Source file src/runtime/stack.go

     1  // Copyright 2013 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  package runtime
     6  
     7  import (
     8  	"internal/abi"
     9  	"internal/cpu"
    10  	"internal/goarch"
    11  	"internal/goos"
    12  	"internal/runtime/atomic"
    13  	"internal/runtime/gc"
    14  	"internal/runtime/sys"
    15  	"unsafe"
    16  )
    17  
    18  /*
    19  Stack layout parameters.
    20  Included both by runtime (compiled via 6c) and linkers (compiled via gcc).
    21  
    22  The per-goroutine g->stackguard is set to point StackGuard bytes
    23  above the bottom of the stack.  Each function compares its stack
    24  pointer against g->stackguard to check for overflow.  To cut one
    25  instruction from the check sequence for functions with tiny frames,
    26  the stack is allowed to protrude StackSmall bytes below the stack
    27  guard.  Functions with large frames don't bother with the check and
    28  always call morestack.  The sequences are (for amd64, others are
    29  similar):
    30  
    31  	guard = g->stackguard
    32  	frame = function's stack frame size
    33  	argsize = size of function arguments (call + return)
    34  
    35  	stack frame size <= StackSmall:
    36  		CMPQ guard, SP
    37  		JHI 3(PC)
    38  		MOVQ m->morearg, $(argsize << 32)
    39  		CALL morestack(SB)
    40  
    41  	stack frame size > StackSmall but < StackBig
    42  		LEAQ (frame-StackSmall)(SP), R0
    43  		CMPQ guard, R0
    44  		JHI 3(PC)
    45  		MOVQ m->morearg, $(argsize << 32)
    46  		CALL morestack(SB)
    47  
    48  	stack frame size >= StackBig:
    49  		MOVQ m->morearg, $((argsize << 32) | frame)
    50  		CALL morestack(SB)
    51  
    52  The bottom StackGuard - StackSmall bytes are important: there has
    53  to be enough room to execute functions that refuse to check for
    54  stack overflow, either because they need to be adjacent to the
    55  actual caller's frame (deferproc) or because they handle the imminent
    56  stack overflow (morestack).
    57  
    58  For example, deferproc might call malloc, which does one of the
    59  above checks (without allocating a full frame), which might trigger
    60  a call to morestack.  This sequence needs to fit in the bottom
    61  section of the stack.  On amd64, morestack's frame is 40 bytes, and
    62  deferproc's frame is 56 bytes.  That fits well within the
    63  StackGuard - StackSmall bytes at the bottom.
    64  The linkers explore all possible call traces involving non-splitting
    65  functions to make sure that this limit cannot be violated.
    66  */
    67  
    68  const (
    69  	// stackSystem is a number of additional bytes to add
    70  	// to each stack below the usual guard area for OS-specific
    71  	// purposes like signal handling. Used on Windows, Plan 9,
    72  	// and iOS because they do not use a separate stack.
    73  	stackSystem = goos.IsWindows*4096 + goos.IsPlan9*512 + goos.IsIos*goarch.IsArm64*1024
    74  
    75  	// The minimum size of stack used by Go code
    76  	stackMin = 2048
    77  
    78  	// The minimum stack size to allocate.
    79  	// The hackery here rounds fixedStack0 up to a power of 2.
    80  	fixedStack0 = stackMin + stackSystem
    81  	fixedStack1 = fixedStack0 - 1
    82  	fixedStack2 = fixedStack1 | (fixedStack1 >> 1)
    83  	fixedStack3 = fixedStack2 | (fixedStack2 >> 2)
    84  	fixedStack4 = fixedStack3 | (fixedStack3 >> 4)
    85  	fixedStack5 = fixedStack4 | (fixedStack4 >> 8)
    86  	fixedStack6 = fixedStack5 | (fixedStack5 >> 16)
    87  	fixedStack  = fixedStack6 + 1
    88  
    89  	// stackNosplit is the maximum number of bytes that a chain of NOSPLIT
    90  	// functions can use.
    91  	// This arithmetic must match that in cmd/internal/objabi/stack.go:StackNosplit.
    92  	stackNosplit = abi.StackNosplitBase * sys.StackGuardMultiplier
    93  
    94  	// The stack guard is a pointer this many bytes above the
    95  	// bottom of the stack.
    96  	//
    97  	// The guard leaves enough room for a stackNosplit chain of NOSPLIT calls
    98  	// plus one stackSmall frame plus stackSystem bytes for the OS.
    99  	// This arithmetic must match that in cmd/internal/objabi/stack.go:StackLimit.
   100  	stackGuard = stackNosplit + stackSystem + abi.StackSmall
   101  )
   102  
   103  const (
   104  	// stackDebug == 0: no logging
   105  	//            == 1: logging of per-stack operations
   106  	//            == 2: logging of per-frame operations
   107  	//            == 3: logging of per-word updates
   108  	//            == 4: logging of per-word reads
   109  	stackDebug       = 0
   110  	stackFromSystem  = 0 // allocate stacks from system memory instead of the heap
   111  	stackFaultOnFree = 0 // old stacks are mapped noaccess to detect use after free
   112  	stackNoCache     = 0 // disable per-P small stack caches
   113  
   114  	// check the BP links during traceback.
   115  	debugCheckBP = false
   116  )
   117  
   118  var (
   119  	stackPoisonCopy = 0 // fill stack that should not be accessed with garbage, to detect bad dereferences during copy
   120  )
   121  
   122  const (
   123  	uintptrMask = 1<<(8*goarch.PtrSize) - 1
   124  
   125  	// The values below can be stored to g.stackguard0 to force
   126  	// the next stack check to fail.
   127  	// These are all larger than any real SP.
   128  
   129  	// Goroutine preemption request.
   130  	// 0xfffffade in hex.
   131  	stackPreempt = uintptrMask & -1314
   132  
   133  	// Thread is forking. Causes a split stack check failure.
   134  	// 0xfffffb2e in hex.
   135  	stackFork = uintptrMask & -1234
   136  
   137  	// Force a stack movement. Used for debugging.
   138  	// 0xfffffeed in hex.
   139  	stackForceMove = uintptrMask & -275
   140  
   141  	// stackPoisonMin is the lowest allowed stack poison value.
   142  	stackPoisonMin = uintptrMask & -4096
   143  )
   144  
   145  // Global pool of spans that have free stacks.
   146  // Stacks are assigned an order according to size.
   147  //
   148  //	order = log_2(size/FixedStack)
   149  //
   150  // There is a free list for each order.
   151  var stackpool [_NumStackOrders]struct {
   152  	item stackpoolItem
   153  	_    [(cpu.CacheLinePadSize - unsafe.Sizeof(stackpoolItem{})%cpu.CacheLinePadSize) % cpu.CacheLinePadSize]byte
   154  }
   155  
   156  type stackpoolItem struct {
   157  	_    sys.NotInHeap
   158  	mu   mutex
   159  	span mSpanList
   160  }
   161  
   162  // Global pool of large stack spans.
   163  var stackLarge struct {
   164  	lock mutex
   165  	free [heapAddrBits - gc.PageShift]mSpanList // free lists by log_2(s.npages)
   166  }
   167  
   168  func stackinit() {
   169  	if _StackCacheSize&pageMask != 0 {
   170  		throw("cache size must be a multiple of page size")
   171  	}
   172  	for i := range stackpool {
   173  		stackpool[i].item.span.init()
   174  		lockInit(&stackpool[i].item.mu, lockRankStackpool)
   175  	}
   176  	for i := range stackLarge.free {
   177  		stackLarge.free[i].init()
   178  		lockInit(&stackLarge.lock, lockRankStackLarge)
   179  	}
   180  }
   181  
   182  // stacklog2 returns ⌊log_2(n)⌋.
