Source file src/runtime/stack.go

  // Copyright 2013 The Go Authors. All rights reserved.
  // Use of this source code is governed by a BSD-style
  // license that can be found in the LICENSE file.
  
  package runtime
  
  import (
  	"runtime/internal/atomic"
  	"runtime/internal/sys"
  	"unsafe"
  )
  
  /*
  Stack layout parameters.
  Included both by runtime (compiled via 6c) and linkers (compiled via gcc).
  
  The per-goroutine g->stackguard is set to point StackGuard bytes
  above the bottom of the stack.  Each function compares its stack
  pointer against g->stackguard to check for overflow.  To cut one
  instruction from the check sequence for functions with tiny frames,
  the stack is allowed to protrude StackSmall bytes below the stack
  guard.  Functions with large frames don't bother with the check and
  always call morestack.  The sequences are (for amd64, others are
  similar):
  
  	guard = g->stackguard
  	frame = function's stack frame size
  	argsize = size of function arguments (call + return)
  
  	stack frame size <= StackSmall:
  		CMPQ guard, SP
  		JHI 3(PC)
  		MOVQ m->morearg, $(argsize << 32)
  		CALL morestack(SB)
  
  	stack frame size > StackSmall but < StackBig
  		LEAQ (frame-StackSmall)(SP), R0
  		CMPQ guard, R0
  		JHI 3(PC)
  		MOVQ m->morearg, $(argsize << 32)
  		CALL morestack(SB)
  
  	stack frame size >= StackBig:
  		MOVQ m->morearg, $((argsize << 32) | frame)
  		CALL morestack(SB)
  
  The bottom StackGuard - StackSmall bytes are important: there has
  to be enough room to execute functions that refuse to check for
  stack overflow, either because they need to be adjacent to the
  actual caller's frame (deferproc) or because they handle the imminent
  stack overflow (morestack).
  
  For example, deferproc might call malloc, which does one of the
  above checks (without allocating a full frame), which might trigger
  a call to morestack.  This sequence needs to fit in the bottom
  section of the stack.  On amd64, morestack's frame is 40 bytes, and
  deferproc's frame is 56 bytes.  That fits well within the
  StackGuard - StackSmall bytes at the bottom.
  The linkers explore all possible call traces involving non-splitting
  functions to make sure that this limit cannot be violated.
  */
  
  const (
  	// StackSystem is a number of additional bytes to add
  	// to each stack below the usual guard area for OS-specific
  	// purposes like signal handling. Used on Windows, Plan 9,
  	// and Darwin/ARM because they do not use a separate stack.
  	_StackSystem = sys.GoosWindows*512*sys.PtrSize + sys.GoosPlan9*512 + sys.GoosDarwin*sys.GoarchArm*1024
  
  	// The minimum size of stack used by Go code
  	_StackMin = 2048
  
  	// The minimum stack size to allocate.
  	// The hackery here rounds FixedStack0 up to a power of 2.
  	_FixedStack0 = _StackMin + _StackSystem
  	_FixedStack1 = _FixedStack0 - 1
  	_FixedStack2 = _FixedStack1 | (_FixedStack1 >> 1)
  	_FixedStack3 = _FixedStack2 | (_FixedStack2 >> 2)
  	_FixedStack4 = _FixedStack3 | (_FixedStack3 >> 4)
  	_FixedStack5 = _FixedStack4 | (_FixedStack4 >> 8)
  	_FixedStack6 = _FixedStack5 | (_FixedStack5 >> 16)
  	_FixedStack  = _FixedStack6 + 1
  
  	// Functions that need frames bigger than this use an extra
  	// instruction to do the stack split check, to avoid overflow
  	// in case SP - framesize wraps below zero.
  	// This value can be no bigger than the size of the unmapped
  	// space at zero.
  	_StackBig = 4096
  
  	// The stack guard is a pointer this many bytes above the
  	// bottom of the stack.
  	_StackGuard = 880*sys.StackGuardMultiplier + _StackSystem
  
  	// After a stack split check the SP is allowed to be this
  	// many bytes below the stack guard. This saves an instruction
  	// in the checking sequence for tiny frames.
  	_StackSmall = 128
  
  	// The maximum number of bytes that a chain of NOSPLIT
  	// functions can use.
  	_StackLimit = _StackGuard - _StackSystem - _StackSmall
  )
  
  // Goroutine preemption request.
  // Stored into g->stackguard0 to cause split stack check failure.
  // Must be greater than any real sp.
  // 0xfffffade in hex.
  const (
  	_StackPreempt = uintptrMask & -1314
  	_StackFork    = uintptrMask & -1234
  )
  
  const (
  	// stackDebug == 0: no logging
  	//            == 1: logging of per-stack operations
  	//            == 2: logging of per-frame operations
  	//            == 3: logging of per-word updates
  	//            == 4: logging of per-word reads
  	stackDebug       = 0
  	stackFromSystem  = 0 // allocate stacks from system memory instead of the heap
  	stackFaultOnFree = 0 // old stacks are mapped noaccess to detect use after free
  	stackPoisonCopy  = 0 // fill stack that should not be accessed with garbage, to detect bad dereferences during copy
  
  	stackCache = 1
  
  	// check the BP links during traceback.
  	debugCheckBP = false
  )
  
  const (
  	uintptrMask = 1<<(8*sys.PtrSize) - 1
  
  	// Goroutine preemption request.
  	// Stored into g->stackguard0 to cause split stack check failure.
  	// Must be greater than any real sp.
  	// 0xfffffade in hex.
  	stackPreempt = uintptrMask & -1314
  
  	// Thread is forking.
  	// Stored into g->stackguard0 to cause split stack check failure.
  	// Must be greater than any real sp.
  	stackFork = uintptrMask & -1234
  )
  
