Merge branch 'master' into pr_lazy_fpsimd_2

13 files changed, 1684 insertions, 0 deletions

diff --git a/pkg/syncevent/BUILD b/pkg/syncevent/BUILD
new file mode 100644
index 000000000..0500a22cf
--- /dev/null
+++ b/pkg/syncevent/BUILD
@@ -0,0 +1,39 @@
+load("//tools:defs.bzl", "go_library", "go_test")
+
+licenses(["notice"])
+
+go_library(
+    name = "syncevent",
+    srcs = [
+        "broadcaster.go",
+        "receiver.go",
+        "source.go",
+        "syncevent.go",
+        "waiter_amd64.s",
+        "waiter_arm64.s",
+        "waiter_asm_unsafe.go",
+        "waiter_noasm_unsafe.go",
+        "waiter_unsafe.go",
+    ],
+    visibility = ["//:sandbox"],
+    deps = [
+        "//pkg/atomicbitops",
+        "//pkg/sync",
+    ],
+)
+
+go_test(
+    name = "syncevent_test",
+    size = "small",
+    srcs = [
+        "broadcaster_test.go",
+        "syncevent_example_test.go",
+        "waiter_test.go",
+    ],
+    library = ":syncevent",
+    deps = [
+        "//pkg/sleep",
+        "//pkg/sync",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/syncevent/broadcaster.go b/pkg/syncevent/broadcaster.go
new file mode 100644
index 000000000..4bff59e7d
--- /dev/null
+++ b/pkg/syncevent/broadcaster.go
@@ -0,0 +1,218 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syncevent
+
+import (
+	"gvisor.dev/gvisor/pkg/sync"
+)
+
+// Broadcaster is an implementation of Source that supports any number of
+// subscribed Receivers.
+//
+// The zero value of Broadcaster is valid and has no subscribed Receivers.
+// Broadcaster is not copyable by value.
+//
+// All Broadcaster methods may be called concurrently from multiple goroutines.
+type Broadcaster struct {
+	// Broadcaster is implemented as a hash table where keys are assigned by
+	// the Broadcaster and returned as SubscriptionIDs, making it safe to use
+	// the identity function for hashing. The hash table resolves collisions
+	// using linear probing and features Robin Hood insertion and backward
+	// shift deletion in order to support a relatively high load factor
+	// efficiently, which matters since the cost of Broadcast is linear in the
+	// size of the table.
+
+	// mu protects the following fields.
+	mu sync.Mutex
+
+	// Invariants: len(table) is 0 or a power of 2.
+	table []broadcasterSlot
+
+	// load is the number of entries in table with receiver != nil.
+	load int
+
+	lastID SubscriptionID
+}
+
+type broadcasterSlot struct {
+	// Invariants: If receiver == nil, then filter == NoEvents and id == 0.
+	// Otherwise, id != 0.
+	receiver *Receiver
+	filter   Set
+	id       SubscriptionID
+}
+
+const (
+	broadcasterMinNonZeroTableSize = 2 // must be a power of 2 > 1
+
+	broadcasterMaxLoadNum = 13
+	broadcasterMaxLoadDen = 16
+)
+
+// SubscribeEvents implements Source.SubscribeEvents.
+func (b *Broadcaster) SubscribeEvents(r *Receiver, filter Set) SubscriptionID {
+	b.mu.Lock()
+
+	// Assign an ID for this subscription.
+	b.lastID++
+	id := b.lastID
+
+	// Expand the table if over the maximum load factor:
+	//
+	//          load / len(b.table) > broadcasterMaxLoadNum / broadcasterMaxLoadDen
+	// load * broadcasterMaxLoadDen > broadcasterMaxLoadNum * len(b.table)
+	b.load++
+	if (b.load * broadcasterMaxLoadDen) > (broadcasterMaxLoadNum * len(b.table)) {
+		// Double the number of slots in the new table.
+		newlen := broadcasterMinNonZeroTableSize
+		if len(b.table) != 0 {
+			newlen = 2 * len(b.table)
+		}
+		if newlen <= cap(b.table) {
+			// Reuse excess capacity in the current table, moving entries not
+			// already in their first-probed positions to better ones.
+			newtable := b.table[:newlen]
+			newmask := uint64(newlen - 1)
+			for i := range b.table {
+				if b.table[i].receiver != nil && uint64(b.table[i].id)&newmask != uint64(i) {
+					entry := b.table[i]
+					b.table[i] = broadcasterSlot{}
+					broadcasterTableInsert(newtable, entry.id, entry.receiver, entry.filter)
+				}
+			}
+			b.table = newtable
+		} else {
+			newtable := make([]broadcasterSlot, newlen)
+			// Copy existing entries to the new table.
+			for i := range b.table {
+				if b.table[i].receiver != nil {
+					broadcasterTableInsert(newtable, b.table[i].id, b.table[i].receiver, b.table[i].filter)
+				}
+			}
+			// Switch to the new table.
+			b.table = newtable
+		}
+	}
+
+	broadcasterTableInsert(b.table, id, r, filter)
+	b.mu.Unlock()
+	return id
+}
+
+// Preconditions: table must not be full. len(table) is a power of 2.
+func broadcasterTableInsert(table []broadcasterSlot, id SubscriptionID, r *Receiver, filter Set) {
+	entry := broadcasterSlot{
+		receiver: r,
+		filter:   filter,
+		id:       id,
+	}
+	mask := uint64(len(table) - 1)
+	i := uint64(id) & mask
+	disp := uint64(0)
+	for {
+		if table[i].receiver == nil {
+			table[i] = entry
+			return
+		}
+		// If we've been displaced farther from our first-probed slot than the
+		// element stored in this one, swap elements and switch to inserting
+		// the replaced one. (This is Robin Hood insertion.)
+		slotDisp := (i - uint64(table[i].id)) & mask
+		if disp > slotDisp {
+			table[i], entry = entry, table[i]
+			disp = slotDisp
+		}
+		i = (i + 1) & mask
+		disp++
+	}
+}
+
+// UnsubscribeEvents implements Source.UnsubscribeEvents.
+func (b *Broadcaster) UnsubscribeEvents(id SubscriptionID) {
+	b.mu.Lock()
+
+	mask := uint64(len(b.table) - 1)
+	i := uint64(id) & mask
+	for {
+		if b.table[i].id == id {
+			// Found the element to remove. Move all subsequent elements
+			// backward until we either find an empty slot, or an element that
+			// is already in its first-probed slot. (This is backward shift
+			// deletion.)
