| 1 | /*++ |
| 2 | |
| 3 | Copyright (c) Microsoft. All rights reserved. |
| 4 | |
| 5 | Module Name: |
| 6 | |
| 7 | WSLCVirtualMachine.cpp |
| 8 | |
| 9 | Abstract: |
| 10 | |
| 11 | Client-side class for WSLC virtual machine operations. |
| 12 | The VM is created via IWSLCVirtualMachine (running in the SYSTEM service). |
| 13 | This class connects to the existing VM for unprivileged operations |
| 14 | and delegates privileged operations back to IWSLCVirtualMachine. |
| 15 | |
| 16 | --*/ |
| 17 | |
| 18 | #include "precomp.h" |
| 19 | #include "WSLCVirtualMachine.h" |
| 20 | #include <format> |
| 21 | #include <filesystem> |
| 22 | #include <nlohmann/json.hpp> |
| 23 | #include "ServiceProcessLauncher.h" |
| 24 | #include "wslutil.h" |
| 25 | #include "lxinitshared.h" |
| 26 | |
| 27 | using namespace wsl::windows::common; |
| 28 | using wsl::windows::common::io::HandleWrapper; |
| 29 | using wsl::windows::service::wslc::TypedHandle; |
| 30 | using wsl::windows::service::wslc::VmPortAllocation; |
| 31 | using wsl::windows::service::wslc::VMPortMapping; |
| 32 | using wsl::windows::service::wslc::WSLCProcess; |
| 33 | using wsl::windows::service::wslc::WSLCVirtualMachine; |
| 34 | namespace wslutil = wsl::windows::common::wslutil; |
| 35 | |
| 36 | constexpr auto CONTAINER_PORT_RANGE = std::pair<uint16_t, uint16_t>(20002, 65535); |
| 37 | |
| 38 | static_assert(c_ephemeralPortRange.second < CONTAINER_PORT_RANGE.first); |
| 39 | |
| 40 | namespace { |
| 41 | |
| 42 | // Escapes regex metacharacters in `input` so a literal hostname can be embedded into a BuildKit |
| 43 | // source-policy regex identifier (e.g. `myreg:5000` stays a literal match). |
| 44 | std::string EscapeRegexMetacharacters(std::string_view input) |
| 45 | { |
| 46 | static constexpr std::string_view c_metacharacters = R"(\.+*?()|[]{}^$)"; |
| 47 | std::string escaped; |
| 48 | escaped.reserve(input.size()); |
| 49 | for (const char ch : input) |
| 50 | { |
| 51 | if (c_metacharacters.find(ch) != std::string_view::npos) |
| 52 | { |
| 53 | escaped.push_back('\\'); |
| 54 | } |
| 55 | escaped.push_back(ch); |
| 56 | } |
| 57 | return escaped; |
| 58 | } |
| 59 | |
| 60 | // DENY-all first, then per-host ALLOW: BuildKit evaluates rules in order and last match wins. |
| 61 | // Hosts are lowercased because BuildKit normalises identifiers before regex matching. |
| 62 | // https://github.com/moby/buildkit/blob/master/docs/sourcepolicy.md |
| 63 | std::string BuildBuildKitSourcePolicyJson(const std::vector<std::string>& allowedHosts) |
| 64 | { |
| 65 | nlohmann::json rules = nlohmann::json::array(); |
| 66 | rules.push_back({{"action", "DENY"}, {"selector", {{"identifier", "docker-image://.*"}, {"match_type", "REGEX"}}}}); |
| 67 | |
| 68 | for (const auto& host : allowedHosts) |
| 69 | { |
| 70 | std::string lowered; |
| 71 | lowered.reserve(host.size()); |
| 72 | std::transform(host.begin(), host.end(), std::back_inserter(lowered), [](unsigned char ch) { |
| 73 | return static_cast<char>(std::tolower(ch)); |
| 74 | }); |
| 75 | |
| 76 | const auto identifier = "docker-image://" + EscapeRegexMetacharacters(lowered) + "/.*"; |
| 77 | rules.push_back({{"action", "ALLOW"}, {"selector", {{"identifier", identifier}, {"match_type", "REGEX"}}}}); |
| 78 | } |
| 79 | |
| 80 | nlohmann::json document; |
| 81 | document["rules"] = std::move(rules); |
| 82 | return document.dump(); |
| 83 | } |
| 84 | |
| 85 | } // namespace |
| 86 | |
| 87 | VmPortAllocation::VmPortAllocation(uint16_t port, int family, int protocol, std::weak_ptr<VmPortReservations> reservations) : |
| 88 | m_port(port), m_family(family), m_protocol(protocol), m_reservations(std::move(reservations)) |
| 89 | { |
| 90 | } |
| 91 | |
| 92 | VmPortAllocation::VmPortAllocation(VmPortAllocation&& Other) |
| 93 | { |
| 94 | *this = std::move(Other); |
| 95 | } |
| 96 | |
| 97 | VmPortAllocation& VmPortAllocation::operator=(VmPortAllocation&& Other) |
| 98 | { |
| 99 | if (this != &Other) |
| 100 | { |
| 101 | Reset(); |
| 102 | m_port = Other.m_port; |
| 103 | m_family = Other.m_family; |
| 104 | m_protocol = Other.m_protocol; |
| 105 | m_reservations = Other.m_reservations; |
| 106 | |
| 107 | Other.Release(); |
| 108 | } |
| 109 | return *this; |
| 110 | } |
| 111 | |
| 112 | VmPortAllocation::~VmPortAllocation() |
| 113 | { |
| 114 | Reset(); |
| 115 | } |
| 116 | |
| 117 | void VmPortAllocation::Reset() |
| 118 | { |
| 119 | // Release the reservation only if the owning VM (and its table) is still alive. If the VM was torn |
| 120 | // down the table is already gone and lock() returns null, so a surviving allocation is a safe no-op. |
| 121 | if (auto reservations = m_reservations.lock()) |
| 122 | { |
| 123 | std::lock_guard lock{reservations->Mutex}; |
| 124 | LOG_HR_IF(E_UNEXPECTED, reservations->Ports.erase(m_port) != 1); |
| 125 | } |
| 126 | |
| 127 | Release(); |
| 128 | } |
| 129 | |
| 130 | void VmPortAllocation::Release() |
| 131 | { |
| 132 | m_reservations.reset(); |
| 133 | m_port = 0; |
| 134 | m_family = 0; |
| 135 | m_protocol = 0; |
| 136 | } |
| 137 | |
| 138 | uint16_t VmPortAllocation::Port() const |
| 139 | { |
| 140 | return m_port; |
| 141 | } |
| 142 | |
| 143 | int VmPortAllocation::Family() const |
| 144 | { |
| 145 | return m_family; |
| 146 | } |
| 147 | |
| 148 | int VmPortAllocation::Protocol() const |
| 149 | { |
| 150 | return m_protocol; |
| 151 | } |
| 152 | |
| 153 | VMPortMapping::VMPortMapping(int protocol, int Family, uint16_t Port, const char* Address) : Protocol(protocol) |
| 154 | { |
| 155 | THROW_HR_IF_MSG(E_INVALIDARG, Protocol != IPPROTO_TCP && Protocol != IPPROTO_UDP, "Invalid protocol: %i", Protocol); |
| 156 | THROW_HR_IF(E_POINTER, Address == nullptr); |
| 157 | if (Family == AF_INET) |
| 158 | { |
| 159 | common::wslutil::ParseIpv4Address(Address, BindAddress.Ipv4.sin_addr); |
| 160 | BindAddress.Ipv4.sin_port = htons(Port); |
| 161 | } |
| 162 | else if (Family == AF_INET6) |
| 163 | { |
| 164 | common::wslutil::ParseIpv6Address(Address, BindAddress.Ipv6.sin6_addr); |
| 165 | BindAddress.Ipv6.sin6_port = htons(Port); |
| 166 | } |
| 167 | else |
| 168 | { |
| 169 | THROW_HR_MSG(E_INVALIDARG, "Invalid address family: %i", Family); |
| 170 | } |
| 171 | |
| 172 | // Must be assigned after parsing is done, since inet_pton writes to the family field as well. |
| 173 | BindAddress.si_family = Family; |
| 174 | } |
| 175 | |
| 176 | VMPortMapping::~VMPortMapping() |
| 177 | { |
| 178 | try |
| 179 | { |
| 180 | Unmap(); |
| 181 | } |
| 182 | CATCH_LOG(); |
| 183 | } |
| 184 | |
| 185 | VMPortMapping::VMPortMapping(VMPortMapping&& Other) |
| 186 | { |
| 187 | *this = std::move(Other); |
| 188 | } |
| 189 | |
| 190 | void VMPortMapping::AssignVmPort(const std::shared_ptr<VmPortAllocation>& Port) |
| 191 | { |
| 192 | WI_ASSERT(!VmPort); |
| 193 | |
| 194 | VmPort = Port; |
| 195 | } |
| 196 | |
| 197 | void VMPortMapping::Unmap() |
| 198 | { |
| 199 | if (Vm) |
| 200 | { |
| 201 | auto clearVm = wil::scope_exit([&] { Vm = nullptr; }); |
| 202 | Vm->UnmapPort(*this); |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | void VMPortMapping::Release() |
| 207 | { |
| 208 | Vm = nullptr; |
| 209 | VmPort.reset(); |
| 210 | } |
| 211 | |
| 212 | bool VMPortMapping::IsLocalhost() const |
| 213 | { |
| 214 | if (BindAddress.si_family == AF_INET6) |
| 215 | { |
| 216 | return IN6_IS_ADDR_LOOPBACK(&BindAddress.Ipv6.sin6_addr); |
| 217 | } |
| 218 | else |
| 219 | { |
| 220 | return IN4ADDR_ISLOOPBACK(&BindAddress.Ipv4); |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | bool VMPortMapping::IsIPv6() const |
| 225 | { |
| 226 | return BindAddress.si_family == AF_INET6; |
| 227 | } |
| 228 | |
| 229 | uint16_t VMPortMapping::HostPort() const |
| 230 | { |
| 231 | if (BindAddress.si_family == AF_INET6) |
| 232 | { |
