Check in gVisor.

PiperOrigin-RevId: 194583126
Change-Id: Ica1d8821a90f74e7e745962d71801c598c652463

3 files changed, 288 insertions, 0 deletions

diff --git a/pkg/waiter/fdnotifier/BUILD b/pkg/waiter/fdnotifier/BUILD
new file mode 100644
index 000000000..d5b5ee82d
--- /dev/null
+++ b/pkg/waiter/fdnotifier/BUILD
@@ -0,0 +1,14 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "fdnotifier",
+    srcs = [
+        "fdnotifier.go",
+        "poll_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier",
+    visibility = ["//:sandbox"],
+    deps = ["//pkg/waiter"],
+)
diff --git a/pkg/waiter/fdnotifier/fdnotifier.go b/pkg/waiter/fdnotifier/fdnotifier.go
new file mode 100644
index 000000000..8bb93e39b
--- /dev/null
+++ b/pkg/waiter/fdnotifier/fdnotifier.go
@@ -0,0 +1,200 @@
+// Copyright 2018 Google Inc.
+//
+// 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 fdnotifier contains an adapter that translates IO events (e.g., a
+// file became readable/writable) from native FDs to the notifications in the
+// waiter package. It uses epoll in edge-triggered mode to receive notifications
+// for registered FDs.
+package fdnotifier
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+type fdInfo struct {
+	queue   *waiter.Queue
+	waiting bool
+}
+
+// notifier holds all the state necessary to issue notifications when IO events
+// occur in the observed FDs.
+type notifier struct {
+	// epFD is the epoll file descriptor used to register for io
+	// notifications.
+	epFD int
+
+	// mu protects fdMap.
+	mu sync.Mutex
+
+	// fdMap maps file descriptors to their notification queues and waiting
+	// status.
+	fdMap map[int32]*fdInfo
+}
+
+// newNotifier creates a new notifier object.
+func newNotifier() (*notifier, error) {
+	epfd, err := syscall.EpollCreate1(0)
+	if err != nil {
+		return nil, err
+	}
+
+	w := &notifier{
+		epFD:  epfd,
+		fdMap: make(map[int32]*fdInfo),
+	}
+
+	go w.waitAndNotify() // S/R-SAFE: no waiter exists during save / load.
+
+	return w, nil
+}
+
+// waitFD waits on mask for fd. The fdMap mutex must be hold.
+func (n *notifier) waitFD(fd int32, fi *fdInfo, mask waiter.EventMask) error {
+	if !fi.waiting && mask == 0 {
+		return nil
+	}
+
+	e := syscall.EpollEvent{
+		Events: uint32(mask) | -syscall.EPOLLET,
+		Fd:     fd,
+	}
+
+	switch {
+	case !fi.waiting && mask != 0:
+		if err := syscall.EpollCtl(n.epFD, syscall.EPOLL_CTL_ADD, int(fd), &e); err != nil {
+			return err
+		}
+		fi.waiting = true
+	case fi.waiting && mask == 0:
+		syscall.EpollCtl(n.epFD, syscall.EPOLL_CTL_DEL, int(fd), nil)
+		fi.waiting = false
+	case fi.waiting && mask != 0:
+		if err := syscall.EpollCtl(n.epFD, syscall.EPOLL_CTL_MOD, int(fd), &e); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// addFD adds an FD to the list of FDs observed by n.
+func (n *notifier) addFD(fd int32, queue *waiter.Queue) {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	// Panic if we're already notifying on this FD.
+	if _, ok := n.fdMap[fd]; ok {
+		panic(fmt.Sprintf("File descriptor %v added twice", fd))
+	}
+
+	// We have nothing to wait for at the moment. Just add it to the map.
+	n.fdMap[fd] = &fdInfo{queue: queue}
+}
+
+// updateFD updates the set of events the fd needs to be notified on.
+func (n *notifier) updateFD(fd int32) error {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	if fi, ok := n.fdMap[fd]; ok {
+		return n.waitFD(fd, fi, fi.queue.Events())
+	}
+
+	return nil
+}
+
+// RemoveFD removes an FD from the list of FDs observed by n.