+			for {
+				next := (i + 1) & mask
+				if b.table[next].receiver == nil {
+					break
+				}
+				if uint64(b.table[next].id)&mask == next {
+					break
+				}
+				b.table[i] = b.table[next]
+				i = next
+			}
+			b.table[i] = broadcasterSlot{}
+			break
+		}
+		i = (i + 1) & mask
+	}
+
+	// If a table 1/4 of the current size would still be at or under the
+	// maximum load factor (i.e. the current table size is at least two
+	// expansions bigger than necessary), halve the size of the table to reduce
+	// the cost of Broadcast. Since we are concerned with iteration time and
+	// not memory usage, reuse the existing slice to reduce future allocations
+	// from table re-expansion.
+	b.load--
+	if len(b.table) > broadcasterMinNonZeroTableSize && (b.load*(4*broadcasterMaxLoadDen)) <= (broadcasterMaxLoadNum*len(b.table)) {
+		newlen := len(b.table) / 2
+		newtable := b.table[:newlen]
+		for i := newlen; i < len(b.table); i++ {
+			if b.table[i].receiver != nil {
+				broadcasterTableInsert(newtable, b.table[i].id, b.table[i].receiver, b.table[i].filter)
+				b.table[i] = broadcasterSlot{}
+			}
+		}
+		b.table = newtable
+	}
+
+	b.mu.Unlock()
+}
+
+// Broadcast notifies all Receivers subscribed to the Broadcaster of the subset
+// of events to which they subscribed. The order in which Receivers are
+// notified is unspecified.
+func (b *Broadcaster) Broadcast(events Set) {
+	b.mu.Lock()
+	for i := range b.table {
+		if intersection := events & b.table[i].filter; intersection != 0 {
+			// We don't need to check if broadcasterSlot.receiver is nil, since
+			// if it is then broadcasterSlot.filter is 0.
+			b.table[i].receiver.Notify(intersection)
+		}
+	}
+	b.mu.Unlock()
+}
+
+// FilteredEvents returns the set of events for which Broadcast will notify at
+// least one Receiver, i.e. the union of filters for all subscribed Receivers.
+func (b *Broadcaster) FilteredEvents() Set {
+	var es Set
+	b.mu.Lock()
+	for i := range b.table {
+		es |= b.table[i].filter
+	}
+	b.mu.Unlock()
+	return es
+}
diff --git a/pkg/syncevent/broadcaster_test.go b/pkg/syncevent/broadcaster_test.go
new file mode 100644
index 000000000..e88779e23
--- /dev/null
+++ b/pkg/syncevent/broadcaster_test.go
@@ -0,0 +1,376 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syncevent
+
+import (
+	"fmt"
+	"math/rand"
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+func TestBroadcasterFilter(t *testing.T) {
+	const numReceivers = 2 * MaxEvents
+
+	var br Broadcaster
+	ws := make([]Waiter, numReceivers)
+	for i := range ws {
+		ws[i].Init()
+		br.SubscribeEvents(ws[i].Receiver(), 1<<(i%MaxEvents))
+	}
+	for ev := 0; ev < MaxEvents; ev++ {
+		br.Broadcast(1 << ev)
+		for i := range ws {
+			want := NoEvents
+			if i%MaxEvents == ev {
+				want = 1 << ev
+			}
+			if got := ws[i].Receiver().PendingAndAckAll(); got != want {
+				t.Errorf("after Broadcast of event %d: waiter %d has pending event set %#x, wanted %#x", ev, i, got, want)
+			}
+		}
+	}
+}
+
+// TestBroadcasterManySubscriptions tests that subscriptions are not lost by
+// table expansion/compaction.
+func TestBroadcasterManySubscriptions(t *testing.T) {
+	const numReceivers = 5000 // arbitrary
+
+	var br Broadcaster
+	ws := make([]Waiter, numReceivers)
+	for i := range ws {
+		ws[i].Init()
+	}
+
+	ids := make([]SubscriptionID, numReceivers)
+	for i := 0; i < numReceivers; i++ {
+		// Subscribe receiver i.
+		ids[i] = br.SubscribeEvents(ws[i].Receiver(), 1)
+		// Check that receivers [0, i] are subscribed.
+		br.Broadcast(1)
+		for j := 0; j <= i; j++ {
+			if ws[j].Pending() != 1 {
+				t.Errorf("receiver %d did not receive an event after subscription of receiver %d", j, i)
+			}
+			ws[j].Ack(1)
+		}
+	}
+
+	// Generate a random order for unsubscriptions.
+	unsub := rand.Perm(numReceivers)
+	for i := 0; i < numReceivers; i++ {
+		// Unsubscribe receiver unsub[i].
+		br.UnsubscribeEvents(ids[unsub[i]])
+		// Check that receivers [unsub[0], unsub[i]] are not subscribed, and that
+		// receivers (unsub[i], unsub[numReceivers]) are still subscribed.
+		br.Broadcast(1)
+		for j := 0; j <= i; j++ {
+			if ws[unsub[j]].Pending() != 0 {
+				t.Errorf("unsub iteration %d: receiver %d received an event after unsubscription of receiver %d", i, unsub[j], unsub[i])
+			}
+		}
+		for j := i + 1; j < numReceivers; j++ {
+			if ws[unsub[j]].Pending() != 1 {
+				t.Errorf("unsub iteration %d: receiver %d did not receive an event after unsubscription of receiver %d", i, unsub[j], unsub[i])
+			}
+			ws[unsub[j]].Ack(1)
+		}
+	}
+}
+
+var (
+	receiverCountsNonZero       = []int{1, 4, 16, 64}
+	receiverCountsIncludingZero = append([]int{0}, receiverCountsNonZero...)
+)
+
+// BenchmarkBroadcasterX, BenchmarkMapX, and BenchmarkQueueX benchmark usage
+// pattern X (described in terms of Broadcaster) with Broadcaster, a
+// Mutex-protected map[*Receiver]Set, and waiter.Queue respectively.
+
+// BenchmarkXxxSubscribeUnsubscribe measures the cost of a Subscribe/Unsubscribe
+// cycle.