| 233 | return ntohs(BindAddress.Ipv6.sin6_port); |
| 234 | } |
| 235 | else |
| 236 | { |
| 237 | WI_ASSERT(BindAddress.si_family == AF_INET); |
| 238 | return ntohs(BindAddress.Ipv4.sin_port); |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | void VMPortMapping::SetHostPort(uint16_t port) |
| 243 | { |
| 244 | if (BindAddress.si_family == AF_INET6) |
| 245 | { |
| 246 | BindAddress.Ipv6.sin6_port = htons(port); |
| 247 | } |
| 248 | else |
| 249 | { |
| 250 | WI_ASSERT(BindAddress.si_family == AF_INET); |
| 251 | BindAddress.Ipv4.sin_port = htons(port); |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | std::string VMPortMapping::BindingAddressString() const |
| 256 | { |
| 257 | char buffer[INET6_ADDRSTRLEN]{}; |
| 258 | if (BindAddress.Ipv4.sin_family == AF_INET6) |
| 259 | { |
| 260 | THROW_LAST_ERROR_IF(inet_ntop(AF_INET6, &BindAddress.Ipv6.sin6_addr, buffer, sizeof(buffer)) == nullptr); |
| 261 | } |
| 262 | else |
| 263 | { |
| 264 | THROW_LAST_ERROR_IF(inet_ntop(AF_INET, &BindAddress.Ipv4.sin_addr, buffer, sizeof(buffer)) == nullptr); |
| 265 | } |
| 266 | |
| 267 | return buffer; |
| 268 | } |
| 269 | |
| 270 | void VMPortMapping::Attach(WSLCVirtualMachine& Vm) |
| 271 | { |
| 272 | WI_ASSERT(this->Vm == nullptr); |
| 273 | |
| 274 | this->Vm = &Vm; |
| 275 | } |
| 276 | |
| 277 | void VMPortMapping::Detach() |
| 278 | { |
| 279 | WI_ASSERT(Vm != nullptr); |
| 280 | |
| 281 | this->Vm = nullptr; |
| 282 | } |
| 283 | |
| 284 | VMPortMapping VMPortMapping::LocalhostTcpMapping(int Family, uint16_t WindowsPort) |
| 285 | { |
| 286 | WI_ASSERT(Family == AF_INET || Family == AF_INET6); |
| 287 | |
| 288 | return VMPortMapping(IPPROTO_TCP, Family, WindowsPort, Family == AF_INET ? "127.0.0.1" : "::1"); |
| 289 | } |
| 290 | |
| 291 | VMPortMapping VMPortMapping::FromWSLCPortMapping(const ::WSLCPortMapping& Mapping) |
| 292 | { |
| 293 | return VMPortMapping(Mapping.Protocol, Mapping.Family, Mapping.HostPort, Mapping.BindingAddress); |
| 294 | } |
| 295 | |
| 296 | VMPortMapping VMPortMapping::FromContainerMetaData(const wslc::WSLCPortMapping& Mapping) |
| 297 | { |
| 298 | return VMPortMapping(Mapping.Protocol, Mapping.Family, Mapping.HostPort, Mapping.BindingAddress.c_str()); |
| 299 | } |
| 300 | |
| 301 | VMPortMapping& VMPortMapping::operator=(VMPortMapping&& Other) |
| 302 | { |
| 303 | if (this != &Other) |
| 304 | { |
| 305 | Unmap(); |
| 306 | Protocol = Other.Protocol; |
| 307 | VmPort = std::move(Other.VmPort); |
| 308 | BindAddress = Other.BindAddress; |
| 309 | Vm = Other.Vm; |
| 310 | |
| 311 | Other.Protocol = 0; |
| 312 | ZeroMemory(&Other.BindAddress, sizeof(Other.BindAddress)); |
| 313 | Other.Vm = nullptr; |
| 314 | } |
| 315 | return *this; |
| 316 | } |
| 317 | |
| 318 | WSLCVirtualMachine::WSLCVirtualMachine( |
| 319 | _In_ IWSLCVirtualMachine* Vm, _In_ const WSLCSessionInitSettings* Settings, _In_ HANDLE SessionTerminatingEvent, _In_ TOnCrashDump&& OnCrashDump) : |
| 320 | m_vm(Vm), |
| 321 | m_featureFlags(static_cast<WSLCFeatureFlags>(Settings->FeatureFlags)), |
| 322 | m_networkingMode(Settings->NetworkingMode), |
| 323 | m_bootTimeoutMs(Settings->BootTimeoutMs), |
| 324 | m_rootVhdType(Settings->RootVhdTypeOverride ? Settings->RootVhdTypeOverride : "ext4"), |
| 325 | m_onCrashDump(std::move(OnCrashDump)), |
| 326 | m_sessionTerminatingEvent(SessionTerminatingEvent) |
| 327 | { |
| 328 | // N.B. The constructor should not run any operation that could throw, so the destructor runs even if the VM fails to boot. |
| 329 | } |
| 330 | |
| 331 | void WSLCVirtualMachine::Initialize() |
| 332 | { |
| 333 | THROW_IF_FAILED(m_vm->GetId(&m_vmId)); |
| 334 | |
| 335 | // Create a job object that will terminate child processes (wslrelay.exe) |
| 336 | // when the VM is destroyed. |
| 337 | m_processJobObject = wsl::windows::common::helpers::CreateKillOnCloseJob(); |
| 338 | |
| 339 | // Start crash dump collection thread. |
| 340 | auto crashDumpSocket = hvsocket::Listen(m_vmId, LX_INIT_UTILITY_VM_CRASH_DUMP_PORT); |
| 341 | THROW_LAST_ERROR_IF(!crashDumpSocket); |
| 342 | |
| 343 | m_crashDumpThread = std::thread{[this, socket = std::move(crashDumpSocket)]() mutable { CollectCrashDumps(std::move(socket)); }}; |
| 344 | |
| 345 | // Establish a socket channel with mini_init in the VM. |
| 346 | wil::unique_socket socket; |
| 347 | THROW_IF_FAILED(m_vm->AcceptConnection(reinterpret_cast<HANDLE*>(&socket))); |
| 348 | |
| 349 | m_initChannel = wsl::shared::SocketChannel{std::move(socket), "mini_init", {m_vmTerminatingEvent.get(), m_sessionTerminatingEvent}}; |
| 350 | |
| 351 | // Create a thread to watch for exited processes. |
| 352 | auto [__, ___, childChannel] = Fork(WSLC_FORK::Thread); |
| 353 | childChannel.SetExitEvents({m_vmTerminatingEvent.get()}); |
| 354 | |
| 355 | WSLC_WATCH_PROCESSES watchMessage{}; |
| 356 | auto watchTransaction = childChannel.StartTransaction(m_initChannelTimeout); |
| 357 | watchTransaction.Send(watchMessage); |
| 358 | |
| 359 | THROW_HR_IF(E_FAIL, watchTransaction.Receive<RESULT_MESSAGE<uint32_t>>().Result != 0); |
| 360 | |
| 361 | m_processExitThread = std::thread(std::bind(&WSLCVirtualMachine::WatchForExitedProcesses, this, std::move(childChannel))); |
| 362 | |
| 363 | // Mount VHDs |
| 364 | const auto rootDevice = GetVhdDevicePath(0); |
| 365 | Mount(m_initChannel, rootDevice.c_str(), "/mnt", m_rootVhdType.c_str(), "ro", WSLC_MOUNT::Chroot | WSLC_MOUNT::OverlayFs); |
| 366 | |
| 367 | const auto modulesDevice = GetVhdDevicePath(1); |
| 368 | MountModules(m_initChannel, modulesDevice.c_str()); |
| 369 | |
| 370 | // Discover the per-VM guest capabilities (currently the hv_pci swiotlb pool) and forward them |
| 371 | // to the service before virtiofs shares or Consomme networking devices are created. |
| 372 | ReadGuestCapabilities(); |
| 373 | |
| 374 | // Configure GPU mounts if enabled |
| 375 | MountGpuLibraries(c_gpuLibrariesPath, c_gpuDriversPath); |
| 376 | |
| 377 | // Snapshot the container-registry allowlist and, if configured, hand the BuildKit source-policy |
| 378 | // JSON to init. Done at boot rather than per build so a compromised user process cannot bypass |
| 379 | // enforcement by racing the write. |
| 380 | ConfigureBuildKitPolicy(); |
| 381 | |
| 382 | // Configure networking. This must happen after all filesystems are mounted since /gns needs to access /sys. |
| 383 | ConfigureNetworking(); |
| 384 | } |
| 385 | |
| 386 | WSLCVirtualMachine::~WSLCVirtualMachine() |
| 387 | { |
| 388 | WSL_LOG("WSLCTerminateVmStart"); |
| 389 | |
| 390 | m_vmTerminatingEvent.SetEvent(); |
| 391 | |
| 392 | m_initChannel.Close(); |
| 393 | |
| 394 | // Terminate the VM. |
| 395 | m_vm.reset(); |
| 396 | |
| 397 | if (m_processExitThread.joinable()) |
| 398 | { |
| 399 | m_processExitThread.join(); |
| 400 | } |
| 401 | |
| 402 | if (m_crashDumpThread.joinable()) |
| 403 | { |
| 404 | m_crashDumpThread.join(); |
| 405 | } |
| 406 | |
| 407 | // Clear the state of all remaining processes now that the VM has exited. |
| 408 | for (auto& e : m_trackedProcesses) |
| 409 | { |
| 410 | if (auto locked = e.lock()) |
| 411 | { |
| 412 | locked->OnVmTerminated(); |
| 413 | } |
| 414 | } |
| 415 | } |
| 416 | |
| 417 | void WSLCVirtualMachine::ConfigureNetworking() |
| 418 | { |
| 419 | if (m_networkingMode == WSLCNetworkingModeNone) |
| 420 | { |
| 421 | return; |
| 422 | } |
| 423 | |
| 424 | // Launch /gns with auto-allocated file descriptors for the GNS channel (and DNS channel if enabled). |
| 425 | std::vector<WSLCProcessFd> fds; |
| 426 | fds.emplace_back(WSLCProcessFd{.Fd = -1, .Type = WSLCFdType::WSLCFdTypeDefault}); |
| 427 | |
| 428 | // Consomme forwards DNS via the host proxy, so the DNS channel and /gns args are only needed for NAT mode. |
