select.go

Documentation: runtime

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package runtime
     6  
     7  // This file contains the implementation of Go select statements.
     8  
     9  import (
    10  	"internal/abi"
    11  	"runtime/internal/atomic"
    12  	"unsafe"
    13  )
    14  
    15  const debugSelect = false
    16  
    17  // Select case descriptor.
    18  // Known to compiler.
    19  // Changes here must also be made in src/cmd/compile/internal/walk/select.go's scasetype.
    20  type scase struct {
    21  	c    *hchan         // chan
    22  	elem unsafe.Pointer // data element
    23  }
    24  
    25  var (
    26  	chansendpc = abi.FuncPCABIInternal(chansend)
    27  	chanrecvpc = abi.FuncPCABIInternal(chanrecv)
    28  )
    29  
    30  func selectsetpc(pc *uintptr) {
    31  	*pc = getcallerpc()
    32  }
    33  
    34  func sellock(scases []scase, lockorder []uint16) {
    35  	var c *hchan
    36  	for _, o := range lockorder {
    37  		c0 := scases[o].c
    38  		if c0 != c {
    39  			c = c0
    40  			lock(&c.lock)
    41  		}
    42  	}
    43  }
    44  
    45  func selunlock(scases []scase, lockorder []uint16) {
    46  	// We must be very careful here to not touch sel after we have unlocked
    47  	// the last lock, because sel can be freed right after the last unlock.
    48  	// Consider the following situation.
    49  	// First M calls runtime·park() in runtime·selectgo() passing the sel.
    50  	// Once runtime·park() has unlocked the last lock, another M makes
    51  	// the G that calls select runnable again and schedules it for execution.
    52  	// When the G runs on another M, it locks all the locks and frees sel.
    53  	// Now if the first M touches sel, it will access freed memory.
    54  	for i := len(lockorder) - 1; i >= 0; i-- {
    55  		c := scases[lockorder[i]].c
    56  		if i > 0 && c == scases[lockorder[i-1]].c {
    57  			continue // will unlock it on the next iteration
    58  		}
    59  		unlock(&c.lock)
    60  	}
    61  }
    62  
    63  func selparkcommit(gp *g, _ unsafe.Pointer) bool {
    64  	// There are unlocked sudogs that point into gp's stack. Stack
    65  	// copying must lock the channels of those sudogs.
    66  	// Set activeStackChans here instead of before we try parking
    67  	// because we could self-deadlock in stack growth on a
    68  	// channel lock.
    69  	gp.activeStackChans = true
    70  	// Mark that it's safe for stack shrinking to occur now,
    71  	// because any thread acquiring this G's stack for shrinking
    72  	// is guaranteed to observe activeStackChans after this store.
    73  	atomic.Store8(&gp.parkingOnChan, 0)
    74  	// Make sure we unlock after setting activeStackChans and
    75  	// unsetting parkingOnChan. The moment we unlock any of the
    76  	// channel locks we risk gp getting readied by a channel operation
    77  	// and so gp could continue running before everything before the
    78  	// unlock is visible (even to gp itself).
    79  
    80  	// This must not access gp's stack (see gopark). In
    81  	// particular, it must not access the *hselect. That's okay,
    82  	// because by the time this is called, gp.waiting has all
    83  	// channels in lock order.
    84  	var lastc *hchan
    85  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
    86  		if sg.c != lastc && lastc != nil {
    87  			// As soon as we unlock the channel, fields in
    88  			// any sudog with that channel may change,
    89  			// including c and waitlink. Since multiple
    90  			// sudogs may have the same channel, we unlock
    91  			// only after we've passed the last instance
    92  			// of a channel.
    93  			unlock(&lastc.lock)
    94  		}
    95  		lastc = sg.c
    96  	}
    97  	if lastc != nil {
    98  		unlock(&lastc.lock)
    99  	}
   100  	return true
   101  }
   102  
   103  func block() {
   104  	gopark(nil, nil, waitReasonSelectNoCases, traceEvGoStop, 1) // forever
   105  }
   106  
   107  // selectgo implements the select statement.
   108  //
   109  // cas0 points to an array of type [ncases]scase, and order0 points to
   110  // an array of type [2*ncases]uint16 where ncases must be <= 65536.
   111  // Both reside on the goroutine's stack (regardless of any escaping in
   112  // selectgo).
   113  //
   114  // For race detector builds, pc0 points to an array of type
   115  // [ncases]uintptr (also on the stack); for other builds, it's set to
   116  // nil.
   117  //
   118  // selectgo returns the index of the chosen scase, which matches the
   119  // ordinal position of its respective select{recv,send,default} call.
