Source file src/runtime/select.go

  // Copyright 2009 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
  
  // This file contains the implementation of Go select statements.
  
  import (
  	"runtime/internal/sys"
  	"unsafe"
  )
  
  const (
  	debugSelect = false
  
  	// scase.kind
  	caseRecv = iota
  	caseSend
  	caseDefault
  )
  
  // Select statement header.
  // Known to compiler.
  // Changes here must also be made in src/cmd/internal/gc/select.go's selecttype.
  type hselect struct {
  	tcase     uint16   // total count of scase[]
  	ncase     uint16   // currently filled scase[]
  	pollorder *uint16  // case poll order
  	lockorder *uint16  // channel lock order
  	scase     [1]scase // one per case (in order of appearance)
  }
  
  // Select case descriptor.
  // Known to compiler.
  // Changes here must also be made in src/cmd/internal/gc/select.go's selecttype.
  type scase struct {
  	elem        unsafe.Pointer // data element
  	c           *hchan         // chan
  	pc          uintptr        // return pc
  	kind        uint16
  	so          uint16 // vararg of selected bool
  	receivedp   *bool  // pointer to received bool (recv2)
  	releasetime int64
  }
  
  var (
  	chansendpc = funcPC(chansend)
  	chanrecvpc = funcPC(chanrecv)
  )
  
  func selectsize(size uintptr) uintptr {
  	selsize := unsafe.Sizeof(hselect{}) +
  		(size-1)*unsafe.Sizeof(hselect{}.scase[0]) +
  		size*unsafe.Sizeof(*hselect{}.lockorder) +
  		size*unsafe.Sizeof(*hselect{}.pollorder)
  	return round(selsize, sys.Int64Align)
  }
  
  func newselect(sel *hselect, selsize int64, size int32) {
  	if selsize != int64(selectsize(uintptr(size))) {
  		print("runtime: bad select size ", selsize, ", want ", selectsize(uintptr(size)), "\n")
  		throw("bad select size")
  	}
  	sel.tcase = uint16(size)
  	sel.ncase = 0
  	sel.lockorder = (*uint16)(add(unsafe.Pointer(&sel.scase), uintptr(size)*unsafe.Sizeof(hselect{}.scase[0])))
  	sel.pollorder = (*uint16)(add(unsafe.Pointer(sel.lockorder), uintptr(size)*unsafe.Sizeof(*hselect{}.lockorder)))
  
  	if debugSelect {
  		print("newselect s=", sel, " size=", size, "\n")
  	}
  }
  
  //go:nosplit
  func selectsend(sel *hselect, c *hchan, elem unsafe.Pointer) (selected bool) {
  	// nil cases do not compete
  	if c != nil {
  		selectsendImpl(sel, c, getcallerpc(unsafe.Pointer(&sel)), elem, uintptr(unsafe.Pointer(&selected))-uintptr(unsafe.Pointer(&sel)))
  	}
  	return
  }
  
  // cut in half to give stack a chance to split
  func selectsendImpl(sel *hselect, c *hchan, pc uintptr, elem unsafe.Pointer, so uintptr) {
  	i := sel.ncase
  	if i >= sel.tcase {
  		throw("selectsend: too many cases")
  	}
  	sel.ncase = i + 1
  	cas := (*scase)(add(unsafe.Pointer(&sel.scase), uintptr(i)*unsafe.Sizeof(sel.scase[0])))
  
  	cas.pc = pc
  	cas.c = c
  	cas.so = uint16(so)
  	cas.kind = caseSend
  	cas.elem = elem
  
  	if debugSelect {
  		print("selectsend s=", sel, " pc=", hex(cas.pc), " chan=", cas.c, " so=", cas.so, "\n")
  	}
  }
  
  //go:nosplit
  func selectrecv(sel *hselect, c *hchan, elem unsafe.Pointer) (selected bool) {
  	// nil cases do not compete
  	if c != nil {
  		selectrecvImpl(sel, c, getcallerpc(unsafe.Pointer(&sel)), elem, nil, uintptr(unsafe.Pointer(&selected))-uintptr(unsafe.Pointer(&sel)))
  	}
  	return
  }
  
