|
| 1 | +package vmnet |
| 2 | + |
| 3 | +/* |
| 4 | +#cgo darwin CFLAGS: -mmacosx-version-min=11 -x objective-c -fno-objc-arc |
| 5 | +#cgo darwin LDFLAGS: -lobjc -framework Foundation -framework vmnet |
| 6 | +# include "vmnet_darwin.h" |
| 7 | +*/ |
| 8 | +import "C" |
| 9 | +import ( |
| 10 | + "errors" |
| 11 | + "fmt" |
| 12 | + "net" |
| 13 | + "os" |
| 14 | + "syscall" |
| 15 | +) |
| 16 | + |
| 17 | +// MARK: - DatagramFileAdaptorForInterface |
| 18 | + |
| 19 | +// DatagramFileAdaptorForInterface returns a file for the given [Network]. |
| 20 | +// - Invoke the returned function in a separate goroutine to start packet forwarding between the vmnet interface and the file. |
| 21 | +// - The context can be used to stop the goroutines and the interface. |
| 22 | +// - The returned error channel can be used to receive errors from the goroutines. |
| 23 | +// - The connection closure is reported as [io.EOF] error or [syscall.ECONNRESET] error in the error channel. |
| 24 | +// |
| 25 | +// The returned file can be used as a datagram file descriptor for QEMU, krunkit, or VZ. |
| 26 | +// |
| 27 | +// QEMU: |
| 28 | +// |
| 29 | +// -netdev datagram,id=net0,addr.type=fd,addr.str=<file descriptor> |
| 30 | +// -netdev tap,id=net1,fd=<file descriptor> |
| 31 | +// |
| 32 | +// krunkit: |
| 33 | +// |
| 34 | +// --device virtio-net,type=unixgram,fd=<file descriptor>,offloading=on // offloading=on is recommended. See krunkit driver in LIMA. |
| 35 | +// |
| 36 | +// VZ: |
| 37 | +// |
| 38 | +// file, errCh, err := DatagramFileAdaptorForInterface(ctx, iface) |
| 39 | +// attachment := NewFileHandleNetworkDeviceAttachment(file) |
| 40 | +var DatagramFileAdaptorForInterface = FileAdaptorForInterface[*DatagramPacketForwarder, net.PacketConn] |
| 41 | + |
| 42 | +// MARK: - DatagramPacketForwarder for datagram file adaptor |
| 43 | + |
| 44 | +// DatagramPacketForwarder implements PacketForwarder for datagram file descriptor. |
| 45 | +type DatagramPacketForwarder struct { |
| 46 | + readPktDescsManager *pktDescsManager |
| 47 | + writePktDescsManager *pktDescsManager |
| 48 | +} |
| 49 | + |
| 50 | +var _ PacketForwarder[net.PacketConn] = (*DatagramPacketForwarder)(nil) |
| 51 | + |
| 52 | +// New creates a new DatagramPacketForwarder. |
| 53 | +func (f *DatagramPacketForwarder) New() PacketForwarder[net.PacketConn] { |
| 54 | + return &DatagramPacketForwarder{} |
| 55 | +} |
| 56 | + |
| 57 | +// Sockopts returns socket options for the given Interface and user desired options. |
| 58 | +// Default values are based on the following references: |
| 59 | +// - https://developer.apple.com/documentation/virtualization/vzfilehandlenetworkdeviceattachment/maximumtransmissionunit?language=objc |
| 60 | +func (*DatagramPacketForwarder) Sockopts(iface *Interface, userOpts Sockopts) Sockopts { |
| 61 | + return sockoptsForPacketConn(iface, userOpts) |
| 62 | +} |
| 63 | + |
| 64 | +// ConnAndFile creates a [net.PacketConn] and *[os.File] pair using [syscall.Socketpair]. |
| 65 | +func (f *DatagramPacketForwarder) ConnAndFile(opts Sockopts) (net.PacketConn, *os.File, error) { |
| 66 | + return packetConnAndFile(opts) |
| 67 | +} |
| 68 | + |
| 69 | +// AllocateBuffers allocates packet descriptor buffers for reading and writing packets. |
| 70 | +func (f *DatagramPacketForwarder) AllocateBuffers(iface *Interface) error { |
| 71 | + maxPacketSize := iface.MaxPacketSize |
| 72 | + if iface.EnableVirtioHeader { |
| 73 | + // Add virtio header size |
| 74 | + maxPacketSize += virtioNetHdrSize |
| 75 | + } |
| 76 | + f.readPktDescsManager = newPktDescsManager(iface.MaxReadPacketCount, maxPacketSize) |
| 77 | + f.writePktDescsManager = newPktDescsManager(iface.MaxWritePacketCount, maxPacketSize) |
| 78 | + return nil |
| 79 | +} |
| 80 | + |
| 81 | +// ReadPacketsFromInterface reads packets from the vmnet Interface. |
| 82 | +func (f *DatagramPacketForwarder) ReadPacketsFromInterface(iface *Interface, estimatedCount int) (int, error) { |
| 83 | + f.readPktDescsManager.reset() |