   120  // Also, if the chosen scase was a receive operation, it reports whether
   121  // a value was received.
   122  func selectgo(cas0 *scase, order0 *uint16, pc0 *uintptr, nsends, nrecvs int, block bool) (int, bool) {
   123  	if debugSelect {
   124  		print("select: cas0=", cas0, "\n")
   125  	}
   126  
   127  	// NOTE: In order to maintain a lean stack size, the number of scases
   128  	// is capped at 65536.
   129  	cas1 := (*[1 << 16]scase)(unsafe.Pointer(cas0))
   130  	order1 := (*[1 << 17]uint16)(unsafe.Pointer(order0))
   131  
   132  	ncases := nsends + nrecvs
   133  	scases := cas1[:ncases:ncases]
   134  	pollorder := order1[:ncases:ncases]
   135  	lockorder := order1[ncases:][:ncases:ncases]
   136  	// NOTE: pollorder/lockorder's underlying array was not zero-initialized by compiler.
   137  
   138  	// Even when raceenabled is true, there might be select
   139  	// statements in packages compiled without -race (e.g.,
   140  	// ensureSigM in runtime/signal_unix.go).
   141  	var pcs []uintptr
   142  	if raceenabled && pc0 != nil {
   143  		pc1 := (*[1 << 16]uintptr)(unsafe.Pointer(pc0))
   144  		pcs = pc1[:ncases:ncases]
   145  	}
   146  	casePC := func(casi int) uintptr {
   147  		if pcs == nil {
   148  			return 0
   149  		}
   150  		return pcs[casi]
   151  	}
   152  
   153  	var t0 int64
   154  	if blockprofilerate > 0 {
   155  		t0 = cputicks()
   156  	}
   157  
   158  	// The compiler rewrites selects that statically have
   159  	// only 0 or 1 cases plus default into simpler constructs.
   160  	// The only way we can end up with such small sel.ncase
   161  	// values here is for a larger select in which most channels
   162  	// have been nilled out. The general code handles those
   163  	// cases correctly, and they are rare enough not to bother
   164  	// optimizing (and needing to test).
   165  
   166  	// generate permuted order
   167  	norder := 0
   168  	for i := range scases {
   169  		cas := &scases[i]
   170  
   171  		// Omit cases without channels from the poll and lock orders.
   172  		if cas.c == nil {
   173  			cas.elem = nil // allow GC
   174  			continue
   175  		}
   176  
   177  		j := fastrandn(uint32(norder + 1))
   178  		pollorder[norder] = pollorder[j]
   179  		pollorder[j] = uint16(i)
   180  		norder++
   181  	}
   182  	pollorder = pollorder[:norder]
   183  	lockorder = lockorder[:norder]
   184  
   185  	// sort the cases by Hchan address to get the locking order.
   186  	// simple heap sort, to guarantee n log n time and constant stack footprint.
   187  	for i := range lockorder {
   188  		j := i
   189  		// Start with the pollorder to permute cases on the same channel.
   190  		c := scases[pollorder[i]].c
   191  		for j > 0 && scases[lockorder[(j-1)/2]].c.sortkey() < c.sortkey() {
   192  			k := (j - 1) / 2
   193  			lockorder[j] = lockorder[k]
   194  			j = k
   195  		}
   196  		lockorder[j] = pollorder[i]
   197  	}
   198  	for i := len(lockorder) - 1; i >= 0; i-- {
   199  		o := lockorder[i]
   200  		c := scases[o].c
   201  		lockorder[i] = lockorder[0]
   202  		j := 0
   203  		for {
   204  			k := j*2 + 1
   205  			if k >= i {
   206  				break
   207  			}
   208  			if k+1 < i && scases[lockorder[k]].c.sortkey() < scases[lockorder[k+1]].c.sortkey() {
   209  				k++
   210  			}
   211  			if c.sortkey() < scases[lockorder[k]].c.sortkey() {
   212  				lockorder[j] = lockorder[k]
   213  				j = k
   214  				continue
   215  			}
   216  			break
   217  		}
   218  		lockorder[j] = o
   219  	}
   220  
   221  	if debugSelect {
   222  		for i := 0; i+1 < len(lockorder); i++ {
   223  			if scases[lockorder[i]].c.sortkey() > scases[lockorder[i+1]].c.sortkey() {
   224  				print("i=", i, " x=", lockorder[i], " y=", lockorder[i+1], "\n")