  //go:nosplit
  func selectrecv2(sel *hselect, c *hchan, elem unsafe.Pointer, received *bool) (selected bool) {
  	// nil cases do not compete
  	if c != nil {
  		selectrecvImpl(sel, c, getcallerpc(unsafe.Pointer(&sel)), elem, received, uintptr(unsafe.Pointer(&selected))-uintptr(unsafe.Pointer(&sel)))
  	}
  	return
  }
  
  func selectrecvImpl(sel *hselect, c *hchan, pc uintptr, elem unsafe.Pointer, received *bool, so uintptr) {
  	i := sel.ncase
  	if i >= sel.tcase {
  		throw("selectrecv: too many cases")
  	}
  	sel.ncase = i + 1
  	cas := (*scase)(add(unsafe.Pointer(&sel.scase), uintptr(i)*unsafe.Sizeof(sel.scase[0])))
  	cas.pc = pc
  	cas.c = c
  	cas.so = uint16(so)
  	cas.kind = caseRecv
  	cas.elem = elem
  	cas.receivedp = received
  
  	if debugSelect {
  		print("selectrecv s=", sel, " pc=", hex(cas.pc), " chan=", cas.c, " so=", cas.so, "\n")
  	}
  }
  
  //go:nosplit
  func selectdefault(sel *hselect) (selected bool) {
  	selectdefaultImpl(sel, getcallerpc(unsafe.Pointer(&sel)), uintptr(unsafe.Pointer(&selected))-uintptr(unsafe.Pointer(&sel)))
  	return
  }
  
  func selectdefaultImpl(sel *hselect, callerpc uintptr, so uintptr) {
  	i := sel.ncase
  	if i >= sel.tcase {
  		throw("selectdefault: too many cases")
  	}
  	sel.ncase = i + 1
  	cas := (*scase)(add(unsafe.Pointer(&sel.scase), uintptr(i)*unsafe.Sizeof(sel.scase[0])))
  	cas.pc = callerpc
  	cas.c = nil
  	cas.so = uint16(so)
  	cas.kind = caseDefault
  
  	if debugSelect {
  		print("selectdefault s=", sel, " pc=", hex(cas.pc), " so=", cas.so, "\n")
  	}
  }
  
  func sellock(scases []scase, lockorder []uint16) {
  	var c *hchan
  	for _, o := range lockorder {
  		c0 := scases[o].c
  		if c0 != nil && c0 != c {
  			c = c0
  			lock(&c.lock)
  		}
  	}
  }
  
  func selunlock(scases []scase, lockorder []uint16) {
  	// We must be very careful here to not touch sel after we have unlocked
  	// the last lock, because sel can be freed right after the last unlock.
  	// Consider the following situation.
  	// First M calls runtime·park() in runtime·selectgo() passing the sel.
  	// Once runtime·park() has unlocked the last lock, another M makes
  	// the G that calls select runnable again and schedules it for execution.
  	// When the G runs on another M, it locks all the locks and frees sel.
  	// Now if the first M touches sel, it will access freed memory.
  	n := len(scases)
  	r := 0
  	// skip the default case
  	if n > 0 && scases[lockorder[0]].c == nil {
  		r = 1
  	}
  	for i := n - 1; i >= r; i-- {
  		c := scases[lockorder[i]].c
  		if i > 0 && c == scases[lockorder[i-1]].c {
  			continue // will unlock it on the next iteration
  		}
  		unlock(&c.lock)
  	}
  }
  
  func selparkcommit(gp *g, _ unsafe.Pointer) bool {
  	// This must not access gp's stack (see gopark). In
  	// particular, it must not access the *hselect. That's okay,
  	// because by the time this is called, gp.waiting has all
  	// channels in lock order.
  	var lastc *hchan
  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
  		if sg.c != lastc && lastc != nil {
  			// As soon as we unlock the channel, fields in
  			// any sudog with that channel may change,
  			// including c and waitlink. Since multiple
  			// sudogs may have the same channel, we unlock
  			// only after we've passed the last instance
  			// of a channel.
  			unlock(&lastc.lock)
  		}
  		lastc = sg.c
  	}
  	if lastc != nil {
  		unlock(&lastc.lock)
  	}
  	return true
  }
  
  func block() {
  	gopark(nil, nil, "select (no cases)", traceEvGoStop, 1) // forever
  }
  