| 84 | + return iface.ReadPackets(f.readPktDescsManager.pktDescs, estimatedCount) |
| 85 | +} |
| 86 | + |
| 87 | +// WritePacketsToConn writes packets to the connection. |
| 88 | +func (f *DatagramPacketForwarder) WritePacketsToConn(conn net.PacketConn, packetCount int) error { |
| 89 | + return f.readPktDescsManager.writePacketsToPacketConn(conn, packetCount) |
| 90 | +} |
| 91 | + |
| 92 | +// ReadPacketsFromConn reads packets from the connection. |
| 93 | +func (f *DatagramPacketForwarder) ReadPacketsFromConn(conn net.PacketConn) (int, error) { |
| 94 | + return f.writePktDescsManager.readPacketsFromPacketConn(conn) |
| 95 | +} |
| 96 | + |
| 97 | +// WritePacketsToInterface writes packets to the vmnet Interface. |
| 98 | +func (f *DatagramPacketForwarder) WritePacketsToInterface(iface *Interface, packetCount int) error { |
| 99 | + return iface.WritePackets(f.writePktDescsManager.pktDescs, packetCount) |
| 100 | +} |
| 101 | + |
| 102 | +// sockoptsForPacketConn returns socket options for the given [Interface] and user desired options for [net.PacketConn]. |
| 103 | +// Default values are based on the following references: |
| 104 | +// - https://developer.apple.com/documentation/virtualization/vzfilehandlenetworkdeviceattachment/maximumtransmissionunit?language=objc |
| 105 | +func sockoptsForPacketConn(iface *Interface, userOpts Sockopts) Sockopts { |
| 106 | + // Calculate minimum buffer sizes based on interface configuration |
| 107 | + packetSize := int(iface.MaxPacketSize) |
| 108 | + if iface.EnableVirtioHeader { |
| 109 | + // Add virtio header size |
| 110 | + packetSize += virtioNetHdrSize |
| 111 | + } |
| 112 | + minPacketCount := max(iface.MaxReadPacketCount, iface.MaxWritePacketCount) |
| 113 | + minSendBufSize := packetSize |
| 114 | + minRecvBufSize := minSendBufSize * minPacketCount |
| 115 | + |
| 116 | + // Default socket options |
| 117 | + sockopts := Sockopts{ |
| 118 | + ReceiveBufferSize: minRecvBufSize * 4 * 10, |
| 119 | + SendBufferSize: packetSize, |
| 120 | + } |
| 121 | + // If user specified options, override with minimums as needed |
| 122 | + if userOpts.ReceiveBufferSize > 0 { |
| 123 | + sockopts.ReceiveBufferSize = max(userOpts.ReceiveBufferSize, minRecvBufSize) |
| 124 | + } |
| 125 | + if userOpts.SendBufferSize > 0 { |
| 126 | + sockopts.SendBufferSize = max(userOpts.SendBufferSize, minSendBufSize) |
| 127 | + } |
| 128 | + return sockopts |
| 129 | +} |
| 130 | + |
| 131 | +// packetConnAndFile creates a [net.PacketConn] and *[os.File] pair using [syscall.Socketpair]. |
| 132 | +func packetConnAndFile(opts Sockopts) (net.PacketConn, *os.File, error) { |
| 133 | + sendBufSize, recvBufSize := opts.SendBufferSize, opts.ReceiveBufferSize |
| 134 | + connFile, file, err := filePair(syscall.SOCK_DGRAM, sendBufSize, recvBufSize) |
| 135 | + if err != nil { |
| 136 | + return nil, nil, fmt.Errorf("ConnAndFile failed: %w", err) |
| 137 | + } |
| 138 | + conn, err := net.FilePacketConn(connFile) |
| 139 | + if err != nil { |
| 140 | + _ = connFile.Close() |
| 141 | + _ = file.Close() |
| 142 | + return nil, nil, fmt.Errorf("net.FilePacketConn failed: %w", err) |
| 143 | + } |
| 144 | + if err = connFile.Close(); err != nil { |
| 145 | + _ = conn.Close() |
| 146 | + _ = file.Close() |
| 147 | + return nil, nil, fmt.Errorf("failed to close connFile: %w", err) |
| 148 | + } |
| 149 | + return conn, file, nil |
| 150 | +} |
| 151 | + |
| 152 | +// MARK: - pktDescsManager methods for datagram file adaptor |
| 153 | + |
| 154 | +// buffersForWritingToPacketConn returns [net.Buffers] to write to the [net.PacketConn] |
| 155 | +// adjusted their buffer sizes based vm_pkt_size in [VMPktDesc]s read from [Interface]. |
| 156 | +// The 4-byte header is excluded. |
| 157 | +func (v *pktDescsManager) buffersForWritingToPacketConn(packetCount int) (net.Buffers, error) { |
| 158 | + for i, vmPktDesc := range v.iter(packetCount) { |
| 159 | + if uint64(vmPktDesc.vm_pkt_size) > v.maxPacketSize { |