+
+func BenchmarkBroadcasterSubscribeUnsubscribe(b *testing.B) {
+	var br Broadcaster
+	var w Waiter
+	w.Init()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		id := br.SubscribeEvents(w.Receiver(), 1)
+		br.UnsubscribeEvents(id)
+	}
+}
+
+func BenchmarkMapSubscribeUnsubscribe(b *testing.B) {
+	var mu sync.Mutex
+	m := make(map[*Receiver]Set)
+	var w Waiter
+	w.Init()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		mu.Lock()
+		m[w.Receiver()] = Set(1)
+		mu.Unlock()
+		mu.Lock()
+		delete(m, w.Receiver())
+		mu.Unlock()
+	}
+}
+
+func BenchmarkQueueSubscribeUnsubscribe(b *testing.B) {
+	var q waiter.Queue
+	e, _ := waiter.NewChannelEntry(nil)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		q.EventRegister(&e, 1)
+		q.EventUnregister(&e)
+	}
+}
+
+// BenchmarkXxxSubscribeUnsubscribeBatch is similar to
+// BenchmarkXxxSubscribeUnsubscribe, but subscribes and unsubscribes a large
+// number of Receivers at a time in order to measure the amortized overhead of
+// table expansion/compaction. (Since waiter.Queue is implemented using a
+// linked list, BenchmarkQueueSubscribeUnsubscribe and
+// BenchmarkQueueSubscribeUnsubscribeBatch should produce nearly the same
+// result.)
+
+const numBatchReceivers = 1000
+
+func BenchmarkBroadcasterSubscribeUnsubscribeBatch(b *testing.B) {
+	var br Broadcaster
+	ws := make([]Waiter, numBatchReceivers)
+	for i := range ws {
+		ws[i].Init()
+	}
+	ids := make([]SubscriptionID, numBatchReceivers)
+
+	// Generate a random order for unsubscriptions.
+	unsub := rand.Perm(numBatchReceivers)
+
+	b.ResetTimer()
+	for i := 0; i < b.N/numBatchReceivers; i++ {
+		for j := 0; j < numBatchReceivers; j++ {
+			ids[j] = br.SubscribeEvents(ws[j].Receiver(), 1)
+		}
+		for j := 0; j < numBatchReceivers; j++ {
+			br.UnsubscribeEvents(ids[unsub[j]])
+		}
+	}
+}
+
+func BenchmarkMapSubscribeUnsubscribeBatch(b *testing.B) {
+	var mu sync.Mutex
+	m := make(map[*Receiver]Set)
+	ws := make([]Waiter, numBatchReceivers)
+	for i := range ws {
+		ws[i].Init()
+	}
+
+	// Generate a random order for unsubscriptions.
+	unsub := rand.Perm(numBatchReceivers)
+
+	b.ResetTimer()
+	for i := 0; i < b.N/numBatchReceivers; i++ {
+		for j := 0; j < numBatchReceivers; j++ {
+			mu.Lock()
+			m[ws[j].Receiver()] = Set(1)
+			mu.Unlock()
+		}
+		for j := 0; j < numBatchReceivers; j++ {
+			mu.Lock()
+			delete(m, ws[unsub[j]].Receiver())
+			mu.Unlock()
+		}
+	}
+}
+
+func BenchmarkQueueSubscribeUnsubscribeBatch(b *testing.B) {
+	var q waiter.Queue
+	es := make([]waiter.Entry, numBatchReceivers)
+	for i := range es {
+		es[i], _ = waiter.NewChannelEntry(nil)
+	}
+
+	// Generate a random order for unsubscriptions.
+	unsub := rand.Perm(numBatchReceivers)
+
+	b.ResetTimer()
+	for i := 0; i < b.N/numBatchReceivers; i++ {
+		for j := 0; j < numBatchReceivers; j++ {
+			q.EventRegister(&es[j], 1)
+		}
+		for j := 0; j < numBatchReceivers; j++ {
+			q.EventUnregister(&es[unsub[j]])
+		}
+	}
+}
+
+// BenchmarkXxxBroadcastRedundant measures how long it takes to Broadcast
+// already-pending events to multiple Receivers.
+
+func BenchmarkBroadcasterBroadcastRedundant(b *testing.B) {
+	for _, n := range receiverCountsIncludingZero {
+		b.Run(fmt.Sprintf("%d", n), func(b *testing.B) {
+			var br Broadcaster
+			ws := make([]Waiter, n)
+			for i := range ws {
+				ws[i].Init()
+				br.SubscribeEvents(ws[i].Receiver(), 1)
+			}
+			br.Broadcast(1)
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				br.Broadcast(1)
+			}
+		})
+	}
+}
+
+func BenchmarkMapBroadcastRedundant(b *testing.B) {
+	for _, n := range receiverCountsIncludingZero {
+		b.Run(fmt.Sprintf("%d", n), func(b *testing.B) {
+			var mu sync.Mutex
+			m := make(map[*Receiver]Set)
+			ws := make([]Waiter, n)
+			for i := range ws {
+				ws[i].Init()
+				m[ws[i].Receiver()] = Set(1)
+			}
+			mu.Lock()
+			for r := range m {
+				r.Notify(1)
+			}
+			mu.Unlock()
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				mu.Lock()
+				for r := range m {
+					r.Notify(1)
+				}
+				mu.Unlock()
+			}
+		})
+	}
+}
+
+func BenchmarkQueueBroadcastRedundant(b *testing.B) {
+	for _, n := range receiverCountsIncludingZero {
+		b.Run(fmt.Sprintf("%d", n), func(b *testing.B) {
+			var q waiter.Queue
+			for i := 0; i < n; i++ {
+				e, _ := waiter.NewChannelEntry(nil)
+				q.EventRegister(&e, 1)
+			}
+			q.Notify(1)
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				q.Notify(1)
+			}
+		})
+	}
+}
+
+// BenchmarkXxxBroadcastAck measures how long it takes to Broadcast events to
+// multiple Receivers, check that all Receivers have received the event, and
+// clear the event from all Receivers.