  // selectgo implements the select statement.
  //
  // *sel is on the current goroutine's stack (regardless of any
  // escaping in selectgo).
  //
  // selectgo does not return. Instead, it overwrites its return PC and
  // returns directly to the triggered select case. Because of this, it
  // cannot appear at the top of a split stack.
  //
  //go:nosplit
  func selectgo(sel *hselect) {
  	pc, offset := selectgoImpl(sel)
  	*(*bool)(add(unsafe.Pointer(&sel), uintptr(offset))) = true
  	setcallerpc(unsafe.Pointer(&sel), pc)
  }
  
  // selectgoImpl returns scase.pc and scase.so for the select
  // case which fired.
  func selectgoImpl(sel *hselect) (uintptr, uint16) {
  	if debugSelect {
  		print("select: sel=", sel, "\n")
  	}
  
  	scaseslice := slice{unsafe.Pointer(&sel.scase), int(sel.ncase), int(sel.ncase)}
  	scases := *(*[]scase)(unsafe.Pointer(&scaseslice))
  
  	var t0 int64
  	if blockprofilerate > 0 {
  		t0 = cputicks()
  		for i := 0; i < int(sel.ncase); i++ {
  			scases[i].releasetime = -1
  		}
  	}
  
  	// The compiler rewrites selects that statically have
  	// only 0 or 1 cases plus default into simpler constructs.
  	// The only way we can end up with such small sel.ncase
  	// values here is for a larger select in which most channels
  	// have been nilled out. The general code handles those
  	// cases correctly, and they are rare enough not to bother
  	// optimizing (and needing to test).
  
  	// generate permuted order
  	pollslice := slice{unsafe.Pointer(sel.pollorder), int(sel.ncase), int(sel.ncase)}
  	pollorder := *(*[]uint16)(unsafe.Pointer(&pollslice))
  	for i := 1; i < int(sel.ncase); i++ {
  		j := int(fastrand()) % (i + 1)
  		pollorder[i] = pollorder[j]
  		pollorder[j] = uint16(i)
  	}
  
  	// sort the cases by Hchan address to get the locking order.
  	// simple heap sort, to guarantee n log n time and constant stack footprint.
  	lockslice := slice{unsafe.Pointer(sel.lockorder), int(sel.ncase), int(sel.ncase)}
  	lockorder := *(*[]uint16)(unsafe.Pointer(&lockslice))
  	for i := 0; i < int(sel.ncase); i++ {
  		j := i
  		// Start with the pollorder to permute cases on the same channel.
  		c := scases[pollorder[i]].c
  		for j > 0 && scases[lockorder[(j-1)/2]].c.sortkey() < c.sortkey() {
  			k := (j - 1) / 2
  			lockorder[j] = lockorder[k]
  			j = k
  		}
  		lockorder[j] = pollorder[i]
  	}
  	for i := int(sel.ncase) - 1; i >= 0; i-- {
  		o := lockorder[i]
  		c := scases[o].c
  		lockorder[i] = lockorder[0]
  		j := 0
  		for {
  			k := j*2 + 1
  			if k >= i {
  				break
  			}
  			if k+1 < i && scases[lockorder[k]].c.sortkey() < scases[lockorder[k+1]].c.sortkey() {
  				k++
  			}
  			if c.sortkey() < scases[lockorder[k]].c.sortkey() {
  				lockorder[j] = lockorder[k]
  				j = k
  				continue
  			}
  			break
  		}
  		lockorder[j] = o
  	}
  	/*
  		for i := 0; i+1 < int(sel.ncase); i++ {
  			if scases[lockorder[i]].c.sortkey() > scases[lockorder[i+1]].c.sortkey() {
  				print("i=", i, " x=", lockorder[i], " y=", lockorder[i+1], "\n")
  				throw("select: broken sort")
  			}
  		}
  	*/
  
  	// lock all the channels involved in the select
  	sellock(scases, lockorder)
  
  	var (
  		gp     *g
  		done   uint32
  		sg     *sudog
  		c      *hchan
  		k      *scase
  		sglist *sudog
  		sgnext *sudog
  		qp     unsafe.Pointer
  		nextp  **sudog
  	)
  
  loop:
  	// pass 1 - look for something already waiting
  	var dfl *scase
  	var cas *scase
  	for i := 0; i < int(sel.ncase); i++ {
  		cas = &scases[pollorder[i]]
  		c = cas.c
  