| 429 | const bool enableDnsTunneling = FeatureEnabled(WslcFeatureFlagsDnsTunneling) && m_networkingMode != WSLCNetworkingModeConsomme; |
| 430 | if (enableDnsTunneling) |
| 431 | { |
| 432 | fds.emplace_back(WSLCProcessFd{.Fd = -1, .Type = WSLCFdType::WSLCFdTypeDefault}); |
| 433 | } |
| 434 | |
| 435 | // Because the file descriptor numbers aren't known in advance, the command line needs to be generated after the |
| 436 | // file descriptors are allocated. |
| 437 | std::vector<const char*> cmd{"/gns", LX_INIT_GNS_SOCKET_ARG}; |
| 438 | std::string gnsSocketFdArg; |
| 439 | std::string dnsSocketFdArg; |
| 440 | int gnsChannelFd = -1; |
| 441 | int dnsChannelFd = -1; |
| 442 | |
| 443 | WSLCProcessOptions options{}; |
| 444 | auto prepareCommandLine = [&](const auto& sockets) { |
| 445 | gnsChannelFd = sockets[0].Fd; |
| 446 | gnsSocketFdArg = std::to_string(gnsChannelFd); |
| 447 | cmd.push_back(gnsSocketFdArg.c_str()); |
| 448 | |
| 449 | if (enableDnsTunneling) |
| 450 | { |
| 451 | dnsChannelFd = sockets[1].Fd; |
| 452 | dnsSocketFdArg = std::to_string(dnsChannelFd); |
| 453 | cmd.push_back(LX_INIT_GNS_DNS_SOCKET_ARG); |
| 454 | cmd.push_back(dnsSocketFdArg.c_str()); |
| 455 | cmd.push_back(LX_INIT_GNS_DNS_TUNNELING_IP); |
| 456 | cmd.push_back(LX_INIT_DNS_TUNNELING_IP_ADDRESS); |
| 457 | } |
| 458 | |
| 459 | options.CommandLine = {.Values = cmd.data(), .Count = static_cast<ULONG>(cmd.size())}; |
| 460 | }; |
| 461 | |
| 462 | auto process = CreateLinuxProcessImpl("/init", options, fds, 0, 0, nullptr, prepareCommandLine); |
| 463 | |
| 464 | // Call back to the service to configure the networking engine. |
| 465 | auto gnsHandle = process->GetStdHandle(gnsChannelFd); |
| 466 | |
| 467 | HandleWrapper dnsHandle; |
| 468 | HANDLE dnsSocketHandle = nullptr; |
| 469 | if (enableDnsTunneling) |
| 470 | { |
| 471 | dnsHandle = process->GetStdHandle(dnsChannelFd); |
| 472 | dnsSocketHandle = dnsHandle.Get(); |
| 473 | } |
| 474 | |
| 475 | THROW_IF_FAILED(m_vm->ConfigureNetworking(gnsHandle.Get(), enableDnsTunneling ? &dnsSocketHandle : nullptr)); |
| 476 | |
| 477 | // Launch port relay for port forwarding |
| 478 | LaunchPortRelay(); |
| 479 | } |
| 480 | |
| 481 | void WSLCVirtualMachine::ReadGuestCapabilities() |
| 482 | { |
| 483 | WSLC_GET_GUEST_CAPABILITIES message{}; |
| 484 | const auto& response = m_initChannel.Transaction(message, nullptr, m_initChannelTimeout); |
| 485 | |
| 486 | m_hvPciSwiotlbBase = response.HvPciSwiotlbBase; |
| 487 | m_hvPciSwiotlbSize = response.HvPciSwiotlbSize; |
| 488 | |
| 489 | WSL_LOG( |
| 490 | "WSLCReadGuestCapabilities", |
| 491 | TraceLoggingValue(m_hvPciSwiotlbBase, "HvPciSwiotlbBase"), |
| 492 | TraceLoggingValue(m_hvPciSwiotlbSize, "HvPciSwiotlbSize")); |
| 493 | |
| 494 | // Forward the values to the service so AddShare and ConfigureNetworking can configure |
| 495 | // wsldevicehost. Passing zero for both means the guest kernel does not support hv_pci swiotlb. |
| 496 | WSLCGuestCapabilities capabilities{}; |
| 497 | capabilities.HvPciSwiotlbBase = m_hvPciSwiotlbBase; |
| 498 | capabilities.HvPciSwiotlbSize = m_hvPciSwiotlbSize; |
| 499 | THROW_IF_FAILED(m_vm->ApplyGuestCapabilities(&capabilities)); |
| 500 | } |
| 501 | |
| 502 | void WSLCVirtualMachine::ConfigureBuildKitPolicy() |
| 503 | { |
| 504 | const auto snapshot = wsl::windows::policies::ReadRegistryAllowlistSnapshotFromPoliciesRoot(); |
| 505 | |
| 506 | if (snapshot.State == wsl::windows::policies::RegistryAllowlistState::NotConfigured) |
| 507 | { |
| 508 | m_buildKitPolicyState = BuildKitPolicyState::NotConfigured; |
| 509 | return; |
| 510 | } |
| 511 | |
| 512 | std::vector<std::string> hosts; |
| 513 | hosts.reserve(snapshot.Hosts.size()); |
| 514 | std::ranges::transform(snapshot.Hosts, std::back_inserter(hosts), [](const std::wstring& host) { |
| 515 | return wsl::shared::string::WideToMultiByte(host); |
| 516 | }); |
| 517 | |
| 518 | const auto policyJson = BuildBuildKitSourcePolicyJson(hosts); |
| 519 | |
| 520 | // Linux <fcntl.h> flags for open(). |
| 521 | constexpr int c_lxOWriteOnly = 0x1; |
| 522 | constexpr int c_lxOCreate = 0x40; |
| 523 | constexpr int c_lxOTruncate = 0x200; |
| 524 | constexpr int c_lxOCloseOnExec = 0x80000; |
| 525 | constexpr int c_lxONoFollow = 0x20000; |
| 526 | |
| 527 | auto message = wsl::shared::MessageWriter<WSLC_WRITE_FILE>{}; |
| 528 | message.WriteString(message->PathIndex, c_buildKitPolicyPath); |
| 529 | message->ContentLength = static_cast<unsigned int>(policyJson.size()); |
| 530 | gsl::copy( |
| 531 | gsl::as_bytes(gsl::make_span(policyJson.data(), policyJson.size())), |
| 532 | message.InsertBuffer(message->ContentIndex, policyJson.size())); |
| 533 | message->OpenFlags = c_lxOWriteOnly | c_lxOCreate | c_lxOTruncate | c_lxOCloseOnExec | c_lxONoFollow; |
| 534 | message->Permissions = 0644; |
| 535 | |
| 536 | const auto& response = m_initChannel.Transaction<WSLC_WRITE_FILE>(message.Span(), nullptr, m_initChannelTimeout); |
| 537 | THROW_HR_IF_MSG(E_FAIL, response.Result != 0, "Guest failed to write %hs: %d", c_buildKitPolicyPath, response.Result); |
| 538 | |
| 539 | m_buildKitPolicyState = BuildKitPolicyState::Configured; |
| 540 | } |
| 541 | |
| 542 | bool WSLCVirtualMachine::FeatureEnabled(WSLCFeatureFlags Value) const |
| 543 | { |
| 544 | return static_cast<ULONG>(m_featureFlags) & static_cast<ULONG>(Value); |
| 545 | } |
| 546 | |
| 547 | WSLCNetworkingMode WSLCVirtualMachine::NetworkingMode() const |
| 548 | { |
| 549 | return m_networkingMode; |
| 550 | } |
| 551 | |
| 552 | bool WSLCVirtualMachine::UseWslRelayPortForwarding() const |
| 553 | { |
| 554 | return m_networkingMode == WSLCNetworkingModeNAT || |
| 555 | (m_networkingMode == WSLCNetworkingModeConsomme && FeatureEnabled(WslcFeatureFlagsPortRelayWslRelay)); |
| 556 | } |
| 557 | |
| 558 | void WSLCVirtualMachine::WatchForExitedProcesses(wsl::shared::SocketChannel& Channel) |
| 559 | try |
| 560 | { |
| 561 | // TODO: Terminate the VM if this thread exits unexpectedly. |
| 562 | while (true) |
| 563 | { |
| 564 | auto [message, _] = Channel.ReceiveMessageOrClosed<WSLC_PROCESS_EXITED>(); |
| 565 | if (message == nullptr) |
| 566 | { |
| 567 | break; // Channel has been closed, exit |
| 568 | } |
| 569 | |
| 570 | WSL_LOG( |
| 571 | "ProcessExited", |
| 572 | TraceLoggingValue(message->Pid, "Pid"), |
| 573 | TraceLoggingValue(message->Code, "Code"), |
| 574 | TraceLoggingValue(message->Signaled, "Signaled")); |
| 575 | |
| 576 | // Signal the exited process, if it's been monitored. |
| 577 | // N.B. Lock weak_ptr under lock, then call OnExited outside it to avoid |
| 578 | // deadlock with the destructor's m_lock -> m_trackedProcessesLock ordering. |
| 579 | std::shared_ptr<VMProcessControl> exited; |
| 580 | { |
| 581 | std::lock_guard lock{m_trackedProcessesLock}; |
| 582 | |
| 583 | for (auto& e : m_trackedProcesses) |
| 584 | { |
| 585 | auto locked = e.lock(); |
| 586 | if (locked && locked->GetPid() == message->Pid) |
| 587 | { |
| 588 | exited = std::move(locked); |
| 589 | break; |
| 590 | } |
| 591 | } |
| 592 | } |
| 593 | |
| 594 | if (exited) |
| 595 | { |
| 596 | try |
| 597 | { |
| 598 | exited->OnExited(message->Signaled ? 128 + message->Code : message->Code); |
| 599 | } |
| 600 | CATCH_LOG(); |
| 601 | } |
| 602 | } |
| 603 | } |
| 604 | CATCH_LOG(); |
| 605 | |
| 606 | std::pair<ULONG, std::string> WSLCVirtualMachine::AttachDisk(_In_ PCWSTR Path, _In_ BOOL ReadOnly) |
| 607 | { |
| 608 | std::lock_guard lock{m_lock}; |
| 609 | |
| 610 | ULONG Lun{}; |
| 611 | std::string Device; |
| 612 | |
| 613 | // Delegate to IWSLCVirtualMachine for the privileged HCS operation |
| 614 | THROW_IF_FAILED(m_vm->AttachDisk(Path, ReadOnly, &Lun)); |
| 615 | |
| 616 | // Detach on failure so the service-side state stays consistent. |
| 617 | auto detachOnFailure = wil::scope_exit_log(WI_DIAGNOSTICS_INFO, [&]() { LOG_IF_FAILED(m_vm->DetachDisk(Lun)); }); |