+
+func BenchmarkBroadcasterBroadcastAck(b *testing.B) {
+	for _, n := range receiverCountsNonZero {
+		b.Run(fmt.Sprintf("%d", n), func(b *testing.B) {
+			var br Broadcaster
+			ws := make([]Waiter, n)
+			for i := range ws {
+				ws[i].Init()
+				br.SubscribeEvents(ws[i].Receiver(), 1)
+			}
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				br.Broadcast(1)
+				for j := range ws {
+					if got, want := ws[j].Pending(), Set(1); got != want {
+						b.Fatalf("Receiver.Pending(): got %#x, wanted %#x", got, want)
+					}
+					ws[j].Ack(1)
+				}
+			}
+		})
+	}
+}
+
+func BenchmarkMapBroadcastAck(b *testing.B) {
+	for _, n := range receiverCountsNonZero {
+		b.Run(fmt.Sprintf("%d", n), func(b *testing.B) {
+			var mu sync.Mutex
+			m := make(map[*Receiver]Set)
+			ws := make([]Waiter, n)
+			for i := range ws {
+				ws[i].Init()
+				m[ws[i].Receiver()] = Set(1)
+			}
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				mu.Lock()
+				for r := range m {
+					r.Notify(1)
+				}
+				mu.Unlock()
+				for j := range ws {
+					if got, want := ws[j].Pending(), Set(1); got != want {
+						b.Fatalf("Receiver.Pending(): got %#x, wanted %#x", got, want)
+					}
+					ws[j].Ack(1)
+				}
+			}
+		})
+	}
+}
+
+func BenchmarkQueueBroadcastAck(b *testing.B) {
+	for _, n := range receiverCountsNonZero {
+		b.Run(fmt.Sprintf("%d", n), func(b *testing.B) {
+			var q waiter.Queue
+			chs := make([]chan struct{}, n)
+			for i := range chs {
+				e, ch := waiter.NewChannelEntry(nil)
+				q.EventRegister(&e, 1)
+				chs[i] = ch
+			}
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				q.Notify(1)
+				for _, ch := range chs {
+					select {
+					case <-ch:
+					default:
+						b.Fatalf("channel did not receive event")
+					}
+				}
+			}
+		})
+	}
+}
diff --git a/pkg/syncevent/receiver.go b/pkg/syncevent/receiver.go
new file mode 100644
index 000000000..5c86e5400
--- /dev/null
+++ b/pkg/syncevent/receiver.go
@@ -0,0 +1,103 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syncevent
+
+import (
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/atomicbitops"
+)
+
+// Receiver is an event sink that holds pending events and invokes a callback
+// whenever new events become pending. Receiver's methods may be called
+// concurrently from multiple goroutines.
+//
+// Receiver.Init() must be called before first use.
+type Receiver struct {
+	// pending is the set of pending events. pending is accessed using atomic
+	// memory operations.
+	pending uint64
+
+	// cb is notified when new events become pending. cb is immutable after
+	// Init().
+	cb ReceiverCallback
+}
+
+// ReceiverCallback receives callbacks from a Receiver.
+type ReceiverCallback interface {
+	// NotifyPending is called when the corresponding Receiver has new pending
+	// events.
+	//
+	// NotifyPending is called synchronously from Receiver.Notify(), so
+	// implementations must not take locks that may be held by callers of
+	// Receiver.Notify(). NotifyPending may be called concurrently from
+	// multiple goroutines.
+	NotifyPending()
+}
+
+// Init must be called before first use of r.
+func (r *Receiver) Init(cb ReceiverCallback) {
+	r.cb = cb
+}
+
+// Pending returns the set of pending events.
+func (r *Receiver) Pending() Set {
+	return Set(atomic.LoadUint64(&r.pending))
+}
+
+// Notify sets the given events as pending.
+func (r *Receiver) Notify(es Set) {
+	p := Set(atomic.LoadUint64(&r.pending))
+	// Optimization: Skip the atomic CAS on r.pending if all events are
+	// already pending.
+	if p&es == es {
+		return
+	}
+	// When this is uncontended (the common case), CAS is faster than
+	// atomic-OR because the former is inlined and the latter (which we
+	// implement in assembly ourselves) is not.
+	if !atomic.CompareAndSwapUint64(&r.pending, uint64(p), uint64(p|es)) {
+		// If the CAS fails, fall back to atomic-OR.
+		atomicbitops.OrUint64(&r.pending, uint64(es))
+	}
+	r.cb.NotifyPending()
+}
+
+// Ack unsets the given events as pending.
+func (r *Receiver) Ack(es Set) {
+	p := Set(atomic.LoadUint64(&r.pending))
+	// Optimization: Skip the atomic CAS on r.pending if all events are
+	// already not pending.
+	if p&es == 0 {
+		return
+	}
+	// When this is uncontended (the common case), CAS is faster than
+	// atomic-AND because the former is inlined and the latter (which we
+	// implement in assembly ourselves) is not.
+	if !atomic.CompareAndSwapUint64(&r.pending, uint64(p), uint64(p&^es)) {
+		// If the CAS fails, fall back to atomic-AND.
+		atomicbitops.AndUint64(&r.pending, ^uint64(es))
+	}
+}
+
+// PendingAndAckAll unsets all events as pending and returns the set of
+// previously-pending events.
+//
+// PendingAndAckAll should only be used in preference to a call to Pending
+// followed by a conditional call to Ack when the caller expects events to be
+// pending (e.g. after a call to ReceiverCallback.NotifyPending()).
+func (r *Receiver) PendingAndAckAll() Set {
+	return Set(atomic.SwapUint64(&r.pending, 0))
+}
diff --git a/pkg/syncevent/source.go b/pkg/syncevent/source.go
new file mode 100644
index 000000000..ddffb171a
--- /dev/null
+++ b/pkg/syncevent/source.go
@@ -0,0 +1,59 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syncevent
+
+// Source represents an event source.
+type Source interface {
+	// SubscribeEvents causes the Source to notify the given Receiver of the
+	// given subset of events.
+	//
+	// Preconditions: r != nil. The ReceiverCallback for r must not take locks
+	// that are ordered prior to the Source; for example, it cannot call any
+	// Source methods.
+	SubscribeEvents(r *Receiver, filter Set) SubscriptionID
+
+	// UnsubscribeEvents causes the Source to stop notifying the Receiver
+	// subscribed by a previous call to SubscribeEvents that returned the given
+	// SubscriptionID.
+	//
+	// Preconditions: UnsubscribeEvents may be called at most once for any
+	// given SubscriptionID.
+	UnsubscribeEvents(id SubscriptionID)
+}
+
+// SubscriptionID identifies a call to Source.SubscribeEvents.
+type SubscriptionID uint64
+
+// UnsubscribeAndAck is a convenience function that unsubscribes r from the
+// given events from src and also clears them from r.
+func UnsubscribeAndAck(src Source, r *Receiver, filter Set, id SubscriptionID) {
+	src.UnsubscribeEvents(id)
+	r.Ack(filter)
+}
+
+// NoopSource implements Source by never sending events to subscribed
+// Receivers.
+type NoopSource struct{}
+
+// SubscribeEvents implements Source.SubscribeEvents.
+func (NoopSource) SubscribeEvents(*Receiver, Set) SubscriptionID {
+	return 0
+}
+
+// UnsubscribeEvents implements Source.UnsubscribeEvents.
