observed
  1329  	// by the GC before pointers into the span are published.
  1330  	publicationBarrier()
  1331  
  1332  	return s
  1333  }
  1334  
  1335  // Try to add at least npage pages of memory to the heap,
  1336  // returning how much the heap grew by and whether it worked.
  1337  //
  1338  // h.lock must be held.
  1339  func (h *mheap) grow(npage uintptr) (uintptr, bool) {
  1340  	assertLockHeld(&h.lock)
  1341  
  1342  	// We must grow the heap in whole palloc chunks.
  1343  	// We call sysMap below but note that because we
  1344  	// round up to pallocChunkPages which is on the order
  1345  	// of MiB (generally >= to the huge page size) we
  1346  	// won't be calling it too much.
  1347  	ask := alignUp(npage, pallocChunkPages) * pageSize
  1348  
  1349  	totalGrowth := uintptr(0)
  1350  	// This may overflow because ask could be very large
  1351  	// and is otherwise unrelated to h.curArena.base.
  1352  	end := h.curArena.base + ask
  1353  	nBase := alignUp(end, physPageSize)
  1354  	if nBase > h.curArena.end || /* overflow */ end < h.curArena.base {
  1355  		// Not enough room in the current arena. Allocate more
  1356  		// arena space. This may not be contiguous with the
  1357  		// current arena, so we have to request the full ask.
  1358  		av, asize := h.sysAlloc(ask)
  1359  		if av == nil {
  1360  			print("runtime: out of memory: cannot allocate ", ask, "-byte block (", memstats.heap_sys, " in use)\n")
  1361  			return 0, false
  1362  		}
  1363  
  1364  		if uintptr(av) == h.curArena.end {
  1365  			// The new space is contiguous with the old
  1366  			// space, so just extend the current space.
  1367  			h.curArena.end = uintptr(av) + asize
  1368  		} else {
  1369  			// The new space is discontiguous. Track what
  1370  			// remains of the current space and switch to
  1371  			// the new space. This should be rare.
  1372  			if size := h.curArena.end - h.curArena.base; size != 0 {
  1373  				// Transition this space from Reserved to Prepared and mark it
  1374  				// as released since we'll be able to start using it after updating
  1375  				// the page allocator and releasing the lock at any time.
  1376  				sysMap(unsafe.Pointer(h.curArena.base), size, &memstats.heap_sys)
  1377  				// Update stats.
  1378  				atomic.Xadd64(&memstats.heap_released, int64(size))
  1379  				stats := memstats.heapStats.acquire()
  1380  				atomic.Xaddint64(&stats.released, int64(size))
  1381  				memstats.heapStats.release()
  1382  				// Update the page allocator's structures to make this
  1383  				// space ready for allocation.
  1384  				h.pages.grow(h.curArena.base, size)
  1385  				totalGrowth += size
  1386  			}
  1387  			// Switch to the new space.
  1388  			h.curArena.base = uintptr(av)
  1389  			h.curArena.end = uintptr(av) + asize
  1390  		}
  1391  
  1392  		// Recalculate nBase.
  1393  		// We know this won't overflow, because sysAlloc returned
  1394  		// a valid region starting at h.curArena.base which is at
  1395  		// least ask bytes in size.
  1396  		nBase = alignUp(h.curArena.base+ask, physPageSize)
  1397  	}
  1398  
  1399  	// Grow into the current arena.
  1400  	v := h.curArena.base
  1401  	h.curArena.base = nBase
  1402  
  1403  	// Transition the space we're going to use from Reserved to Prepared.
  1404  	sysMap(unsafe.Pointer(v), nBase-v, &memstats.heap_sys)
  1405  
  1406  	// The memory just allocated counts as both released
  1407  	// and idle, even though it's not yet backed by spans.
  1408  	//
  1409  	// The allocation is always aligned to the heap arena
  1410  	// size which is always > physPageSize, so its safe to
  1411  	// just add directly to heap_released.
  1412  	atomic.Xadd64(&memstats.heap_released, int64(nBase-v))
  1413  	stats := memstats.heapStats.acquire()
  1414  	atomic.Xaddint64(&stats.released, int64(nBase-v))
  1415  	memstats.heapStats.release()
  1416  
  1417  	// Update the page allocator's structures to make this
  1418  	// space ready for allocation.