  // Global pool of spans that have free stacks.
  // Stacks are assigned an order according to size.
  //     order = log_2(size/FixedStack)
  // There is a free list for each order.
  // TODO: one lock per order?
  var stackpool [_NumStackOrders]mSpanList
  var stackpoolmu mutex
  
  // Global pool of large stack spans.
  var stackLarge struct {
  	lock mutex
  	free [_MHeapMap_Bits]mSpanList // free lists by log_2(s.npages)
  }
  
  func stackinit() {
  	if _StackCacheSize&_PageMask != 0 {
  		throw("cache size must be a multiple of page size")
  	}
  	for i := range stackpool {
  		stackpool[i].init()
  	}
  	for i := range stackLarge.free {
  		stackLarge.free[i].init()
  	}
  }
  
  // stacklog2 returns ⌊log_2(n)⌋.
  func stacklog2(n uintptr) int {
  	log2 := 0
  	for n > 1 {
  		n >>= 1
  		log2++
  	}
  	return log2
  }
  
  // Allocates a stack from the free pool. Must be called with
  // stackpoolmu held.
  func stackpoolalloc(order uint8) gclinkptr {
  	list := &stackpool[order]
  	s := list.first
  	if s == nil {
  		// no free stacks. Allocate another span worth.
  		s = mheap_.allocStack(_StackCacheSize >> _PageShift)
  		if s == nil {
  			throw("out of memory")
  		}
  		if s.allocCount != 0 {
  			throw("bad allocCount")
  		}
  		if s.stackfreelist.ptr() != nil {
  			throw("bad stackfreelist")
  		}
  		for i := uintptr(0); i < _StackCacheSize; i += _FixedStack << order {
  			x := gclinkptr(s.base() + i)
  			x.ptr().next = s.stackfreelist
  			s.stackfreelist = x
  		}
  		list.insert(s)
  	}
  	x := s.stackfreelist
  	if x.ptr() == nil {
  		throw("span has no free stacks")
  	}
  	s.stackfreelist = x.ptr().next
  	s.allocCount++
  	if s.stackfreelist.ptr() == nil {
  		// all stacks in s are allocated.
  		list.remove(s)
  	}
  	return x
  }
  
  // Adds stack x to the free pool. Must be called with stackpoolmu held.
  func stackpoolfree(x gclinkptr, order uint8) {
  	s := mheap_.lookup(unsafe.Pointer(x))
  	if s.state != _MSpanStack {
  		throw("freeing stack not in a stack span")
  	}
  	if s.stackfreelist.ptr() == nil {
  		// s will now have a free stack
  		stackpool[order].insert(s)
  	}
  	x.ptr().next = s.stackfreelist
  	s.stackfreelist = x
  	s.allocCount--
  	if gcphase == _GCoff && s.allocCount == 0 {
  		// Span is completely free. Return it to the heap
  		// immediately if we're sweeping.
  		//
  		// If GC is active, we delay the free until the end of
  		// GC to avoid the following type of situation:
  		//
  		// 1) GC starts, scans a SudoG but does not yet mark the SudoG.elem pointer
  		// 2) The stack that pointer points to is copied
  		// 3) The old stack is freed
  		// 4) The containing span is marked free
  		// 5) GC attempts to mark the SudoG.elem pointer. The
  		//    marking fails because the pointer looks like a
  		//    pointer into a free span.
  		//
  		// By not freeing, we prevent step #4 until GC is done.
  		stackpool[order].remove(s)
  		s.stackfreelist = 0
  		mheap_.freeStack(s)
  	}
  }
  
  // stackcacherefill/stackcacherelease implement a global pool of stack segments.
  // The pool is required to prevent unlimited growth of per-thread caches.
  //
  //go:systemstack
  func stackcacherefill(c *mcache, order uint8) {
  	if stackDebug >= 1 {
  		print("stackcacherefill order=", order, "\n")
  	}
  
  	// Grab some stacks from the global cache.
  	// Grab half of the allowed capacity (to prevent thrashing).
  	var list gclinkptr
  	var size uintptr
  	lock(&stackpoolmu)
  	for size < _StackCacheSize/2 {
  		x := stackpoolalloc(order)
  		x.ptr().next = list
  		list = x
  		size += _FixedStack << order
  	}
  	unlock(&stackpoolmu)
  	c.stackcache[order].list = list
  	c.stackcache[order].size = size
  }
  
  //go:systemstack
  func stackcacherelease(c *mcache, order uint8) {
  	if stackDebug >= 1 {
  		print("stackcacherelease order=", order, "\n")
  	}
  	x := c.stackcache[order].list
  	size := c.stackcache[order].size
  	lock(&stackpoolmu)
  	for size > _StackCacheSize/2 {
  		y := x.ptr().next
  		stackpoolfree(x, order)
  		x = y
  		size -= _FixedStack << order
  	}
  	unlock(&stackpoolmu)
  	c.stackcache[order].list = x
  	c.stackcache[order].size = size
  }
  
  //go:systemstack
  func stackcache_clear(c *mcache) {
  	if stackDebug >= 1 {
  		print("stackcache clear\n")
  	}
  	lock(&stackpoolmu)
  	for order := uint8(0); order < _NumStackOrders; order++ {
  		x := c.stackcache[order].list
  		for x.ptr() != nil {
  			y := x.ptr().next
  			stackpoolfree(x, order)
  			x = y
  		}
  		c.stackcache[order].list = 0
  		c.stackcache[order].size = 0
  	}
  	unlock(&stackpoolmu)
  }
  