+func (NoopSource) UnsubscribeEvents(SubscriptionID) {
+}
+
+// See Broadcaster for a non-noop implementations of Source.
diff --git a/pkg/syncevent/syncevent.go b/pkg/syncevent/syncevent.go
new file mode 100644
index 000000000..9fb6a06de
--- /dev/null
+++ b/pkg/syncevent/syncevent.go
@@ -0,0 +1,32 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package syncevent provides efficient primitives for goroutine
+// synchronization based on event bitmasks.
+package syncevent
+
+// Set is a bitmask where each bit represents a distinct user-defined event.
+// The event package does not treat any bits in Set specially.
+type Set uint64
+
+const (
+	// NoEvents is a Set containing no events.
+	NoEvents = Set(0)
+
+	// AllEvents is a Set containing all possible events.
+	AllEvents = ^Set(0)
+
+	// MaxEvents is the number of distinct events that can be represented by a Set.
+	MaxEvents = 64
+)
diff --git a/pkg/syncevent/syncevent_example_test.go b/pkg/syncevent/syncevent_example_test.go
new file mode 100644
index 000000000..bfb18e2ea
--- /dev/null
+++ b/pkg/syncevent/syncevent_example_test.go
@@ -0,0 +1,108 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syncevent
+
+import (
+	"fmt"
+	"sync/atomic"
+	"time"
+)
+
+func Example_ioReadinessInterrputible() {
+	const (
+		evReady = Set(1 << iota)
+		evInterrupt
+	)
+	errNotReady := fmt.Errorf("not ready for I/O")
+
+	// State of some I/O object.
+	var (
+		br    Broadcaster
+		ready uint32
+	)
+	doIO := func() error {
+		if atomic.LoadUint32(&ready) == 0 {
+			return errNotReady
+		}
+		return nil
+	}
+	go func() {
+		// The I/O object eventually becomes ready for I/O.
+		time.Sleep(100 * time.Millisecond)
+		// When it does, it first ensures that future calls to isReady() return
+		// true, then broadcasts the readiness event to Receivers.
+		atomic.StoreUint32(&ready, 1)
+		br.Broadcast(evReady)
+	}()
+
+	// Each user of the I/O object owns a Waiter.
+	var w Waiter
+	w.Init()
+	// The Waiter may be asynchronously interruptible, e.g. for signal
+	// handling in the sentry.
+	go func() {
+		time.Sleep(200 * time.Millisecond)
+		w.Receiver().Notify(evInterrupt)
+	}()
+
+	// To use the I/O object:
+	//
+	// Optionally, if the I/O object is likely to be ready, attempt I/O first.
+	err := doIO()
+	if err == nil {
+		// Success, we're done.
+		return /* nil */
+	}
+	if err != errNotReady {
+		// Failure, I/O failed for some reason other than readiness.
+		return /* err */
+	}
+	// Subscribe for readiness events from the I/O object.
+	id := br.SubscribeEvents(w.Receiver(), evReady)
+	// When we are finished blocking, unsubscribe from readiness events and
+	// remove readiness events from the pending event set.
+	defer UnsubscribeAndAck(&br, w.Receiver(), evReady, id)
+	for {
+		// Attempt I/O again. This must be done after the call to SubscribeEvents,
+		// since the I/O object might have become ready between the previous call
+		// to doIO and the call to SubscribeEvents.
+		err = doIO()
+		if err == nil {
+			return /* nil */
+		}
+		if err != errNotReady {
+			return /* err */
+		}
+		// Block until either the I/O object indicates it is ready, or we are
+		// interrupted.
+		events := w.Wait()
+		if events&evInterrupt != 0 {
+			// In the specific case of sentry signal handling, signal delivery
+			// is handled by another system, so we aren't responsible for
+			// acknowledging evInterrupt.
+			return /* errInterrupted */
+		}
+		// Note that, in a concurrent context, the I/O object might become
+		// ready and then not ready again. To handle this:
+		//
+		// - evReady must be acknowledged before calling doIO() again (rather
+		// than after), so that if the I/O object becomes ready *again* after
+		// the call to doIO(), the readiness event is not lost.
+		//
+		// - We must loop instead of just calling doIO() once after receiving
+		// evReady.
+		w.Ack(evReady)
+	}
+}
diff --git a/pkg/syncevent/waiter_amd64.s b/pkg/syncevent/waiter_amd64.s
new file mode 100644
index 000000000..985b56ae5
--- /dev/null
+++ b/pkg/syncevent/waiter_amd64.s
@@ -0,0 +1,32 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// See waiter_noasm_unsafe.go for a description of waiterUnlock.
+//
+// func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool
+TEXT ·waiterUnlock(SB),NOSPLIT,$0-24
+	MOVQ g+0(FP), DI
+	MOVQ wg+8(FP), SI
+
+	MOVQ $·preparingG(SB), AX
+	LOCK
+	CMPXCHGQ DI, 0(SI)
+
+	SETEQ AX
+	MOVB AX, ret+16(FP)
+
+	RET
+
diff --git a/pkg/syncevent/waiter_arm64.s b/pkg/syncevent/waiter_arm64.s
new file mode 100644
index 000000000..20d7ac23b
--- /dev/null
+++ b/pkg/syncevent/waiter_arm64.s
@@ -0,0 +1,34 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// See waiter_noasm_unsafe.go for a description of waiterUnlock.
+//
+// func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool
+TEXT ·waiterUnlock(SB),NOSPLIT,$0-24
+	MOVD wg+8(FP), R0
+	MOVD $·preparingG(SB), R1
+	MOVD g+0(FP), R2
+again:
+	LDAXR (R0), R3
+	CMP R1, R3
+	BNE ok
+	STLXR R2, (R0), R3
+	CBNZ R3, again
+ok:
+	CSET EQ, R0
+	MOVB R0, ret+16(FP)
+	RET
+
diff --git a/pkg/syncevent/waiter_asm_unsafe.go b/pkg/syncevent/waiter_asm_unsafe.go
new file mode 100644
index 000000000..0995e9053
--- /dev/null
+++ b/pkg/syncevent/waiter_asm_unsafe.go
@@ -0,0 +1,24 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64 arm64
+
+package syncevent
+
+import (
+	"unsafe"
+)
+
+// See waiter_noasm_unsafe.go for a description of waiterUnlock.
+func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool
diff --git a/pkg/syncevent/waiter_noasm_unsafe.go b/pkg/syncevent/waiter_noasm_unsafe.go
new file mode 100644
index 000000000..1c4b0e39a
--- /dev/null
+++ b/pkg/syncevent/waiter_noasm_unsafe.go
@@ -0,0 +1,39 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// waiterUnlock is called from g0, so when the race detector is enabled,
+// waiterUnlock must be implemented in assembly since no race context is
+// available.
