| 1 | /*++ |
| 2 | |
| 3 | Copyright (c) Microsoft. All rights reserved. |
| 4 | |
| 5 | Module Name: |
| 6 | |
| 7 | WSLCSessionRuntime.cpp |
| 8 | |
| 9 | Abstract: |
| 10 | |
| 11 | Contains the implementation for WSLCSessionRuntime. |
| 12 | |
| 13 | --*/ |
| 14 | |
| 15 | #include "precomp.h" |
| 16 | #include "WSLCSessionRuntime.h" |
| 17 | #include "WSLCSession.h" |
| 18 | #include "WSLCSessionDefaults.h" |
| 19 | |
| 20 | using wsl::windows::service::wslc::WSLCSessionRuntime; |
| 21 | |
| 22 | namespace { |
| 23 | |
| 24 | constexpr auto c_dockerdReadyLogLine = "API listen on /var/run/docker.sock"; |
| 25 | |
| 26 | // How long a VM lease waits for an announced stop to complete before logging that it is still |
| 27 | // blocked. Purely diagnostic: the wait itself is unbounded (see VmLease). |
| 28 | constexpr DWORD c_vmStopWaitLogIntervalMs = 30 * 1000; |
| 29 | |
| 30 | } // namespace |
| 31 | |
| 32 | namespace wsl::windows::service::wslc { |
| 33 | |
| 34 | WSLCSessionRuntime::WSLCSessionRuntime(WSLCSession& Session) noexcept : m_session(&Session) |
| 35 | { |
| 36 | } |
| 37 | |
| 38 | void WSLCSessionRuntime::Initialize( |
| 39 | DWORD vmFactoryGitCookie, |
| 40 | wil::com_ptr<IGlobalInterfaceTable> git, |
| 41 | const WSLCSessionInitSettings* settings, |
| 42 | std::chrono::milliseconds idleGrace, |
| 43 | SessionContext sessionContext, |
| 44 | RuntimeHooks hooks) |
| 45 | { |
| 46 | m_vmFactoryGitCookie = vmFactoryGitCookie; |
| 47 | m_git = std::move(git); |
| 48 | m_settings = settings; |
| 49 | m_id = sessionContext.Id; |
| 50 | m_displayName = std::move(sessionContext.DisplayName); |
| 51 | m_terminating = sessionContext.Terminating; |
| 52 | m_sessionTerminatingEvent = sessionContext.SessionTerminatingEvent; |
| 53 | m_sessionTerminatedEvent = sessionContext.SessionTerminatedEvent; |
| 54 | m_hooks = std::move(hooks); |
| 55 | |
| 56 | m_idleState->Initialize(idleGrace, [this]() { OnIdleTimer(); }); |
| 57 | |
| 58 | // Session-scoped: subscriptions must survive VM restarts. Rebound per-VM in InitializeDockerRuntime. |
| 59 | m_eventTracker.emplace(*m_session); |
| 60 | |
| 61 | m_initialized = true; |
| 62 | } |
| 63 | |
| 64 | WSLCVirtualMachine& WSLCSessionRuntime::Vm() |
| 65 | { |
| 66 | WI_ASSERT(m_virtualMachine.has_value()); |
| 67 | return m_virtualMachine.value(); |
| 68 | } |
| 69 | |
| 70 | bool WSLCSessionRuntime::HasVm() const noexcept |
| 71 | { |
| 72 | return m_hasVm.load(); |
| 73 | } |
| 74 | |
| 75 | IORelay* WSLCSessionRuntime::Relay() |
| 76 | { |
| 77 | return m_ioRelay ? &m_ioRelay.value() : nullptr; |
| 78 | } |
| 79 | |
| 80 | bool WSLCSessionRuntime::HasRelay() const noexcept |
| 81 | { |
| 82 | return m_ioRelay.has_value(); |
| 83 | } |
| 84 | |
| 85 | DockerHTTPClient& WSLCSessionRuntime::Docker() |
| 86 | { |
| 87 | WI_ASSERT(m_dockerClient.has_value()); |
| 88 | return m_dockerClient.value(); |
| 89 | } |
| 90 | |
| 91 | bool WSLCSessionRuntime::HasDocker() const noexcept |
| 92 | { |
| 93 | return m_dockerClient.has_value(); |
| 94 | } |
| 95 | |