  		switch cas.kind {
  		case caseRecv:
  			sg = c.sendq.dequeue()
  			if sg != nil {
  				goto recv
  			}
  			if c.qcount > 0 {
  				goto bufrecv
  			}
  			if c.closed != 0 {
  				goto rclose
  			}
  
  		case caseSend:
  			if raceenabled {
  				racereadpc(unsafe.Pointer(c), cas.pc, chansendpc)
  			}
  			if c.closed != 0 {
  				goto sclose
  			}
  			sg = c.recvq.dequeue()
  			if sg != nil {
  				goto send
  			}
  			if c.qcount < c.dataqsiz {
  				goto bufsend
  			}
  
  		case caseDefault:
  			dfl = cas
  		}
  	}
  
  	if dfl != nil {
  		selunlock(scases, lockorder)
  		cas = dfl
  		goto retc
  	}
  
  	// pass 2 - enqueue on all chans
  	gp = getg()
  	done = 0
  	if gp.waiting != nil {
  		throw("gp.waiting != nil")
  	}
  	nextp = &gp.waiting
  	for _, casei := range lockorder {
  		cas = &scases[casei]
  		c = cas.c
  		sg := acquireSudog()
  		sg.g = gp
  		// Note: selectdone is adjusted for stack copies in stack1.go:adjustsudogs
  		sg.selectdone = (*uint32)(noescape(unsafe.Pointer(&done)))
  		// No stack splits between assigning elem and enqueuing
  		// sg on gp.waiting where copystack can find it.
  		sg.elem = cas.elem
  		sg.releasetime = 0
  		if t0 != 0 {
  			sg.releasetime = -1
  		}
  		sg.c = c
  		// Construct waiting list in lock order.
  		*nextp = sg
  		nextp = &sg.waitlink
  
  		switch cas.kind {
  		case caseRecv:
  			c.recvq.enqueue(sg)
  
  		case caseSend:
  			c.sendq.enqueue(sg)
  		}
  	}
  
  	// wait for someone to wake us up
  	gp.param = nil
  	gopark(selparkcommit, nil, "select", traceEvGoBlockSelect, 2)
  
  	// While we were asleep, some goroutine came along and completed
  	// one of the cases in the select and woke us up (called ready).
  	// As part of that process, the goroutine did a cas on done above
  	// (aka *sg.selectdone for all queued sg) to win the right to
  	// complete the select. Now done = 1.
  	//
  	// If we copy (grow) our own stack, we will update the
  	// selectdone pointers inside the gp.waiting sudog list to point
  	// at the new stack. Another goroutine attempting to
  	// complete one of our (still linked in) select cases might
  	// see the new selectdone pointer (pointing at the new stack)
  	// before the new stack has real data; if the new stack has done = 0
  	// (before the old values are copied over), the goroutine might
  	// do a cas via sg.selectdone and incorrectly believe that it has
  	// won the right to complete the select, executing a second
  	// communication and attempting to wake us (call ready) again.
  	//
  	// Then things break.
  	//
  	// The best break is that the goroutine doing ready sees the
  	// _Gcopystack status and throws, as in #17007.
  	// A worse break would be for us to continue on, start running real code,
  	// block in a semaphore acquisition (sema.go), and have the other
  	// goroutine wake us up without having really acquired the semaphore.
  	// That would result in the goroutine spuriously running and then
  	// queue up another spurious wakeup when the semaphore really is ready.
  	// In general the situation can cascade until something notices the
  	// problem and causes a crash.
  	//
  	// A stack shrink does not have this problem, because it locks
  	// all the channels that are involved first, blocking out the
  	// possibility of a cas on selectdone.
  	//
  	// A stack growth before gopark above does not have this
  	// problem, because we hold those channel locks (released by
  	// selparkcommit).
  	//
  	// A stack growth after sellock below does not have this
  	// problem, because again we hold those channel locks.
  	//
  	// The only problem is a stack growth during sellock.
  	// To keep that from happening, run sellock on the system stack.
  	//
  	// It might be that we could avoid this if copystack copied the
  	// stack before calling adjustsudogs. In that case,
  	// syncadjustsudogs would need to recopy the tiny part that
  	// it copies today, resulting in a little bit of extra copying.
  	//
  	// An even better fix, not for the week before a release candidate,
  	// would be to put space in every sudog and make selectdone
  	// point at (say) the space in the first sudog.
  