  // stackalloc allocates an n byte stack.
  //
  // stackalloc must run on the system stack because it uses per-P
  // resources and must not split the stack.
  //
  //go:systemstack
  func stackalloc(n uint32) (stack, []stkbar) {
  	// Stackalloc must be called on scheduler stack, so that we
  	// never try to grow the stack during the code that stackalloc runs.
  	// Doing so would cause a deadlock (issue 1547).
  	thisg := getg()
  	if thisg != thisg.m.g0 {
  		throw("stackalloc not on scheduler stack")
  	}
  	if n&(n-1) != 0 {
  		throw("stack size not a power of 2")
  	}
  	if stackDebug >= 1 {
  		print("stackalloc ", n, "\n")
  	}
  
  	// Compute the size of stack barrier array.
  	maxstkbar := gcMaxStackBarriers(int(n))
  	nstkbar := unsafe.Sizeof(stkbar{}) * uintptr(maxstkbar)
  	var stkbarSlice slice
  
  	if debug.efence != 0 || stackFromSystem != 0 {
  		v := sysAlloc(round(uintptr(n), _PageSize), &memstats.stacks_sys)
  		if v == nil {
  			throw("out of memory (stackalloc)")
  		}
  		top := uintptr(n) - nstkbar
  		if maxstkbar != 0 {
  			stkbarSlice = slice{add(v, top), 0, maxstkbar}
  		}
  		return stack{uintptr(v), uintptr(v) + top}, *(*[]stkbar)(unsafe.Pointer(&stkbarSlice))
  	}
  
  	// Small stacks are allocated with a fixed-size free-list allocator.
  	// If we need a stack of a bigger size, we fall back on allocating
  	// a dedicated span.
  	var v unsafe.Pointer
  	if stackCache != 0 && n < _FixedStack<<_NumStackOrders && n < _StackCacheSize {
  		order := uint8(0)
  		n2 := n
  		for n2 > _FixedStack {
  			order++
  			n2 >>= 1
  		}
  		var x gclinkptr
  		c := thisg.m.mcache
  		if c == nil || thisg.m.preemptoff != "" || thisg.m.helpgc != 0 {
  			// c == nil can happen in the guts of exitsyscall or
  			// procresize. Just get a stack from the global pool.
  			// Also don't touch stackcache during gc
  			// as it's flushed concurrently.
  			lock(&stackpoolmu)
  			x = stackpoolalloc(order)
  			unlock(&stackpoolmu)
  		} else {
  			x = c.stackcache[order].list
  			if x.ptr() == nil {
  				stackcacherefill(c, order)
  				x = c.stackcache[order].list
  			}
  			c.stackcache[order].list = x.ptr().next
  			c.stackcache[order].size -= uintptr(n)
  		}
  		v = unsafe.Pointer(x)
  	} else {
  		var s *mspan
  		npage := uintptr(n) >> _PageShift
  		log2npage := stacklog2(npage)
  
  		// Try to get a stack from the large stack cache.
  		lock(&stackLarge.lock)
  		if !stackLarge.free[log2npage].isEmpty() {
  			s = stackLarge.free[log2npage].first
  			stackLarge.free[log2npage].remove(s)
  		}
  		unlock(&stackLarge.lock)
  
  		if s == nil {
  			// Allocate a new stack from the heap.
  			s = mheap_.allocStack(npage)
  			if s == nil {
  				throw("out of memory")
  			}
  		}
  		v = unsafe.Pointer(s.base())
  	}
  
  	if raceenabled {
  		racemalloc(v, uintptr(n))
  	}
  	if msanenabled {
  		msanmalloc(v, uintptr(n))
  	}
  	if stackDebug >= 1 {
  		print("  allocated ", v, "\n")
  	}
  	top := uintptr(n) - nstkbar
  	if maxstkbar != 0 {
  		stkbarSlice = slice{add(v, top), 0, maxstkbar}
  	}
  	return stack{uintptr(v), uintptr(v) + top}, *(*[]stkbar)(unsafe.Pointer(&stkbarSlice))
  }
  
  // stackfree frees an n byte stack allocation at stk.
  //
  // stackfree must run on the system stack because it uses per-P
  // resources and must not split the stack.
  //
  //go:systemstack
  func stackfree(stk stack, n uintptr) {
  	gp := getg()
  	v := unsafe.Pointer(stk.lo)
  	if n&(n-1) != 0 {
  		throw("stack not a power of 2")
  	}
  	if stk.lo+n < stk.hi {
  		throw("bad stack size")
  	}
  	if stackDebug >= 1 {
  		println("stackfree", v, n)
  		memclrNoHeapPointers(v, n) // for testing, clobber stack data
  	}
  	if debug.efence != 0 || stackFromSystem != 0 {
  		if debug.efence != 0 || stackFaultOnFree != 0 {
  			sysFault(v, n)
  		} else {
  			sysFree(v, n, &memstats.stacks_sys)
  		}
  		return
  	}
  	if msanenabled {
  		msanfree(v, n)
  	}
  	if stackCache != 0 && n < _FixedStack<<_NumStackOrders && n < _StackCacheSize {
  		order := uint8(0)
  		n2 := n
  		for n2 > _FixedStack {
  			order++
  			n2 >>= 1
  		}
  		x := gclinkptr(v)
  		c := gp.m.mcache
  		if c == nil || gp.m.preemptoff != "" || gp.m.helpgc != 0 {
  			lock(&stackpoolmu)
  			stackpoolfree(x, order)
  			unlock(&stackpoolmu)
  		} else {
  			if c.stackcache[order].size >= _StackCacheSize {
  				stackcacherelease(c, order)
  			}
  			x.ptr().next = c.stackcache[order].list
  			c.stackcache[order].list = x
  			c.stackcache[order].size += n
  		}
  	} else {
  		s := mheap_.lookup(v)
  		if s.state != _MSpanStack {
  			println(hex(s.base()), v)
  			throw("bad span state")
  		}
  		if gcphase == _GCoff {
  			// Free the stack immediately if we're
  			// sweeping.
  			mheap_.freeStack(s)
  		} else {
  			// If the GC is running, we can't return a
  			// stack span to the heap because it could be
  			// reused as a heap span, and this state
  			// change would race with GC. Add it to the
  			// large stack cache instead.
  			log2npage := stacklog2(s.npages)
  			lock(&stackLarge.lock)
  			stackLarge.free[log2npage].insert(s)
  			unlock(&stackLarge.lock)
  		}
  	}
  }
  