   225  				throw("select: broken sort")
   226  			}
   227  		}
   228  	}
   229  
   230  	// lock all the channels involved in the select
   231  	sellock(scases, lockorder)
   232  
   233  	var (
   234  		gp     *g
   235  		sg     *sudog
   236  		c      *hchan
   237  		k      *scase
   238  		sglist *sudog
   239  		sgnext *sudog
   240  		qp     unsafe.Pointer
   241  		nextp  **sudog
   242  	)
   243  
   244  	// pass 1 - look for something already waiting
   245  	var casi int
   246  	var cas *scase
   247  	var caseSuccess bool
   248  	var caseReleaseTime int64 = -1
   249  	var recvOK bool
   250  	for _, casei := range pollorder {
   251  		casi = int(casei)
   252  		cas = &scases[casi]
   253  		c = cas.c
   254  
   255  		if casi >= nsends {
   256  			sg = c.sendq.dequeue()
   257  			if sg != nil {
   258  				goto recv
   259  			}
   260  			if c.qcount > 0 {
   261  				goto bufrecv
   262  			}
   263  			if c.closed != 0 {
   264  				goto rclose
   265  			}
   266  		} else {
   267  			if raceenabled {
   268  				racereadpc(c.raceaddr(), casePC(casi), chansendpc)
   269  			}
   270  			if c.closed != 0 {
   271  				goto sclose
   272  			}
   273  			sg = c.recvq.dequeue()
   274  			if sg != nil {
   275  				goto send
   276  			}
   277  			if c.qcount < c.dataqsiz {
   278  				goto bufsend
   279  			}
   280  		}
   281  	}
   282  
   283  	if !block {
   284  		selunlock(scases, lockorder)
   285  		casi = -1
   286  		goto retc
   287  	}
   288  
   289  	// pass 2 - enqueue on all chans
   290  	gp = getg()
   291  	if gp.waiting != nil {
   292  		throw("gp.waiting != nil")
   293  	}
   294  	nextp = &gp.waiting
   295  	for _, casei := range lockorder {
   296  		casi = int(casei)
   297  		cas = &scases[casi]
   298  		c = cas.c
   299  		sg := acquireSudog()
   300  		sg.g = gp
   301  		sg.isSelect = true
   302  		// No stack splits between assigning elem and enqueuing
   303  		// sg on gp.waiting where copystack can find it.
   304  		sg.elem = cas.elem
   305  		sg.releasetime = 0
   306  		if t0 != 0 {
   307  			sg.releasetime = -1
   308  		}
   309  		sg.c = c
   310  		// Construct waiting list in lock order.
   311  		*nextp = sg
   312  		nextp = &sg.waitlink
   313  
   314  		if casi < nsends {
   315  			c.sendq.enqueue(sg)
   316  		} else {
   317  			c.recvq.enqueue(sg)
   318  		}
   319  	}
   320  
   321  	// wait for someone to wake us up
   322  	gp.param = nil
   323  	// Signal to anyone trying to shrink our stack that we're about
   324  	// to park on a channel. The window between when this G's status
   325  	// changes and when we set gp.activeStackChans is not safe for
   326  	// stack shrinking.
   327  	atomic.Store8(&gp.parkingOnChan, 1)
   328  	gopark(selparkcommit, nil, waitReasonSelect, traceEvGoBlockSelect, 1)
   329  	gp.activeStackChans = false
   330  
   331  	sellock(scases, lockorder)
   332  
   333  	gp.selectDone = 0
   334  	sg = (*sudog)(gp.param)
   335  	gp.param = nil
   336  
   337  	// pass 3 - dequeue from unsuccessful chans
   338  	// otherwise they stack up on quiet channels
   339  	// record the successful case, if any.
   340  	// We singly-linked up the SudoGs in lock order.
   341  	casi = -1
   342  	cas = nil
   343  	caseSuccess = false
   344  	sglist = gp.waiting
   345  	// Clear all elem before unlinking from gp.waiting.
   346  	for sg1 := gp.waiting; sg1 != nil; sg1 = sg1.waitlink {
   347  		sg1.isSelect = false
   348  		sg1.elem = nil
   349  		sg1.c = nil
   350  	}
   351  	gp.waiting = nil
   352  
   353  	for _, casei := range lockorder {
   354  		k = &scases[casei]
   355  		if sg == sglist {
   356  			// sg has already been dequeued by the G that woke us up.