| 618 | |
| 619 | // Query the guest for the device path |
| 620 | Device = GetVhdDevicePath(Lun); |
| 621 | |
| 622 | WSL_LOG( |
| 623 | "WSLCAttachDisk", |
| 624 | TraceLoggingValue(Path, "Path"), |
| 625 | TraceLoggingValue(ReadOnly, "ReadOnly"), |
| 626 | TraceLoggingValue(Device.c_str(), "Device"), |
| 627 | TraceLoggingValue(Lun, "Lun")); |
| 628 | |
| 629 | m_attachedDisks.emplace(Lun, AttachedDisk{Path, Device}); |
| 630 | |
| 631 | detachOnFailure.release(); |
| 632 | |
| 633 | return {Lun, Device}; |
| 634 | } |
| 635 | |
| 636 | void WSLCVirtualMachine::Ext4Format(const std::string& Device, std::optional<uint32_t> Uid, std::optional<uint32_t> Gid) |
| 637 | { |
| 638 | constexpr auto mkfsPath = "/usr/sbin/mkfs.ext4"; |
| 639 | |
| 640 | // Uid/Gid must be paired; the named-volume parser enforces this for user |
| 641 | // input — this guards future internal callers that bypass it. |
| 642 | THROW_HR_IF(E_UNEXPECTED, Uid.has_value() != Gid.has_value()); |
| 643 | |
| 644 | std::vector<std::string> args = {mkfsPath}; |
| 645 | std::string rootOwner; |
| 646 | if (Uid.has_value() && Gid.has_value()) |
| 647 | { |
| 648 | rootOwner = std::format("root_owner={}:{}", *Uid, *Gid); |
| 649 | args.push_back("-E"); |
| 650 | args.push_back(rootOwner); |
| 651 | } |
| 652 | args.push_back(Device); |
| 653 | |
| 654 | ServiceProcessLauncher launcher(mkfsPath, args); |
| 655 | auto result = launcher.Launch(*this).WaitAndCaptureOutput(); |
| 656 | |
| 657 | THROW_HR_IF_MSG(E_FAIL, result.Code != 0, "%hs", launcher.FormatResult(result).c_str()); |
| 658 | } |
| 659 | |
| 660 | void WSLCVirtualMachine::RemoveDirectory(const std::string& Path) |
| 661 | { |
| 662 | // rmdir only removes an empty directory, so callers can rely on it to leave |
| 663 | // a non-empty directory untouched. |
| 664 | constexpr auto rmdirPath = "/bin/rmdir"; |
| 665 | |
| 666 | std::vector<std::string> args = {rmdirPath, Path}; |
| 667 | |
| 668 | ServiceProcessLauncher launcher(rmdirPath, args); |
| 669 | auto result = launcher.Launch(*this).WaitAndCaptureOutput(); |
| 670 | |
| 671 | THROW_HR_IF_MSG(E_FAIL, result.Code != 0, "%hs", launcher.FormatResult(result).c_str()); |
| 672 | } |
| 673 | |
| 674 | std::vector<std::string> WSLCVirtualMachine::ListDirectory(const std::string& Path) |
| 675 | { |
| 676 | wsl::shared::MessageWriter<WSLC_LISTDIR> message; |
| 677 | message.WriteString(Path); |
| 678 | |
| 679 | gsl::span<gsl::byte> responseSpan; |
| 680 | const auto& response = m_initChannel.Transaction<WSLC_LISTDIR>(message.Span(), &responseSpan, m_initChannelTimeout); |
| 681 | |
| 682 | THROW_HR_IF_MSG(E_FAIL, response.Result != 0, "Failed to list directory '%hs', init returned: %d", Path.c_str(), response.Result); |
| 683 | |
| 684 | return wsl::shared::string::ArrayFromSpan(responseSpan, response.EntriesIndex); |
| 685 | } |
| 686 | |
| 687 | void WSLCVirtualMachine::Unmount(_In_ const char* Path) |
| 688 | { |
| 689 | auto [pid, _, subChannel] = Fork(WSLC_FORK::Thread); |
| 690 | |
| 691 | wsl::shared::MessageWriter<WSLC_UNMOUNT> message; |
| 692 | message.WriteString(Path); |
| 693 | |
| 694 | const auto& response = subChannel.Transaction<WSLC_UNMOUNT>(message.Span(), nullptr, m_initChannelTimeout); |
| 695 | |
| 696 | // TODO: Return errno to caller |
| 697 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_NOT_FOUND), response.Result == EINVAL); |
| 698 | THROW_HR_IF(E_FAIL, response.Result != 0); |
| 699 | } |
| 700 | |
| 701 | void WSLCVirtualMachine::DetachDisk(_In_ ULONG Lun) |
| 702 | { |
| 703 | std::lock_guard lock{m_lock}; |
| 704 | |
| 705 | // Find the disk |
| 706 | auto it = m_attachedDisks.find(Lun); |
| 707 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_NOT_FOUND), it == m_attachedDisks.end()); |
| 708 | |
| 709 | // Detach it from the guest |
| 710 | WSLC_DETACH message; |
| 711 | message.Lun = Lun; |
| 712 | const auto& response = m_initChannel.Transaction(message, nullptr, m_initChannelTimeout); |
| 713 | |
| 714 | // TODO: Return errno to caller |
| 715 | THROW_HR_IF(E_FAIL, response.Result != 0); |
| 716 | |
| 717 | // Remove it from the VM |
| 718 | THROW_IF_FAILED(m_vm->DetachDisk(Lun)); |
| 719 | |
| 720 | m_attachedDisks.erase(it); |
| 721 | } |
| 722 | |
| 723 | std::tuple<int32_t, int32_t, wsl::shared::SocketChannel> WSLCVirtualMachine::Fork(enum WSLC_FORK::ForkType Type) |
| 724 | { |
| 725 | std::lock_guard lock{m_lock}; |
| 726 | return Fork(m_initChannel, Type); |
| 727 | } |
| 728 | |
| 729 | std::tuple<int32_t, int32_t, wsl::shared::SocketChannel> WSLCVirtualMachine::Fork( |
| 730 | wsl::shared::SocketChannel& Channel, enum WSLC_FORK::ForkType Type, ULONG TtyRows, ULONG TtyColumns) |
| 731 | { |
| 732 | uint32_t port{}; |
| 733 | int32_t pid{}; |
| 734 | int32_t ptyMaster{}; |
| 735 | { |
| 736 | WSLC_FORK message; |
| 737 | message.ForkType = Type; |
| 738 | message.TtyColumns = static_cast<uint16_t>(TtyColumns); |
| 739 | message.TtyRows = static_cast<uint16_t>(TtyRows); |
| 740 | const auto& response = Channel.Transaction(message, nullptr, m_initChannelTimeout); |
| 741 | port = response.Port; |
| 742 | pid = response.Pid; |
| 743 | ptyMaster = response.PtyMasterFd; |
| 744 | } |
| 745 | |
| 746 | THROW_HR_IF_MSG(E_FAIL, pid <= 0, "fork() returned %i", pid); |
| 747 | |
| 748 | auto socket = wsl::windows::common::hvsocket::Connect(m_vmId, port, m_vmTerminatingEvent.get(), m_initChannelTimeout); |
| 749 | |
| 750 | return std::make_tuple( |
| 751 | pid, ptyMaster, wsl::shared::SocketChannel{std::move(socket), std::to_string(pid), std::vector<HANDLE>(Channel.GetExitEvents())}); |
| 752 | } |
| 753 | |
| 754 | WSLCVirtualMachine::ConnectedSocket WSLCVirtualMachine::ConnectSocket(wsl::shared::SocketChannel& Channel, int32_t Fd) |
| 755 | { |
| 756 | WSLC_ACCEPT message{}; |
| 757 | message.Fd = Fd; |
| 758 | |
| 759 | auto transaction = Channel.StartTransaction(m_initChannelTimeout); |
| 760 | transaction.Send(message); |
| 761 | const auto& response = transaction.Receive<WSLC_ACCEPT::TResponse>(); |
| 762 | |
| 763 | ConnectedSocket socket; |
| 764 | socket.Socket = wsl::windows::common::hvsocket::Connect(m_vmId, response.Result, m_vmTerminatingEvent.get(), m_initChannelTimeout); |
| 765 | |
| 766 | // If the FD was unspecified, read the Linux file descriptor from the guest. |
| 767 | if (Fd == -1) |
| 768 | { |
| 769 | socket.Fd = transaction.Receive<RESULT_MESSAGE<int32_t>>().Result; |
| 770 | } |
| 771 | else |
| 772 | { |
| 773 | socket.Fd = Fd; |
| 774 | } |
| 775 | |
| 776 | return socket; |
| 777 | } |
| 778 | |
| 779 | std::string WSLCVirtualMachine::GetVhdDevicePath(ULONG Lun) |
| 780 | { |
| 781 | WSLC_GET_DISK message{}; |
| 782 | message.Header.MessageSize = sizeof(message); |
| 783 | message.Header.MessageType = WSLC_GET_DISK::Type; |
| 784 | message.ScsiLun = Lun; |
| 785 | const auto& response = m_initChannel.Transaction(message, nullptr, m_initChannelTimeout); |
| 786 | THROW_HR_IF_MSG(E_FAIL, response.Result != 0, "Failed to get disk path, init returned: %lu", response.Result); |
| 787 | |
| 788 | return response.Buffer; |
| 789 | } |
| 790 | |
| 791 | Microsoft::WRL::ComPtr<WSLCProcess> WSLCVirtualMachine::CreateLinuxProcess( |
| 792 | _In_ LPCSTR Executable, _In_ const WSLCProcessOptions& Options, ULONG TtyRows, ULONG TtyColumns, int* Errno, const TPrepareCommandLine& PrepareCommandLine) |
| 793 | { |
| 794 | // Check if this is a tty or not |
| 795 | std::vector<WSLCProcessFd> fds; |
| 796 | if (WI_IsFlagSet(Options.Flags, WSLCProcessFlagsTty)) |
| 797 | { |
| 798 | fds.emplace_back(WSLCProcessFd{.Fd = WSLCFDTty, .Type = WSLCFdType::WSLCFdTypeTty}); |
| 799 | fds.emplace_back(WSLCProcessFd{.Fd = 0, .Type = WSLCFdType::WSLCFdTypeTtyControl}); |
| 800 | } |
| 801 | else |
| 802 | { |