| 96 | DockerEventTracker& WSLCSessionRuntime::Events() |
| 97 | { |
| 98 | WI_ASSERT(m_eventTracker.has_value()); |
| 99 | return m_eventTracker.value(); |
| 100 | } |
| 101 | |
| 102 | bool WSLCSessionRuntime::HasEvents() const noexcept |
| 103 | { |
| 104 | return m_eventTracker.has_value(); |
| 105 | } |
| 106 | |
| 107 | WSLCVolumes& WSLCSessionRuntime::Volumes() |
| 108 | { |
| 109 | WI_ASSERT(m_volumes.has_value()); |
| 110 | return m_volumes.value(); |
| 111 | } |
| 112 | |
| 113 | bool WSLCSessionRuntime::HasVolumes() const noexcept |
| 114 | { |
| 115 | return m_volumes.has_value(); |
| 116 | } |
| 117 | |
| 118 | wil::rwlock_release_exclusive_scope_exit WSLCSessionRuntime::TryLockExclusive() noexcept |
| 119 | { |
| 120 | return m_lock.try_lock_exclusive(); |
| 121 | } |
| 122 | |
| 123 | IdleState& WSLCSessionRuntime::Idle() noexcept |
| 124 | { |
| 125 | return *m_idleState; |
| 126 | } |
| 127 | |
| 128 | std::shared_ptr<IdleState> WSLCSessionRuntime::IdleStateShared() const noexcept |
| 129 | { |
| 130 | return m_idleState; |
| 131 | } |
| 132 | |
| 133 | WSLCSessionRuntime::VmState WSLCSessionRuntime::State() const noexcept |
| 134 | { |
| 135 | return m_vmState.load(); |
| 136 | } |
| 137 | |
| 138 | WSLCSessionRuntime::VmExitDisposition WSLCSessionRuntime::ExitDisposition() const noexcept |
| 139 | { |
| 140 | return m_vmExitDisposition.load(); |
| 141 | } |
| 142 | |
| 143 | bool WSLCSessionRuntime::VmExited() const noexcept |
| 144 | { |
| 145 | return m_vmExited.load(); |
| 146 | } |
| 147 | |
| 148 | void WSLCSessionRuntime::ResetDockerdReady() noexcept |
| 149 | { |
| 150 | m_dockerdReadyEvent.ResetEvent(); |
| 151 | } |
| 152 | |
| 153 | void WSLCSessionRuntime::OnProcessLog(const gsl::span<char>& buffer, PCSTR source) noexcept |
| 154 | try |
| 155 | { |
| 156 | if (buffer.empty()) |
| 157 | { |
| 158 | return; |
| 159 | } |
| 160 | |
| 161 | std::string entry{buffer.begin(), buffer.end()}; |
| 162 | WSL_LOG( |
| 163 | "ContainerdLog", |
| 164 | TraceLoggingValue(source, "Source"), |
| 165 | TraceLoggingValue(entry.c_str(), "Content"), |
| 166 | TraceLoggingValue(m_displayName.c_str(), "Name")); |
| 167 | |
| 168 | if (!m_dockerdReadyEvent.is_signaled() && entry.find(c_dockerdReadyLogLine) != std::string::npos) |
| 169 | { |
| 170 | m_dockerdReadyEvent.SetEvent(); |
| 171 | } |
| 172 | } |
| 173 | CATCH_LOG() |
| 174 | |
| 175 | void WSLCSessionRuntime::SetContainerdProcess(ServiceRunningProcess&& process) |
| 176 | { |
| 177 | m_containerdProcess = std::move(process); |
| 178 | } |
| 179 | |
| 180 | void WSLCSessionRuntime::SetDockerdProcess(ServiceRunningProcess&& process) |
| 181 | { |
| 182 | m_dockerdProcess = std::move(process); |
| 183 | } |
| 184 | |
| 185 | void WSLCSessionRuntime::SetSwapVhdPath(std::filesystem::path path) |
| 186 | { |
| 187 | m_swapVhdPath = std::move(path); |
| 188 | } |
| 189 | |
| 190 | void WSLCSessionRuntime::SetStorageMounted(bool value) noexcept |
| 191 | { |
| 192 | m_storageMounted = value; |
| 193 | } |
| 194 | |
| 195 | std::mutex& WSLCSessionRuntime::AllocatedPortsLock() noexcept |