  	systemstack(func() {
  		sellock(scases, lockorder)
  	})
  
  	sg = (*sudog)(gp.param)
  	gp.param = nil
  
  	// pass 3 - dequeue from unsuccessful chans
  	// otherwise they stack up on quiet channels
  	// record the successful case, if any.
  	// We singly-linked up the SudoGs in lock order.
  	cas = nil
  	sglist = gp.waiting
  	// Clear all elem before unlinking from gp.waiting.
  	for sg1 := gp.waiting; sg1 != nil; sg1 = sg1.waitlink {
  		sg1.selectdone = nil
  		sg1.elem = nil
  		sg1.c = nil
  	}
  	gp.waiting = nil
  
  	for _, casei := range lockorder {
  		k = &scases[casei]
  		if sglist.releasetime > 0 {
  			k.releasetime = sglist.releasetime
  		}
  		if sg == sglist {
  			// sg has already been dequeued by the G that woke us up.
  			cas = k
  		} else {
  			c = k.c
  			if k.kind == caseSend {
  				c.sendq.dequeueSudoG(sglist)
  			} else {
  				c.recvq.dequeueSudoG(sglist)
  			}
  		}
  		sgnext = sglist.waitlink
  		sglist.waitlink = nil
  		releaseSudog(sglist)
  		sglist = sgnext
  	}
  
  	if cas == nil {
  		// We can wake up with gp.param == nil (so cas == nil)
  		// when a channel involved in the select has been closed.
  		// It is easiest to loop and re-run the operation;
  		// we'll see that it's now closed.
  		// Maybe some day we can signal the close explicitly,
  		// but we'd have to distinguish close-on-reader from close-on-writer.
  		// It's easiest not to duplicate the code and just recheck above.
  		// We know that something closed, and things never un-close,
  		// so we won't block again.
  		goto loop
  	}
  
  	c = cas.c
  
  	if debugSelect {
  		print("wait-return: sel=", sel, " c=", c, " cas=", cas, " kind=", cas.kind, "\n")
  	}
  
  	if cas.kind == caseRecv {
  		if cas.receivedp != nil {
  			*cas.receivedp = true
  		}
  	}
  
  	if raceenabled {
  		if cas.kind == caseRecv && cas.elem != nil {
  			raceWriteObjectPC(c.elemtype, cas.elem, cas.pc, chanrecvpc)
  		} else if cas.kind == caseSend {
  			raceReadObjectPC(c.elemtype, cas.elem, cas.pc, chansendpc)
  		}
  	}
  	if msanenabled {
  		if cas.kind == caseRecv && cas.elem != nil {
  			msanwrite(cas.elem, c.elemtype.size)
  		} else if cas.kind == caseSend {
  			msanread(cas.elem, c.elemtype.size)
  		}
  	}
  
  	selunlock(scases, lockorder)
  	goto retc
  
  bufrecv:
  	// can receive from buffer
  	if raceenabled {
  		if cas.elem != nil {
  			raceWriteObjectPC(c.elemtype, cas.elem, cas.pc, chanrecvpc)
  		}
  		raceacquire(chanbuf(c, c.recvx))
  		racerelease(chanbuf(c, c.recvx))
  	}
  	if msanenabled && cas.elem != nil {
  		msanwrite(cas.elem, c.elemtype.size)
  	}
  	if cas.receivedp != nil {
  		*cas.receivedp = true
  	}
  	qp = chanbuf(c, c.recvx)
  	if cas.elem != nil {
  		typedmemmove(c.elemtype, cas.elem, qp)
  	}
  	typedmemclr(c.elemtype, qp)
  	c.recvx++
  	if c.recvx == c.dataqsiz {
  		c.recvx = 0
  	}
  	c.qcount--
  	selunlock(scases, lockorder)
  	goto retc
  
  bufsend:
  	// can send to buffer
  	if raceenabled {
  		raceacquire(chanbuf(c, c.sendx))
  		racerelease(chanbuf(c, c.sendx))
  		raceReadObjectPC(c.elemtype, cas.elem, cas.pc, chansendpc)
  	}
  	if msanenabled {
  		msanread(cas.elem, c.elemtype.size)
  	}
  	typedmemmove(c.elemtype, chanbuf(c, c.sendx), cas.elem)
  	c.sendx++
  	if c.sendx == c.dataqsiz {
  		c.sendx = 0
  	}
  	c.qcount++
  	selunlock(scases, lockorder)
  	goto retc
  