   357  			casi = int(casei)
   358  			cas = k
   359  			caseSuccess = sglist.success
   360  			if sglist.releasetime > 0 {
   361  				caseReleaseTime = sglist.releasetime
   362  			}
   363  		} else {
   364  			c = k.c
   365  			if int(casei) < nsends {
   366  				c.sendq.dequeueSudoG(sglist)
   367  			} else {
   368  				c.recvq.dequeueSudoG(sglist)
   369  			}
   370  		}
   371  		sgnext = sglist.waitlink
   372  		sglist.waitlink = nil
   373  		releaseSudog(sglist)
   374  		sglist = sgnext
   375  	}
   376  
   377  	if cas == nil {
   378  		throw("selectgo: bad wakeup")
   379  	}
   380  
   381  	c = cas.c
   382  
   383  	if debugSelect {
   384  		print("wait-return: cas0=", cas0, " c=", c, " cas=", cas, " send=", casi < nsends, "\n")
   385  	}
   386  
   387  	if casi < nsends {
   388  		if !caseSuccess {
   389  			goto sclose
   390  		}
   391  	} else {
   392  		recvOK = caseSuccess
   393  	}
   394  
   395  	if raceenabled {
   396  		if casi < nsends {
   397  			raceReadObjectPC(c.elemtype, cas.elem, casePC(casi), chansendpc)
   398  		} else if cas.elem != nil {
   399  			raceWriteObjectPC(c.elemtype, cas.elem, casePC(casi), chanrecvpc)
   400  		}
   401  	}
   402  	if msanenabled {
   403  		if casi < nsends {
   404  			msanread(cas.elem, c.elemtype.size)
   405  		} else if cas.elem != nil {
   406  			msanwrite(cas.elem, c.elemtype.size)
   407  		}
   408  	}
   409  	if asanenabled {
   410  		if casi < nsends {
   411  			asanread(cas.elem, c.elemtype.size)
   412  		} else if cas.elem != nil {
   413  			asanwrite(cas.elem, c.elemtype.size)
   414  		}
   415  	}
   416  
   417  	selunlock(scases, lockorder)
   418  	goto retc
   419  
   420  bufrecv:
   421  	// can receive from buffer
   422  	if raceenabled {
   423  		if cas.elem != nil {
   424  			raceWriteObjectPC(c.elemtype, cas.elem, casePC(casi), chanrecvpc)
   425  		}
   426  		racenotify(c, c.recvx, nil)
   427  	}
   428  	if msanenabled && cas.elem != nil {
   429  		msanwrite(cas.elem, c.elemtype.size)
   430  	}
   431  	if asanenabled && cas.elem != nil {
   432  		asanwrite(cas.elem, c.elemtype.size)
   433  	}
   434  	recvOK = true
   435  	qp = chanbuf(c, c.recvx)
   436  	if cas.elem != nil {
   437  		typedmemmove(c.elemtype, cas.elem, qp)
   438  	}
   439  	typedmemclr(c.elemtype, qp)
   440  	c.recvx++
   441  	if c.recvx == c.dataqsiz {
   442  		c.recvx = 0
   443  	}
   444  	c.qcount--
   445  	selunlock(scases, lockorder)
   446  	goto retc
   447  
   448  bufsend:
   449  	// can send to buffer
   450  	if raceenabled {
   451  		racenotify(c, c.sendx, nil)
   452  		raceReadObjectPC(c.elemtype, cas.elem, casePC(casi), chansendpc)
   453  	}
   454  	if msanenabled {
   455  		msanread(cas.elem, c.elemtype.size)
   456  	}
   457  	if asanenabled {
   458  		asanread(cas.elem, c.elemtype.size)
   459  	}
   460  	typedmemmove(c.elemtype, chanbuf(c, c.sendx), cas.elem)
   461  	c.sendx++
   462  	if c.sendx == c.dataqsiz {
   463  		c.sendx = 0
   464  	}
   465  	c.qcount++
   466  	selunlock(scases, lockorder)
   467  	goto retc
   468  
   469  recv:
   470  	// can receive from sleeping sender (sg)
   471  	recv(c, sg, cas.elem, func() { selunlock(scases, lockorder) }, 2)
   472  	if debugSelect {
   473  		print("syncrecv: cas0=", cas0, " c=", c, "\n")
   474  	}
   475  	recvOK = true
   476  	goto retc
   477  
   478  rclose:
   479  	// read at end of closed channel
   480  	selunlock(scases, lockorder)
   481  	recvOK = false
   482  	if cas.elem != nil {
   483  		typedmemclr(c.elemtype, cas.elem)
   484  	}
   485  	if raceenabled {
   486  		raceacquire(c.raceaddr())
   487  	}
   488  	goto retc
   489  
   490  send:
   491  	// can send to a sleeping receiver (sg)
   492  	if raceenabled {
   493  		raceReadObjectPC(c.elemtype, cas.elem, casePC(casi), chansendpc)
   494  	}
   495  	if msanenabled {
   496  		msanread(cas.elem, c.elemtype.size)
   497  	}
   498  	if asanenabled {
   499  		asanread(cas.elem, c.elemtype.size)
   500  	}
   501  	send(c, sg, cas.elem, func() { selunlock(scases, lockorder) }, 2)
   502  	if debugSelect {
   503  		print("syncsend: cas0=", cas0, " c=", c, "\n")
   504  	}
   505  	goto retc
   506  
   507  retc:
   508  	if caseReleaseTime > 0 {
   509  		blockevent(caseReleaseTime-t0, 1)
   510  	}
   511  	return casi, recvOK
   512  
   513  sclose:
   514  	// send on closed channel
   515  	selunlock(scases, lockorder)
   516  	panic(plainError("send on closed channel"))
   517  }
   518  
   519  func (c *hchan) sortkey() uintptr {
   520  	return uintptr(unsafe.Pointer(c))
   521  }
   522  
   523  // A runtimeSelect is a single case passed to rselect.