  1419  	h.pages.grow(v, nBase-v)
  1420  	totalGrowth += nBase - v
  1421  	return totalGrowth, true
  1422  }
  1423  
  1424  // Free the span back into the heap.
  1425  func (h *mheap) freeSpan(s *mspan) {
  1426  	systemstack(func() {
  1427  		lock(&h.lock)
  1428  		if msanenabled {
  1429  			// Tell msan that this entire span is no longer in use.
  1430  			base := unsafe.Pointer(s.base())
  1431  			bytes := s.npages << _PageShift
  1432  			msanfree(base, bytes)
  1433  		}
  1434  		if asanenabled {
  1435  			// Tell asan that this entire span is no longer in use.
  1436  			base := unsafe.Pointer(s.base())
  1437  			bytes := s.npages << _PageShift
  1438  			asanpoison(base, bytes)
  1439  		}
  1440  		h.freeSpanLocked(s, spanAllocHeap)
  1441  		unlock(&h.lock)
  1442  	})
  1443  }
  1444  
  1445  // freeManual frees a manually-managed span returned by allocManual.
  1446  // typ must be the same as the spanAllocType passed to the allocManual that
  1447  // allocated s.
  1448  //
  1449  // This must only be called when gcphase == _GCoff. See mSpanState for
  1450  // an explanation.
  1451  //
  1452  // freeManual must be called on the system stack because it acquires
  1453  // the heap lock. See mheap for details.
  1454  //
  1455  //go:systemstack
  1456  func (h *mheap) freeManual(s *mspan, typ spanAllocType) {
  1457  	s.needzero = 1
  1458  	lock(&h.lock)
  1459  	h.freeSpanLocked(s, typ)
  1460  	unlock(&h.lock)
  1461  }
  1462  
  1463  func (h *mheap) freeSpanLocked(s *mspan, typ spanAllocType) {
  1464  	assertLockHeld(&h.lock)
  1465  
  1466  	switch s.state.get() {
  1467  	case mSpanManual:
  1468  		if s.allocCount != 0 {
  1469  			throw("mheap.freeSpanLocked - invalid stack free")
  1470  		}
  1471  	case mSpanInUse:
  1472  		if s.allocCount != 0 || s.sweepgen != h.sweepgen {
  1473  			print("mheap.freeSpanLocked - span ", s, " ptr ", hex(s.base()), " allocCount ", s.allocCount, " sweepgen ", s.sweepgen, "/", h.sweepgen, "\n")
  1474  			throw("mheap.freeSpanLocked - invalid free")
  1475  		}
  1476  		h.pagesInUse.Add(-int64(s.npages))
  1477  
  1478  		// Clear in-use bit in arena page bitmap.
  1479  		arena, pageIdx, pageMask := pageIndexOf(s.base())
  1480  		atomic.And8(&arena.pageInUse[pageIdx], ^pageMask)
  1481  	default:
  1482  		throw("mheap.freeSpanLocked - invalid span state")
  1483  	}
  1484  
  1485  	// Update stats.
  1486  	//
  1487  	// Mirrors the code in allocSpan.
  1488  	nbytes := s.npages * pageSize
  1489  	if typ == spanAllocHeap {
  1490  		atomic.Xadd64(&memstats.heap_inuse, -int64(nbytes))
  1491  	}
  1492  	if typ.manual() {
  1493  		// Manually managed memory doesn't count toward heap_sys, so add it back.
  1494  		memstats.heap_sys.add(int64(nbytes))
  1495  	}
  1496  	// Update consistent stats.
  1497  	stats := memstats.heapStats.acquire()
  1498  	switch typ {
  1499  	case spanAllocHeap:
  1500  		atomic.Xaddint64(&stats.inHeap, -int64(nbytes))
  1501  	case spanAllocStack:
  1502  		atomic.Xaddint64(&stats.inStacks, -int64(nbytes))
  1503  	case spanAllocPtrScalarBits:
  1504  		atomic.Xaddint64(&stats.inPtrScalarBits, -int64(nbytes))
  1505  	case spanAllocWorkBuf:
  1506  		atomic.Xaddint64(&stats.inWorkBufs, -int64(nbytes))
  1507  	}
  1508  	memstats.heapStats.release()
  1509  
  1510  	// Mark the space as free.