  recv:
  	// can receive from sleeping sender (sg)
  	recv(c, sg, cas.elem, func() { selunlock(scases, lockorder) })
  	if debugSelect {
  		print("syncrecv: sel=", sel, " c=", c, "\n")
  	}
  	if cas.receivedp != nil {
  		*cas.receivedp = true
  	}
  	goto retc
  
  rclose:
  	// read at end of closed channel
  	selunlock(scases, lockorder)
  	if cas.receivedp != nil {
  		*cas.receivedp = false
  	}
  	if cas.elem != nil {
  		typedmemclr(c.elemtype, cas.elem)
  	}
  	if raceenabled {
  		raceacquire(unsafe.Pointer(c))
  	}
  	goto retc
  
  send:
  	// can send to a sleeping receiver (sg)
  	if raceenabled {
  		raceReadObjectPC(c.elemtype, cas.elem, cas.pc, chansendpc)
  	}
  	if msanenabled {
  		msanread(cas.elem, c.elemtype.size)
  	}
  	send(c, sg, cas.elem, func() { selunlock(scases, lockorder) })
  	if debugSelect {
  		print("syncsend: sel=", sel, " c=", c, "\n")
  	}
  	goto retc
  
  retc:
  	if cas.releasetime > 0 {
  		blockevent(cas.releasetime-t0, 2)
  	}
  	return cas.pc, cas.so
  
  sclose:
  	// send on closed channel
  	selunlock(scases, lockorder)
  	panic(plainError("send on closed channel"))
  }
  
  func (c *hchan) sortkey() uintptr {
  	// TODO(khr): if we have a moving garbage collector, we'll need to
  	// change this function.
  	return uintptr(unsafe.Pointer(c))
  }
  
  // A runtimeSelect is a single case passed to rselect.
  // This must match ../reflect/value.go:/runtimeSelect
  type runtimeSelect struct {
  	dir selectDir
  	typ unsafe.Pointer // channel type (not used here)
  	ch  *hchan         // channel
  	val unsafe.Pointer // ptr to data (SendDir) or ptr to receive buffer (RecvDir)
  }
  
  // These values must match ../reflect/value.go:/SelectDir.
  type selectDir int
  
  const (
  	_             selectDir = iota
  	selectSend              // case Chan <- Send
  	selectRecv              // case <-Chan:
  	selectDefault           // default
  )
  
  //go:linkname reflect_rselect reflect.rselect
  func reflect_rselect(cases []runtimeSelect) (chosen int, recvOK bool) {
  	// flagNoScan is safe here, because all objects are also referenced from cases.
  	size := selectsize(uintptr(len(cases)))
  	sel := (*hselect)(mallocgc(size, nil, true))
  	newselect(sel, int64(size), int32(len(cases)))
  	r := new(bool)
  	for i := range cases {
  		rc := &cases[i]
  		switch rc.dir {
  		case selectDefault:
  			selectdefaultImpl(sel, uintptr(i), 0)
  		case selectSend:
  			if rc.ch == nil {
  				break
  			}
  			selectsendImpl(sel, rc.ch, uintptr(i), rc.val, 0)
  		case selectRecv:
  			if rc.ch == nil {
  				break
  			}
  			selectrecvImpl(sel, rc.ch, uintptr(i), rc.val, r, 0)
  		}
  	}
  
  	pc, _ := selectgoImpl(sel)
  	chosen = int(pc)
  	recvOK = *r
  	return
  }
  
  func (q *waitq) dequeueSudoG(sgp *sudog) {
  	x := sgp.prev
  	y := sgp.next
  	if x != nil {
  		if y != nil {
  			// middle of queue
  			x.next = y
  			y.prev = x
  			sgp.next = nil
  			sgp.prev = nil
  			return
  		}
  		// end of queue
  		x.next = nil
  		q.last = x
  		sgp.prev = nil
  		return
  	}
  	if y != nil {
  		// start of queue
  		y.prev = nil
  		q.first = y
  		sgp.next = nil
  		return
  	}
  
  	// x==y==nil. Either sgp is the only element in the queue,
  	// or it has already been removed. Use q.first to disambiguate.
  	if q.first == sgp {
  		q.first = nil
  		q.last = nil
  	}
  }
  

View as plain text