   524  // This must match ../reflect/value.go:/runtimeSelect
   525  type runtimeSelect struct {
   526  	dir selectDir
   527  	typ unsafe.Pointer // channel type (not used here)
   528  	ch  *hchan         // channel
   529  	val unsafe.Pointer // ptr to data (SendDir) or ptr to receive buffer (RecvDir)
   530  }
   531  
   532  // These values must match ../reflect/value.go:/SelectDir.
   533  type selectDir int
   534  
   535  const (
   536  	_             selectDir = iota
   537  	selectSend              // case Chan <- Send
   538  	selectRecv              // case <-Chan:
   539  	selectDefault           // default
   540  )
   541  
   542  //go:linkname reflect_rselect reflect.rselect
   543  func reflect_rselect(cases []runtimeSelect) (int, bool) {
   544  	if len(cases) == 0 {
   545  		block()
   546  	}
   547  	sel := make([]scase, len(cases))
   548  	orig := make([]int, len(cases))
   549  	nsends, nrecvs := 0, 0
   550  	dflt := -1
   551  	for i, rc := range cases {
   552  		var j int
   553  		switch rc.dir {
   554  		case selectDefault:
   555  			dflt = i
   556  			continue
   557  		case selectSend:
   558  			j = nsends
   559  			nsends++
   560  		case selectRecv:
   561  			nrecvs++
   562  			j = len(cases) - nrecvs
   563  		}
   564  
   565  		sel[j] = scase{c: rc.ch, elem: rc.val}
   566  		orig[j] = i
   567  	}
   568  
   569  	// Only a default case.
   570  	if nsends+nrecvs == 0 {
   571  		return dflt, false
   572  	}
   573  
   574  	// Compact sel and orig if necessary.
   575  	if nsends+nrecvs < len(cases) {
   576  		copy(sel[nsends:], sel[len(cases)-nrecvs:])
   577  		copy(orig[nsends:], orig[len(cases)-nrecvs:])
   578  	}
   579  
   580  	order := make([]uint16, 2*(nsends+nrecvs))
   581  	var pc0 *uintptr
   582  	if raceenabled {
   583  		pcs := make([]uintptr, nsends+nrecvs)
   584  		for i := range pcs {
   585  			selectsetpc(&pcs[i])
   586  		}
   587  		pc0 = &pcs[0]
   588  	}
   589  
   590  	chosen, recvOK := selectgo(&sel[0], &order[0], pc0, nsends, nrecvs, dflt == -1)
   591  
   592  	// Translate chosen back to caller's ordering.
   593  	if chosen < 0 {
   594  		chosen = dflt
   595  	} else {
   596  		chosen = orig[chosen]
   597  	}
   598  	return chosen, recvOK
   599  }
   600  
   601  func (q *waitq) dequeueSudoG(sgp *sudog) {
   602  	x := sgp.prev
   603  	y := sgp.next
   604  	if x != nil {
   605  		if y != nil {
   606  			// middle of queue
   607  			x.next = y
   608  			y.prev = x
   609  			sgp.next = nil
   610  			sgp.prev = nil
   611  			return
   612  		}
   613  		// end of queue
   614  		x.next = nil
   615  		q.last = x
   616  		sgp.prev = nil
   617  		return
   618  	}
   619  	if y != nil {
   620  		// start of queue
   621  		y.prev = nil
   622  		q.first = y
   623  		sgp.next = nil
   624  		return
   625  	}
   626  
   627  	// x==y==nil. Either sgp is the only element in the queue,
   628  	// or it has already been removed. Use q.first to disambiguate.
   629  	if q.first == sgp {
   630  		q.first = nil
   631  		q.last = nil
   632  	}
   633  }
   634
View as plain text