+//
+// +build !race
+// +build !amd64,!arm64
+
+package syncevent
+
+import (
+	"sync/atomic"
+	"unsafe"
+)
+
+// waiterUnlock is the "unlock function" passed to runtime.gopark by
+// Waiter.Wait*. wg is &Waiter.g, and g is a pointer to the calling runtime.g.
+// waiterUnlock returns true if Waiter.Wait should sleep and false if sleeping
+// should be aborted.
+//
+//go:nosplit
+func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool {
+	// The only way this CAS can fail is if a call to Waiter.NotifyPending()
+	// has replaced *wg with nil, in which case we should not sleep.
+	return atomic.CompareAndSwapPointer(wg, (unsafe.Pointer)(&preparingG), g)
+}
diff --git a/pkg/syncevent/waiter_test.go b/pkg/syncevent/waiter_test.go
new file mode 100644
index 000000000..3c8cbcdd8
--- /dev/null
+++ b/pkg/syncevent/waiter_test.go
@@ -0,0 +1,414 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syncevent
+
+import (
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gvisor.dev/gvisor/pkg/sleep"
+	"gvisor.dev/gvisor/pkg/sync"
+)
+
+func TestWaiterAlreadyPending(t *testing.T) {
+	var w Waiter
+	w.Init()
+	want := Set(1)
+	w.Notify(want)
+	if got := w.Wait(); got != want {
+		t.Errorf("Waiter.Wait: got %#x, wanted %#x", got, want)
+	}
+}
+
+func TestWaiterAsyncNotify(t *testing.T) {
+	var w Waiter
+	w.Init()
+	want := Set(1)
+	go func() {
+		time.Sleep(100 * time.Millisecond)
+		w.Notify(want)
+	}()
+	if got := w.Wait(); got != want {
+		t.Errorf("Waiter.Wait: got %#x, wanted %#x", got, want)
+	}
+}
+
+func TestWaiterWaitFor(t *testing.T) {
+	var w Waiter
+	w.Init()
+	evWaited := Set(1)
+	evOther := Set(2)
+	w.Notify(evOther)
+	notifiedEvent := uint32(0)
+	go func() {
+		time.Sleep(100 * time.Millisecond)
+		atomic.StoreUint32(&notifiedEvent, 1)
+		w.Notify(evWaited)
+	}()
+	if got, want := w.WaitFor(evWaited), evWaited|evOther; got != want {
+		t.Errorf("Waiter.WaitFor: got %#x, wanted %#x", got, want)
+	}
+	if atomic.LoadUint32(&notifiedEvent) == 0 {
+		t.Errorf("Waiter.WaitFor returned before goroutine notified waited-for event")
+	}
+}
+
+func TestWaiterWaitAndAckAll(t *testing.T) {
+	var w Waiter
+	w.Init()
+	w.Notify(AllEvents)
+	if got := w.WaitAndAckAll(); got != AllEvents {
+		t.Errorf("Waiter.WaitAndAckAll: got %#x, wanted %#x", got, AllEvents)
+	}
+	if got := w.Pending(); got != NoEvents {
+		t.Errorf("Waiter.WaitAndAckAll did not ack all events: got %#x, wanted 0", got)
+	}
+}
+
+// BenchmarkWaiterX, BenchmarkSleeperX, and BenchmarkChannelX benchmark usage
+// pattern X (described in terms of Waiter) with Waiter, sleep.Sleeper, and
+// buffered chan struct{} respectively. When the maximum number of event
+// sources is relevant, we use 3 event sources because this is representative
+// of the kernel.Task.block() use case: an interrupt source, a timeout source,
+// and the actual event source being waited on.
+
+// Event set used by most benchmarks.
+const evBench Set = 1
+
+// BenchmarkXxxNotifyRedundant measures how long it takes to notify a Waiter of
+// an event that is already pending.
+
+func BenchmarkWaiterNotifyRedundant(b *testing.B) {
+	var w Waiter
+	w.Init()
+	w.Notify(evBench)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w.Notify(evBench)
+	}
+}
+
+func BenchmarkSleeperNotifyRedundant(b *testing.B) {
+	var s sleep.Sleeper
+	var w sleep.Waker
+	s.AddWaker(&w, 0)
+	w.Assert()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w.Assert()
+	}
+}
+
+func BenchmarkChannelNotifyRedundant(b *testing.B) {
+	ch := make(chan struct{}, 1)
+	ch <- struct{}{}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		select {
+		case ch <- struct{}{}:
+		default:
+		}
+	}
+}
+
+// BenchmarkXxxNotifyWaitAck measures how long it takes to notify a Waiter an
+// event, return that event using a blocking check, and then unset the event as
+// pending.
+
+func BenchmarkWaiterNotifyWaitAck(b *testing.B) {
+	var w Waiter
+	w.Init()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w.Notify(evBench)
+		w.Wait()
+		w.Ack(evBench)
+	}
+}
+
+func BenchmarkSleeperNotifyWaitAck(b *testing.B) {
+	var s sleep.Sleeper
+	var w sleep.Waker
+	s.AddWaker(&w, 0)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w.Assert()
+		s.Fetch(true)
+	}
+}
+
+func BenchmarkChannelNotifyWaitAck(b *testing.B) {
+	ch := make(chan struct{}, 1)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		// notify
+		select {
+		case ch <- struct{}{}:
+		default:
+		}
+
+		// wait + ack
+		<-ch
+	}
+}
+
+// BenchmarkSleeperMultiNotifyWaitAck is equivalent to
+// BenchmarkSleeperNotifyWaitAck, but also includes allocation of a
+// temporary sleep.Waker. This is necessary when multiple goroutines may wait
+// for the same event, since each sleep.Waker can wake only a single
+// sleep.Sleeper.
+//
+// The syncevent package does not require a distinct object for each
+// waiter-waker relationship, so BenchmarkWaiterNotifyWaitAck and
+// BenchmarkWaiterMultiNotifyWaitAck would be identical. The analogous state
+// for channels, runtime.sudog, is inescapably runtime-allocated, so
+// BenchmarkChannelNotifyWaitAck and BenchmarkChannelMultiNotifyWaitAck would
+// also be identical.
+
+func BenchmarkSleeperMultiNotifyWaitAck(b *testing.B) {
+	var s sleep.Sleeper
+	// The sleep package doesn't provide sync.Pool allocation of Wakers;
+	// we do for a fairer comparison.