  1511  	h.pages.free(s.base(), s.npages, false)
  1512  
  1513  	// Free the span structure. We no longer have a use for it.
  1514  	s.state.set(mSpanDead)
  1515  	h.freeMSpanLocked(s)
  1516  }
  1517  
  1518  // scavengeAll acquires the heap lock (blocking any additional
  1519  // manipulation of the page allocator) and iterates over the whole
  1520  // heap, scavenging every free page available.
  1521  func (h *mheap) scavengeAll() {
  1522  	// Disallow malloc or panic while holding the heap lock. We do
  1523  	// this here because this is a non-mallocgc entry-point to
  1524  	// the mheap API.
  1525  	gp := getg()
  1526  	gp.m.mallocing++
  1527  
  1528  	lock(&h.lock)
  1529  	// Start a new scavenge generation so we have a chance to walk
  1530  	// over the whole heap.
  1531  	h.pages.scavengeStartGen()
  1532  	unlock(&h.lock)
  1533  
  1534  	released := h.pages.scavenge(^uintptr(0))
  1535  
  1536  	lock(&h.pages.scav.lock)
  1537  	gen := h.pages.scav.gen
  1538  	unlock(&h.pages.scav.lock)
  1539  
  1540  	gp.m.mallocing--
  1541  
  1542  	if debug.scavtrace > 0 {
  1543  		printScavTrace(gen, released, true)
  1544  	}
  1545  }
  1546  
  1547  //go:linkname runtime_debug_freeOSMemory runtime/debug.freeOSMemory
  1548  func runtime_debug_freeOSMemory() {
  1549  	GC()
  1550  	systemstack(func() { mheap_.scavengeAll() })
  1551  }
  1552  
  1553  // Initialize a new span with the given start and npages.
  1554  func (span *mspan) init(base uintptr, npages uintptr) {
  1555  	// span is *not* zeroed.
  1556  	span.next = nil
  1557  	span.prev = nil
  1558  	span.list = nil
  1559  	span.startAddr = base
  1560  	span.npages = npages
  1561  	span.allocCount = 0
  1562  	span.spanclass = 0
  1563  	span.elemsize = 0
  1564  	span.speciallock.key = 0
  1565  	span.specials = nil
  1566  	span.needzero = 0
  1567  	span.freeindex = 0
  1568  	span.allocBits = nil
  1569  	span.gcmarkBits = nil
  1570  	span.state.set(mSpanDead)
  1571  	lockInit(&span.speciallock, lockRankMspanSpecial)
  1572  }
  1573  
  1574  func (span *mspan) inList() bool {
  1575  	return span.list != nil
  1576  }
  1577  
  1578  // Initialize an empty doubly-linked list.
  1579  func (list *mSpanList) init() {
  1580  	list.first = nil
  1581  	list.last = nil
  1582  }
  1583  
  1584  func (list *mSpanList) remove(span *mspan) {
  1585  	if span.list != list {
  1586  		print("runtime: failed mSpanList.remove span.npages=", span.npages,
  1587  			" span=", span, " prev=", span.prev, " span.list=", span.list, " list=", list, "\n")
  1588  		throw("mSpanList.remove")
  1589  	}
  1590  	if list.first == span {
  1591  		list.first = span.next
  1592  	} else {
  1593  		span.prev.next = span.next
  1594  	}
  1595  	if list.last == span {
  1596  		list.last = span.prev
  1597  	} else {
  1598  		span.next.prev = span.prev
  1599  	}
  1600  	span.next = nil
  1601  	span.prev = nil
  1602  	span.list = nil
  1603  }
  1604  
  1605  func (list *mSpanList) isEmpty() bool {
  1606  	return list.first == nil
  1607  }
  1608  
  1609  func (list *mSpanList) insert(span *mspan) {
  1610  	if span.next != nil || span.prev != nil || span.list != nil {
  1611  		println("runtime: failed mSpanList.insert", span, span.next, span.prev, span.list)
  1612  		throw("mSpanList.insert")
  1613  	}
  1614  	span.next = list.first
  1615  	if list.first != nil {
  1616  		// The list contains at least one span; link it in.
  1617  		// The last span in the list doesn't change.
  1618  		list.first.prev = span
  1619  	} else {
  1620  		// The list contains no spans, so this is also the last span.
