Black Lives Matter. Support the Equal Justice Initiative.

Source file src/runtime/sema.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  // Semaphore implementation exposed to Go.
     6  // Intended use is provide a sleep and wakeup
     7  // primitive that can be used in the contended case
     8  // of other synchronization primitives.
     9  // Thus it targets the same goal as Linux's futex,
    10  // but it has much simpler semantics.
    11  //
    12  // That is, don't think of these as semaphores.
    13  // Think of them as a way to implement sleep and wakeup
    14  // such that every sleep is paired with a single wakeup,
    15  // even if, due to races, the wakeup happens before the sleep.
    16  //
    17  // See Mullender and Cox, ``Semaphores in Plan 9,''
    18  // https://swtch.com/semaphore.pdf
    19  
    20  package runtime
    21  
    22  import (
    23  	"internal/cpu"
    24  	"runtime/internal/atomic"
    25  	"unsafe"
    26  )
    27  
    28  // Asynchronous semaphore for sync.Mutex.
    29  
    30  // A semaRoot holds a balanced tree of sudog with distinct addresses (s.elem).
    31  // Each of those sudog may in turn point (through s.waitlink) to a list
    32  // of other sudogs waiting on the same address.
    33  // The operations on the inner lists of sudogs with the same address
    34  // are all O(1). The scanning of the top-level semaRoot list is O(log n),
    35  // where n is the number of distinct addresses with goroutines blocked
    36  // on them that hash to the given semaRoot.
    37  // See golang.org/issue/17953 for a program that worked badly
    38  // before we introduced the second level of list, and test/locklinear.go
    39  // for a test that exercises this.
    40  type semaRoot struct {
    41  	lock  mutex
    42  	treap *sudog // root of balanced tree of unique waiters.
    43  	nwait uint32 // Number of waiters. Read w/o the lock.
    44  }
    45  
    46  // Prime to not correlate with any user patterns.
    47  const semTabSize = 251
    48  
    49  var semtable [semTabSize]struct {
    50  	root semaRoot
    51  	pad  [cpu.CacheLinePadSize - unsafe.Sizeof(semaRoot{})]byte
    52  }
    53  
    54  //go:linkname sync_runtime_Semacquire sync.runtime_Semacquire
    55  func sync_runtime_Semacquire(addr *uint32) {
    56  	semacquire1(addr, false, semaBlockProfile, 0)
    57  }
    58  
    59  //go:linkname poll_runtime_Semacquire internal/poll.runtime_Semacquire
    60  func poll_runtime_Semacquire(addr *uint32) {
    61  	semacquire1(addr, false, semaBlockProfile, 0)
    62  }
    63  
    64  //go:linkname sync_runtime_Semrelease sync.runtime_Semrelease
    65  func sync_runtime_Semrelease(addr *uint32, handoff bool, skipframes int) {
    66  	semrelease1(addr, handoff, skipframes)
    67  }
    68  
    69  //go:linkname sync_runtime_SemacquireMutex sync.runtime_SemacquireMutex
    70  func sync_runtime_SemacquireMutex(addr *uint32, lifo bool, skipframes int) {
    71  	semacquire1(addr, lifo, semaBlockProfile|semaMutexProfile, skipframes)
    72  }
    73  
    74  //go:linkname poll_runtime_Semrelease internal/poll.runtime_Semrelease
    75  func poll_runtime_Semrelease(addr *uint32) {
    76  	semrelease(addr)
    77  }
    78  
    79  func readyWithTime(s *sudog, traceskip int) {
    80  	if s.releasetime != 0 {
    81  		s.releasetime = cputicks()
    82  	}
    83  	goready(s.g, traceskip)
    84  }
    85  
    86  type semaProfileFlags int
    87  
    88  const (
    89  	semaBlockProfile semaProfileFlags = 1 << iota
    90  	semaMutexProfile
    91  )
    92  
    93  // Called from runtime.