+func (n *notifier) removeFD(fd int32) {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	// Remove from map, then from epoll object.
+	n.waitFD(fd, n.fdMap[fd], 0)
+	delete(n.fdMap, fd)
+}
+
+// hasFD returns true if the fd is in the list of observed FDs.
+func (n *notifier) hasFD(fd int32) bool {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	_, ok := n.fdMap[fd]
+	return ok
+}
+
+// waitAndNotify run is its own goroutine and loops waiting for io event
+// notifications from the epoll object. Once notifications arrive, they are
+// dispatched to the registered queue.
+func (n *notifier) waitAndNotify() error {
+	e := make([]syscall.EpollEvent, 100)
+	for {
+		v, err := epollWait(n.epFD, e, -1)
+		if err == syscall.EINTR {
+			continue
+		}
+
+		if err != nil {
+			return err
+		}
+
+		n.mu.Lock()
+		for i := 0; i < v; i++ {
+			if fi, ok := n.fdMap[e[i].Fd]; ok {
+				fi.queue.Notify(waiter.EventMask(e[i].Events))
+			}
+		}
+		n.mu.Unlock()
+	}
+}
+
+var shared struct {
+	notifier *notifier
+	once     sync.Once
+	initErr  error
+}
+
+// AddFD adds an FD to the list of observed FDs.
+func AddFD(fd int32, queue *waiter.Queue) error {
+	shared.once.Do(func() {
+		shared.notifier, shared.initErr = newNotifier()
+	})
+
+	if shared.initErr != nil {
+		return shared.initErr
+	}
+
+	shared.notifier.addFD(fd, queue)
+	return nil
+}
+
+// UpdateFD updates the set of events the fd needs to be notified on.
+func UpdateFD(fd int32) error {
+	return shared.notifier.updateFD(fd)
+}
+
+// RemoveFD removes an FD from the list of observed FDs.
+func RemoveFD(fd int32) {
+	shared.notifier.removeFD(fd)
+}
+
+// HasFD returns true if the FD is in the list of observed FDs.
+//
+// This should only be used by tests to assert that FDs are correctly registered.
+func HasFD(fd int32) bool {
+	return shared.notifier.hasFD(fd)
+}
diff --git a/pkg/waiter/fdnotifier/poll_unsafe.go b/pkg/waiter/fdnotifier/poll_unsafe.go
new file mode 100644
index 000000000..26bca2b53
--- /dev/null
+++ b/pkg/waiter/fdnotifier/poll_unsafe.go
@@ -0,0 +1,74 @@
+// Copyright 2018 Google Inc.
+//
+// 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 fdnotifier
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// NonBlockingPoll polls the given FD in non-blocking fashion. It is used just
+// to query the FD's current state.
+func NonBlockingPoll(fd int32, mask waiter.EventMask) waiter.EventMask {
+	e := struct {
+		fd      int32
+		events  int16
+		revents int16
+	}{
+		fd:     fd,
+		events: int16(mask),
+	}
+
+	for {
+		n, _, err := syscall.RawSyscall(syscall.SYS_POLL, uintptr(unsafe.Pointer(&e)), 1, 0)
+		// Interrupted by signal, try again.
+		if err == syscall.EINTR {
+			continue
+		}
+		// If an error occur we'll conservatively say the FD is ready for
+		// whatever is being checked.
+		if err != 0 {
+			return mask
+		}
+
+		// If no FDs were returned, it wasn't ready for anything.
+		if n == 0 {
+			return 0
+		}
+
+		// Otherwise we got the ready events in the revents field.
+		return waiter.EventMask(e.revents)
+	}
+}
+
+// epollWait performs a blocking wait on epfd.
+//
+// Preconditions:
+//  * len(events) > 0
+func epollWait(epfd int, events []syscall.EpollEvent, msec int) (int, error) {
+	if len(events) == 0 {
+		panic("Empty events passed to EpollWait")
+	}
+
+	// We actually use epoll_pwait with NULL sigmask instead of epoll_wait
+	// since that is what the Go >= 1.11 runtime prefers.
+	r, _, e := syscall.Syscall6(syscall.SYS_EPOLL_PWAIT, uintptr(epfd), uintptr(unsafe.Pointer(&events[0])), uintptr(len(events)), uintptr(msec), 0, 0)
+	if e != 0 {
+		return 0, e
+	}
+	return int(r), nil
+}