   183  func stacklog2(n uintptr) int {
   184  	log2 := 0
   185  	for n > 1 {
   186  		n >>= 1
   187  		log2++
   188  	}
   189  	return log2
   190  }
   191  
   192  // Allocates a stack from the free pool. Must be called with
   193  // stackpool[order].item.mu held.
   194  func stackpoolalloc(order uint8) gclinkptr {
   195  	list := &stackpool[order].item.span
   196  	s := list.first
   197  	lockWithRankMayAcquire(&mheap_.lock, lockRankMheap)
   198  	if s == nil {
   199  		// no free stacks. Allocate another span worth.
   200  		s = mheap_.allocManual(_StackCacheSize>>gc.PageShift, spanAllocStack)
   201  		if s == nil {
   202  			throw("out of memory")
   203  		}
   204  		if s.allocCount != 0 {
   205  			throw("bad allocCount")
   206  		}
   207  		if s.manualFreeList.ptr() != nil {
   208  			throw("bad manualFreeList")
   209  		}
   210  		osStackAlloc(s)
   211  		s.elemsize = fixedStack << order
   212  		for i := uintptr(0); i < _StackCacheSize; i += s.elemsize {
   213  			x := gclinkptr(s.base() + i)
   214  			if valgrindenabled {
   215  				// The address of x.ptr() becomes the base of stacks. We need to
   216  				// mark it allocated here and in stackfree and stackpoolfree, and free'd in
   217  				// stackalloc in order to avoid overlapping allocations and
   218  				// uninitialized memory errors in valgrind.
   219  				valgrindMalloc(unsafe.Pointer(x.ptr()), unsafe.Sizeof(x.ptr()))
   220  			}
   221  			x.ptr().next = s.manualFreeList
   222  			s.manualFreeList = x
   223  		}
   224  		list.insert(s)
   225  	}
   226  	x := s.manualFreeList
   227  	if x.ptr() == nil {
   228  		throw("span has no free stacks")
   229  	}
   230  	s.manualFreeList = x.ptr().next
   231  	s.allocCount++
   232  	if s.manualFreeList.ptr() == nil {
   233  		// all stacks in s are allocated.
   234  		list.remove(s)
   235  	}
   236  	return x
   237  }
   238  
   239  // Adds stack x to the free pool. Must be called with stackpool[order].item.mu held.
   240  func stackpoolfree(x gclinkptr, order uint8) {
   241  	s := spanOfUnchecked(uintptr(x))
   242  	if s.state.get() != mSpanManual {
   243  		throw("freeing stack not in a stack span")
   244  	}
   245  	if s.manualFreeList.ptr() == nil {
   246  		// s will now have a free stack
   247  		stackpool[order].item.span.insert(s)
   248  	}
   249  	x.ptr().next = s.manualFreeList
   250  	s.manualFreeList = x
   251  	s.allocCount--
   252  	if gcphase == _GCoff && s.allocCount == 0 {
   253  		// Span is completely free. Return it to the heap
   254  		// immediately if we're sweeping.
   255  		//
   256  		// If GC is active, we delay the free until the end of
   257  		// GC to avoid the following type of situation:
   258  		//
   259  		// 1) GC starts, scans a SudoG but does not yet mark the SudoG.elem pointer
   260  		// 2) The stack that pointer points to is copied
   261  		// 3) The old stack is freed
   262  		// 4) The containing span is marked free
   263  		// 5) GC attempts to mark the SudoG.elem pointer. The
   264  		//    marking fails because the pointer looks like a
   265  		//    pointer into a free span.
   266  		//
   267  		// By not freeing, we prevent step #4 until GC is done.
   268  		stackpool[order].item.span.remove(s)
   269  		s.manualFreeList = 0
   270  		osStackFree(s)
   271  		mheap_.freeManual(s, spanAllocStack)
   272  	}
   273  }
   274  
   275  // stackcacherefill/stackcacherelease implement a global pool of stack segments.
   276  // The pool is required to prevent unlimited growth of per-thread caches.
   277  //
   278  //go:systemstack
   279  func stackcacherefill(c *mcache, order uint8) {
   280  	if stackDebug >= 1 {
   281  		print("stackcacherefill order=", order, "\n")
   282  	}
   283  
   284  	// Grab some stacks from the global cache.
   285  	// Grab half of the allowed capacity (to prevent thrashing).
   286  	var list gclinkptr
   287  	var size uintptr
   288  	lock(&stackpool[order].item.mu)
   289  	for size < _StackCacheSize/2 {
   290  		x := stackpoolalloc(order)
   291  		x.ptr().next = list
   292  		list = x
   293  		size += fixedStack << order
   294  	}
   295  	unlock(&stackpool[order].item.mu)
   296  	c.stackcache[order].list = list
   297  	c.stackcache[order].size = size
   298  }
   299  
   300  //go:systemstack
   301  func stackcacherelease(c *mcache, order uint8) {
   302  	if stackDebug >= 1 {
   303  		print("stackcacherelease order=", order, "\n")
   304  	}
   305  	x := c.stackcache[order].list
   306  	size := c.stackcache[order].size
   307  	lock(&stackpool[order].item.mu)
   308  	for size > _StackCacheSize/2 {
   309  		y := x.ptr().next
   310  		stackpoolfree(x, order)
   311  		x = y
   312  		size -= fixedStack << order
   313  	}
   314  	unlock(&stackpool[order].item.mu)
   315  	c.stackcache[order].list = x
   316  	c.stackcache[order].size = size
   317  }
   318  
   319  //go:systemstack
   320  func stackcache_clear(c *mcache) {
   321  	if stackDebug >= 1 {
   322  		print("stackcache clear\n")
   323  	}
   324  	for order := uint8(0); order < _NumStackOrders; order++ {
   325  		lock(&stackpool[order].item.mu)
   326  		x := c.stackcache[order].list
   327  		for x.ptr() != nil {
   328  			y := x.ptr().next
   329  			stackpoolfree(x, order)
   330  			x = y
   331  		}
   332  		c.stackcache[order].list = 0
   333  		c.stackcache[order].size = 0
   334  		unlock(&stackpool[order].item.mu)
   335  	}
   336  }
   337  
   338  // stackalloc allocates an n byte stack.
   339  //
   340  // stackalloc must run on the system stack because it uses per-P
   341  // resources and must not split the stack.
   342  //
   343  //go:systemstack
   344  func stackalloc(n uint32) stack {
   345  	// Stackalloc must be called on scheduler stack, so that we
   346  	// never try to grow the stack during the code that stackalloc runs.
   347  	// Doing so would cause a deadlock (issue 1547).
   348  	thisg := getg()
   349  	if thisg != thisg.m.g0 {
   350  		throw("stackalloc not on scheduler stack")
   351  	}
   352  	if n&(n-1) != 0 {
   353  		throw("stack size not a power of 2")
   354  	}
   355  	if stackDebug >= 1 {
   356  		print("stackalloc ", n, "\n")
   357  	}
   358  
   359  	if debug.efence != 0 || stackFromSystem != 0 {
   360  		n = uint32(alignUp(uintptr(n), physPageSize))
   361  		v := sysAlloc(uintptr(n), &memstats.stacks_sys, "goroutine stack (system)")
   362  		if v == nil {
   363  			throw("out of memory (stackalloc)")
   364  		}
   365  		return stack{uintptr(v), uintptr(v) + uintptr(n)}
   366  	}
   367  
   368  	// Small stacks are allocated with a fixed-size free-list allocator.
   369  	// If we need a stack of a bigger size, we fall back on allocating
   370  	// a dedicated span.