| 803 | if (WI_IsFlagSet(Options.Flags, WSLCProcessFlagsStdin)) |
| 804 | { |
| 805 | fds.emplace_back(WSLCProcessFd{.Fd = WSLCFDStdin, .Type = WSLCFdType::WSLCFdTypeDefault}); |
| 806 | } |
| 807 | |
| 808 | fds.emplace_back(WSLCProcessFd{.Fd = WSLCFDStdout, .Type = WSLCFdType::WSLCFdTypeDefault}); |
| 809 | fds.emplace_back(WSLCProcessFd{.Fd = WSLCFDStderr, .Type = WSLCFdType::WSLCFdTypeDefault}); |
| 810 | } |
| 811 | |
| 812 | return CreateLinuxProcessImpl(Executable, Options, fds, TtyRows, TtyColumns, Errno, PrepareCommandLine); |
| 813 | } |
| 814 | |
| 815 | Microsoft::WRL::ComPtr<WSLCProcess> WSLCVirtualMachine::CreateLinuxProcessImpl( |
| 816 | LPCSTR Executable, const WSLCProcessOptions& Options, const std::vector<WSLCProcessFd>& Fds, ULONG TtyRows, ULONG TtyColumns, int* Errno, const TPrepareCommandLine& PrepareCommandLine) |
| 817 | { |
| 818 | // N.B This check is there to prevent processes from being started before the VM is done initializing. |
| 819 | // to avoid potential deadlocks, since the processExitThread is required to signal the process exit events. |
| 820 | // std::thread::joinable() is const, so this can be called without acquiring the lock. |
| 821 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_INVALID_STATE), !m_processExitThread.joinable()); |
| 822 | |
| 823 | THROW_WIN32_IF_MSG( |
| 824 | ERROR_NOT_SUPPORTED, Options.User != nullptr, "Custom users are not supported for root namespace processes"); |
| 825 | |
| 826 | auto setErrno = [Errno](int Error) { |
| 827 | if (Errno != nullptr) |
| 828 | { |
| 829 | *Errno = Error; |
| 830 | } |
| 831 | }; |
| 832 | |
| 833 | // Check if this is a tty or not |
| 834 | const WSLCProcessFd* tty = nullptr; |
| 835 | auto [pid, _, childChannel] = Fork(WSLC_FORK::Process); |
| 836 | |
| 837 | std::vector<WSLCVirtualMachine::ConnectedSocket> sockets; |
| 838 | ConnectedSocket ttyControlhandle; |
| 839 | for (const auto& e : Fds) |
| 840 | { |
| 841 | |
| 842 | if (e.Type == WSLCFdTypeTtyControl) |
| 843 | { |
| 844 | THROW_HR_IF_MSG(E_INVALIDARG, ttyControlhandle.Fd != -1, "Multiple terminal control fds specified"); |
| 845 | |
| 846 | ttyControlhandle = ConnectSocket(childChannel, e.Fd); |
| 847 | } |
| 848 | else |
| 849 | { |
| 850 | if (e.Type == WSLCFdTypeTty) |
| 851 | { |
| 852 | THROW_HR_IF_MSG(E_INVALIDARG, tty != nullptr, "Multiple terminal fds specified"); |
| 853 | tty = &e; |
| 854 | } |
| 855 | |
| 856 | sockets.emplace_back(ConnectSocket(childChannel, e.Fd)); |
| 857 | } |
| 858 | } |
| 859 | |
| 860 | PrepareCommandLine(sockets); |
| 861 | |
| 862 | wsl::shared::MessageWriter<WSLC_EXEC> Message; |
| 863 | |
| 864 | Message.WriteString(Message->ExecutableIndex, Executable); |
| 865 | Message.WriteString(Message->CurrentDirectoryIndex, Options.CurrentDirectory ? Options.CurrentDirectory : "/"); |
| 866 | Message.WriteStringArray(Message->CommandLineIndex, Options.CommandLine.Values, Options.CommandLine.Count); |
| 867 | Message.WriteStringArray(Message->EnvironmentIndex, Options.Environment.Values, Options.Environment.Count); |
| 868 | |
| 869 | // N.B. The process control needs to be registered before the actual exec message is sent. Otherwise, if the process exits quickly, we might receive the exit notification before registering it. |
| 870 | std::shared_ptr<VMProcessControl> control; |
| 871 | auto registerProcess = [&](int processPid) { |
| 872 | control = std::make_shared<VMProcessControl>(*this, processPid, std::move(ttyControlhandle.Socket)); |
| 873 | { |
| 874 | std::lock_guard lock{m_trackedProcessesLock}; |
| 875 | m_trackedProcesses.emplace_back(control); |
| 876 | } |
| 877 | }; |
| 878 | |
| 879 | // If this is an interactive tty, we need a relay process |
| 880 | if (tty != nullptr) |
| 881 | { |
| 882 | auto [grandChildPid, ptyMaster, grandChildChannel] = Fork(childChannel, WSLC_FORK::Pty, TtyRows, TtyColumns); |
| 883 | WSLC_TTY_RELAY relayMessage{}; |
| 884 | relayMessage.TtyMaster = ptyMaster; |
| 885 | relayMessage.Socket = tty->Fd; |
| 886 | relayMessage.TtyControl = ttyControlhandle.Fd; // N.B. Fd is set to -1 if unset. |
| 887 | { |
| 888 | auto relayTransaction = childChannel.StartTransaction(m_initChannelTimeout); |
| 889 | relayTransaction.Send(relayMessage); |
| 890 | } |
| 891 | |
| 892 | auto result = ExpectClosedChannelOrError(childChannel); |
| 893 | if (result != 0) |
| 894 | { |
| 895 | setErrno(result); |
| 896 | THROW_HR_MSG(E_FAIL, "errno: %i", result); |
| 897 | } |
| 898 | |
| 899 | registerProcess(grandChildPid); |
| 900 | |
| 901 | { |
| 902 | auto execTransaction = grandChildChannel.StartTransaction(m_initChannelTimeout); |
| 903 | execTransaction.Send<WSLC_EXEC>(Message.Span()); |
| 904 | auto [execResponse, execSpan] = execTransaction.ReceiveOrClosed<RESULT_MESSAGE<int32_t>>(); |
| 905 | result = execResponse != nullptr ? execResponse->Result : 0; |
| 906 | } |
| 907 | if (result != 0) |
| 908 | { |
| 909 | setErrno(result); |
| 910 | THROW_HR_MSG(E_FAIL, "errno: %i", result); |
| 911 | } |
| 912 | |
| 913 | pid = grandChildPid; |
| 914 | } |
| 915 | else |
| 916 | { |
| 917 | registerProcess(pid); |
| 918 | |
| 919 | auto execTransaction = childChannel.StartTransaction(m_initChannelTimeout); |
| 920 | execTransaction.Send<WSLC_EXEC>(Message.Span()); |
| 921 | auto [execResponse, execSpan] = execTransaction.ReceiveOrClosed<RESULT_MESSAGE<int32_t>>(); |
| 922 | auto result = execResponse != nullptr ? execResponse->Result : 0; |
| 923 | if (result != 0) |
| 924 | { |
| 925 | setErrno(result); |
| 926 | THROW_HR_MSG(E_FAIL, "errno: %i", result); |
| 927 | } |
| 928 | } |
| 929 | |
| 930 | std::map<ULONG, TypedHandle> stdHandles; |
| 931 | for (auto& [fd, handle] : sockets) |
| 932 | { |
| 933 | stdHandles.emplace(fd, TypedHandle{std::move(handle), WSLCHandleTypeSocket}); |
| 934 | } |
| 935 | |
| 936 | auto io = std::make_unique<VMProcessIO>(std::move(stdHandles)); |
| 937 | |
| 938 | auto process = wil::MakeOrThrow<WSLCProcess>(std::move(control), std::move(io), Options.Flags); |
| 939 | |
| 940 | setErrno(0); |
| 941 | |
| 942 | return process; |
| 943 | } |
| 944 | |
| 945 | void WSLCVirtualMachine::Mount(LPCSTR Source, LPCSTR Target, LPCSTR Type, LPCSTR Options, ULONG Flags) |
| 946 | { |
| 947 | std::lock_guard lock{m_lock}; |
| 948 | |
| 949 | Mount(m_initChannel, Source, Target, Type, Options, Flags); |
| 950 | } |
| 951 | |
| 952 | void WSLCVirtualMachine::Mount(shared::SocketChannel& Channel, LPCSTR Source, LPCSTR Target, LPCSTR Type, LPCSTR Options, ULONG Flags) |
| 953 | { |
| 954 | static_assert(WSLCMountFlagsNone == WSLC_MOUNT::None); |
| 955 | static_assert(WSLCMountFlagsReadOnly == WSLC_MOUNT::ReadOnly); |
| 956 | static_assert(WSLCMountFlagsChroot == WSLC_MOUNT::Chroot); |
| 957 | static_assert(WSLCMountFlagsWriteableOverlayFs == WSLC_MOUNT::OverlayFs); |
| 958 | |
| 959 | wsl::shared::MessageWriter<WSLC_MOUNT> message; |
| 960 | |
| 961 | auto optionalAdd = [&](auto value, unsigned int& index) { |
| 962 | if (value != nullptr) |
| 963 | { |
| 964 | message.WriteString(index, value); |
| 965 | } |
| 966 | }; |
| 967 | |
| 968 | optionalAdd(Source, message->SourceIndex); |
| 969 | optionalAdd(Target, message->DestinationIndex); |
| 970 | optionalAdd(Type, message->TypeIndex); |
| 971 | optionalAdd(Options, message->OptionsIndex); |
| 972 | message->Flags = Flags; |
| 973 | |
| 974 | const auto& response = Channel.Transaction<WSLC_MOUNT>(message.Span()); |
| 975 | |
| 976 | WSL_LOG( |
| 977 | "WSLCMount", |
| 978 | TraceLoggingValue(Source == nullptr ? "<null>" : Source, "Source"), |
| 979 | TraceLoggingValue(Target == nullptr ? "<null>" : Target, "Target"), |
| 980 | TraceLoggingValue(Type == nullptr ? "<null>" : Type, "Type"), |
| 981 | TraceLoggingValue(Options == nullptr ? "<null>" : Options, "Options"), |
| 982 | TraceLoggingValue(Flags, "Flags"), |