    94  func semacquire(addr *uint32) {
    95  	semacquire1(addr, false, 0, 0)
    96  }
    97  
    98  func semacquire1(addr *uint32, lifo bool, profile semaProfileFlags, skipframes int) {
    99  	gp := getg()
   100  	if gp != gp.m.curg {
   101  		throw("semacquire not on the G stack")
   102  	}
   103  
   104  	// Easy case.
   105  	if cansemacquire(addr) {
   106  		return
   107  	}
   108  
   109  	// Harder case:
   110  	//	increment waiter count
   111  	//	try cansemacquire one more time, return if succeeded
   112  	//	enqueue itself as a waiter
   113  	//	sleep
   114  	//	(waiter descriptor is dequeued by signaler)
   115  	s := acquireSudog()
   116  	root := semroot(addr)
   117  	t0 := int64(0)
   118  	s.releasetime = 0
   119  	s.acquiretime = 0
   120  	s.ticket = 0
   121  	if profile&semaBlockProfile != 0 && blockprofilerate > 0 {
   122  		t0 = cputicks()
   123  		s.releasetime = -1
   124  	}
   125  	if profile&semaMutexProfile != 0 && mutexprofilerate > 0 {
   126  		if t0 == 0 {
   127  			t0 = cputicks()
   128  		}
   129  		s.acquiretime = t0
   130  	}
   131  	for {
   132  		lockWithRank(&root.lock, lockRankRoot)
   133  		// Add ourselves to nwait to disable "easy case" in semrelease.
   134  		atomic.Xadd(&root.nwait, 1)
   135  		// Check cansemacquire to avoid missed wakeup.
   136  		if cansemacquire(addr) {
   137  			atomic.Xadd(&root.nwait, -1)
   138  			unlock(&root.lock)
   139  			break
   140  		}
   141  		// Any semrelease after the cansemacquire knows we're waiting
   142  		// (we set nwait above), so go to sleep.
   143  		root.queue(addr, s, lifo)
   144  		goparkunlock(&root.lock, waitReasonSemacquire, traceEvGoBlockSync, 4+skipframes)
   145  		if s.ticket != 0 || cansemacquire(addr) {
   146  			break
   147  		}
   148  	}
   149  	if s.releasetime > 0 {
   150  		blockevent(s.releasetime-t0, 3+skipframes)
   151  	}
   152  	releaseSudog(s)
   153  }
   154  
   155  func semrelease(addr *uint32) {
   156  	semrelease1(addr, false, 0)
   157  }
   158  
   159  func semrelease1(addr *uint32, handoff bool, skipframes int) {
   160  	root := semroot(addr)
   161  	atomic.Xadd(addr, 1)
   162  
   163  	// Easy case: no waiters?
   164  	// This check must happen after the xadd, to avoid a missed wakeup
   165  	// (see loop in semacquire).
   166  	if atomic.Load(&root.nwait) == 0 {
   167  		return
   168  	}
   169  
   170  	// Harder case: search for a waiter and wake it.
   171  	lockWithRank(&root.lock, lockRankRoot)
   172  	if atomic.Load(&root.nwait) == 0 {
   173  		// The count is already consumed by another goroutine,
   174  		// so no need to wake up another goroutine.
   175  		unlock(&root.lock)
   176  		return
   177  	}
   178  	s, t0 := root.dequeue(addr)
   179  	if s != nil {
   180  		atomic.Xadd(&root.nwait, -1)
   181  	}
   182  	unlock(&root.lock)
   183  	if s != nil { // May be slow or even yield, so unlock first
   184  		acquiretime := s.acquiretime
   185  		if acquiretime != 0 {
   186  			mutexevent(t0-acquiretime, 3+skipframes)
   187  		}
   188  		if s.ticket != 0 {
   189  			throw("corrupted semaphore ticket")
   190  		}
   191  		if handoff && cansemacquire(addr) {
   192  			s.ticket = 1
   193  		}
   194  		readyWithTime(s, 5+skipframes)
   195  		if s.ticket == 1 && getg().m.locks == 0 {
   196  			// Direct G handoff
   197  			// readyWithTime has added the waiter G as runnext in the
   198  			// current P; we now call the scheduler so that we start running
   199  			// the waiter G immediately.