   371  	var v unsafe.Pointer
   372  	if n < fixedStack<<_NumStackOrders && n < _StackCacheSize {
   373  		order := uint8(0)
   374  		n2 := n
   375  		for n2 > fixedStack {
   376  			order++
   377  			n2 >>= 1
   378  		}
   379  		var x gclinkptr
   380  		if stackNoCache != 0 || thisg.m.p == 0 || thisg.m.preemptoff != "" {
   381  			// thisg.m.p == 0 can happen in the guts of exitsyscall
   382  			// or procresize. Just get a stack from the global pool.
   383  			// Also don't touch stackcache during gc
   384  			// as it's flushed concurrently.
   385  			lock(&stackpool[order].item.mu)
   386  			x = stackpoolalloc(order)
   387  			unlock(&stackpool[order].item.mu)
   388  		} else {
   389  			c := thisg.m.p.ptr().mcache
   390  			x = c.stackcache[order].list
   391  			if x.ptr() == nil {
   392  				stackcacherefill(c, order)
   393  				x = c.stackcache[order].list
   394  			}
   395  			c.stackcache[order].list = x.ptr().next
   396  			c.stackcache[order].size -= uintptr(n)
   397  		}
   398  		if valgrindenabled {
   399  			// We're about to allocate the stack region starting at x.ptr().
   400  			// To prevent valgrind from complaining about overlapping allocations,
   401  			// we need to mark the (previously allocated) memory as free'd.
   402  			valgrindFree(unsafe.Pointer(x.ptr()))
   403  		}
   404  		v = unsafe.Pointer(x)
   405  	} else {
   406  		var s *mspan
   407  		npage := uintptr(n) >> gc.PageShift
   408  		log2npage := stacklog2(npage)
   409  
   410  		// Try to get a stack from the large stack cache.
   411  		lock(&stackLarge.lock)
   412  		if !stackLarge.free[log2npage].isEmpty() {
   413  			s = stackLarge.free[log2npage].first
   414  			stackLarge.free[log2npage].remove(s)
   415  		}
   416  		unlock(&stackLarge.lock)
   417  
   418  		lockWithRankMayAcquire(&mheap_.lock, lockRankMheap)
   419  
   420  		if s == nil {
   421  			// Allocate a new stack from the heap.
   422  			s = mheap_.allocManual(npage, spanAllocStack)
   423  			if s == nil {
   424  				throw("out of memory")
   425  			}
   426  			osStackAlloc(s)
   427  			s.elemsize = uintptr(n)
   428  		}
   429  		v = unsafe.Pointer(s.base())
   430  	}
   431  
   432  	if traceAllocFreeEnabled() {
   433  		trace := traceAcquire()
   434  		if trace.ok() {
   435  			trace.GoroutineStackAlloc(uintptr(v), uintptr(n))
   436  			traceRelease(trace)
   437  		}
   438  	}
   439  	if raceenabled {
   440  		racemalloc(v, uintptr(n))
   441  	}
   442  	if msanenabled {
   443  		msanmalloc(v, uintptr(n))
   444  	}
   445  	if asanenabled {
   446  		asanunpoison(v, uintptr(n))
   447  	}
   448  	if valgrindenabled {
   449  		valgrindMalloc(v, uintptr(n))
   450  	}
   451  	if stackDebug >= 1 {
   452  		print("  allocated ", v, "\n")
   453  	}
   454  	return stack{uintptr(v), uintptr(v) + uintptr(n)}
   455  }
   456  
   457  // stackfree frees an n byte stack allocation at stk.
   458  //
   459  // stackfree must run on the system stack because it uses per-P
   460  // resources and must not split the stack.
   461  //
   462  //go:systemstack
   463  func stackfree(stk stack) {
   464  	gp := getg()
   465  	v := unsafe.Pointer(stk.lo)
   466  	n := stk.hi - stk.lo
   467  	if n&(n-1) != 0 {
   468  		throw("stack not a power of 2")
   469  	}
   470  	if stk.lo+n < stk.hi {
   471  		throw("bad stack size")
   472  	}
   473  	if stackDebug >= 1 {
   474  		println("stackfree", v, n)
   475  		memclrNoHeapPointers(v, n) // for testing, clobber stack data
   476  	}
   477  	if debug.efence != 0 || stackFromSystem != 0 {
   478  		if debug.efence != 0 || stackFaultOnFree != 0 {
   479  			sysFault(v, n)
   480  		} else {
   481  			sysFree(v, n, &memstats.stacks_sys)
   482  		}
   483  		return
   484  	}
   485  	if traceAllocFreeEnabled() {
   486  		trace := traceAcquire()
   487  		if trace.ok() {
   488  			trace.GoroutineStackFree(uintptr(v))
   489  			traceRelease(trace)
   490  		}
   491  	}
   492  	if msanenabled {
   493  		msanfree(v, n)
   494  	}
   495  	if asanenabled {
   496  		asanpoison(v, n)
   497  	}
   498  	if valgrindenabled {
   499  		valgrindFree(v)
   500  	}
   501  	if n < fixedStack<<_NumStackOrders && n < _StackCacheSize {
   502  		order := uint8(0)
   503  		n2 := n
   504  		for n2 > fixedStack {
   505  			order++
   506  			n2 >>= 1
   507  		}
   508  		x := gclinkptr(v)
   509  		if stackNoCache != 0 || gp.m.p == 0 || gp.m.preemptoff != "" {
   510  			lock(&stackpool[order].item.mu)
   511  			if valgrindenabled {
   512  				// x.ptr() is the head of the list of free stacks, and will be used
   513  				// when allocating a new stack, so it has to be marked allocated.
   514  				valgrindMalloc(unsafe.Pointer(x.ptr()), unsafe.Sizeof(x.ptr()))
   515  			}
   516  			stackpoolfree(x, order)
   517  			unlock(&stackpool[order].item.mu)
   518  		} else {
   519  			c := gp.m.p.ptr().mcache
   520  			if c.stackcache[order].size >= _StackCacheSize {
   521  				stackcacherelease(c, order)
   522  			}
   523  			if valgrindenabled {
   524  				// x.ptr() is the head of the list of free stacks, and will
   525  				// be used when allocating a new stack, so it has to be
   526  				// marked allocated.
   527  				valgrindMalloc(unsafe.Pointer(x.ptr()), unsafe.Sizeof(x.ptr()))
   528  			}
   529  			x.ptr().next = c.stackcache[order].list
   530  			c.stackcache[order].list = x
   531  			c.stackcache[order].size += n
   532  		}
   533  	} else {
   534  		s := spanOfUnchecked(uintptr(v))
   535  		if s.state.get() != mSpanManual {
   536  			println(hex(s.base()), v)
   537  			throw("bad span state")
   538  		}
   539  		if gcphase == _GCoff {
   540  			// Free the stack immediately if we're
   541  			// sweeping.
   542  			osStackFree(s)
   543  			mheap_.freeManual(s, spanAllocStack)
   544  		} else {
   545  			// If the GC is running, we can't return a
   546  			// stack span to the heap because it could be
   547  			// reused as a heap span, and this state
   548  			// change would race with GC. Add it to the
   549  			// large stack cache instead.