| 196 | { |
| 197 | return m_allocatedPortsLock; |
| 198 | } |
| 199 | |
| 200 | std::map<uint16_t, std::pair<std::shared_ptr<VmPortAllocation>, size_t>>& WSLCSessionRuntime::AllocatedPorts() noexcept |
| 201 | { |
| 202 | return m_allocatedPorts; |
| 203 | } |
| 204 | |
| 205 | bool WSLCSessionRuntime::IdleTerminationEnabled() const noexcept |
| 206 | { |
| 207 | // Only tear the VM down when there is persistent storage to recover from. A tmpfs-backed |
| 208 | // session would lose all image/container state on teardown, so its VM is kept alive once started. |
| 209 | return m_settings->StoragePath != nullptr; |
| 210 | } |
| 211 | |
| 212 | int WSLCSessionRuntime::StopProcess(ServiceRunningProcess& Process, DWORD TerminateTimeoutMs, DWORD KillTimeoutMs) |
| 213 | { |
| 214 | auto signalResult = Process.Get().Signal(WSLCSignalSIGTERM); |
| 215 | if (FAILED(signalResult)) |
| 216 | { |
| 217 | LOG_HR_MSG(signalResult, "Failed to terminate process %i", Process.Get().GetPid()); |
| 218 | return -1; |
| 219 | } |
| 220 | |
| 221 | try |
| 222 | { |
| 223 | return Process.Wait(TerminateTimeoutMs); |
| 224 | } |
| 225 | catch (...) |
| 226 | { |
| 227 | LOG_CAUGHT_EXCEPTION(); |
| 228 | try |
| 229 | { |
| 230 | LOG_IF_FAILED(Process.Get().Signal(WSLCSignalSIGKILL)); |
| 231 | return Process.Wait(KillTimeoutMs); |
| 232 | } |
| 233 | CATCH_LOG(); |
| 234 | } |
| 235 | |
| 236 | return -1; |
| 237 | } |
| 238 | |
| 239 | void WSLCSessionRuntime::EnsureVmRunning() |
| 240 | { |
| 241 | // Reject leases once the session is terminating/terminated, including on the running fast path: |
| 242 | // Shutdown() drops the lock to fire OnVmStopping, and a reentrant lease that finds the VM still |
| 243 | // Running must fail here rather than run work against a VM being permanently torn down. |
| 244 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_INVALID_STATE), m_terminating->load() || m_sessionTerminatedEvent.is_signaled()); |
| 245 | |
| 246 | if (m_vmState.load() == VmState::Running) |
| 247 | { |
| 248 | return; |
| 249 | } |
| 250 | |
| 251 | bool started = false; |
| 252 | uint64_t generation = 0; |
| 253 | { |
| 254 | auto lock = m_lock.lock_exclusive(); |
| 255 | |
| 256 | // Re-check under the lock: terminating may have been set since the check above. This also |
| 257 | // bounds VmLease's retry loop: a lease that races with Terminate() fails here instead of |
| 258 | // restarting a VM that is being permanently torn down. |
| 259 | THROW_HR_IF(HRESULT_FROM_WIN32(ERROR_INVALID_STATE), m_terminating->load() || m_sessionTerminatedEvent.is_signaled()); |
| 260 | |
| 261 | if (m_vmState.load() != VmState::Running) |
| 262 | { |
| 263 | StartVmLockHeld(); |
| 264 | started = true; |
| 265 | } |
| 266 | |
| 267 | generation = m_vmGeneration.load(); |
| 268 | } |
| 269 | |
| 270 | // Notify plugins that a VM has started, outside the exclusive lock: the handler forwards to the |
| 271 | // plugin, which may call back into the session (e.g. WSLCCreateProcess acquires a VM lease and |