   200  			// Note that waiter inherits our time slice: this is desirable
   201  			// to avoid having a highly contended semaphore hog the P
   202  			// indefinitely. goyield is like Gosched, but it emits a
   203  			// "preempted" trace event instead and, more importantly, puts
   204  			// the current G on the local runq instead of the global one.
   205  			// We only do this in the starving regime (handoff=true), as in
   206  			// the non-starving case it is possible for a different waiter
   207  			// to acquire the semaphore while we are yielding/scheduling,
   208  			// and this would be wasteful. We wait instead to enter starving
   209  			// regime, and then we start to do direct handoffs of ticket and
   210  			// P.
   211  			// See issue 33747 for discussion.
   212  			goyield()
   213  		}
   214  	}
   215  }
   216  
   217  func semroot(addr *uint32) *semaRoot {
   218  	return &semtable[(uintptr(unsafe.Pointer(addr))>>3)%semTabSize].root
   219  }
   220  
   221  func cansemacquire(addr *uint32) bool {
   222  	for {
   223  		v := atomic.Load(addr)
   224  		if v == 0 {
   225  			return false
   226  		}
   227  		if atomic.Cas(addr, v, v-1) {
   228  			return true
   229  		}
   230  	}
   231  }
   232  
   233  // queue adds s to the blocked goroutines in semaRoot.
   234  func (root *semaRoot) queue(addr *uint32, s *sudog, lifo bool) {
   235  	s.g = getg()
   236  	s.elem = unsafe.Pointer(addr)
   237  	s.next = nil
   238  	s.prev = nil
   239  
   240  	var last *sudog
   241  	pt := &root.treap
   242  	for t := *pt; t != nil; t = *pt {
   243  		if t.elem == unsafe.Pointer(addr) {
   244  			// Already have addr in list.
   245  			if lifo {
   246  				// Substitute s in t's place in treap.
   247  				*pt = s
   248  				s.ticket = t.ticket
   249  				s.acquiretime = t.acquiretime
   250  				s.parent = t.parent
   251  				s.prev = t.prev
   252  				s.next = t.next
   253  				if s.prev != nil {
   254  					s.prev.parent = s
   255  				}
   256  				if s.next != nil {
   257  					s.next.parent = s
   258  				}
   259  				// Add t first in s's wait list.
   260  				s.waitlink = t
   261  				s.waittail = t.waittail
   262  				if s.waittail == nil {
   263  					s.waittail = t
   264  				}
   265  				t.parent = nil
   266  				t.prev = nil
   267  				t.next = nil
   268  				t.waittail = nil
   269  			} else {
   270  				// Add s to end of t's wait list.
   271  				if t.waittail == nil {
   272  					t.waitlink = s
   273  				} else {
   274  					t.waittail.waitlink = s
   275  				}
   276  				t.waittail = s
   277  				s.waitlink = nil
   278  			}
   279  			return
   280  		}
   281  		last = t
   282  		if uintptr(unsafe.Pointer(addr)) < uintptr(t.elem) {
   283  			pt = &t.prev
   284  		} else {
   285  			pt = &t.next
   286  		}
   287  	}
   288  
   289  	// Add s as new leaf in tree of unique addrs.
   290  	// The balanced tree is a treap using ticket as the random heap priority.
   291  	// That is, it is a binary tree ordered according to the elem addresses,
   292  	// but then among the space of possible binary trees respecting those
   293  	// addresses, it is kept balanced on average by maintaining a heap ordering
   294  	// on the ticket: s.ticket <= both s.prev.ticket and s.next.ticket.
   295  	// https://en.wikipedia.org/wiki/Treap
   296  	// https://faculty.washington.edu/aragon/pubs/rst89.pdf
   297  	//
   298  	// s.ticket compared with zero in couple of places, therefore set lowest bit.