   550  			log2npage := stacklog2(s.npages)
   551  			lock(&stackLarge.lock)
   552  			stackLarge.free[log2npage].insert(s)
   553  			unlock(&stackLarge.lock)
   554  		}
   555  	}
   556  }
   557  
   558  var maxstacksize uintptr = 1 << 20 // enough until runtime.main sets it for real
   559  
   560  var maxstackceiling = maxstacksize
   561  
   562  var ptrnames = []string{
   563  	0: "scalar",
   564  	1: "ptr",
   565  }
   566  
   567  // Stack frame layout
   568  //
   569  // (x86)
   570  // +------------------+
   571  // | args from caller |
   572  // +------------------+ <- frame->argp
   573  // |  return address  |
   574  // +------------------+
   575  // |  caller's BP (*) | (*) if framepointer_enabled && varp > sp
   576  // +------------------+ <- frame->varp
   577  // |     locals       |
   578  // +------------------+
   579  // |  args to callee  |
   580  // +------------------+ <- frame->sp
   581  //
   582  // (arm)
   583  // +------------------+
   584  // | args from caller |
   585  // +------------------+ <- frame->argp
   586  // | caller's retaddr |
   587  // +------------------+
   588  // |  caller's FP (*) | (*) on ARM64, if framepointer_enabled && varp > sp
   589  // +------------------+ <- frame->varp
   590  // |     locals       |
   591  // +------------------+
   592  // |  args to callee  |
   593  // +------------------+
   594  // |  return address  |
   595  // +------------------+ <- frame->sp
   596  //
   597  // varp > sp means that the function has a frame;
   598  // varp == sp means frameless function.
   599  
   600  type adjustinfo struct {
   601  	old   stack
   602  	delta uintptr // ptr distance from old to new stack (newbase - oldbase)
   603  
   604  	// sghi is the highest sudog.elem on the stack.
   605  	sghi uintptr
   606  }
   607  
   608  // adjustpointer checks whether *vpp is in the old stack described by adjinfo.
   609  // If so, it rewrites *vpp to point into the new stack.
   610  func adjustpointer(adjinfo *adjustinfo, vpp unsafe.Pointer) {
   611  	pp := (*uintptr)(vpp)
   612  	p := *pp
   613  	if stackDebug >= 4 {
   614  		print("        ", pp, ":", hex(p), "\n")
   615  	}
   616  	if valgrindenabled {
   617  		// p is a pointer on a stack, it is inherently initialized, as
   618  		// everything on the stack is, but valgrind for _some unknown reason_
   619  		// sometimes thinks it's uninitialized, and flags operations on p below
   620  		// as uninitialized. We just initialize it if valgrind thinks its
   621  		// uninitialized.
   622  		//
   623  		// See go.dev/issues/73801.
   624  		valgrindMakeMemDefined(unsafe.Pointer(&p), unsafe.Sizeof(&p))
   625  	}
   626  	if adjinfo.old.lo <= p && p < adjinfo.old.hi {
   627  		*pp = p + adjinfo.delta
   628  		if stackDebug >= 3 {
   629  			print("        adjust ptr ", pp, ":", hex(p), " -> ", hex(*pp), "\n")
   630  		}
   631  	}
   632  }
   633  
   634  // Information from the compiler about the layout of stack frames.
   635  // Note: this type must agree with reflect.bitVector.
   636  type bitvector struct {
   637  	n        int32 // # of bits
   638  	bytedata *uint8
   639  }
   640  
   641  // ptrbit returns the i'th bit in bv.
   642  // ptrbit is less efficient than iterating directly over bitvector bits,
   643  // and should only be used in non-performance-critical code.
   644  // See adjustpointers for an example of a high-efficiency walk of a bitvector.
   645  func (bv *bitvector) ptrbit(i uintptr) uint8 {
   646  	b := *(addb(bv.bytedata, i/8))
   647  	return (b >> (i % 8)) & 1
   648  }
   649  
   650  // bv describes the memory starting at address scanp.
   651  // Adjust any pointers contained therein.
   652  func adjustpointers(scanp unsafe.Pointer, bv *bitvector, adjinfo *adjustinfo, f funcInfo) {
   653  	minp := adjinfo.old.lo
   654  	maxp := adjinfo.old.hi
   655  	delta := adjinfo.delta
   656  	num := uintptr(bv.n)
   657  	// If this frame might contain channel receive slots, use CAS
   658  	// to adjust pointers. If the slot hasn't been received into
   659  	// yet, it may contain stack pointers and a concurrent send
   660  	// could race with adjusting those pointers. (The sent value
   661  	// itself can never contain stack pointers.)
   662  	useCAS := uintptr(scanp) < adjinfo.sghi
   663  	for i := uintptr(0); i < num; i += 8 {
   664  		if stackDebug >= 4 {
   665  			for j := uintptr(0); j < 8; j++ {
   666  				print("        ", add(scanp, (i+j)*goarch.PtrSize), ":", ptrnames[bv.ptrbit(i+j)], ":", hex(*(*uintptr)(add(scanp, (i+j)*goarch.PtrSize))), " # ", i, " ", *addb(bv.bytedata, i/8), "\n")
   667  			}
   668  		}
   669  		b := *(addb(bv.bytedata, i/8))
   670  		for b != 0 {
   671  			j := uintptr(sys.TrailingZeros8(b))
   672  			b &= b - 1
   673  			pp := (*uintptr)(add(scanp, (i+j)*goarch.PtrSize))
   674  		retry:
   675  			p := *pp
   676  			if f.valid() && 0 < p && p < minLegalPointer && debug.invalidptr != 0 {
   677  				// Looks like a junk value in a pointer slot.
   678  				// Live analysis wrong?
   679  				getg().m.traceback = 2
   680  				print("runtime: bad pointer in frame ", funcname(f), " at ", pp, ": ", hex(p), "\n")
   681  				throw("invalid pointer found on stack")
   682  			}
   683  			if minp <= p && p < maxp {
   684  				if stackDebug >= 3 {
   685  					print("adjust ptr ", hex(p), " ", funcname(f), "\n")
   686  				}
   687  				if useCAS {
   688  					ppu := (*unsafe.Pointer)(unsafe.Pointer(pp))
   689  					if !atomic.Casp1(ppu, unsafe.Pointer(p), unsafe.Pointer(p+delta)) {
   690  						goto retry
   691  					}
   692  				} else {
   693  					*pp = p + delta
   694  				}
   695  			}
   696  		}
   697  	}
   698  }
   699  
   700  // Note: the argument/return area is adjusted by the callee.
   701  func adjustframe(frame *stkframe, adjinfo *adjustinfo) {
   702  	if frame.continpc == 0 {
   703  		// Frame is dead.
   704  		return
   705  	}
   706  	f := frame.fn
   707  	if stackDebug >= 2 {
   708  		print("    adjusting ", funcname(f), " frame=[", hex(frame.sp), ",", hex(frame.fp), "] pc=", hex(frame.pc), " continpc=", hex(frame.continpc), "\n")
   709  	}
   710  
   711  	// Adjust saved frame pointer if there is one.
   712  	if (goarch.ArchFamily == goarch.AMD64 || goarch.ArchFamily == goarch.ARM64) && frame.argp-frame.varp == 2*goarch.PtrSize {
   713  		if stackDebug >= 3 {
   714  			print("      saved bp\n")
   715  		}
   716  		if debugCheckBP {
   717  			// Frame pointers should always point to the next higher frame on
   718  			// the Go stack (or be nil, for the top frame on the stack).
   719  			bp := *(*uintptr)(unsafe.Pointer(frame.varp))
   720  			if bp != 0 && (bp < adjinfo.old.lo || bp >= adjinfo.old.hi) {
   721  				println("runtime: found invalid frame pointer")
   722  				print("bp=", hex(bp), " min=", hex(adjinfo.old.lo), " max=", hex(adjinfo.old.hi), "\n")
   723  				throw("bad frame pointer")
   724  			}
   725  		}
   726  		// On AMD64, this is the caller's frame pointer saved in the current
   727  		// frame.
   728  		// On ARM64, this is the frame pointer of the caller's caller saved
   729  		// by the caller in its frame (one word below its SP).
   730  		adjustpointer(adjinfo, unsafe.Pointer(frame.varp))
   731  	}
   732  
   733  	locals, args, objs := frame.getStackMap(true)
   734  
   735  	// Adjust local variables if stack frame has been allocated.