  1621  		list.last = span
  1622  	}
  1623  	list.first = span
  1624  	span.list = list
  1625  }
  1626  
  1627  func (list *mSpanList) insertBack(span *mspan) {
  1628  	if span.next != nil || span.prev != nil || span.list != nil {
  1629  		println("runtime: failed mSpanList.insertBack", span, span.next, span.prev, span.list)
  1630  		throw("mSpanList.insertBack")
  1631  	}
  1632  	span.prev = list.last
  1633  	if list.last != nil {
  1634  		// The list contains at least one span.
  1635  		list.last.next = span
  1636  	} else {
  1637  		// The list contains no spans, so this is also the first span.
  1638  		list.first = span
  1639  	}
  1640  	list.last = span
  1641  	span.list = list
  1642  }
  1643  
  1644  // takeAll removes all spans from other and inserts them at the front
  1645  // of list.
  1646  func (list *mSpanList) takeAll(other *mSpanList) {
  1647  	if other.isEmpty() {
  1648  		return
  1649  	}
  1650  
  1651  	// Reparent everything in other to list.
  1652  	for s := other.first; s != nil; s = s.next {
  1653  		s.list = list
  1654  	}
  1655  
  1656  	// Concatenate the lists.
  1657  	if list.isEmpty() {
  1658  		*list = *other
  1659  	} else {
  1660  		// Neither list is empty. Put other before list.
  1661  		other.last.next = list.first
  1662  		list.first.prev = other.last
  1663  		list.first = other.first
  1664  	}
  1665  
  1666  	other.first, other.last = nil, nil
  1667  }
  1668  
  1669  const (
  1670  	_KindSpecialFinalizer = 1
  1671  	_KindSpecialProfile   = 2
  1672  	// _KindSpecialReachable is a special used for tracking
  1673  	// reachability during testing.
  1674  	_KindSpecialReachable = 3
  1675  	// Note: The finalizer special must be first because if we're freeing
  1676  	// an object, a finalizer special will cause the freeing operation
  1677  	// to abort, and we want to keep the other special records around
  1678  	// if that happens.
  1679  )
  1680  
  1681  //go:notinheap
  1682  type special struct {
  1683  	next   *special // linked list in span
  1684  	offset uint16   // span offset of object
  1685  	kind   byte     // kind of special
  1686  }
  1687  
  1688  // spanHasSpecials marks a span as having specials in the arena bitmap.
  1689  func spanHasSpecials(s *mspan) {
  1690  	arenaPage := (s.base() / pageSize) % pagesPerArena
  1691  	ai := arenaIndex(s.base())
  1692  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  1693  	atomic.Or8(&ha.pageSpecials[arenaPage/8], uint8(1)<<(arenaPage%8))
  1694  }
  1695  
  1696  // spanHasNoSpecials marks a span as having no specials in the arena bitmap.
  1697  func spanHasNoSpecials(s *mspan) {
  1698  	arenaPage := (s.base() / pageSize) % pagesPerArena
  1699  	ai := arenaIndex(s.base())
  1700  	ha := mheap_.arenas[ai.l1()][ai.l2()]
  1701  	atomic.And8(&ha.pageSpecials[arenaPage/8], ^(uint8(1) << (arenaPage % 8)))
  1702  }
  1703  
  1704  // Adds the special record s to the list of special records for
  1705  // the object p. All fields of s should be filled in except for
  1706  // offset & next, which this routine will fill in.
  1707  // Returns true if the special was successfully added, false otherwise.
  1708  // (The add will fail only if a record with the same p and s->kind
  1709  //  already exists.)
  1710  func addspecial(p unsafe.Pointer, s *special) bool {
  1711  	span := spanOfHeap(uintptr(p))
  1712  	if span == nil {
  1713  		throw("addspecial on invalid pointer")
  1714  	}
  1715  
  1716  	// Ensure that the span is swept.
  1717  	// Sweeping accesses the specials list w/o locks, so we have
  1718  	// to synchronize with it. And it's just much safer.
  1719  	mp := acquirem()
  1720  	span.ensureSwept()
  1721  
  1722  	offset := uintptr(p) - span.base()
  1723  	kind := s.kind
  1724  
  1725  	lock(&span.speciallock)
  1726  
  1727  	// Find splice point, check for existing record.
