the definitive tun offloads branch (#1704)

overlay/tio/queueset_gso_linux.go

@@ -0,0 +1,101 @@
+//go:build linux && !android
+// +build linux,!android
+
+package tio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"log/slog"
+	"sync/atomic"
+
+	"golang.org/x/sys/unix"
+)
+
+type offloadQueueSet struct {
+	pq []*Offload
+	// pqi is exactly the same as pq, but stored as the interface type
+	pqi        []Queue
+	shutdownFd int
+	// usoEnabled is true when newTun successfully negotiated TUN_F_USO4|6 with the kernel.
+	// Queues created by Add inherit this and surface it via Offload.USOSupported so coalescers can gate USO emission.
+	usoEnabled bool
+	closed     atomic.Bool
+	// l is handed to each queue for its bad-vnet-header drop logging.
+	l *slog.Logger
+}
+
+// NewOffloadQueueSet creates a QueueSet that uses virtio_net_hdr to do TSO segmentation.
+// usoEnabled tells downstream queues whether the kernel agreed to deliver/accept GSO_UDP_L4 superpackets.
+func NewOffloadQueueSet(usoEnabled bool, l *slog.Logger) (QueueSet, error) {
+	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create eventfd: %w", err)
+	}
+
+	out := &offloadQueueSet{
+		pq:         []*Offload{},
+		pqi:        []Queue{},
+		shutdownFd: shutdownFd,
+		usoEnabled: usoEnabled,
+		l:          l,
+	}
+
+	return out, nil
+}
+
+func (c *offloadQueueSet) Queues() []Queue {
+	return c.pqi
+}
+
+func (c *offloadQueueSet) Add(fd int) error {
+	if c.closed.Load() {
+		return errors.New("queue set already closed")
+	}
+	x, err := newOffload(fd, c.shutdownFd, c.usoEnabled, c.l)
+	if err != nil {
+		return err
+	}
+	c.pq = append(c.pq, x)
+	c.pqi = append(c.pqi, x)
+
+	return nil
+}
+
+func (c *offloadQueueSet) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(c.shutdownFd, buf[:])
+	return err
+}
+
+func (c *offloadQueueSet) Close() error {
+	if c.closed.Swap(true) {
+		return nil
+	}
+
+	errs := []error{}
+
+	// Signal all readers blocked in poll to wake up and exit.
+	// They observe POLLIN on the shutdown eventfd and return os.ErrClosed.
+	if err := c.wakeForShutdown(); err != nil {
+		errs = append(errs, err)
+	}
+
+	// Close the per-queue tun fds; this also unblocks any in-flight reads.
+	for _, x := range c.pq {
+		if err := x.Close(); err != nil {
+			errs = append(errs, err)
+		}
+	}
+
+	// Close the shutdown eventfd last: every reader's pollfd set references it,
+	// so it must outlive the wake + per-queue teardown above.
+	if err := unix.Close(c.shutdownFd); err != nil {
+		errs = append(errs, err)
+	}
+	c.shutdownFd = -1
+
+	return errors.Join(errs...)
+}