   736  	if locals.n > 0 {
   737  		size := uintptr(locals.n) * goarch.PtrSize
   738  		adjustpointers(unsafe.Pointer(frame.varp-size), &locals, adjinfo, f)
   739  	}
   740  
   741  	// Adjust arguments.
   742  	if args.n > 0 {
   743  		if stackDebug >= 3 {
   744  			print("      args\n")
   745  		}
   746  		adjustpointers(unsafe.Pointer(frame.argp), &args, adjinfo, funcInfo{})
   747  	}
   748  
   749  	// Adjust pointers in all stack objects (whether they are live or not).
   750  	// See comments in mgcmark.go:scanframeworker.
   751  	if frame.varp != 0 {
   752  		for i := range objs {
   753  			obj := &objs[i]
   754  			off := obj.off
   755  			base := frame.varp // locals base pointer
   756  			if off >= 0 {
   757  				base = frame.argp // arguments and return values base pointer
   758  			}
   759  			p := base + uintptr(off)
   760  			if p < frame.sp {
   761  				// Object hasn't been allocated in the frame yet.
   762  				// (Happens when the stack bounds check fails and
   763  				// we call into morestack.)
   764  				continue
   765  			}
   766  			ptrBytes, gcData := obj.gcdata()
   767  			for i := uintptr(0); i < ptrBytes; i += goarch.PtrSize {
   768  				if *addb(gcData, i/(8*goarch.PtrSize))>>(i/goarch.PtrSize&7)&1 != 0 {
   769  					adjustpointer(adjinfo, unsafe.Pointer(p+i))
   770  				}
   771  			}
   772  		}
   773  	}
   774  }
   775  
   776  func adjustctxt(gp *g, adjinfo *adjustinfo) {
   777  	adjustpointer(adjinfo, unsafe.Pointer(&gp.sched.ctxt))
   778  	if !framepointer_enabled {
   779  		return
   780  	}
   781  	if debugCheckBP {
   782  		bp := gp.sched.bp
   783  		if bp != 0 && (bp < adjinfo.old.lo || bp >= adjinfo.old.hi) {
   784  			println("runtime: found invalid top frame pointer")
   785  			print("bp=", hex(bp), " min=", hex(adjinfo.old.lo), " max=", hex(adjinfo.old.hi), "\n")
   786  			throw("bad top frame pointer")
   787  		}
   788  	}
   789  	oldfp := gp.sched.bp
   790  	adjustpointer(adjinfo, unsafe.Pointer(&gp.sched.bp))
   791  	if GOARCH == "arm64" {
   792  		// On ARM64, the frame pointer is saved one word *below* the SP,
   793  		// which is not copied or adjusted in any frame. Do it explicitly
   794  		// here.
   795  		if oldfp == gp.sched.sp-goarch.PtrSize {
   796  			memmove(unsafe.Pointer(gp.sched.bp), unsafe.Pointer(oldfp), goarch.PtrSize)
   797  			adjustpointer(adjinfo, unsafe.Pointer(gp.sched.bp))
   798  		}
   799  	}
   800  }
   801  
   802  func adjustdefers(gp *g, adjinfo *adjustinfo) {
   803  	// Adjust pointers in the Defer structs.
   804  	// We need to do this first because we need to adjust the
   805  	// defer.link fields so we always work on the new stack.
   806  	adjustpointer(adjinfo, unsafe.Pointer(&gp._defer))
   807  	for d := gp._defer; d != nil; d = d.link {
   808  		adjustpointer(adjinfo, unsafe.Pointer(&d.fn))
   809  		adjustpointer(adjinfo, unsafe.Pointer(&d.sp))
   810  		adjustpointer(adjinfo, unsafe.Pointer(&d.link))
   811  	}
   812  }
   813  
   814  func adjustpanics(gp *g, adjinfo *adjustinfo) {
   815  	// Panics are on stack and already adjusted.
   816  	// Update pointer to head of list in G.
   817  	adjustpointer(adjinfo, unsafe.Pointer(&gp._panic))
   818  }
   819  
   820  func adjustsudogs(gp *g, adjinfo *adjustinfo) {
   821  	// the data elements pointed to by a SudoG structure
   822  	// might be in the stack.
   823  	for s := gp.waiting; s != nil; s = s.waitlink {
   824  		adjustpointer(adjinfo, unsafe.Pointer(&s.elem))
   825  	}
   826  }
   827  
   828  func fillstack(stk stack, b byte) {
   829  	for p := stk.lo; p < stk.hi; p++ {
   830  		*(*byte)(unsafe.Pointer(p)) = b
   831  	}
   832  }
   833  
   834  func findsghi(gp *g, stk stack) uintptr {
   835  	var sghi uintptr
   836  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
   837  		p := uintptr(sg.elem) + uintptr(sg.c.elemsize)
   838  		if stk.lo <= p && p < stk.hi && p > sghi {
   839  			sghi = p
   840  		}
   841  	}
   842  	return sghi
   843  }
   844  
   845  // syncadjustsudogs adjusts gp's sudogs and copies the part of gp's
   846  // stack they refer to while synchronizing with concurrent channel
   847  // operations. It returns the number of bytes of stack copied.
   848  func syncadjustsudogs(gp *g, used uintptr, adjinfo *adjustinfo) uintptr {
   849  	if gp.waiting == nil {
   850  		return 0
   851  	}
   852  
   853  	// Lock channels to prevent concurrent send/receive.
   854  	var lastc *hchan
   855  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
   856  		if sg.c != lastc {
   857  			// There is a ranking cycle here between gscan bit and
   858  			// hchan locks. Normally, we only allow acquiring hchan
   859  			// locks and then getting a gscan bit. In this case, we
   860  			// already have the gscan bit. We allow acquiring hchan
   861  			// locks here as a special case, since a deadlock can't
   862  			// happen because the G involved must already be
   863  			// suspended. So, we get a special hchan lock rank here
   864  			// that is lower than gscan, but doesn't allow acquiring
   865  			// any other locks other than hchan.
   866  			lockWithRank(&sg.c.lock, lockRankHchanLeaf)
   867  		}
   868  		lastc = sg.c
   869  	}
   870  
   871  	// Adjust sudogs.
   872  	adjustsudogs(gp, adjinfo)
   873  
   874  	// Copy the part of the stack the sudogs point in to
   875  	// while holding the lock to prevent races on
   876  	// send/receive slots.
   877  	var sgsize uintptr
   878  	if adjinfo.sghi != 0 {
   879  		oldBot := adjinfo.old.hi - used
   880  		newBot := oldBot + adjinfo.delta
   881  		sgsize = adjinfo.sghi - oldBot
   882  		memmove(unsafe.Pointer(newBot), unsafe.Pointer(oldBot), sgsize)
   883  	}
   884  
   885  	// Unlock channels.
   886  	lastc = nil
   887  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
   888  		if sg.c != lastc {
   889  			unlock(&sg.c.lock)
   890  		}
   891  		lastc = sg.c
   892  	}
   893  
   894  	return sgsize
   895  }
   896  
   897  // Copies gp's stack to a new stack of a different size.
   898  // Caller must have changed gp status to Gcopystack.
   899  func copystack(gp *g, newsize uintptr) {
   900  	if gp.syscallsp != 0 {
   901  		throw("stack growth not allowed in system call")
   902  	}
   903  	old := gp.stack
   904  	if old.lo == 0 {
   905  		throw("nil stackbase")
   906  	}
   907  	used := old.hi - gp.sched.sp
   908  	// Add just the difference to gcController.addScannableStack.
   909  	// g0 stacks never move, so this will never account for them.
   910  	// It's also fine if we have no P, addScannableStack can deal with
   911  	// that case.