   299  	// It will not affect treap's quality noticeably.
   300  	s.ticket = fastrand() | 1
   301  	s.parent = last
   302  	*pt = s
   303  
   304  	// Rotate up into tree according to ticket (priority).
   305  	for s.parent != nil && s.parent.ticket > s.ticket {
   306  		if s.parent.prev == s {
   307  			root.rotateRight(s.parent)
   308  		} else {
   309  			if s.parent.next != s {
   310  				panic("semaRoot queue")
   311  			}
   312  			root.rotateLeft(s.parent)
   313  		}
   314  	}
   315  }
   316  
   317  // dequeue searches for and finds the first goroutine
   318  // in semaRoot blocked on addr.
   319  // If the sudog was being profiled, dequeue returns the time
   320  // at which it was woken up as now. Otherwise now is 0.
   321  func (root *semaRoot) dequeue(addr *uint32) (found *sudog, now int64) {
   322  	ps := &root.treap
   323  	s := *ps
   324  	for ; s != nil; s = *ps {
   325  		if s.elem == unsafe.Pointer(addr) {
   326  			goto Found
   327  		}
   328  		if uintptr(unsafe.Pointer(addr)) < uintptr(s.elem) {
   329  			ps = &s.prev
   330  		} else {
   331  			ps = &s.next
   332  		}
   333  	}
   334  	return nil, 0
   335  
   336  Found:
   337  	now = int64(0)
   338  	if s.acquiretime != 0 {
   339  		now = cputicks()
   340  	}
   341  	if t := s.waitlink; t != nil {
   342  		// Substitute t, also waiting on addr, for s in root tree of unique addrs.
   343  		*ps = t
   344  		t.ticket = s.ticket
   345  		t.parent = s.parent
   346  		t.prev = s.prev
   347  		if t.prev != nil {
   348  			t.prev.parent = t
   349  		}
   350  		t.next = s.next
   351  		if t.next != nil {
   352  			t.next.parent = t
   353  		}
   354  		if t.waitlink != nil {
   355  			t.waittail = s.waittail
   356  		} else {
   357  			t.waittail = nil
   358  		}
   359  		t.acquiretime = now
   360  		s.waitlink = nil
   361  		s.waittail = nil
   362  	} else {
   363  		// Rotate s down to be leaf of tree for removal, respecting priorities.
   364  		for s.next != nil || s.prev != nil {
   365  			if s.next == nil || s.prev != nil && s.prev.ticket < s.next.ticket {
   366  				root.rotateRight(s)
   367  			} else {
   368  				root.rotateLeft(s)
   369  			}
   370  		}
   371  		// Remove s, now a leaf.
   372  		if s.parent != nil {
   373  			if s.parent.prev == s {
   374  				s.parent.prev = nil
   375  			} else {
   376  				s.parent.next = nil
   377  			}
   378  		} else {
   379  			root.treap = nil
   380  		}
   381  	}
   382  	s.parent = nil
   383  	s.elem = nil
   384  	s.next = nil
   385  	s.prev = nil
   386  	s.ticket = 0
   387  	return s, now
   388  }
   389  
   390  // rotateLeft rotates the tree rooted at node x.
   391  // turning (x a (y b c)) into (y (x a b) c).
   392  func (root *semaRoot) rotateLeft(x *sudog) {
   393  	// p -> (x a (y b c))
   394  	p := x.parent
   395  	y := x.next
   396  	b := y.prev
   397  
   398  	y.prev = x
   399  	x.parent = y
   400  	x.next = b
   401  	if b != nil {
   402  		b.parent = x
   403  	}
   404  
   405  	y.parent = p
   406  	if p == nil {
   407  		root.treap = y
   408  	} else if p.prev == x {
   409  		p.prev = y
   410  	} else {
   411  		if p.next != x {
   412  			throw("semaRoot rotateLeft")
   413  		}
   414  		p.next = y
   415  	}
   416  }
   417  
   418  // rotateRight rotates the tree rooted at node y.