  var maxstacksize uintptr = 1 << 20 // enough until runtime.main sets it for real
  
  var ptrnames = []string{
  	0: "scalar",
  	1: "ptr",
  }
  
  // Stack frame layout
  //
  // (x86)
  // +------------------+
  // | args from caller |
  // +------------------+ <- frame->argp
  // |  return address  |
  // +------------------+
  // |  caller's BP (*) | (*) if framepointer_enabled && varp < sp
  // +------------------+ <- frame->varp
  // |     locals       |
  // +------------------+
  // |  args to callee  |
  // +------------------+ <- frame->sp
  //
  // (arm)
  // +------------------+
  // | args from caller |
  // +------------------+ <- frame->argp
  // | caller's retaddr |
  // +------------------+ <- frame->varp
  // |     locals       |
  // +------------------+
  // |  args to callee  |
  // +------------------+
  // |  return address  |
  // +------------------+ <- frame->sp
  
  type adjustinfo struct {
  	old   stack
  	delta uintptr // ptr distance from old to new stack (newbase - oldbase)
  	cache pcvalueCache
  
  	// sghi is the highest sudog.elem on the stack.
  	sghi uintptr
  }
  
  // Adjustpointer checks whether *vpp is in the old stack described by adjinfo.
  // If so, it rewrites *vpp to point into the new stack.
  func adjustpointer(adjinfo *adjustinfo, vpp unsafe.Pointer) {
  	pp := (*uintptr)(vpp)
  	p := *pp
  	if stackDebug >= 4 {
  		print("        ", pp, ":", hex(p), "\n")
  	}
  	if adjinfo.old.lo <= p && p < adjinfo.old.hi {
  		*pp = p + adjinfo.delta
  		if stackDebug >= 3 {
  			print("        adjust ptr ", pp, ":", hex(p), " -> ", hex(*pp), "\n")
  		}
  	}
  }
  
  // Information from the compiler about the layout of stack frames.
  type bitvector struct {
  	n        int32 // # of bits
  	bytedata *uint8
  }
  
  type gobitvector struct {
  	n        uintptr
  	bytedata []uint8
  }
  
  func gobv(bv bitvector) gobitvector {
  	return gobitvector{
  		uintptr(bv.n),
  		(*[1 << 30]byte)(unsafe.Pointer(bv.bytedata))[:(bv.n+7)/8],
  	}
  }
  
  func ptrbit(bv *gobitvector, i uintptr) uint8 {
  	return (bv.bytedata[i/8] >> (i % 8)) & 1
  }
  
  // bv describes the memory starting at address scanp.
  // Adjust any pointers contained therein.
  func adjustpointers(scanp unsafe.Pointer, cbv *bitvector, adjinfo *adjustinfo, f *_func) {
  	bv := gobv(*cbv)
  	minp := adjinfo.old.lo
  	maxp := adjinfo.old.hi
  	delta := adjinfo.delta
  	num := bv.n
  	// If this frame might contain channel receive slots, use CAS
  	// to adjust pointers. If the slot hasn't been received into
  	// yet, it may contain stack pointers and a concurrent send
  	// could race with adjusting those pointers. (The sent value
  	// itself can never contain stack pointers.)
  	useCAS := uintptr(scanp) < adjinfo.sghi
  	for i := uintptr(0); i < num; i++ {
  		if stackDebug >= 4 {
  			print("        ", add(scanp, i*sys.PtrSize), ":", ptrnames[ptrbit(&bv, i)], ":", hex(*(*uintptr)(add(scanp, i*sys.PtrSize))), " # ", i, " ", bv.bytedata[i/8], "\n")
  		}
  		if ptrbit(&bv, i) == 1 {
  			pp := (*uintptr)(add(scanp, i*sys.PtrSize))
  		retry:
  			p := *pp
  			if f != nil && 0 < p && p < minLegalPointer && debug.invalidptr != 0 {
  				// Looks like a junk value in a pointer slot.
  				// Live analysis wrong?
  				getg().m.traceback = 2
  				print("runtime: bad pointer in frame ", funcname(f), " at ", pp, ": ", hex(p), "\n")
  				throw("invalid pointer found on stack")
  			}
  			if minp <= p && p < maxp {
  				if stackDebug >= 3 {
  					print("adjust ptr ", hex(p), " ", funcname(f), "\n")
  				}
  				if useCAS {
  					ppu := (*unsafe.Pointer)(unsafe.Pointer(pp))
  					if !atomic.Casp1(ppu, unsafe.Pointer(p), unsafe.Pointer(p+delta)) {
  						goto retry
  					}
  				} else {
  					*pp = p + delta
  				}
  			}
  		}
  	}
  }
  
  // Note: the argument/return area is adjusted by the callee.
  func adjustframe(frame *stkframe, arg unsafe.Pointer) bool {
  	adjinfo := (*adjustinfo)(arg)
  	targetpc := frame.continpc
  	if targetpc == 0 {
  		// Frame is dead.
  		return true
  	}
  	f := frame.fn
  	if stackDebug >= 2 {
  		print("    adjusting ", funcname(f), " frame=[", hex(frame.sp), ",", hex(frame.fp), "] pc=", hex(frame.pc), " continpc=", hex(frame.continpc), "\n")
  	}
  	if f.entry == systemstack_switchPC {
  		// A special routine at the bottom of stack of a goroutine that does an systemstack call.
  		// We will allow it to be copied even though we don't
  		// have full GC info for it (because it is written in asm).
  		return true
  	}
  	if targetpc != f.entry {
  		targetpc--
  	}
  	pcdata := pcdatavalue(f, _PCDATA_StackMapIndex, targetpc, &adjinfo.cache)
  	if pcdata == -1 {
  		pcdata = 0 // in prologue
  	}
  