+	wakerPool := sync.Pool{
+		New: func() interface{} {
+			return &sleep.Waker{}
+		},
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w := wakerPool.Get().(*sleep.Waker)
+		s.AddWaker(w, 0)
+		w.Assert()
+		s.Fetch(true)
+		s.Done()
+		wakerPool.Put(w)
+	}
+}
+
+// BenchmarkXxxTempNotifyWaitAck is equivalent to NotifyWaitAck, but also
+// includes allocation of a temporary Waiter. This models the case where a
+// goroutine not already associated with a Waiter needs one in order to block.
+//
+// The analogous state for channels is built into runtime.g, so
+// BenchmarkChannelNotifyWaitAck and BenchmarkChannelTempNotifyWaitAck would be
+// identical.
+
+func BenchmarkWaiterTempNotifyWaitAck(b *testing.B) {
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w := GetWaiter()
+		w.Notify(evBench)
+		w.Wait()
+		w.Ack(evBench)
+		PutWaiter(w)
+	}
+}
+
+func BenchmarkSleeperTempNotifyWaitAck(b *testing.B) {
+	// The sleep package doesn't provide sync.Pool allocation of Sleepers;
+	// we do for a fairer comparison.
+	sleeperPool := sync.Pool{
+		New: func() interface{} {
+			return &sleep.Sleeper{}
+		},
+	}
+	var w sleep.Waker
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		s := sleeperPool.Get().(*sleep.Sleeper)
+		s.AddWaker(&w, 0)
+		w.Assert()
+		s.Fetch(true)
+		s.Done()
+		sleeperPool.Put(s)
+	}
+}
+
+// BenchmarkXxxNotifyWaitMultiAck is equivalent to NotifyWaitAck, but allows
+// for multiple event sources.
+
+func BenchmarkWaiterNotifyWaitMultiAck(b *testing.B) {
+	var w Waiter
+	w.Init()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		w.Notify(evBench)
+		if e := w.Wait(); e != evBench {
+			b.Fatalf("Wait: got %#x, wanted %#x", e, evBench)
+		}
+		w.Ack(evBench)
+	}
+}
+
+func BenchmarkSleeperNotifyWaitMultiAck(b *testing.B) {
+	var s sleep.Sleeper
+	var ws [3]sleep.Waker
+	for i := range ws {
+		s.AddWaker(&ws[i], i)
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		ws[0].Assert()
+		if id, _ := s.Fetch(true); id != 0 {
+			b.Fatalf("Fetch: got %d, wanted 0", id)
+		}
+	}
+}
+
+func BenchmarkChannelNotifyWaitMultiAck(b *testing.B) {
+	ch0 := make(chan struct{}, 1)
+	ch1 := make(chan struct{}, 1)
+	ch2 := make(chan struct{}, 1)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		// notify
+		select {
+		case ch0 <- struct{}{}:
+		default:
+		}
+
+		// wait + clear
+		select {
+		case <-ch0:
+			// ok
+		case <-ch1:
+			b.Fatalf("received from ch1")
+		case <-ch2:
+			b.Fatalf("received from ch2")
+		}
+	}
+}
+
+// BenchmarkXxxNotifyAsyncWaitAck measures how long it takes to wait for an
+// event while another goroutine signals the event. This assumes that a new
+// goroutine doesn't run immediately (i.e. the creator of a new goroutine is
+// allowed to go to sleep before the new goroutine has a chance to run).
+
+func BenchmarkWaiterNotifyAsyncWaitAck(b *testing.B) {
+	var w Waiter
+	w.Init()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		go func() {
+			w.Notify(1)
+		}()
+		w.Wait()
+		w.Ack(evBench)
+	}
+}
+
+func BenchmarkSleeperNotifyAsyncWaitAck(b *testing.B) {
+	var s sleep.Sleeper
+	var w sleep.Waker
+	s.AddWaker(&w, 0)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		go func() {
+			w.Assert()
+		}()
+		s.Fetch(true)
+	}
+}
+
+func BenchmarkChannelNotifyAsyncWaitAck(b *testing.B) {
+	ch := make(chan struct{}, 1)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		go func() {
+			select {
+			case ch <- struct{}{}:
+			default:
+			}
+		}()
+		<-ch
+	}
+}
+
+// BenchmarkXxxNotifyAsyncWaitMultiAck is equivalent to NotifyAsyncWaitAck, but
+// allows for multiple event sources.
+
+func BenchmarkWaiterNotifyAsyncWaitMultiAck(b *testing.B) {
+	var w Waiter
+	w.Init()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		go func() {
+			w.Notify(evBench)
+		}()
+		if e := w.Wait(); e != evBench {
+			b.Fatalf("Wait: got %#x, wanted %#x", e, evBench)
+		}
+		w.Ack(evBench)
+	}
+}
+
+func BenchmarkSleeperNotifyAsyncWaitMultiAck(b *testing.B) {
+	var s sleep.Sleeper
+	var ws [3]sleep.Waker
+	for i := range ws {
+		s.AddWaker(&ws[i], i)
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		go func() {
+			ws[0].Assert()
+		}()
+		if id, _ := s.Fetch(true); id != 0 {
+			b.Fatalf("Fetch: got %d, expected 0", id)
+		}
+	}
+}
+
+func BenchmarkChannelNotifyAsyncWaitMultiAck(b *testing.B) {
+	ch0 := make(chan struct{}, 1)
+	ch1 := make(chan struct{}, 1)
+	ch2 := make(chan struct{}, 1)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		go func() {
+			select {
+			case ch0 <- struct{}{}:
+			default:
+			}
+		}()
+
+		select {
+		case <-ch0:
+			// ok
+		case <-ch1:
+			b.Fatalf("received from ch1")
+		case <-ch2:
+			b.Fatalf("received from ch2")
+		}
+	}
+}
diff --git a/pkg/syncevent/waiter_unsafe.go b/pkg/syncevent/waiter_unsafe.go
new file mode 100644
index 000000000..112e0e604
--- /dev/null
+++ b/pkg/syncevent/waiter_unsafe.go
@@ -0,0 +1,206 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build go1.11
+// +build !go1.15
+
+// Check go:linkname function signatures when updating Go version.
+
+package syncevent
+
+import (
+	"sync/atomic"
+	"unsafe"
+
+	"gvisor.dev/gvisor/pkg/sync"
+)
+
+//go:linkname gopark runtime.gopark
+func gopark(unlockf func(unsafe.Pointer, *unsafe.Pointer) bool, wg *unsafe.Pointer, reason uint8, traceEv byte, traceskip int)
+
+//go:linkname goready runtime.goready
+func goready(g unsafe.Pointer, traceskip int)
+
+const (
+	waitReasonSelect     = 9  // Go: src/runtime/runtime2.go
+	traceEvGoBlockSelect = 24 // Go: src/runtime/trace.go
+)
+
+// Waiter allows a goroutine to block on pending events received by a Receiver.