   912  	gcController.addScannableStack(getg().m.p.ptr(), int64(newsize)-int64(old.hi-old.lo))
   913  
   914  	// allocate new stack
   915  	new := stackalloc(uint32(newsize))
   916  	if stackPoisonCopy != 0 {
   917  		fillstack(new, 0xfd)
   918  	}
   919  	if stackDebug >= 1 {
   920  		print("copystack gp=", gp, " [", hex(old.lo), " ", hex(old.hi-used), " ", hex(old.hi), "]", " -> [", hex(new.lo), " ", hex(new.hi-used), " ", hex(new.hi), "]/", newsize, "\n")
   921  	}
   922  
   923  	// Compute adjustment.
   924  	var adjinfo adjustinfo
   925  	adjinfo.old = old
   926  	adjinfo.delta = new.hi - old.hi
   927  
   928  	// Adjust sudogs, synchronizing with channel ops if necessary.
   929  	ncopy := used
   930  	if !gp.activeStackChans {
   931  		if newsize < old.hi-old.lo && gp.parkingOnChan.Load() {
   932  			// It's not safe for someone to shrink this stack while we're actively
   933  			// parking on a channel, but it is safe to grow since we do that
   934  			// ourselves and explicitly don't want to synchronize with channels
   935  			// since we could self-deadlock.
   936  			throw("racy sudog adjustment due to parking on channel")
   937  		}
   938  		adjustsudogs(gp, &adjinfo)
   939  	} else {
   940  		// sudogs may be pointing in to the stack and gp has
   941  		// released channel locks, so other goroutines could
   942  		// be writing to gp's stack. Find the highest such
   943  		// pointer so we can handle everything there and below
   944  		// carefully. (This shouldn't be far from the bottom
   945  		// of the stack, so there's little cost in handling
   946  		// everything below it carefully.)
   947  		adjinfo.sghi = findsghi(gp, old)
   948  
   949  		// Synchronize with channel ops and copy the part of
   950  		// the stack they may interact with.
   951  		ncopy -= syncadjustsudogs(gp, used, &adjinfo)
   952  	}
   953  
   954  	// Copy the stack (or the rest of it) to the new location
   955  	memmove(unsafe.Pointer(new.hi-ncopy), unsafe.Pointer(old.hi-ncopy), ncopy)
   956  
   957  	// Adjust remaining structures that have pointers into stacks.
   958  	// We have to do most of these before we traceback the new
   959  	// stack because gentraceback uses them.
   960  	adjustctxt(gp, &adjinfo)
   961  	adjustdefers(gp, &adjinfo)
   962  	adjustpanics(gp, &adjinfo)
   963  	if adjinfo.sghi != 0 {
   964  		adjinfo.sghi += adjinfo.delta
   965  	}
   966  
   967  	// Swap out old stack for new one
   968  	gp.stack = new
   969  	gp.stackguard0 = new.lo + stackGuard // NOTE: might clobber a preempt request
   970  	gp.sched.sp = new.hi - used
   971  	gp.stktopsp += adjinfo.delta
   972  
   973  	// Adjust pointers in the new stack.
   974  	var u unwinder
   975  	for u.init(gp, 0); u.valid(); u.next() {
   976  		adjustframe(&u.frame, &adjinfo)
   977  	}
   978  
   979  	if valgrindenabled {
   980  		if gp.valgrindStackID == 0 {
   981  			gp.valgrindStackID = valgrindRegisterStack(unsafe.Pointer(new.lo), unsafe.Pointer(new.hi))
   982  		} else {
   983  			valgrindChangeStack(gp.valgrindStackID, unsafe.Pointer(new.lo), unsafe.Pointer(new.hi))
   984  		}
   985  	}
   986  
   987  	// free old stack
   988  	if stackPoisonCopy != 0 {
   989  		fillstack(old, 0xfc)
   990  	}
   991  	stackfree(old)
   992  }
   993  
   994  // round x up to a power of 2.
   995  func round2(x int32) int32 {
   996  	s := uint(0)
   997  	for 1<<s < x {
   998  		s++
   999  	}
  1000  	return 1 << s
  1001  }
  1002  
  1003  // Called from runtime·morestack when more stack is needed.
  1004  // Allocate larger stack and relocate to new stack.
  1005  // Stack growth is multiplicative, for constant amortized cost.
  1006  //
  1007  // g->atomicstatus will be Grunning or Gscanrunning upon entry.
  1008  // If the scheduler is trying to stop this g, then it will set preemptStop.
  1009  //
  1010  // This must be nowritebarrierrec because it can be called as part of
  1011  // stack growth from other nowritebarrierrec functions, but the
  1012  // compiler doesn't check this.
  1013  //
  1014  //go:nowritebarrierrec
  1015  func newstack() {
  1016  	thisg := getg()
  1017  	// TODO: double check all gp. shouldn't be getg().
  1018  	if thisg.m.morebuf.g.ptr().stackguard0 == stackFork {
  1019  		throw("stack growth after fork")
  1020  	}
  1021  	if thisg.m.morebuf.g.ptr() != thisg.m.curg {
  1022  		print("runtime: newstack called from g=", hex(thisg.m.morebuf.g), "\n"+"\tm=", thisg.m, " m->curg=", thisg.m.curg, " m->g0=", thisg.m.g0, " m->gsignal=", thisg.m.gsignal, "\n")
  1023  		morebuf := thisg.m.morebuf
  1024  		traceback(morebuf.pc, morebuf.sp, morebuf.lr, morebuf.g.ptr())
  1025  		throw("runtime: wrong goroutine in newstack")
  1026  	}
  1027  
  1028  	gp := thisg.m.curg
  1029  
  1030  	if thisg.m.curg.throwsplit {
  1031  		// Update syscallsp, syscallpc in case traceback uses them.
  1032  		morebuf := thisg.m.morebuf
  1033  		gp.syscallsp = morebuf.sp
  1034  		gp.syscallpc = morebuf.pc
  1035  		pcname, pcoff := "(unknown)", uintptr(0)
  1036  		f := findfunc(gp.sched.pc)
  1037  		if f.valid() {
  1038  			pcname = funcname(f)
  1039  			pcoff = gp.sched.pc - f.entry()
  1040  		}
  1041  		print("runtime: newstack at ", pcname, "+", hex(pcoff),
  1042  			" sp=", hex(gp.sched.sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n",
  1043  			"\tmorebuf={pc:", hex(morebuf.pc), " sp:", hex(morebuf.sp), " lr:", hex(morebuf.lr), "}\n",
  1044  			"\tsched={pc:", hex(gp.sched.pc), " sp:", hex(gp.sched.sp), " lr:", hex(gp.sched.lr), " ctxt:", gp.sched.ctxt, "}\n")
  1045  
  1046  		thisg.m.traceback = 2 // Include runtime frames
  1047  		traceback(morebuf.pc, morebuf.sp, morebuf.lr, gp)
  1048  		throw("runtime: stack split at bad time")
  1049  	}
  1050  
  1051  	morebuf := thisg.m.morebuf
  1052  	thisg.m.morebuf.pc = 0
  1053  	thisg.m.morebuf.lr = 0
  1054  	thisg.m.morebuf.sp = 0
  1055  	thisg.m.morebuf.g = 0
  1056  
  1057  	// NOTE: stackguard0 may change underfoot, if another thread
  1058  	// is about to try to preempt gp. Read it just once and use that same
  1059  	// value now and below.
  1060  	stackguard0 := atomic.Loaduintptr(&gp.stackguard0)
  1061  
  1062  	// Be conservative about where we preempt.
  1063  	// We are interested in preempting user Go code, not runtime code.
  1064  	// If we're holding locks, mallocing, or preemption is disabled, don't
  1065  	// preempt.
  1066  	// This check is very early in newstack so that even the status change
  1067  	// from Grunning to Gwaiting and back doesn't happen in this case.