| 160 | + return nil, fmt.Errorf("vm_pkt_size %d exceeds maxPacketSize %d", vmPktDesc.vm_pkt_size, v.maxPacketSize) |
| 161 | + } |
| 162 | + // Resize buffer to exclude the 4-byte header |
| 163 | + v.writingBuffers[i] = v.backingBuffers[i][headerSize : headerSize+uintptr(vmPktDesc.vm_pkt_size)] |
| 164 | + } |
| 165 | + return v.writingBuffers[:packetCount], nil |
| 166 | +} |
| 167 | + |
| 168 | +// writePacketsToPacketConn writes packets from [VMPktDesc]s to the [net.PacketConn]. |
| 169 | +// - It returns an error if any occurs during sending packets. |
| 170 | +func (v *pktDescsManager) writePacketsToPacketConn(conn net.PacketConn, packetCount int) error { |
| 171 | + buffers, err := v.buffersForWritingToPacketConn(packetCount) |
| 172 | + if err != nil { |
| 173 | + return fmt.Errorf("buffersForWritingToPacketConn failed: %w", err) |
| 174 | + } |
| 175 | + // Get rawConn for syscall.Sendmsg |
| 176 | + rawConn, _ := conn.(syscall.Conn).SyscallConn() |
| 177 | + var sentCount int |
| 178 | + var sendErr error |
| 179 | + rawConnWriteErr := rawConn.Write(func(fd uintptr) (done bool) { |
| 180 | + for sentCount < packetCount { |
| 181 | + // send packet from buffer |
| 182 | + if err := syscall.Sendmsg(int(fd), buffers[sentCount], nil, nil, 0); err != nil { |
| 183 | + if errors.Is(err, syscall.EAGAIN) || errors.Is(err, syscall.ENOBUFS) { |
| 184 | + return false // try again later |
| 185 | + } |
| 186 | + sendErr = fmt.Errorf("syscall.Sendmsg failed: %w", err) |
| 187 | + return true |
| 188 | + } |
| 189 | + sentCount++ |
| 190 | + } |
| 191 | + return true |
| 192 | + }) |
| 193 | + if rawConnWriteErr != nil { |
| 194 | + return fmt.Errorf("rawConn.Write failed: %w", rawConnWriteErr) |
| 195 | + } |
| 196 | + if sendErr != nil { |
| 197 | + return sendErr |
| 198 | + } |
| 199 | + return nil |
| 200 | +} |
| 201 | + |
| 202 | +// readPacketsFromPacketConn reads packets from the [net.PacketConn] into [VMPktDesc]s. |
| 203 | +// - It returns the number of packets read. |
| 204 | +// - The packets are expected to come one by one. |
| 205 | +// - It receives all available packets until no more packets are available, packetCount reaches maxPacketCount, or an error occurs. |
| 206 | +// - It waits for the connection to be ready for initial packet. |
| 207 | +func (v *pktDescsManager) readPacketsFromPacketConn(conn net.PacketConn) (int, error) { |
| 208 | + var packetCount int |
| 209 | + // Read the first packet (blocking) |
| 210 | + n, _, err := conn.ReadFrom(v.backingBuffers[packetCount][headerSize:]) |
| 211 | + if n == 0 { |
| 212 | + // normal closure. Will this happen in datagram socket? |
| 213 | + return 0, errors.New("conn.ReadFrom: use of closed network connection") |
| 214 | + } |
| 215 | + if err != nil { |
| 216 | + return 0, fmt.Errorf("conn.ReadFrom failed: %w", err) |
| 217 | + } |
| 218 | + v.at(packetCount).SetPacketSize(n) |
| 219 | + packetCount++ |
| 220 | + // Get rawConn for syscall.Recvfrom |
| 221 | + rawConn, _ := conn.(syscall.Conn).SyscallConn() |
| 222 | + var recvErr error |
| 223 | + rawConnReadErr := rawConn.Read(func(fd uintptr) (done bool) { |
| 224 | + // Read available packets until no more packets are available or packetCount reaches maxPacketCount |
| 225 | + for packetCount < v.maxPacketCount { |
| 226 | + // receive packet into buffer |
| 227 | + n, _, err := syscall.Recvfrom(int(fd), v.backingBuffers[packetCount][headerSize:], 0) |
| 228 | + if err != nil { |
| 229 | + if !errors.Is(err, syscall.EAGAIN) { |
| 230 | + recvErr = fmt.Errorf("syscall.Recvfrom failed: %w", err) |
| 231 | + } |
| 232 | + return true // Do not retry on error |
| 233 | + } |
| 234 | + v.at(packetCount).SetPacketSize(n) |
| 235 | + packetCount++ |
| 236 | + } |
| 237 | + return true |
| 238 | + }) |
| 239 | + if rawConnReadErr != nil { |
| 240 | + return 0, fmt.Errorf("rawConn.Read failed: %w", rawConnReadErr) |
| 241 | + } |
| 242 | + if recvErr != nil { |
| 243 | + return 0, recvErr |
| 244 | + } |
| 245 | + return packetCount, nil |
| 246 | +} |
0 commit comments