+//
+// Waiter.Init() must be called before first use.
+type Waiter struct {
+	r Receiver
+
+	// g is one of:
+	//
+	// - nil: No goroutine is blocking in Wait.
+	//
+	// - &preparingG: A goroutine is in Wait preparing to sleep, but hasn't yet
+	// completed waiterUnlock(). Thus the wait can only be interrupted by
+	// replacing the value of g with nil (the G may not be in state Gwaiting
+	// yet, so we can't call goready.)
+	//
+	// - Otherwise: g is a pointer to the runtime.g in state Gwaiting for the
+	// goroutine blocked in Wait, which can only be woken by calling goready.
+	g unsafe.Pointer `state:"zerovalue"`
+}
+
+// Sentinel object for Waiter.g.
+var preparingG struct{}
+
+// Init must be called before first use of w.
+func (w *Waiter) Init() {
+	w.r.Init(w)
+}
+
+// Receiver returns the Receiver that receives events that unblock calls to
+// w.Wait().
+func (w *Waiter) Receiver() *Receiver {
+	return &w.r
+}
+
+// Pending returns the set of pending events.
+func (w *Waiter) Pending() Set {
+	return w.r.Pending()
+}
+
+// Wait blocks until at least one event is pending, then returns the set of
+// pending events. It does not affect the set of pending events; callers must
+// call w.Ack() to do so, or use w.WaitAndAck() instead.
+//
+// Precondition: Only one goroutine may call any Wait* method at a time.
+func (w *Waiter) Wait() Set {
+	return w.WaitFor(AllEvents)
+}
+
+// WaitFor blocks until at least one event in es is pending, then returns the
+// set of pending events (including those not in es). It does not affect the
+// set of pending events; callers must call w.Ack() to do so.
+//
+// Precondition: Only one goroutine may call any Wait* method at a time.
+func (w *Waiter) WaitFor(es Set) Set {
+	for {
+		// Optimization: Skip the atomic store to w.g if an event is already
+		// pending.
+		if p := w.r.Pending(); p&es != NoEvents {
+			return p
+		}
+
+		// Indicate that we're preparing to go to sleep.
+		atomic.StorePointer(&w.g, (unsafe.Pointer)(&preparingG))
+
+		// If an event is pending, abort the sleep.
+		if p := w.r.Pending(); p&es != NoEvents {
+			atomic.StorePointer(&w.g, nil)
+			return p
+		}
+
+		// If w.g is still preparingG (i.e. w.NotifyPending() has not been
+		// called or has not reached atomic.SwapPointer()), go to sleep until
+		// w.NotifyPending() => goready().
+		gopark(waiterUnlock, &w.g, waitReasonSelect, traceEvGoBlockSelect, 0)
+	}
+}
+
+// Ack marks the given events as not pending.
+func (w *Waiter) Ack(es Set) {
+	w.r.Ack(es)
+}
+
+// WaitAndAckAll blocks until at least one event is pending, then marks all
+// events as not pending and returns the set of previously-pending events.
+//
+// Precondition: Only one goroutine may call any Wait* method at a time.
+func (w *Waiter) WaitAndAckAll() Set {
+	// Optimization: Skip the atomic store to w.g if an event is already
+	// pending. Call Pending() first since, in the common case that events are
+	// not yet pending, this skips an atomic swap on w.r.pending.
+	if w.r.Pending() != NoEvents {
+		if p := w.r.PendingAndAckAll(); p != NoEvents {
+			return p
+		}
+	}
+
+	for {
+		// Indicate that we're preparing to go to sleep.
+		atomic.StorePointer(&w.g, (unsafe.Pointer)(&preparingG))
+
+		// If an event is pending, abort the sleep.
+		if w.r.Pending() != NoEvents {
+			if p := w.r.PendingAndAckAll(); p != NoEvents {
+				atomic.StorePointer(&w.g, nil)
+				return p
+			}
+		}
+
+		// If w.g is still preparingG (i.e. w.NotifyPending() has not been
+		// called or has not reached atomic.SwapPointer()), go to sleep until
+		// w.NotifyPending() => goready().
+		gopark(waiterUnlock, &w.g, waitReasonSelect, traceEvGoBlockSelect, 0)
+
+		// Check for pending events. We call PendingAndAckAll() directly now since
+		// we only expect to be woken after events become pending.
+		if p := w.r.PendingAndAckAll(); p != NoEvents {
+			return p
+		}
+	}
+}
+
+// Notify marks the given events as pending, possibly unblocking concurrent
+// calls to w.Wait() or w.WaitFor().
+func (w *Waiter) Notify(es Set) {
+	w.r.Notify(es)
+}
+
+// NotifyPending implements ReceiverCallback.NotifyPending. Users of Waiter
+// should not call NotifyPending.
+func (w *Waiter) NotifyPending() {
+	// Optimization: Skip the atomic swap on w.g if there is no sleeping
+	// goroutine. NotifyPending is called after w.r.Pending() is updated, so
+	// concurrent and future calls to w.Wait() will observe pending events and
+	// abort sleeping.
+	if atomic.LoadPointer(&w.g) == nil {
+		return
+	}
+	// Wake a sleeping G, or prevent a G that is preparing to sleep from doing
+	// so. Swap is needed here to ensure that only one call to NotifyPending
+	// calls goready.
+	if g := atomic.SwapPointer(&w.g, nil); g != nil && g != (unsafe.Pointer)(&preparingG) {
+		goready(g, 0)
+	}
+}
+
+var waiterPool = sync.Pool{
+	New: func() interface{} {
+		w := &Waiter{}
+		w.Init()
+		return w
+	},
+}
+
+// GetWaiter returns an unused Waiter. PutWaiter should be called to release
+// the Waiter once it is no longer needed.
+//
+// Where possible, users should prefer to associate each goroutine that calls
+// Waiter.Wait() with a distinct pre-allocated Waiter to avoid allocation of
+// Waiters in hot paths.
+func GetWaiter() *Waiter {
+	return waiterPool.Get().(*Waiter)
+}
+
+// PutWaiter releases an unused Waiter previously returned by GetWaiter.
+func PutWaiter(w *Waiter) {
+	waiterPool.Put(w)
+}