  	// Adjust local variables if stack frame has been allocated.
  	size := frame.varp - frame.sp
  	var minsize uintptr
  	switch sys.ArchFamily {
  	case sys.ARM64:
  		minsize = sys.SpAlign
  	default:
  		minsize = sys.MinFrameSize
  	}
  	if size > minsize {
  		var bv bitvector
  		stackmap := (*stackmap)(funcdata(f, _FUNCDATA_LocalsPointerMaps))
  		if stackmap == nil || stackmap.n <= 0 {
  			print("runtime: frame ", funcname(f), " untyped locals ", hex(frame.varp-size), "+", hex(size), "\n")
  			throw("missing stackmap")
  		}
  		// Locals bitmap information, scan just the pointers in locals.
  		if pcdata < 0 || pcdata >= stackmap.n {
  			// don't know where we are
  			print("runtime: pcdata is ", pcdata, " and ", stackmap.n, " locals stack map entries for ", funcname(f), " (targetpc=", targetpc, ")\n")
  			throw("bad symbol table")
  		}
  		bv = stackmapdata(stackmap, pcdata)
  		size = uintptr(bv.n) * sys.PtrSize
  		if stackDebug >= 3 {
  			print("      locals ", pcdata, "/", stackmap.n, " ", size/sys.PtrSize, " words ", bv.bytedata, "\n")
  		}
  		adjustpointers(unsafe.Pointer(frame.varp-size), &bv, adjinfo, f)
  	}
  
  	// Adjust saved base pointer if there is one.
  	if sys.ArchFamily == sys.AMD64 && frame.argp-frame.varp == 2*sys.RegSize {
  		if !framepointer_enabled {
  			print("runtime: found space for saved base pointer, but no framepointer experiment\n")
  			print("argp=", hex(frame.argp), " varp=", hex(frame.varp), "\n")
  			throw("bad frame layout")
  		}
  		if stackDebug >= 3 {
  			print("      saved bp\n")
  		}
  		if debugCheckBP {
  			// Frame pointers should always point to the next higher frame on
  			// the Go stack (or be nil, for the top frame on the stack).
  			bp := *(*uintptr)(unsafe.Pointer(frame.varp))
  			if bp != 0 && (bp < adjinfo.old.lo || bp >= adjinfo.old.hi) {
  				println("runtime: found invalid frame pointer")
  				print("bp=", hex(bp), " min=", hex(adjinfo.old.lo), " max=", hex(adjinfo.old.hi), "\n")
  				throw("bad frame pointer")
  			}
  		}
  		adjustpointer(adjinfo, unsafe.Pointer(frame.varp))
  	}
  
  	// Adjust arguments.
  	if frame.arglen > 0 {
  		var bv bitvector
  		if frame.argmap != nil {
  			bv = *frame.argmap
  		} else {
  			stackmap := (*stackmap)(funcdata(f, _FUNCDATA_ArgsPointerMaps))
  			if stackmap == nil || stackmap.n <= 0 {
  				print("runtime: frame ", funcname(f), " untyped args ", frame.argp, "+", frame.arglen, "\n")
  				throw("missing stackmap")
  			}
  			if pcdata < 0 || pcdata >= stackmap.n {
  				// don't know where we are
  				print("runtime: pcdata is ", pcdata, " and ", stackmap.n, " args stack map entries for ", funcname(f), " (targetpc=", targetpc, ")\n")
  				throw("bad symbol table")
  			}
  			bv = stackmapdata(stackmap, pcdata)
  		}
  		if stackDebug >= 3 {
  			print("      args\n")
  		}
  		adjustpointers(unsafe.Pointer(frame.argp), &bv, adjinfo, nil)
  	}
  	return true
  }
  
  func adjustctxt(gp *g, adjinfo *adjustinfo) {
  	adjustpointer(adjinfo, unsafe.Pointer(&gp.sched.ctxt))
  	if !framepointer_enabled {
  		return
  	}
  	if debugCheckBP {
  		bp := gp.sched.bp
  		if bp != 0 && (bp < adjinfo.old.lo || bp >= adjinfo.old.hi) {
  			println("runtime: found invalid top frame pointer")
  			print("bp=", hex(bp), " min=", hex(adjinfo.old.lo), " max=", hex(adjinfo.old.hi), "\n")
  			throw("bad top frame pointer")
  		}
  	}
  	adjustpointer(adjinfo, unsafe.Pointer(&gp.sched.bp))
  }
  
  func adjustdefers(gp *g, adjinfo *adjustinfo) {
  	// Adjust defer argument blocks the same way we adjust active stack frames.
  	tracebackdefers(gp, adjustframe, noescape(unsafe.Pointer(adjinfo)))
  
  	// Adjust pointers in the Defer structs.
  	// Defer structs themselves are never on the stack.
  	for d := gp._defer; d != nil; d = d.link {
  		adjustpointer(adjinfo, unsafe.Pointer(&d.fn))
  		adjustpointer(adjinfo, unsafe.Pointer(&d.sp))
  		adjustpointer(adjinfo, unsafe.Pointer(&d._panic))
  	}
  }
  
  func adjustpanics(gp *g, adjinfo *adjustinfo) {
  	// Panics are on stack and already adjusted.
  	// Update pointer to head of list in G.
  	adjustpointer(adjinfo, unsafe.Pointer(&gp._panic))
  }
  
  func adjustsudogs(gp *g, adjinfo *adjustinfo) {
  	// the data elements pointed to by a SudoG structure
  	// might be in the stack.
  	for s := gp.waiting; s != nil; s = s.waitlink {
  		adjustpointer(adjinfo, unsafe.Pointer(&s.elem))
  		adjustpointer(adjinfo, unsafe.Pointer(&s.selectdone))
  	}
  }
  