| 983 | TraceLoggingValue(response.Result, "Result")); |
| 984 | |
| 985 | THROW_HR_IF(E_FAIL, response.Result != 0); |
| 986 | } |
| 987 | |
| 988 | void WSLCVirtualMachine::MountModules(shared::SocketChannel& Channel, LPCSTR Source) |
| 989 | { |
| 990 | wsl::shared::MessageWriter<WSLC_MOUNT_MODULES> message; |
| 991 | message.WriteString(message->SourceIndex, Source); |
| 992 | |
| 993 | const auto& response = Channel.Transaction<WSLC_MOUNT_MODULES>(message.Span()); |
| 994 | |
| 995 | WSL_LOG("WSLCMountModules", TraceLoggingValue(Source, "Source"), TraceLoggingValue(response.Result, "Result")); |
| 996 | |
| 997 | THROW_HR_IF(E_FAIL, response.Result != 0); |
| 998 | } |
| 999 | |
| 1000 | void WSLCVirtualMachine::MountVirtioFsChild(shared::SocketChannel& Channel, LPCSTR Source, LPCSTR ChildName, LPCSTR Target, LPCSTR Options, ULONG Flags) |
| 1001 | { |
| 1002 | wsl::shared::MessageWriter<WSLC_MOUNT_VIRTIOFS> message; |
| 1003 | message.WriteString(message->SourceIndex, Source); |
| 1004 | message.WriteString(message->ChildNameIndex, ChildName); |
| 1005 | message.WriteString(message->DestinationIndex, Target); |
| 1006 | message.WriteString(message->TypeIndex, "virtiofs"); |
| 1007 | message.WriteString(message->OptionsIndex, Options); |
| 1008 | message->Flags = Flags; |
| 1009 | |
| 1010 | const auto& response = Channel.Transaction<WSLC_MOUNT_VIRTIOFS>(message.Span()); |
| 1011 | |
| 1012 | WSL_LOG( |
| 1013 | "WSLCMountVirtioFsChild", |
| 1014 | TraceLoggingValue(Source, "Source"), |
| 1015 | TraceLoggingValue(ChildName, "ChildName"), |
| 1016 | TraceLoggingValue(Target, "Target"), |
| 1017 | TraceLoggingValue(Options, "Options"), |
| 1018 | TraceLoggingValue(Flags, "Flags"), |
| 1019 | TraceLoggingValue(response.Result, "Result")); |
| 1020 | |
| 1021 | THROW_HR_IF(E_FAIL, response.Result != 0); |
| 1022 | } |
| 1023 | |
| 1024 | int32_t WSLCVirtualMachine::ExpectClosedChannelOrError(wsl::shared::SocketChannel& Channel) |
| 1025 | { |
| 1026 | auto [response, span] = Channel.ReceiveMessageOrClosed<RESULT_MESSAGE<int32_t>>(); |
| 1027 | if (response != nullptr) |
| 1028 | { |
| 1029 | return response->Result; |
| 1030 | } |
| 1031 | else |
| 1032 | { |
| 1033 | return 0; |
| 1034 | } |
| 1035 | } |
| 1036 | |
| 1037 | void WSLCVirtualMachine::Signal(_In_ LONG Pid, _In_ int Signal) |
| 1038 | { |
| 1039 | std::lock_guard lock(m_lock); |
| 1040 | |
| 1041 | WSLC_SIGNAL message; |
| 1042 | message.Pid = Pid; |
| 1043 | message.Signal = Signal; |
| 1044 | const auto& response = m_initChannel.Transaction(message, nullptr, m_initChannelTimeout); |
| 1045 | |
| 1046 | THROW_HR_IF(E_FAIL, response.Result != 0); |
| 1047 | } |
| 1048 | |
| 1049 | void WSLCVirtualMachine::LaunchPortRelay() |
| 1050 | { |
| 1051 | WI_ASSERT(!m_portRelayChannelRead); |
| 1052 | |
| 1053 | auto [_, __, channel] = Fork(WSLC_FORK::ForkType::Process); |
| 1054 | |
| 1055 | std::lock_guard lock(m_portRelaylock); |
| 1056 | auto relayPort = channel.Transaction<WSLC_PORT_RELAY>(); |
| 1057 | |
| 1058 | wil::unique_handle readPipe; |
| 1059 | wil::unique_handle writePipe; |
| 1060 | THROW_IF_WIN32_BOOL_FALSE(CreatePipe(&readPipe, &m_portRelayChannelWrite, nullptr, 0)); |
| 1061 | THROW_IF_WIN32_BOOL_FALSE(CreatePipe(&m_portRelayChannelRead, &writePipe, nullptr, 0)); |
| 1062 | |
| 1063 | // TODO: move the port relay infra into this process. Create a thread, pass handle ownership to thread, refactor and remove |
| 1064 | // wslrelaymode. wsl::windows::wslrelay::localhost::RunWSLCPortRelay( |
| 1065 | // readPipe.release(), writePipe.release(), m_vmId, relayPort.Result, m_vmTerminatingEvent.get()); |
| 1066 | |
| 1067 | wsl::windows::common::helpers::SetHandleInheritable(readPipe.get()); |
| 1068 | wsl::windows::common::helpers::SetHandleInheritable(writePipe.get()); |
| 1069 | wsl::windows::common::helpers::SetHandleInheritable(m_vmTerminatingEvent.get()); |
| 1070 | |
| 1071 | auto path = wsl::windows::common::wslutil::GetBasePath() / L"wslrelay.exe"; |
| 1072 | |
| 1073 | auto cmd = std::format( |
| 1074 | L"\"{}\" {} {} {} {} {} {} {} {}", |
| 1075 | path, |
| 1076 | wslrelay::mode_option, |
| 1077 | static_cast<int>(wslrelay::RelayMode::WSLCPortRelay), |
| 1078 | wslrelay::exit_event_option, |
| 1079 | HandleToUlong(m_vmTerminatingEvent.get()), |
| 1080 | wslrelay::port_option, |
| 1081 | relayPort.Result, |
| 1082 | wslrelay::vm_id_option, |
| 1083 | m_vmId); |
| 1084 | |
| 1085 | WSL_LOG("LaunchWslRelay", TraceLoggingValue(cmd.c_str(), "cmd")); |
| 1086 | |
| 1087 | wsl::windows::common::SubProcess process{nullptr, cmd.c_str()}; |
| 1088 | process.SetStdHandles(readPipe.get(), writePipe.get(), nullptr); |
| 1089 | process.InheritHandle(m_vmTerminatingEvent.get()); |
| 1090 | process.SetJobObject(m_processJobObject.get()); |
| 1091 | process.Start(); |
| 1092 | |
| 1093 | readPipe.release(); |
| 1094 | writePipe.release(); |
| 1095 | } |
| 1096 | |
| 1097 | void WSLCVirtualMachine::MapRelayPort(_In_ int Family, _In_ unsigned short WindowsPort, _In_ unsigned short LinuxPort, _In_ bool Remove) |
| 1098 | { |
| 1099 | std::lock_guard lock(m_portRelaylock); |
| 1100 | |
| 1101 | THROW_HR_IF(E_ILLEGAL_STATE_CHANGE, !m_portRelayChannelWrite); |
| 1102 | |
| 1103 | WSLC_MAP_PORT message; |
| 1104 | message.WindowsPort = WindowsPort; |
| 1105 | message.LinuxPort = LinuxPort; |
| 1106 | message.AddressFamily = Family; |
| 1107 | message.Stop = Remove; |
| 1108 | |
| 1109 | DWORD bytesTransfered{}; |
| 1110 | THROW_IF_WIN32_BOOL_FALSE(WriteFile(m_portRelayChannelWrite.get(), &message, sizeof(message), &bytesTransfered, nullptr)); |
| 1111 | THROW_HR_IF_MSG(E_UNEXPECTED, bytesTransfered != sizeof(message), "%u bytes transfered", bytesTransfered); |
| 1112 | |
| 1113 | HRESULT result = E_UNEXPECTED; |
| 1114 | THROW_IF_WIN32_BOOL_FALSE(ReadFile(m_portRelayChannelRead.get(), &result, sizeof(result), &bytesTransfered, nullptr)); |
| 1115 | |
| 1116 | THROW_HR_IF(E_UNEXPECTED, bytesTransfered != sizeof(result)); |
| 1117 | THROW_IF_FAILED_MSG(result, "Failed to map port: WindowsPort=%d, LinuxPort=%d, Family=%d, Remove=%d", WindowsPort, LinuxPort, Family, Remove); |
| 1118 | } |
| 1119 | |
| 1120 | void WSLCVirtualMachine::MapPort(VMPortMapping& Mapping) |
| 1121 | { |
| 1122 | THROW_HR_IF_MSG(E_INVALIDARG, !Mapping.VmPort, "Can't map a VM port without an allocated port"); |
| 1123 | |
| 1124 | if (m_networkingMode == WSLCNetworkingModeNone) |
| 1125 | { |
| 1126 | THROW_HR_MSG(E_ILLEGAL_STATE_CHANGE, "Port mapping is not supported with the current networking mode"); |
| 1127 | } |
| 1128 | else if (UseWslRelayPortForwarding()) |
| 1129 | { |
| 1130 | THROW_HR_IF_MSG( |
| 1131 | HRESULT_FROM_WIN32(ERROR_NOT_SUPPORTED), |
| 1132 | !Mapping.IsLocalhost() || Mapping.Protocol != IPPROTO_TCP, |
| 1133 | "Unsupported port mapping for the wslrelay port relay: %hs, protocol: %i", |
| 1134 | Mapping.BindingAddressString().c_str(), |
| 1135 | Mapping.Protocol); |
| 1136 | |
| 1137 | MapRelayPort(Mapping.BindAddress.si_family, Mapping.HostPort(), Mapping.VmPort->Port(), false); |
| 1138 | } |
| 1139 | else if (m_networkingMode == WSLCNetworkingModeConsomme) |
| 1140 | { |
| 1141 | USHORT allocatedHostPort = 0; |
| 1142 | auto result = m_vm->MapVirtioNetPort( |
| 1143 | Mapping.HostPort(), Mapping.VmPort->Port(), Mapping.Protocol, Mapping.BindingAddressString().c_str(), &allocatedHostPort); |
| 1144 | |
| 1145 | if (FAILED(result)) |
| 1146 | { |
| 1147 | auto portString = std::format( |
| 1148 | "{}:{}/{}", |
| 1149 | Mapping.IsIPv6() ? std::format("[{}]", Mapping.BindingAddressString()) : Mapping.BindingAddressString(), |
| 1150 | Mapping.HostPort(), |
| 1151 | Mapping.Protocol == IPPROTO_TCP ? "tcp" : "udp"); |
| 1152 | |
| 1153 | THROW_HR_WITH_USER_ERROR(result, shared::Localization::MessageFailedToMapPort(portString, common::wslutil::GetErrorString(result))); |