   419  // turning (y (x a b) c) into (x a (y b c)).
   420  func (root *semaRoot) rotateRight(y *sudog) {
   421  	// p -> (y (x a b) c)
   422  	p := y.parent
   423  	x := y.prev
   424  	b := x.next
   425  
   426  	x.next = y
   427  	y.parent = x
   428  	y.prev = b
   429  	if b != nil {
   430  		b.parent = y
   431  	}
   432  
   433  	x.parent = p
   434  	if p == nil {
   435  		root.treap = x
   436  	} else if p.prev == y {
   437  		p.prev = x
   438  	} else {
   439  		if p.next != y {
   440  			throw("semaRoot rotateRight")
   441  		}
   442  		p.next = x
   443  	}
   444  }
   445  
   446  // notifyList is a ticket-based notification list used to implement sync.Cond.
   447  //
   448  // It must be kept in sync with the sync package.
   449  type notifyList struct {
   450  	// wait is the ticket number of the next waiter. It is atomically
   451  	// incremented outside the lock.
   452  	wait uint32
   453  
   454  	// notify is the ticket number of the next waiter to be notified. It can
   455  	// be read outside the lock, but is only written to with lock held.
   456  	//
   457  	// Both wait & notify can wrap around, and such cases will be correctly
   458  	// handled as long as their "unwrapped" difference is bounded by 2^31.
   459  	// For this not to be the case, we'd need to have 2^31+ goroutines
   460  	// blocked on the same condvar, which is currently not possible.
   461  	notify uint32
   462  
   463  	// List of parked waiters.
   464  	lock mutex
   465  	head *sudog
   466  	tail *sudog
   467  }
   468  
   469  // less checks if a < b, considering a & b running counts that may overflow the
   470  // 32-bit range, and that their "unwrapped" difference is always less than 2^31.
   471  func less(a, b uint32) bool {
   472  	return int32(a-b) < 0
   473  }
   474  
   475  // notifyListAdd adds the caller to a notify list such that it can receive
   476  // notifications. The caller must eventually call notifyListWait to wait for
   477  // such a notification, passing the returned ticket number.
   478  //go:linkname notifyListAdd sync.runtime_notifyListAdd
   479  func notifyListAdd(l *notifyList) uint32 {
   480  	// This may be called concurrently, for example, when called from
   481  	// sync.Cond.Wait while holding a RWMutex in read mode.
   482  	return atomic.Xadd(&l.wait, 1) - 1
   483  }
   484  
   485  // notifyListWait waits for a notification. If one has been sent since
   486  // notifyListAdd was called, it returns immediately. Otherwise, it blocks.
   487  //go:linkname notifyListWait sync.runtime_notifyListWait
   488  func notifyListWait(l *notifyList, t uint32) {
   489  	lockWithRank(&l.lock, lockRankNotifyList)
   490  
   491  	// Return right away if this ticket has already been notified.
   492  	if less(t, l.notify) {
   493  		unlock(&l.lock)
   494  		return
   495  	}
   496  
   497  	// Enqueue itself.
   498  	s := acquireSudog()
   499  	s.g = getg()
   500  	s.ticket = t
   501  	s.releasetime = 0
   502  	t0 := int64(0)
   503  	if blockprofilerate > 0 {
   504  		t0 = cputicks()
   505  		s.releasetime = -1
   506  	}
   507  	if l.tail == nil {
   508  		l.head = s
   509  	} else {
   510  		l.tail.next = s
   511  	}
   512  	l.tail = s
   513  	goparkunlock(&l.lock, waitReasonSyncCondWait, traceEvGoBlockCond, 3)
   514  	if t0 != 0 {
   515  		blockevent(s.releasetime-t0, 2)
   516  	}
   517  	releaseSudog(s)
   518  }
   519  
   520  // notifyListNotifyAll notifies all entries in the list.