  func adjuststkbar(gp *g, adjinfo *adjustinfo) {
  	for i := int(gp.stkbarPos); i < len(gp.stkbar); i++ {
  		adjustpointer(adjinfo, unsafe.Pointer(&gp.stkbar[i].savedLRPtr))
  	}
  }
  
  func fillstack(stk stack, b byte) {
  	for p := stk.lo; p < stk.hi; p++ {
  		*(*byte)(unsafe.Pointer(p)) = b
  	}
  }
  
  func findsghi(gp *g, stk stack) uintptr {
  	var sghi uintptr
  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
  		p := uintptr(sg.elem) + uintptr(sg.c.elemsize)
  		if stk.lo <= p && p < stk.hi && p > sghi {
  			sghi = p
  		}
  		p = uintptr(unsafe.Pointer(sg.selectdone)) + unsafe.Sizeof(sg.selectdone)
  		if stk.lo <= p && p < stk.hi && p > sghi {
  			sghi = p
  		}
  	}
  	return sghi
  }
  
  // syncadjustsudogs adjusts gp's sudogs and copies the part of gp's
  // stack they refer to while synchronizing with concurrent channel
  // operations. It returns the number of bytes of stack copied.
  func syncadjustsudogs(gp *g, used uintptr, adjinfo *adjustinfo) uintptr {
  	if gp.waiting == nil {
  		return 0
  	}
  
  	// Lock channels to prevent concurrent send/receive.
  	// It's important that we *only* do this for async
  	// copystack; otherwise, gp may be in the middle of
  	// putting itself on wait queues and this would
  	// self-deadlock.
  	var lastc *hchan
  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
  		if sg.c != lastc {
  			lock(&sg.c.lock)
  		}
  		lastc = sg.c
  	}
  
  	// Adjust sudogs.
  	adjustsudogs(gp, adjinfo)
  
  	// Copy the part of the stack the sudogs point in to
  	// while holding the lock to prevent races on
  	// send/receive slots.
  	var sgsize uintptr
  	if adjinfo.sghi != 0 {
  		oldBot := adjinfo.old.hi - used
  		newBot := oldBot + adjinfo.delta
  		sgsize = adjinfo.sghi - oldBot
  		memmove(unsafe.Pointer(newBot), unsafe.Pointer(oldBot), sgsize)
  	}
  
  	// Unlock channels.
  	lastc = nil
  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
  		if sg.c != lastc {
  			unlock(&sg.c.lock)
  		}
  		lastc = sg.c
  	}
  
  	return sgsize
  }
  
  // Copies gp's stack to a new stack of a different size.
  // Caller must have changed gp status to Gcopystack.
  //
  // If sync is true, this is a self-triggered stack growth and, in
  // particular, no other G may be writing to gp's stack (e.g., via a
  // channel operation). If sync is false, copystack protects against
  // concurrent channel operations.
  func copystack(gp *g, newsize uintptr, sync bool) {
  	if gp.syscallsp != 0 {
  		throw("stack growth not allowed in system call")
  	}
  	old := gp.stack
  	if old.lo == 0 {
  		throw("nil stackbase")
  	}
  	used := old.hi - gp.sched.sp
  
  	// allocate new stack
  	new, newstkbar := stackalloc(uint32(newsize))
  	if stackPoisonCopy != 0 {
  		fillstack(new, 0xfd)
  	}
  	if stackDebug >= 1 {
  		print("copystack gp=", gp, " [", hex(old.lo), " ", hex(old.hi-used), " ", hex(old.hi), "]/", gp.stackAlloc, " -> [", hex(new.lo), " ", hex(new.hi-used), " ", hex(new.hi), "]/", newsize, "\n")
  	}
  
  	// Compute adjustment.
  	var adjinfo adjustinfo
  	adjinfo.old = old
  	adjinfo.delta = new.hi - old.hi
  
  	// Adjust sudogs, synchronizing with channel ops if necessary.
  	ncopy := used
  	if sync {
  		adjustsudogs(gp, &adjinfo)
  	} else {
  		// sudogs can point in to the stack. During concurrent
  		// shrinking, these areas may be written to. Find the
  		// highest such pointer so we can handle everything
  		// there and below carefully. (This shouldn't be far
  		// from the bottom of the stack, so there's little
  		// cost in handling everything below it carefully.)
  		adjinfo.sghi = findsghi(gp, old)
  
  		// Synchronize with channel ops and copy the part of
  		// the stack they may interact with.
  		ncopy -= syncadjustsudogs(gp, used, &adjinfo)
  	}
  
  	// Copy the stack (or the rest of it) to the new location
  	memmove(unsafe.Pointer(new.hi-ncopy), unsafe.Pointer(old.hi-ncopy), ncopy)
  
  	// Disallow sigprof scans of this stack and block if there's
  	// one in progress.
  	gcLockStackBarriers(gp)
  
  	// Adjust remaining structures that have pointers into stacks.
  	// We have to do most of these before we traceback the new
  	// stack because gentraceback uses them.
  	adjustctxt(gp, &adjinfo)
  	adjustdefers(gp, &adjinfo)
  	adjustpanics(gp, &adjinfo)
  	adjuststkbar(gp, &adjinfo)
  	if adjinfo.sghi != 0 {
  		adjinfo.sghi += adjinfo.delta
  	}
  
  	// copy old stack barriers to new stack barrier array
  	newstkbar = newstkbar[:len(gp.stkbar)]
  	copy(newstkbar, gp.stkbar)
  