  1728  	t := &span.specials
  1729  	for {
  1730  		x := *t
  1731  		if x == nil {
  1732  			break
  1733  		}
  1734  		if offset == uintptr(x.offset) && kind == x.kind {
  1735  			unlock(&span.speciallock)
  1736  			releasem(mp)
  1737  			return false // already exists
  1738  		}
  1739  		if offset < uintptr(x.offset) || (offset == uintptr(x.offset) && kind < x.kind) {
  1740  			break
  1741  		}
  1742  		t = &x.next
  1743  	}
  1744  
  1745  	// Splice in record, fill in offset.
  1746  	s.offset = uint16(offset)
  1747  	s.next = *t
  1748  	*t = s
  1749  	spanHasSpecials(span)
  1750  	unlock(&span.speciallock)
  1751  	releasem(mp)
  1752  
  1753  	return true
  1754  }
  1755  
  1756  // Removes the Special record of the given kind for the object p.
  1757  // Returns the record if the record existed, nil otherwise.
  1758  // The caller must FixAlloc_Free the result.
  1759  func removespecial(p unsafe.Pointer, kind uint8) *special {
  1760  	span := spanOfHeap(uintptr(p))
  1761  	if span == nil {
  1762  		throw("removespecial on invalid pointer")
  1763  	}
  1764  
  1765  	// Ensure that the span is swept.
  1766  	// Sweeping accesses the specials list w/o locks, so we have
  1767  	// to synchronize with it. And it's just much safer.
  1768  	mp := acquirem()
  1769  	span.ensureSwept()
  1770  
  1771  	offset := uintptr(p) - span.base()
  1772  
  1773  	var result *special
  1774  	lock(&span.speciallock)
  1775  	t := &span.specials
  1776  	for {
  1777  		s := *t
  1778  		if s == nil {
  1779  			break
  1780  		}
  1781  		// This function is used for finalizers only, so we don't check for
  1782  		// "interior" specials (p must be exactly equal to s->offset).
  1783  		if offset == uintptr(s.offset) && kind == s.kind {
  1784  			*t = s.next
  1785  			result = s
  1786  			break
  1787  		}
  1788  		t = &s.next
  1789  	}
  1790  	if span.specials == nil {
  1791  		spanHasNoSpecials(span)
  1792  	}
  1793  	unlock(&span.speciallock)
  1794  	releasem(mp)
  1795  	return result
  1796  }
  1797  
  1798  // The described object has a finalizer set for it.
  1799  //
  1800  // specialfinalizer is allocated from non-GC'd memory, so any heap
  1801  // pointers must be specially handled.
  1802  //
  1803  //go:notinheap
  1804  type specialfinalizer struct {
  1805  	special special
  1806  	fn      *funcval // May be a heap pointer.
  1807  	nret    uintptr
  1808  	fint    *_type   // May be a heap pointer, but always live.
  1809  	ot      *ptrtype // May be a heap pointer, but always live.
  1810  }
  1811  
  1812  // Adds a finalizer to the object p. Returns true if it succeeded.
  1813  func addfinalizer(p unsafe.Pointer, f *funcval, nret uintptr, fint *_type, ot *ptrtype) bool {
  1814  	lock(&mheap_.speciallock)
  1815  	s := (*specialfinalizer)(mheap_.specialfinalizeralloc.alloc())
  1816  	unlock(&mheap_.speciallock)
  1817  	s.special.kind = _KindSpecialFinalizer
  1818  	s.fn = f
  1819  	s.nret = nret
  1820  	s.fint = fint
  1821  	s.ot = ot
  1822  	if addspecial(p, &s.special) {
  1823  		// This is responsible for maintaining the same
  1824  		// GC-related invariants as markrootSpans in any
  1825  		// situation where it's possible that markrootSpans
  1826  		// has already run but mark termination hasn't yet.
  1827  		if gcphase != _GCoff {
  1828  			base, _, _ := findObject(uintptr(p), 0, 0)
  1829  			mp := acquirem()
  1830  			gcw := &mp.p.ptr().gcw
  1831  			// Mark everything reachable from the object
  1832  			// so it's retained for the finalizer.
  1833  			scanobject(base, gcw)
  1834  			// Mark the finalizer itself, since the
  1835  			// special isn't part of the GC'd heap.