| 1154 | } |
| 1155 | |
| 1156 | // For anonymous binds, write back the allocated host port. |
| 1157 | if (Mapping.HostPort() == WSLC_EPHEMERAL_PORT) |
| 1158 | { |
| 1159 | WSL_LOG( |
| 1160 | "AllocatedHostPort", |
| 1161 | TraceLoggingValue(allocatedHostPort, "HostPort"), |
| 1162 | TraceLoggingValue(Mapping.VmPort->Port(), "GuestPort")); |
| 1163 | Mapping.SetHostPort(allocatedHostPort); |
| 1164 | } |
| 1165 | } |
| 1166 | else |
| 1167 | { |
| 1168 | THROW_HR_MSG(E_UNEXPECTED, "Unexpected networking mode: %i", m_networkingMode); |
| 1169 | } |
| 1170 | |
| 1171 | Mapping.Attach(*this); |
| 1172 | } |
| 1173 | |
| 1174 | void WSLCVirtualMachine::UnmapPort(VMPortMapping& Mapping) |
| 1175 | { |
| 1176 | THROW_HR_IF_MSG(E_INVALIDARG, !Mapping.VmPort, "Can't unmap a VM port without an allocated port"); |
| 1177 | |
| 1178 | if (m_networkingMode == WSLCNetworkingModeNone) |
| 1179 | { |
| 1180 | THROW_HR_MSG(E_ILLEGAL_STATE_CHANGE, "Port mapping is not supported with the current networking mode"); |
| 1181 | } |
| 1182 | else if (UseWslRelayPortForwarding()) |
| 1183 | { |
| 1184 | MapRelayPort(Mapping.BindAddress.si_family, Mapping.HostPort(), Mapping.VmPort->Port(), true); |
| 1185 | } |
| 1186 | else if (m_networkingMode == WSLCNetworkingModeConsomme) |
| 1187 | { |
| 1188 | THROW_IF_FAILED(m_vm->UnmapVirtioNetPort( |
| 1189 | Mapping.HostPort(), Mapping.VmPort->Port(), Mapping.Protocol, Mapping.BindingAddressString().c_str())); |
| 1190 | } |
| 1191 | else |
| 1192 | { |
| 1193 | THROW_HR_MSG(E_UNEXPECTED, "Unexpected networking mode: %i", m_networkingMode); |
| 1194 | } |
| 1195 | |
| 1196 | Mapping.Detach(); |
| 1197 | } |
| 1198 | |
| 1199 | HRESULT WSLCVirtualMachine::MountWindowsFolder(_In_ LPCWSTR WindowsPath, _In_ LPCSTR LinuxPath, _In_ BOOL ReadOnly) |
| 1200 | { |
| 1201 | return MountWindowsFolderImpl(WindowsPath, LinuxPath, ReadOnly ? WSLCMountFlagsReadOnly : WSLCMountFlagsNone); |
| 1202 | } |
| 1203 | |
| 1204 | HRESULT WSLCVirtualMachine::MountWindowsFolderImpl(_In_ LPCWSTR WindowsPath, _In_ LPCSTR LinuxPath, _In_ WSLCMountFlags Flags) |
| 1205 | try |
| 1206 | { |
| 1207 | std::filesystem::path path(WindowsPath); |
| 1208 | THROW_HR_IF_MSG(E_INVALIDARG, !path.is_absolute(), "Path is not absolute: '%ls'", WindowsPath); |
| 1209 | THROW_HR_IF_MSG( |
| 1210 | HRESULT_FROM_WIN32(ERROR_PATH_NOT_FOUND), !std::filesystem::is_directory(path), "Path is not a directory: '%ls'", WindowsPath); |
| 1211 | |
| 1212 | THROW_HR_IF_MSG(E_INVALIDARG, LinuxPath[0] != '/', "Mountpoint is not absolute: '%hs'", LinuxPath); |
| 1213 | |
| 1214 | const bool readOnly = WI_IsFlagSet(Flags, WSLCMountFlagsReadOnly); |
| 1215 | GUID shareGuid{}; |
| 1216 | |
| 1217 | { |
| 1218 | std::lock_guard lock(m_lock); |
| 1219 | |
| 1220 | // Verify that this folder isn't already mounted. |
| 1221 | auto it = m_mountedWindowsFolders.find(LinuxPath); |
| 1222 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_ALREADY_EXISTS), it != m_mountedWindowsFolders.end()); |
| 1223 | |
| 1224 | // Delegate to IWSLCVirtualMachine for the privileged share creation. |
| 1225 | THROW_IF_FAILED(m_vm->AddShare(WindowsPath, readOnly, &shareGuid)); |
| 1226 | |
| 1227 | m_mountedWindowsFolders.emplace(LinuxPath, shareGuid); |
| 1228 | } |
| 1229 | |
| 1230 | auto deleteOnFailure = wil::scope_exit_log(WI_DIAGNOSTICS_INFO, [&]() { |
| 1231 | std::lock_guard lock(m_lock); |
| 1232 | auto mountIt = m_mountedWindowsFolders.find(LinuxPath); |
| 1233 | if (WI_VERIFY(mountIt != m_mountedWindowsFolders.end())) |
| 1234 | { |
| 1235 | m_mountedWindowsFolders.erase(mountIt); |
| 1236 | LOG_IF_FAILED(m_vm->RemoveShare(shareGuid)); |
| 1237 | } |
| 1238 | }); |
| 1239 | |
| 1240 | // Create the guest mount |
| 1241 | auto shareName = shared::string::GuidToString<char>(shareGuid, shared::string::None); |
| 1242 | if (!FeatureEnabled(WslcFeatureFlagsVirtioFs)) |
| 1243 | { |
| 1244 | auto [_, __, channel] = Fork(WSLC_FORK::Process); |
| 1245 | |
| 1246 | WSLC_CONNECT message; |
| 1247 | message.HostPort = LX_INIT_UTILITY_VM_PLAN9_PORT; |
| 1248 | |
| 1249 | auto fd = channel.Transaction(message).Result; |
| 1250 | THROW_HR_IF_MSG(E_FAIL, fd < 0, "WSLC_CONNECT failed with %i", fd); |
| 1251 | |
| 1252 | auto mountOptions = std::format( |
| 1253 | "{},msize={},trans=fd,rfdno={},wfdno={},aname={},cache=mmap", readOnly ? "ro" : "rw", LX_INIT_UTILITY_VM_PLAN9_BUFFER_SIZE, fd, fd, shareName); |
| 1254 | |
| 1255 | Mount(channel, shareName.c_str(), LinuxPath, "9p", mountOptions.c_str(), Flags); |
| 1256 | } |
| 1257 | else |
| 1258 | { |
| 1259 | std::string options = readOnly ? "ro" : "rw"; |
| 1260 | MountVirtioFsChild(m_initChannel, LX_INIT_DRVFS_VIRTIO_TAG, shareName.c_str(), LinuxPath, options.c_str(), Flags); |
| 1261 | } |
| 1262 | |
| 1263 | deleteOnFailure.release(); |
| 1264 | |
| 1265 | return S_OK; |
| 1266 | } |
| 1267 | CATCH_RETURN(); |
| 1268 | |
| 1269 | HRESULT WSLCVirtualMachine::UnmountWindowsFolder(_In_ LPCSTR LinuxPath) |
| 1270 | try |
| 1271 | { |
| 1272 | std::lock_guard lock(m_lock); |
| 1273 | |
| 1274 | // Verify that this folder is mounted. |
| 1275 | auto it = m_mountedWindowsFolders.find(LinuxPath); |
| 1276 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_NOT_FOUND), it == m_mountedWindowsFolders.end()); |
| 1277 | |
| 1278 | // Unmount the folder from the guest. |
| 1279 | auto result = wil::ResultFromException([&]() { Unmount(LinuxPath); }); |
| 1280 | THROW_HR_IF(result, FAILED(result) && result != HRESULT_FROM_WIN32(ERROR_NOT_FOUND)); |
| 1281 | |
| 1282 | auto shareId = it->second; |
| 1283 | |
| 1284 | // Delegate to IWSLCVirtualMachine for the privileged share removal. |
| 1285 | THROW_IF_FAILED(m_vm->RemoveShare(shareId)); |
| 1286 | |
| 1287 | m_mountedWindowsFolders.erase(it); |
| 1288 | |
| 1289 | return S_OK; |
| 1290 | } |
| 1291 | CATCH_RETURN(); |
| 1292 | |
| 1293 | void WSLCVirtualMachine::MountGpuLibraries(_In_ LPCSTR LibrariesMountPoint, _In_ LPCSTR DriversMountpoint) |
| 1294 | { |
| 1295 | if (!FeatureEnabled(WslcFeatureFlagsGPU)) |
| 1296 | { |
| 1297 | return; |
| 1298 | } |
| 1299 | |
| 1300 | auto windowsPath = wil::GetWindowsDirectoryW<std::wstring>(); |
| 1301 | |
| 1302 | // Mount drivers. |
| 1303 | THROW_IF_FAILED(MountWindowsFolderImpl( |
| 1304 | std::format(L"{}\\System32\\DriverStore\\FileRepository", windowsPath).c_str(), DriversMountpoint, WSLCMountFlagsReadOnly)); |
| 1305 | |
| 1306 | // Mount the inbox libraries. |
| 1307 | auto inboxLibPath = std::format(L"{}\\System32\\lxss\\lib", windowsPath); |
| 1308 | std::optional<std::string> inboxLibMountPoint; |
| 1309 | if (std::filesystem::is_directory(inboxLibPath)) |
| 1310 | { |
| 1311 | inboxLibMountPoint = std::format("{}/inbox", LibrariesMountPoint); |
| 1312 | THROW_IF_FAILED(MountWindowsFolderImpl(inboxLibPath.c_str(), inboxLibMountPoint->c_str(), WSLCMountFlagsReadOnly)); |
| 1313 | } |
| 1314 | |
| 1315 | // Mount the packaged libraries. |
| 1316 | #ifdef WSL_GPU_LIB_PATH |
| 1317 | |
| 1318 | auto packagedLibPath = std::filesystem::path(TEXT(WSL_GPU_LIB_PATH)); |
| 1319 | |
| 1320 | #else |
| 1321 | |
| 1322 | auto packagedLibPath = wslutil::GetBasePath() / L"lib"; |
| 1323 | |
| 1324 | #endif |
| 1325 | |
| 1326 | if (inboxLibMountPoint.has_value()) |
| 1327 | { |
| 1328 | // Mount an overlay containing both inbox and packaged libraries (the packaged mount takes precedence). |
| 1329 | auto packagedLibMountPoint = std::format("{}/packaged", LibrariesMountPoint); |
| 1330 | THROW_IF_FAILED(MountWindowsFolderImpl(packagedLibPath.c_str(), packagedLibMountPoint.c_str(), WSLCMountFlagsReadOnly)); |
| 1331 | |
| 1332 | Mount( |
| 1333 | m_initChannel, |
| 1334 | "none", |
| 1335 | LibrariesMountPoint, |
| 1336 | "overlay", |