  	// Swap out old stack for new one
  	gp.stack = new
  	gp.stackguard0 = new.lo + _StackGuard // NOTE: might clobber a preempt request
  	gp.sched.sp = new.hi - used
  	oldsize := gp.stackAlloc
  	gp.stackAlloc = newsize
  	gp.stkbar = newstkbar
  	gp.stktopsp += adjinfo.delta
  
  	// Adjust pointers in the new stack.
  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, adjustframe, noescape(unsafe.Pointer(&adjinfo)), 0)
  
  	gcUnlockStackBarriers(gp)
  
  	// free old stack
  	if stackPoisonCopy != 0 {
  		fillstack(old, 0xfc)
  	}
  	stackfree(old, oldsize)
  }
  
  // round x up to a power of 2.
  func round2(x int32) int32 {
  	s := uint(0)
  	for 1<<s < x {
  		s++
  	}
  	return 1 << s
  }
  
  // Called from runtime·morestack when more stack is needed.
  // Allocate larger stack and relocate to new stack.
  // Stack growth is multiplicative, for constant amortized cost.
  //
  // g->atomicstatus will be Grunning or Gscanrunning upon entry.
  // If the GC is trying to stop this g then it will set preemptscan to true.
  //
  // ctxt is the value of the context register on morestack. newstack
  // will write it to g.sched.ctxt.
  func newstack(ctxt unsafe.Pointer) {
  	thisg := getg()
  	// TODO: double check all gp. shouldn't be getg().
  	if thisg.m.morebuf.g.ptr().stackguard0 == stackFork {
  		throw("stack growth after fork")
  	}
  	if thisg.m.morebuf.g.ptr() != thisg.m.curg {
  		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")
  		morebuf := thisg.m.morebuf
  		traceback(morebuf.pc, morebuf.sp, morebuf.lr, morebuf.g.ptr())
  		throw("runtime: wrong goroutine in newstack")
  	}
  
  	gp := thisg.m.curg
  	// Write ctxt to gp.sched. We do this here instead of in
  	// morestack so it has the necessary write barrier.
  	gp.sched.ctxt = ctxt
  
  	if thisg.m.curg.throwsplit {
  		// Update syscallsp, syscallpc in case traceback uses them.
  		morebuf := thisg.m.morebuf
  		gp.syscallsp = morebuf.sp
  		gp.syscallpc = morebuf.pc
  		print("runtime: newstack sp=", hex(gp.sched.sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n",
  			"\tmorebuf={pc:", hex(morebuf.pc), " sp:", hex(morebuf.sp), " lr:", hex(morebuf.lr), "}\n",
  			"\tsched={pc:", hex(gp.sched.pc), " sp:", hex(gp.sched.sp), " lr:", hex(gp.sched.lr), " ctxt:", gp.sched.ctxt, "}\n")
  
  		traceback(morebuf.pc, morebuf.sp, morebuf.lr, gp)
  		throw("runtime: stack split at bad time")
  	}
  
  	morebuf := thisg.m.morebuf
  	thisg.m.morebuf.pc = 0
  	thisg.m.morebuf.lr = 0
  	thisg.m.morebuf.sp = 0
  	thisg.m.morebuf.g = 0
  
  	// NOTE: stackguard0 may change underfoot, if another thread
  	// is about to try to preempt gp. Read it just once and use that same
  	// value now and below.
  	preempt := atomic.Loaduintptr(&gp.stackguard0) == stackPreempt
  
  	// Be conservative about where we preempt.
  	// We are interested in preempting user Go code, not runtime code.
  	// If we're holding locks, mallocing, or preemption is disabled, don't
  	// preempt.
  	// This check is very early in newstack so that even the status change
  	// from Grunning to Gwaiting and back doesn't happen in this case.
  	// That status change by itself can be viewed as a small preemption,
  	// because the GC might change Gwaiting to Gscanwaiting, and then
  	// this goroutine has to wait for the GC to finish before continuing.
  	// If the GC is in some way dependent on this goroutine (for example,
  	// it needs a lock held by the goroutine), that small preemption turns
  	// into a real deadlock.
  	if preempt {
  		if thisg.m.locks != 0 || thisg.m.mallocing != 0 || thisg.m.preemptoff != "" || thisg.m.p.ptr().status != _Prunning {
  			// Let the goroutine keep running for now.
  			// gp->preempt is set, so it will be preempted next time.
  			gp.stackguard0 = gp.stack.lo + _StackGuard
  			gogo(&gp.sched) // never return
  		}
  	}
  
  	if gp.stack.lo == 0 {
  		throw("missing stack in newstack")
  	}
  	sp := gp.sched.sp
  	if sys.ArchFamily == sys.AMD64 || sys.ArchFamily == sys.I386 {
  		// The call to morestack cost a word.
  		sp -= sys.PtrSize
  	}
  	if stackDebug >= 1 || sp < gp.stack.lo {
  		print("runtime: newstack sp=", hex(sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n",
  			"\tmorebuf={pc:", hex(morebuf.pc), " sp:", hex(morebuf.sp), " lr:", hex(morebuf.lr), "}\n",
  			"\tsched={pc:", hex(gp.sched.pc), " sp:", hex(gp.sched.sp), " lr:", hex(gp.sched.lr), " ctxt:", gp.sched.ctxt, "}\n")
  	}
  	if sp < gp.stack.lo {
  		print("runtime: gp=", gp, ", gp->status=", hex(readgstatus(gp)), "\n ")
  		print("runtime: split stack overflow: ", hex(sp), " < ", hex(gp.stack.lo), "\n")
  		throw("runtime: split stack overflow")
  	}
  