  1068  	// That status change by itself can be viewed as a small preemption,
  1069  	// because the GC might change Gwaiting to Gscanwaiting, and then
  1070  	// this goroutine has to wait for the GC to finish before continuing.
  1071  	// If the GC is in some way dependent on this goroutine (for example,
  1072  	// it needs a lock held by the goroutine), that small preemption turns
  1073  	// into a real deadlock.
  1074  	preempt := stackguard0 == stackPreempt
  1075  	if preempt {
  1076  		if !canPreemptM(thisg.m) {
  1077  			// Let the goroutine keep running for now.
  1078  			// gp->preempt is set, so it will be preempted next time.
  1079  			gp.stackguard0 = gp.stack.lo + stackGuard
  1080  			gogo(&gp.sched) // never return
  1081  		}
  1082  	}
  1083  
  1084  	if gp.stack.lo == 0 {
  1085  		throw("missing stack in newstack")
  1086  	}
  1087  	sp := gp.sched.sp
  1088  	if goarch.ArchFamily == goarch.AMD64 || goarch.ArchFamily == goarch.I386 || goarch.ArchFamily == goarch.WASM {
  1089  		// The call to morestack cost a word.
  1090  		sp -= goarch.PtrSize
  1091  	}
  1092  	if stackDebug >= 1 || sp < gp.stack.lo {
  1093  		print("runtime: newstack sp=", hex(sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n",
  1094  			"\tmorebuf={pc:", hex(morebuf.pc), " sp:", hex(morebuf.sp), " lr:", hex(morebuf.lr), "}\n",
  1095  			"\tsched={pc:", hex(gp.sched.pc), " sp:", hex(gp.sched.sp), " lr:", hex(gp.sched.lr), " ctxt:", gp.sched.ctxt, "}\n")
  1096  	}
  1097  	if sp < gp.stack.lo {
  1098  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->status=", hex(readgstatus(gp)), "\n ")
  1099  		print("runtime: split stack overflow: ", hex(sp), " < ", hex(gp.stack.lo), "\n")
  1100  		throw("runtime: split stack overflow")
  1101  	}
  1102  
  1103  	if preempt {
  1104  		if gp == thisg.m.g0 {
  1105  			throw("runtime: preempt g0")
  1106  		}
  1107  		if thisg.m.p == 0 && thisg.m.locks == 0 {
  1108  			throw("runtime: g is running but p is not")
  1109  		}
  1110  
  1111  		if gp.preemptShrink {
  1112  			// We're at a synchronous safe point now, so
  1113  			// do the pending stack shrink.
  1114  			gp.preemptShrink = false
  1115  			shrinkstack(gp)
  1116  		}
  1117  
  1118  		// Set a flag indicated that we've been synchronously preempted.
  1119  		gp.syncSafePoint = true
  1120  
  1121  		if gp.preemptStop {
  1122  			preemptPark(gp) // never returns
  1123  		}
  1124  
  1125  		// Act like goroutine called runtime.Gosched.
  1126  		gopreempt_m(gp) // never return
  1127  	}
  1128  
  1129  	// Allocate a bigger segment and move the stack.
  1130  	oldsize := gp.stack.hi - gp.stack.lo
  1131  	newsize := oldsize * 2
  1132  
  1133  	// Make sure we grow at least as much as needed to fit the new frame.
  1134  	// (This is just an optimization - the caller of morestack will
  1135  	// recheck the bounds on return.)
  1136  	if f := findfunc(gp.sched.pc); f.valid() {
  1137  		max := uintptr(funcMaxSPDelta(f))
  1138  		needed := max + stackGuard
  1139  		used := gp.stack.hi - gp.sched.sp
  1140  		for newsize-used < needed {
  1141  			newsize *= 2
  1142  		}
  1143  	}
  1144  
  1145  	if stackguard0 == stackForceMove {
  1146  		// Forced stack movement used for debugging.
  1147  		// Don't double the stack (or we may quickly run out
  1148  		// if this is done repeatedly).
  1149  		newsize = oldsize
  1150  	}
  1151  
  1152  	if newsize > maxstacksize || newsize > maxstackceiling {
  1153  		if maxstacksize < maxstackceiling {
  1154  			print("runtime: goroutine stack exceeds ", maxstacksize, "-byte limit\n")
  1155  		} else {
  1156  			print("runtime: goroutine stack exceeds ", maxstackceiling, "-byte limit\n")
  1157  		}
  1158  		print("runtime: sp=", hex(sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n")
  1159  		throw("stack overflow")
  1160  	}
  1161  
  1162  	// The goroutine must be executing in order to call newstack,
  1163  	// so it must be Grunning (or Gscanrunning).
  1164  	casgstatus(gp, _Grunning, _Gcopystack)
  1165  
  1166  	// The concurrent GC will not scan the stack while we are doing the copy since
  1167  	// the gp is in a Gcopystack status.
  1168  	copystack(gp, newsize)
  1169  	if stackDebug >= 1 {
  1170  		print("stack grow done\n")
  1171  	}
  1172  	casgstatus(gp, _Gcopystack, _Grunning)
  1173  	gogo(&gp.sched)
  1174  }
  1175  
  1176  //go:nosplit
  1177  func nilfunc() {
  1178  	*(*uint8)(nil) = 0
  1179  }
  1180  
  1181  // adjust Gobuf as if it executed a call to fn
  1182  // and then stopped before the first instruction in fn.
  1183  func gostartcallfn(gobuf *gobuf, fv *funcval) {
  1184  	var fn unsafe.Pointer
  1185  	if fv != nil {
  1186  		fn = unsafe.Pointer(fv.fn)
  1187  	} else {
  1188  		fn = unsafe.Pointer(abi.FuncPCABIInternal(nilfunc))
  1189  	}
  1190  	gostartcall(gobuf, fn, unsafe.Pointer(fv))
  1191  }
  1192  
  1193  // isShrinkStackSafe returns whether it's safe to attempt to shrink
  1194  // gp's stack. Shrinking the stack is only safe when we have precise
  1195  // pointer maps for all frames on the stack. The caller must hold the
  1196  // _Gscan bit for gp or must be running gp itself.
  1197  func isShrinkStackSafe(gp *g) bool {
  1198  	// We can't copy the stack if we're in a syscall.
  1199  	// The syscall might have pointers into the stack and
  1200  	// often we don't have precise pointer maps for the innermost
  1201  	// frames.
  1202  	if gp.syscallsp != 0 {
  1203  		return false
  1204  	}
  1205  	// We also can't copy the stack if we're at an asynchronous
  1206  	// safe-point because we don't have precise pointer maps for
  1207  	// all frames.
  1208  	if gp.asyncSafePoint {
  1209  		return false
  1210  	}
  1211  	// We also can't *shrink* the stack in the window between the
  1212  	// goroutine calling gopark to park on a channel and
  1213  	// gp.activeStackChans being set.
  1214  	if gp.parkingOnChan.Load() {
  1215  		return false
  1216  	}
  1217  	// We also can't copy the stack while a gp is in _Gwaiting solely
  1218  	// to make itself available to suspendG.
  1219  	//
  1220  	// In these cases, the G is actually executing on the system
  1221  	// stack, and the execution tracer, mutex profiler, etc. may want
  1222  	// to take a stack trace of the G's stack.
  1223  	//
  1224  	// Note: it's safe to access gp.waitreason here.
  1225  	// We're only calling isShrinkStackSafe if we took ownership of the
  1226  	// G with the _Gscan bit. This prevents the goroutine from transitioning,
  1227  	// which prevents gp.waitreason from changing.
  1228  	if readgstatus(gp)&^_Gscan == _Gwaiting && gp.waitreason.isWaitingForSuspendG() {
  1229  		return false
  1230  	}
  1231  	return true
  1232  }
  1233  
  1234  // Maybe shrink the stack being used by gp.