  1836  			scanblock(uintptr(unsafe.Pointer(&s.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
  1837  			releasem(mp)
  1838  		}
  1839  		return true
  1840  	}
  1841  
  1842  	// There was an old finalizer
  1843  	lock(&mheap_.speciallock)
  1844  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  1845  	unlock(&mheap_.speciallock)
  1846  	return false
  1847  }
  1848  
  1849  // Removes the finalizer (if any) from the object p.
  1850  func removefinalizer(p unsafe.Pointer) {
  1851  	s := (*specialfinalizer)(unsafe.Pointer(removespecial(p, _KindSpecialFinalizer)))
  1852  	if s == nil {
  1853  		return // there wasn't a finalizer to remove
  1854  	}
  1855  	lock(&mheap_.speciallock)
  1856  	mheap_.specialfinalizeralloc.free(unsafe.Pointer(s))
  1857  	unlock(&mheap_.speciallock)
  1858  }
  1859  
  1860  // The described object is being heap profiled.
  1861  //
  1862  //go:notinheap
  1863  type specialprofile struct {
  1864  	special special
  1865  	b       *bucket
  1866  }
  1867  
  1868  // Set the heap profile bucket associated with addr to b.
  1869  func setprofilebucket(p unsafe.Pointer, b *bucket) {
  1870  	lock(&mheap_.speciallock)
  1871  	s := (*specialprofile)(mheap_.specialprofilealloc.alloc())
  1872  	unlock(&mheap_.speciallock)
  1873  	s.special.kind = _KindSpecialProfile
  1874  	s.b = b
  1875  	if !addspecial(p, &s.special) {
  1876  		throw("setprofilebucket: profile already set")
  1877  	}
  1878  }
  1879  
  1880  // specialReachable tracks whether an object is reachable on the next
  1881  // GC cycle. This is used by testing.
  1882  type specialReachable struct {
  1883  	special   special
  1884  	done      bool
  1885  	reachable bool
  1886  }
  1887  
  1888  // specialsIter helps iterate over specials lists.
  1889  type specialsIter struct {
  1890  	pprev **special
  1891  	s     *special
  1892  }
  1893  
  1894  func newSpecialsIter(span *mspan) specialsIter {
  1895  	return specialsIter{&span.specials, span.specials}
  1896  }
  1897  
  1898  func (i *specialsIter) valid() bool {
  1899  	return i.s != nil
  1900  }
  1901  
  1902  func (i *specialsIter) next() {
  1903  	i.pprev = &i.s.next
  1904  	i.s = *i.pprev
  1905  }
  1906  
  1907  // unlinkAndNext removes the current special from the list and moves
  1908  // the iterator to the next special. It returns the unlinked special.
  1909  func (i *specialsIter) unlinkAndNext() *special {
  1910  	cur := i.s
  1911  	i.s = cur.next
  1912  	*i.pprev = i.s
  1913  	return cur
  1914  }
  1915  
  1916  // freeSpecial performs any cleanup on special s and deallocates it.
  1917  // s must already be unlinked from the specials list.
  1918  func freeSpecial(s *special, p unsafe.Pointer, size uintptr) {
  1919  	switch s.kind {
  1920  	case _KindSpecialFinalizer:
  1921  		sf := (*specialfinalizer)(unsafe.Pointer(s))
  1922  		queuefinalizer(p, sf.fn, sf.nret, sf.fint, sf.ot)
  1923  		lock(&mheap_.speciallock)
  1924  		mheap_.specialfinalizeralloc.free(unsafe.Pointer(sf))
  1925  		unlock(&mheap_.speciallock)
  1926  	case _KindSpecialProfile:
  1927  		sp := (*specialprofile)(unsafe.Pointer(s))
  1928  		mProf_Free(sp.b, size)
  1929  		lock(&mheap_.speciallock)
  1930  		mheap_.specialprofilealloc.free(unsafe.Pointer(sp))
  1931  		unlock(&mheap_.speciallock)
  1932  	case _KindSpecialReachable:
  1933  		sp := (*specialReachable)(unsafe.Pointer(s))
  1934  		sp.done = true
  1935  		// The creator frees these.
  1936  	default:
  1937  		throw("bad special kind")
  1938  		panic("not reached")
  1939  	}
  1940  }
  1941  
  1942  // gcBits is an alloc/mark bitmap. This is always used as *gcBits.
  1943  //
  1944  //go:notinheap
  1945  type gcBits uint8
  1946  
  1947  // bytep returns a pointer to the n'th byte of b.