   521  //go:linkname notifyListNotifyAll sync.runtime_notifyListNotifyAll
   522  func notifyListNotifyAll(l *notifyList) {
   523  	// Fast-path: if there are no new waiters since the last notification
   524  	// we don't need to acquire the lock.
   525  	if atomic.Load(&l.wait) == atomic.Load(&l.notify) {
   526  		return
   527  	}
   528  
   529  	// Pull the list out into a local variable, waiters will be readied
   530  	// outside the lock.
   531  	lockWithRank(&l.lock, lockRankNotifyList)
   532  	s := l.head
   533  	l.head = nil
   534  	l.tail = nil
   535  
   536  	// Update the next ticket to be notified. We can set it to the current
   537  	// value of wait because any previous waiters are already in the list
   538  	// or will notice that they have already been notified when trying to
   539  	// add themselves to the list.
   540  	atomic.Store(&l.notify, atomic.Load(&l.wait))
   541  	unlock(&l.lock)
   542  
   543  	// Go through the local list and ready all waiters.
   544  	for s != nil {
   545  		next := s.next
   546  		s.next = nil
   547  		readyWithTime(s, 4)
   548  		s = next
   549  	}
   550  }
   551  
   552  // notifyListNotifyOne notifies one entry in the list.
   553  //go:linkname notifyListNotifyOne sync.runtime_notifyListNotifyOne
   554  func notifyListNotifyOne(l *notifyList) {
   555  	// Fast-path: if there are no new waiters since the last notification
   556  	// we don't need to acquire the lock at all.
   557  	if atomic.Load(&l.wait) == atomic.Load(&l.notify) {
   558  		return
   559  	}
   560  
   561  	lockWithRank(&l.lock, lockRankNotifyList)
   562  
   563  	// Re-check under the lock if we need to do anything.
   564  	t := l.notify
   565  	if t == atomic.Load(&l.wait) {
   566  		unlock(&l.lock)
   567  		return
   568  	}
   569  
   570  	// Update the next notify ticket number.
   571  	atomic.Store(&l.notify, t+1)
   572  
   573  	// Try to find the g that needs to be notified.
   574  	// If it hasn't made it to the list yet we won't find it,
   575  	// but it won't park itself once it sees the new notify number.
   576  	//
   577  	// This scan looks linear but essentially always stops quickly.
   578  	// Because g's queue separately from taking numbers,
   579  	// there may be minor reorderings in the list, but we
   580  	// expect the g we're looking for to be near the front.
   581  	// The g has others in front of it on the list only to the
   582  	// extent that it lost the race, so the iteration will not
   583  	// be too long. This applies even when the g is missing:
   584  	// it hasn't yet gotten to sleep and has lost the race to
   585  	// the (few) other g's that we find on the list.
   586  	for p, s := (*sudog)(nil), l.head; s != nil; p, s = s, s.next {
   587  		if s.ticket == t {
   588  			n := s.next
   589  			if p != nil {
   590  				p.next = n
   591  			} else {
   592  				l.head = n
   593  			}
   594  			if n == nil {
   595  				l.tail = p
   596  			}
   597  			unlock(&l.lock)
   598  			s.next = nil
   599  			readyWithTime(s, 4)
   600  			return
   601  		}
   602  	}
   603  	unlock(&l.lock)
   604  }
   605  
   606  //go:linkname notifyListCheck sync.runtime_notifyListCheck
   607  func notifyListCheck(sz uintptr) {
   608  	if sz != unsafe.Sizeof(notifyList{}) {
   609  		print("runtime: bad notifyList size - sync=", sz, " runtime=", unsafe.Sizeof(notifyList{}), "\n")
   610  		throw("bad notifyList size")
   611  	}
   612  }
   613  
   614  //go:linkname sync_nanotime sync.runtime_nanotime
   615  func sync_nanotime() int64 {
   616  	return nanotime()
   617  }
   618  

View as plain text