  1235  //
  1236  // gp must be stopped and we must own its stack. It may be in
  1237  // _Grunning, but only if this is our own user G.
  1238  func shrinkstack(gp *g) {
  1239  	if gp.stack.lo == 0 {
  1240  		throw("missing stack in shrinkstack")
  1241  	}
  1242  	if s := readgstatus(gp); s&_Gscan == 0 {
  1243  		// We don't own the stack via _Gscan. We could still
  1244  		// own it if this is our own user G and we're on the
  1245  		// system stack.
  1246  		if !(gp == getg().m.curg && getg() != getg().m.curg && s == _Grunning) {
  1247  			// We don't own the stack.
  1248  			throw("bad status in shrinkstack")
  1249  		}
  1250  	}
  1251  	if !isShrinkStackSafe(gp) {
  1252  		throw("shrinkstack at bad time")
  1253  	}
  1254  	// Check for self-shrinks while in a libcall. These may have
  1255  	// pointers into the stack disguised as uintptrs, but these
  1256  	// code paths should all be nosplit.
  1257  	if gp == getg().m.curg && gp.m.libcallsp != 0 {
  1258  		throw("shrinking stack in libcall")
  1259  	}
  1260  
  1261  	if debug.gcshrinkstackoff > 0 {
  1262  		return
  1263  	}
  1264  
  1265  	oldsize := gp.stack.hi - gp.stack.lo
  1266  	newsize := oldsize / 2
  1267  	// Don't shrink the allocation below the minimum-sized stack
  1268  	// allocation.
  1269  	if newsize < fixedStack {
  1270  		return
  1271  	}
  1272  	// Compute how much of the stack is currently in use and only
  1273  	// shrink the stack if gp is using less than a quarter of its
  1274  	// current stack. The currently used stack includes everything
  1275  	// down to the SP plus the stack guard space that ensures
  1276  	// there's room for nosplit functions.
  1277  	avail := gp.stack.hi - gp.stack.lo
  1278  	if used := gp.stack.hi - gp.sched.sp + stackNosplit; used >= avail/4 {
  1279  		return
  1280  	}
  1281  
  1282  	if stackDebug > 0 {
  1283  		print("shrinking stack ", oldsize, "->", newsize, "\n")
  1284  	}
  1285  
  1286  	copystack(gp, newsize)
  1287  }
  1288  
  1289  // freeStackSpans frees unused stack spans at the end of GC.
  1290  func freeStackSpans() {
  1291  	// Scan stack pools for empty stack spans.
  1292  	for order := range stackpool {
  1293  		lock(&stackpool[order].item.mu)
  1294  		list := &stackpool[order].item.span
  1295  		for s := list.first; s != nil; {
  1296  			next := s.next
  1297  			if s.allocCount == 0 {
  1298  				list.remove(s)
  1299  				s.manualFreeList = 0
  1300  				osStackFree(s)
  1301  				mheap_.freeManual(s, spanAllocStack)
  1302  			}
  1303  			s = next
  1304  		}
  1305  		unlock(&stackpool[order].item.mu)
  1306  	}
  1307  
  1308  	// Free large stack spans.
  1309  	lock(&stackLarge.lock)
  1310  	for i := range stackLarge.free {
  1311  		for s := stackLarge.free[i].first; s != nil; {
  1312  			next := s.next
  1313  			stackLarge.free[i].remove(s)
  1314  			osStackFree(s)
  1315  			mheap_.freeManual(s, spanAllocStack)
  1316  			s = next
  1317  		}
  1318  	}
  1319  	unlock(&stackLarge.lock)
  1320  }
  1321  
  1322  // A stackObjectRecord is generated by the compiler for each stack object in a stack frame.
  1323  // This record must match the generator code in cmd/compile/internal/liveness/plive.go:emitStackObjects.
  1324  type stackObjectRecord struct {
  1325  	// offset in frame
  1326  	// if negative, offset from varp
  1327  	// if non-negative, offset from argp
  1328  	off       int32
  1329  	size      int32
  1330  	ptrBytes  int32
  1331  	gcdataoff uint32 // offset to gcdata from moduledata.rodata
  1332  }
  1333  
  1334  // gcdata returns the number of bytes that contain pointers, and
  1335  // a ptr/nonptr bitmask covering those bytes.
  1336  // Note that this bitmask might be larger than internal/abi.MaxPtrmaskBytes.
  1337  func (r *stackObjectRecord) gcdata() (uintptr, *byte) {
  1338  	ptr := uintptr(unsafe.Pointer(r))
  1339  	var mod *moduledata
  1340  	for datap := &firstmoduledata; datap != nil; datap = datap.next {
  1341  		if datap.gofunc <= ptr && ptr < datap.end {
  1342  			mod = datap
  1343  			break
  1344  		}
  1345  	}
  1346  	// If you get a panic here due to a nil mod,
  1347  	// you may have made a copy of a stackObjectRecord.
  1348  	// You must use the original pointer.
  1349  	res := mod.rodata + uintptr(r.gcdataoff)
  1350  	return uintptr(r.ptrBytes), (*byte)(unsafe.Pointer(res))
  1351  }
  1352  
  1353  // This is exported as ABI0 via linkname so obj can call it.
  1354  //
  1355  //go:nosplit
  1356  //go:linkname morestackc
  1357  func morestackc() {
  1358  	throw("attempt to execute system stack code on user stack")
  1359  }
  1360  
  1361  // startingStackSize is the amount of stack that new goroutines start with.
  1362  // It is a power of 2, and between fixedStack and maxstacksize, inclusive.
  1363  // startingStackSize is updated every GC by tracking the average size of
  1364  // stacks scanned during the GC.
  1365  var startingStackSize uint32 = fixedStack
  1366  
  1367  func gcComputeStartingStackSize() {
  1368  	if debug.adaptivestackstart == 0 {
  1369  		return
  1370  	}
  1371  	// For details, see the design doc at
  1372  	// https://docs.google.com/document/d/1YDlGIdVTPnmUiTAavlZxBI1d9pwGQgZT7IKFKlIXohQ/edit?usp=sharing
  1373  	// The basic algorithm is to track the average size of stacks
  1374  	// and start goroutines with stack equal to that average size.
  1375  	// Starting at the average size uses at most 2x the space that
  1376  	// an ideal algorithm would have used.
  1377  	// This is just a heuristic to avoid excessive stack growth work
  1378  	// early in a goroutine's lifetime. See issue 18138. Stacks that
  1379  	// are allocated too small can still grow, and stacks allocated
  1380  	// too large can still shrink.
  1381  	var scannedStackSize uint64
  1382  	var scannedStacks uint64
  1383  	for _, p := range allp {
  1384  		scannedStackSize += p.scannedStackSize
  1385  		scannedStacks += p.scannedStacks
  1386  		// Reset for next time
  1387  		p.scannedStackSize = 0
  1388  		p.scannedStacks = 0
  1389  	}
  1390  	if scannedStacks == 0 {
  1391  		startingStackSize = fixedStack
  1392  		return
  1393  	}
  1394  	avg := scannedStackSize/scannedStacks + stackGuard
  1395  	// Note: we add stackGuard to ensure that a goroutine that
  1396  	// uses the average space will not trigger a growth.
  1397  	if avg > uint64(maxstacksize) {
  1398  		avg = uint64(maxstacksize)
  1399  	}
  1400  	if avg < fixedStack {
  1401  		avg = fixedStack
  1402  	}
  1403  	// Note: maxstacksize fits in 30 bits, so avg also does.
  1404  	startingStackSize = uint32(round2(int32(avg)))
  1405  }
  1406  

View as plain text