  1948  func (b *gcBits) bytep(n uintptr) *uint8 {
  1949  	return addb((*uint8)(b), n)
  1950  }
  1951  
  1952  // bitp returns a pointer to the byte containing bit n and a mask for
  1953  // selecting that bit from *bytep.
  1954  func (b *gcBits) bitp(n uintptr) (bytep *uint8, mask uint8) {
  1955  	return b.bytep(n / 8), 1 << (n % 8)
  1956  }
  1957  
  1958  const gcBitsChunkBytes = uintptr(64 << 10)
  1959  const gcBitsHeaderBytes = unsafe.Sizeof(gcBitsHeader{})
  1960  
  1961  type gcBitsHeader struct {
  1962  	free uintptr // free is the index into bits of the next free byte.
  1963  	next uintptr // *gcBits triggers recursive type bug. (issue 14620)
  1964  }
  1965  
  1966  //go:notinheap
  1967  type gcBitsArena struct {
  1968  	// gcBitsHeader // side step recursive type bug (issue 14620) by including fields by hand.
  1969  	free uintptr // free is the index into bits of the next free byte; read/write atomically
  1970  	next *gcBitsArena
  1971  	bits [gcBitsChunkBytes - gcBitsHeaderBytes]gcBits
  1972  }
  1973  
  1974  var gcBitsArenas struct {
  1975  	lock     mutex
  1976  	free     *gcBitsArena
  1977  	next     *gcBitsArena // Read atomically. Write atomically under lock.
  1978  	current  *gcBitsArena
  1979  	previous *gcBitsArena
  1980  }
  1981  
  1982  // tryAlloc allocates from b or returns nil if b does not have enough room.
  1983  // This is safe to call concurrently.
  1984  func (b *gcBitsArena) tryAlloc(bytes uintptr) *gcBits {
  1985  	if b == nil || atomic.Loaduintptr(&b.free)+bytes > uintptr(len(b.bits)) {
  1986  		return nil
  1987  	}
  1988  	// Try to allocate from this block.
  1989  	end := atomic.Xadduintptr(&b.free, bytes)
  1990  	if end > uintptr(len(b.bits)) {
  1991  		return nil
  1992  	}
  1993  	// There was enough room.
  1994  	start := end - bytes
  1995  	return &b.bits[start]
  1996  }
  1997  
  1998  // newMarkBits returns a pointer to 8 byte aligned bytes
  1999  // to be used for a span's mark bits.
  2000  func newMarkBits(nelems uintptr) *gcBits {
  2001  	blocksNeeded := uintptr((nelems + 63) / 64)
  2002  	bytesNeeded := blocksNeeded * 8
  2003  
  2004  	// Try directly allocating from the current head arena.
  2005  	head := (*gcBitsArena)(atomic.Loadp(unsafe.Pointer(&gcBitsArenas.next)))
  2006  	if p := head.tryAlloc(bytesNeeded); p != nil {
  2007  		return p
  2008  	}
  2009  
  2010  	// There's not enough room in the head arena. We may need to
  2011  	// allocate a new arena.
  2012  	lock(&gcBitsArenas.lock)
  2013  	// Try the head arena again, since it may have changed. Now
  2014  	// that we hold the lock, the list head can't change, but its
  2015  	// free position still can.
  2016  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2017  		unlock(&gcBitsArenas.lock)
  2018  		return p
  2019  	}
  2020  
  2021  	// Allocate a new arena. This may temporarily drop the lock.
  2022  	fresh := newArenaMayUnlock()
  2023  	// If newArenaMayUnlock dropped the lock, another thread may
  2024  	// have put a fresh arena on the "next" list. Try allocating
  2025  	// from next again.
  2026  	if p := gcBitsArenas.next.tryAlloc(bytesNeeded); p != nil {
  2027  		// Put fresh back on the free list.
  2028  		// TODO: Mark it "already zeroed"
  2029  		fresh.next = gcBitsArenas.free
  2030  		gcBitsArenas.free = fresh
  2031  		unlock(&gcBitsArenas.lock)
  2032  		return p
  2033  	}
  2034  
  2035  	// Allocate from the fresh arena. We haven't linked it in yet, so
  2036  	// this cannot race and is guaranteed to succeed.
