runtime: use two-level list for semaphore address search in semaRoot

If there are many goroutines contending for two different locks
and both locks hash to the same semaRoot, the scans to find the
goroutines for a particular lock can end up being O(n), making
n lock acquisitions quadratic.

As long as only one actively-used lock hashes to each semaRoot
there's no problem, since the list operations in that case are O(1).
But when the second actively-used lock hits the same semaRoot,
then scans for entries with for a given lock have to scan over the
entries for the other lock.

Fix this problem by changing the semaRoot to hold only one sudog
per unique address. In the running example, this drops the length of
that list from O(n) to 2. Then attach other goroutines waiting on the
same address to a separate list headed by the sudog in the semaRoot list.
Those "same address list" operations are still O(1), so now the
example from above works much better.

There is still an assumption here that in real programs you don't have
many many goroutines queueing up on many many distinct addresses.
If we end up with that problem, we can replace the top-level list with
a treap.

Fixes #17953.

Change-Id: I78c5b1a5053845275ab31686038aa4f6db5720b2
Reviewed-on: https://go-review.googlesource.com/36792
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

src/runtime/runtime2.go

@@ -270,7 +270,7 @@ type gobuf struct {
 type sudog struct {
 	// The following fields are protected by the hchan.lock of the
 	// channel this sudog is blocking on. shrinkstack depends on
-	// this.
+	// this for sudogs involved in channel ops.
 
 	g          *g
 	selectdone *uint32 // CAS to 1 to win select race (may point to stack)
@@ -279,12 +279,15 @@ type sudog struct {
 	elem       unsafe.Pointer // data element (may point to stack)
 
 	// The following fields are never accessed concurrently.
-	// waitlink is only accessed by g.
+	// For channels, waitlink is only accessed by g.
+	// For semaphores, all fields (including the ones above)
+	// are only accessed when holding a semaRoot lock.
 
 	acquiretime int64
 	releasetime int64
 	ticket      uint32
-	waitlink    *sudog // g.waiting list
+	waitlink    *sudog // g.waiting list or semaRoot
+	waittail    *sudog // semaRoot
 	c           *hchan // channel
 }
 

src/runtime/sema.go

@@ -27,6 +27,20 @@ import (
 
 // Asynchronous semaphore for sync.Mutex.
 
+// A semaRoot holds a linked list of sudog with distinct addresses (s.elem).
+// Each of those sudog may in turn point (through s.waitlink) to a list
+// of other sudogs waiting on the same address.
+// The operations on the inner lists of sudogs with the same address
+// are all O(1). Only the scanning of the top-level semaRoot list is O(n),
+// where n is the number of distinct addresses with goroutines blocked
+// on them that hash to the given semaRoot.
+// In systems with many goroutines, most queue up on a few addresses,
+// so the linear search across unique addresses is probably OK.
+// At least, we'll use this until it's not.
+// The next step is probably to make the top-level list a treap instead
+// of a linked list.
+// See golang.org/issue/17953 for a program that worked badly
+// before we introduced the second level of list.
 type semaRoot struct {
 	lock  mutex
 	head  *sudog
@@ -157,22 +171,10 @@ func semrelease(addr *uint32) {
 		unlock(&root.lock)
 		return
 	}
-	s := root.head
-	for ; s != nil; s = s.next {
-		if s.elem == unsafe.Pointer(addr) {
-			atomic.Xadd(&root.nwait, -1)
-			root.dequeue(s)
-			break
-		}
-	}
+	s, t0 := root.dequeue(addr)
 	if s != nil {
+		atomic.Xadd(&root.nwait, -1)
 		if s.acquiretime != 0 {
-			t0 := cputicks()
-			for x := root.head; x != nil; x = x.next {
-				if x.elem == unsafe.Pointer(addr) {
-					x.acquiretime = t0
-				}
-			}
 			mutexevent(t0-s.acquiretime, 3)
 		}
 	}
@@ -198,9 +200,26 @@ func cansemacquire(addr *uint32) bool {
 	}
 }
 
+// queue adds s to the blocked goroutines in semaRoot.
 func (root *semaRoot) queue(addr *uint32, s *sudog) {
 	s.g = getg()
 	s.elem = unsafe.Pointer(addr)
+
+	for t := root.head; t != nil; t = t.next {
+		if t.elem == unsafe.Pointer(addr) {
+			// Already have addr in list; add s to end of per-addr list.
+			if t.waittail == nil {
+				t.waitlink = s
+			} else {
+				t.waittail.waitlink = s
+			}
+			t.waittail = s
+			s.waitlink = nil
+			return
+		}
+	}
+
+	// Add s as new entry in list of unique addrs.
 	s.next = nil
 	s.prev = root.tail
 	if root.tail != nil {
@@ -211,7 +230,48 @@ func (root *semaRoot) queue(addr *uint32, s *sudog) {
 	root.tail = s
 }
 
-func (root *semaRoot) dequeue(s *sudog) {
+// dequeue searches for and finds the first goroutine
+// in semaRoot blocked on addr.
+// If the sudog was being profiled, dequeue returns the time
+// at which it was woken up as now. Otherwise now is 0.
+func (root *semaRoot) dequeue(addr *uint32) (found *sudog, now int64) {
+	s := root.head
+	for ; s != nil; s = s.next {
+		if s.elem == unsafe.Pointer(addr) {
+			goto Found
+		}
+	}
+	return nil, 0
+
+Found:
+	now = int64(0)
+	if s.acquiretime != 0 {
+		now = cputicks()
+	}
+	if t := s.waitlink; t != nil {
+		// Substitute t, also waiting on addr, for s in root list of unique addrs.
+		t.prev = s.prev
+		t.next = s.next
+		if t.prev != nil {
+			t.prev.next = t
+		} else {
+			root.head = t
+		}
+		if t.next != nil {
+			t.next.prev = t
+		} else {
+			root.tail = t
+		}
+		if t.waitlink != nil {
+			t.waittail = s.waittail
+		} else {
+			t.waittail = nil
+		}
+		t.acquiretime = now
+		s.waitlink = nil
+		s.waittail = nil
+	} else {
+		// Remove s from list.
 		if s.next != nil {
 			s.next.prev = s.prev
 		} else {
@@ -222,9 +282,11 @@ func (root *semaRoot) dequeue(s *sudog) {
 		} else {
 			root.head = s.next
 		}
+	}
 	s.elem = nil
 	s.next = nil
 	s.prev = nil
+	return s, now
 }
 
 // notifyList is a ticket-based notification list used to implement sync.Cond.