  	if preempt {
  		if gp == thisg.m.g0 {
  			throw("runtime: preempt g0")
  		}
  		if thisg.m.p == 0 && thisg.m.locks == 0 {
  			throw("runtime: g is running but p is not")
  		}
  		// Synchronize with scang.
  		casgstatus(gp, _Grunning, _Gwaiting)
  		if gp.preemptscan {
  			for !castogscanstatus(gp, _Gwaiting, _Gscanwaiting) {
  				// Likely to be racing with the GC as
  				// it sees a _Gwaiting and does the
  				// stack scan. If so, gcworkdone will
  				// be set and gcphasework will simply
  				// return.
  			}
  			if !gp.gcscandone {
  				// gcw is safe because we're on the
  				// system stack.
  				gcw := &gp.m.p.ptr().gcw
  				scanstack(gp, gcw)
  				if gcBlackenPromptly {
  					gcw.dispose()
  				}
  				gp.gcscandone = true
  			}
  			gp.preemptscan = false
  			gp.preempt = false
  			casfrom_Gscanstatus(gp, _Gscanwaiting, _Gwaiting)
  			// This clears gcscanvalid.
  			casgstatus(gp, _Gwaiting, _Grunning)
  			gp.stackguard0 = gp.stack.lo + _StackGuard
  			gogo(&gp.sched) // never return
  		}
  
  		// Act like goroutine called runtime.Gosched.
  		casgstatus(gp, _Gwaiting, _Grunning)
  		gopreempt_m(gp) // never return
  	}
  
  	// Allocate a bigger segment and move the stack.
  	oldsize := int(gp.stackAlloc)
  	newsize := oldsize * 2
  	if uintptr(newsize) > maxstacksize {
  		print("runtime: goroutine stack exceeds ", maxstacksize, "-byte limit\n")
  		throw("stack overflow")
  	}
  
  	// The goroutine must be executing in order to call newstack,
  	// so it must be Grunning (or Gscanrunning).
  	casgstatus(gp, _Grunning, _Gcopystack)
  
  	// The concurrent GC will not scan the stack while we are doing the copy since
  	// the gp is in a Gcopystack status.
  	copystack(gp, uintptr(newsize), true)
  	if stackDebug >= 1 {
  		print("stack grow done\n")
  	}
  	casgstatus(gp, _Gcopystack, _Grunning)
  	gogo(&gp.sched)
  }
  
  //go:nosplit
  func nilfunc() {
  	*(*uint8)(nil) = 0
  }
  
  // adjust Gobuf as if it executed a call to fn
  // and then did an immediate gosave.
  func gostartcallfn(gobuf *gobuf, fv *funcval) {
  	var fn unsafe.Pointer
  	if fv != nil {
  		fn = unsafe.Pointer(fv.fn)
  	} else {
  		fn = unsafe.Pointer(funcPC(nilfunc))
  	}
  	gostartcall(gobuf, fn, unsafe.Pointer(fv))
  }
  
  // Maybe shrink the stack being used by gp.
  // Called at garbage collection time.
  // gp must be stopped, but the world need not be.
  func shrinkstack(gp *g) {
  	gstatus := readgstatus(gp)
  	if gstatus&^_Gscan == _Gdead {
  		if gp.stack.lo != 0 {
  			// Free whole stack - it will get reallocated
  			// if G is used again.
  			stackfree(gp.stack, gp.stackAlloc)
  			gp.stack.lo = 0
  			gp.stack.hi = 0
  			gp.stkbar = nil
  			gp.stkbarPos = 0
  		}
  		return
  	}
  	if gp.stack.lo == 0 {
  		throw("missing stack in shrinkstack")
  	}
  	if gstatus&_Gscan == 0 {
  		throw("bad status in shrinkstack")
  	}
  
  	if debug.gcshrinkstackoff > 0 {
  		return
  	}
  	if gp.startpc == gcBgMarkWorkerPC {
  		// We're not allowed to shrink the gcBgMarkWorker
  		// stack (see gcBgMarkWorker for explanation).
  		return
  	}
  
  	oldsize := gp.stackAlloc
  	newsize := oldsize / 2
  	// Don't shrink the allocation below the minimum-sized stack
  	// allocation.
  	if newsize < _FixedStack {
  		return
  	}
  	// Compute how much of the stack is currently in use and only
  	// shrink the stack if gp is using less than a quarter of its
  	// current stack. The currently used stack includes everything
  	// down to the SP plus the stack guard space that ensures
  	// there's room for nosplit functions.
  	avail := gp.stack.hi - gp.stack.lo
  	if used := gp.stack.hi - gp.sched.sp + _StackLimit; used >= avail/4 {
  		return
  	}
  
  	// We can't copy the stack if we're in a syscall.
  	// The syscall might have pointers into the stack.
  	if gp.syscallsp != 0 {
  		return
  	}
  	if sys.GoosWindows != 0 && gp.m != nil && gp.m.libcallsp != 0 {
  		return
  	}
  
  	if stackDebug > 0 {
  		print("shrinking stack ", oldsize, "->", newsize, "\n")
  	}
  
  	copystack(gp, newsize, false)
  }
  
  // freeStackSpans frees unused stack spans at the end of GC.
  func freeStackSpans() {
  	lock(&stackpoolmu)
  
  	// Scan stack pools for empty stack spans.
  	for order := range stackpool {
  		list := &stackpool[order]
  		for s := list.first; s != nil; {
  			next := s.next
  			if s.allocCount == 0 {
  				list.remove(s)
  				s.stackfreelist = 0
  				mheap_.freeStack(s)
  			}
  			s = next
  		}
  	}
  
  	unlock(&stackpoolmu)
  
  	// Free large stack spans.
  	lock(&stackLarge.lock)
  	for i := range stackLarge.free {
  		for s := stackLarge.free[i].first; s != nil; {
  			next := s.next
  			stackLarge.free[i].remove(s)
  			mheap_.freeStack(s)
  			s = next
  		}
  	}
  	unlock(&stackLarge.lock)
  }
  
  //go:nosplit
  func morestackc() {
  	systemstack(func() {
  		throw("attempt to execute C code on Go stack")
  	})
  }
  

View as plain text