  2037  	p := fresh.tryAlloc(bytesNeeded)
  2038  	if p == nil {
  2039  		throw("markBits overflow")
  2040  	}
  2041  
  2042  	// Add the fresh arena to the "next" list.
  2043  	fresh.next = gcBitsArenas.next
  2044  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), unsafe.Pointer(fresh))
  2045  
  2046  	unlock(&gcBitsArenas.lock)
  2047  	return p
  2048  }
  2049  
  2050  // newAllocBits returns a pointer to 8 byte aligned bytes
  2051  // to be used for this span's alloc bits.
  2052  // newAllocBits is used to provide newly initialized spans
  2053  // allocation bits. For spans not being initialized the
  2054  // mark bits are repurposed as allocation bits when
  2055  // the span is swept.
  2056  func newAllocBits(nelems uintptr) *gcBits {
  2057  	return newMarkBits(nelems)
  2058  }
  2059  
  2060  // nextMarkBitArenaEpoch establishes a new epoch for the arenas
  2061  // holding the mark bits. The arenas are named relative to the
  2062  // current GC cycle which is demarcated by the call to finishweep_m.
  2063  //
  2064  // All current spans have been swept.
  2065  // During that sweep each span allocated room for its gcmarkBits in
  2066  // gcBitsArenas.next block. gcBitsArenas.next becomes the gcBitsArenas.current
  2067  // where the GC will mark objects and after each span is swept these bits
  2068  // will be used to allocate objects.
  2069  // gcBitsArenas.current becomes gcBitsArenas.previous where the span's
  2070  // gcAllocBits live until all the spans have been swept during this GC cycle.
  2071  // The span's sweep extinguishes all the references to gcBitsArenas.previous
  2072  // by pointing gcAllocBits into the gcBitsArenas.current.
  2073  // The gcBitsArenas.previous is released to the gcBitsArenas.free list.
  2074  func nextMarkBitArenaEpoch() {
  2075  	lock(&gcBitsArenas.lock)
  2076  	if gcBitsArenas.previous != nil {
  2077  		if gcBitsArenas.free == nil {
  2078  			gcBitsArenas.free = gcBitsArenas.previous
  2079  		} else {
  2080  			// Find end of previous arenas.
  2081  			last := gcBitsArenas.previous
  2082  			for last = gcBitsArenas.previous; last.next != nil; last = last.next {
  2083  			}
  2084  			last.next = gcBitsArenas.free
  2085  			gcBitsArenas.free = gcBitsArenas.previous
  2086  		}
  2087  	}
  2088  	gcBitsArenas.previous = gcBitsArenas.current
  2089  	gcBitsArenas.current = gcBitsArenas.next
  2090  	atomic.StorepNoWB(unsafe.Pointer(&gcBitsArenas.next), nil) // newMarkBits calls newArena when needed
  2091  	unlock(&gcBitsArenas.lock)
  2092  }
  2093  
  2094  // newArenaMayUnlock allocates and zeroes a gcBits arena.
  2095  // The caller must hold gcBitsArena.lock. This may temporarily release it.
  2096  func newArenaMayUnlock() *gcBitsArena {
  2097  	var result *gcBitsArena
  2098  	if gcBitsArenas.free == nil {
  2099  		unlock(&gcBitsArenas.lock)
  2100  		result = (*gcBitsArena)(sysAlloc(gcBitsChunkBytes, &memstats.gcMiscSys))
  2101  		if result == nil {
  2102  			throw("runtime: cannot allocate memory")
  2103  		}
  2104  		lock(&gcBitsArenas.lock)
  2105  	} else {
  2106  		result = gcBitsArenas.free
  2107  		gcBitsArenas.free = gcBitsArenas.free.next
  2108  		memclrNoHeapPointers(unsafe.Pointer(result), gcBitsChunkBytes)
  2109  	}
  2110  	result.next = nil
  2111  	// If result.bits is not 8 byte aligned adjust index so
  2112  	// that &result.bits[result.free] is 8 byte aligned.
  2113  	if uintptr(unsafe.Offsetof(gcBitsArena{}.bits))&7 == 0 {
  2114  		result.free = 0
  2115  	} else {
  2116  		result.free = 8 - (uintptr(unsafe.Pointer(&result.bits[0])) & 7)
  2117  	}
  2118  	return result
  2119  }
  2120  

View as plain text