| 1337 | std::format("lowerdir={}:{}", packagedLibMountPoint, inboxLibMountPoint.value()).c_str(), |
| 1338 | 0); |
| 1339 | } |
| 1340 | else |
| 1341 | { |
| 1342 | // If the inbox libraries are not present, mount the packaged libraries directly at final location (no overlay needed). |
| 1343 | THROW_IF_FAILED(MountWindowsFolderImpl(packagedLibPath.c_str(), LibrariesMountPoint, WSLCMountFlagsReadOnly)); |
| 1344 | } |
| 1345 | } |
| 1346 | void WSLCVirtualMachine::OnProcessReleased(int Pid) |
| 1347 | { |
| 1348 | std::lock_guard lock{m_trackedProcessesLock}; |
| 1349 | |
| 1350 | std::erase_if(m_trackedProcesses, [Pid](const auto& e) { |
| 1351 | auto locked = e.lock(); |
| 1352 | return !locked || locked->GetPid() == Pid; |
| 1353 | }); |
| 1354 | } |
| 1355 | |
| 1356 | void WSLCVirtualMachine::OnSessionTerminated() |
| 1357 | { |
| 1358 | std::lock_guard lock{m_lock}; |
| 1359 | |
| 1360 | // Don't cancel init transactions on the session termination event, since that event is set. |
| 1361 | m_initChannel.SetExitEvents({m_vmTerminatingEvent.get()}); |
| 1362 | |
| 1363 | // Set a lower timeout for init transactions since we're terminating. |
| 1364 | m_initChannelTimeout = 15 * 1000; |
| 1365 | } |
| 1366 | |
| 1367 | std::shared_ptr<VmPortAllocation> WSLCVirtualMachine::TryAllocatePort(uint16_t Port, int Family, int Protocol) |
| 1368 | { |
| 1369 | std::lock_guard lock{m_reservations->Mutex}; |
| 1370 | |
| 1371 | WSL_LOG("AllocatePort", TraceLoggingValue(Port, "Port")); |
| 1372 | |
| 1373 | if (!m_reservations->Ports.insert(Port).second) |
| 1374 | { |
| 1375 | return {}; |
| 1376 | } |
| 1377 | |
| 1378 | // Roll the reservation back if the allocation object can't be created: nothing owns the port |
| 1379 | // until the shared_ptr exists, so it would otherwise stay marked in use for the VM's lifetime. |
| 1380 | auto reservationCleanup = wil::scope_exit([&]() { m_reservations->Ports.erase(Port); }); |
| 1381 | auto allocation = std::make_shared<VmPortAllocation>(Port, Family, Protocol, m_reservations); |
| 1382 | reservationCleanup.release(); |
| 1383 | |
| 1384 | return allocation; |
| 1385 | } |
| 1386 | |
| 1387 | std::shared_ptr<VmPortAllocation> WSLCVirtualMachine::AllocatePort(int Family, int Protocol) |
| 1388 | { |
| 1389 | std::lock_guard lock{m_reservations->Mutex}; |
| 1390 | |
| 1391 | for (uint32_t i = CONTAINER_PORT_RANGE.first; i <= CONTAINER_PORT_RANGE.second; i++) |
| 1392 | { |
| 1393 | uint16_t port = static_cast<uint16_t>(i); |
| 1394 | if (!m_reservations->Ports.contains(port)) |
| 1395 | { |
| 1396 | WI_VERIFY(m_reservations->Ports.insert(port).second); |
| 1397 | |
| 1398 | auto reservationCleanup = wil::scope_exit([&]() { m_reservations->Ports.erase(port); }); |
| 1399 | auto allocation = std::make_shared<VmPortAllocation>(port, Family, Protocol, m_reservations); |
| 1400 | reservationCleanup.release(); |
| 1401 | |
| 1402 | return allocation; |
| 1403 | } |
| 1404 | } |
| 1405 | |
| 1406 | // Fail if we couldn't find a port. |
| 1407 | THROW_HR_MSG(HRESULT_FROM_WIN32(ERROR_NO_SYSTEM_RESOURCES), "Failed to allocate port"); |
| 1408 | } |
| 1409 | |
| 1410 | wil::unique_socket WSLCVirtualMachine::ConnectUnixSocket(const char* Path) |
| 1411 | { |
| 1412 | auto [_, __, channel] = Fork(WSLC_FORK::Thread); |
| 1413 | |
| 1414 | shared::MessageWriter<WSLC_UNIX_CONNECT> message; |
| 1415 | message.WriteString(message->PathOffset, Path); |
| 1416 | |
| 1417 | auto result = channel.Transaction<WSLC_UNIX_CONNECT>(message.Span()); |
| 1418 | |
| 1419 | THROW_HR_IF_MSG(E_FAIL, result.Result < 0, "Failed to connect to unix socket: '%hs', %i", Path, result.Result); |
| 1420 | |
| 1421 | return channel.Release(); |
| 1422 | } |
| 1423 | |
| 1424 | void WSLCVirtualMachine::CollectCrashDumps(wil::unique_socket&& listenSocket) |
| 1425 | { |
| 1426 | // No impersonation needed - the session process already runs as the user. |
| 1427 | wslutil::SetThreadDescription(L"CrashDumpCollection"); |
| 1428 | |
| 1429 | const auto coInit = wil::CoInitializeEx(COINIT_MULTITHREADED); |
| 1430 | |
| 1431 | const auto crashDumpFolder = filesystem::GetTempFolderPath(GetCurrentProcessToken()) / L"wslc-crashes"; |
| 1432 | |
| 1433 | while (!m_vmTerminatingEvent.is_signaled()) |
| 1434 | { |
| 1435 | try |
| 1436 | { |
| 1437 | auto socket = wsl::windows::common::socket::CancellableAccept(listenSocket.get(), INFINITE, m_vmTerminatingEvent.get()); |
| 1438 | if (!socket) |
| 1439 | { |
| 1440 | // VM is exiting. |
| 1441 | break; |
| 1442 | } |
| 1443 | |
| 1444 | constexpr DWORD timeout = 30 * 1000; |
| 1445 | THROW_LAST_ERROR_IF(setsockopt(socket->get(), SOL_SOCKET, SO_RCVTIMEO, (const char*)&timeout, sizeof(timeout)) == SOCKET_ERROR); |
| 1446 | |
| 1447 | auto channel = wsl::shared::SocketChannel{ |
| 1448 | std::move(socket.value()), "crash_dump", {m_vmTerminatingEvent.get(), m_sessionTerminatingEvent}}; |
| 1449 | |
| 1450 | auto transaction = channel.ReceiveTransaction(); |
| 1451 | gsl::span<gsl::byte> responseSpan; |
| 1452 | const auto& message = transaction.Receive<LX_PROCESS_CRASH>(&responseSpan); |
| 1453 | |
| 1454 | const auto bufferSize = responseSpan.size_bytes() - offsetof(LX_PROCESS_CRASH, Buffer); |
| 1455 | const std::string process(message.Buffer, strnlen(message.Buffer, bufferSize)); |
| 1456 | |
| 1457 | const auto crashPid = message.Pid; |
| 1458 | const auto crashSignal = message.Signal; |
| 1459 | const auto crashTimestamp = message.Timestamp; |
| 1460 | |
| 1461 | constexpr auto dumpExtension = ".dmp"; |
| 1462 | constexpr auto dumpPrefix = "wsl-crash"; |
| 1463 | |
| 1464 | auto filename = std::format("{}-{}-{}-{}-{}{}", dumpPrefix, crashTimestamp, crashPid, process, crashSignal, dumpExtension); |
| 1465 | |
| 1466 | std::replace_if( |
| 1467 | filename.begin(), |
| 1468 | filename.end(), |
| 1469 | [](char e) { return !std::isalnum(static_cast<unsigned char>(e)) && e != '.' && e != '-'; }, |
| 1470 | '_'); |
| 1471 | |
| 1472 | auto fullPath = crashDumpFolder / filename; |
| 1473 | |
| 1474 | WSL_LOG( |
| 1475 | "WSLCLinuxCrash", |
| 1476 | TraceLoggingValue(fullPath.c_str(), "FullPath"), |
| 1477 | TraceLoggingValue(crashPid, "Pid"), |
| 1478 | TraceLoggingValue(crashSignal, "Signal"), |
| 1479 | TraceLoggingValue(process.c_str(), "process")); |
| 1480 | |
| 1481 | filesystem::EnsureDirectory(crashDumpFolder.c_str()); |
| 1482 | |
| 1483 | // Only delete files that: |
| 1484 | // - have the temporary flag set |
| 1485 | // - start with 'wsl-crash' |
| 1486 | // - end in .dmp |
| 1487 | // |
| 1488 | // This logic is here to prevent accidental user file deletion |
| 1489 | auto pred = [&dumpExtension, &dumpPrefix](const auto& e) { |
| 1490 | return WI_IsFlagSet(GetFileAttributes(e.path().c_str()), FILE_ATTRIBUTE_TEMPORARY) && e.path().has_extension() && |
| 1491 | e.path().extension() == dumpExtension && e.path().has_filename() && |
| 1492 | e.path().filename().string().find(dumpPrefix) == 0; |
| 1493 | }; |
| 1494 | |
| 1495 | wslutil::EnforceFileLimit(crashDumpFolder.c_str(), 10, pred); |
| 1496 | |
| 1497 | wil::unique_hfile file{CreateFileW(fullPath.c_str(), GENERIC_WRITE, 0, nullptr, CREATE_NEW, FILE_ATTRIBUTE_TEMPORARY, nullptr)}; |
| 1498 | THROW_LAST_ERROR_IF(!file); |
| 1499 | |
| 1500 | transaction.SendResultMessage<std::int32_t>(0); |
| 1501 | relay::InterruptableRelay(reinterpret_cast<HANDLE>(channel.Socket()), file.get(), nullptr); |
| 1502 | |
| 1503 | file.reset(); |
| 1504 | |
| 1505 | // Notify the session that a crash dump has been fully written. The session fans out |
| 1506 | // to any registered ICrashDumpCallback subscribers. Failures are caller-handled. |
| 1507 | if (m_onCrashDump) |
| 1508 | { |
| 1509 | m_onCrashDump(fullPath.wstring(), process, crashPid, crashSignal, crashTimestamp); |
| 1510 | } |
| 1511 | } |
| 1512 | CATCH_LOG() |
| 1513 | } |
| 1514 | } |