| 272 | // the exclusive lock), so firing under the lock would deadlock. EnsureVmRunning's only caller |
| 273 | // (VmLease) holds an activity reference across this call, so idle teardown cannot race the VM |
| 274 | // down in the gap between releasing the lock and notifying. |
| 275 | if (started) |
| 276 | { |
| 277 | NotifyVmStarted(generation); |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | void WSLCSessionRuntime::NotifyVmStarted(uint64_t Generation) |
| 282 | { |
| 283 | // Hold m_notifyLock across both the pairing-state flip and the hook invocation, so a concurrent |
| 284 | // NotifyVmStopping on another thread cannot deliver its OnVmStopping in between and observe this |
| 285 | // OnVmStarted arrive late (which would otherwise leave the plugin with a trailing "started" for a |
| 286 | // VM that already stopped). m_notifyLock is recursive: the handler may reentrantly restart the VM |
| 287 | // (e.g. WSLCCreateProcess -> NotifyVmStarted/Stopping) on this same thread, which would self-deadlock |
| 288 | // under a plain mutex. Suppressed once terminating, else a racing Shutdown leaves OnVmStarted unpaired. |
| 289 | // m_notifiedGeneration tracks the VM lifecycle, not whether OnVmStarted is installed, so a hooks |
| 290 | // user that sets only OnVmStopping still gets paired stop notifications. |
| 291 | auto lock = std::lock_guard(m_notifyLock); |
| 292 | |
| 293 | // Drop the notification if this instance is already gone: between releasing the runtime lock and |
| 294 | // getting here it may have been torn down and replaced, and announcing it now would misattribute |
| 295 | // the start to whichever instance is running. |
| 296 | if (m_terminating->load() || m_vmGeneration.load() != Generation) |
| 297 | { |
| 298 | return; |
| 299 | } |
| 300 | |
| 301 | m_notifiedGeneration.store(Generation); |
| 302 | if (m_hooks.OnVmStarted) |
| 303 | { |
| 304 | m_hooks.OnVmStarted(); |
| 305 | } |
| 306 | } |
| 307 | |
| 308 | void WSLCSessionRuntime::NotifyVmStopping(uint64_t Generation) |
| 309 | { |
| 310 | // See NotifyVmStarted: hold m_notifyLock across both the pairing check and the hook invocation, so |
| 311 | // the two notifications cannot interleave across threads. Recursive because the handler may |
| 312 | // reentrantly restart the VM and call NotifyVmStarted on this same thread. |
| 313 | // |
| 314 | // N.B. The CAS below retires the generation as the stop is announced, so a second stop for the |
| 315 | // same VM (e.g. Terminate() racing an idle teardown that has dropped the lock across its |
| 316 | // notification) cannot deliver OnVmStopping twice. This is safe because the stop is committed |
| 317 | // before it is announced: the VM cannot come back, so nothing needs the generation afterwards. |
| 318 | // |
| 319 | // Generation 0 is the sentinel for "no VM has been announced": both counters start there, so a |
| 320 | // session that is torn down without ever starting a VM must not match, or Shutdown would deliver |
| 321 | // an OnVmStopping that no OnVmStarted ever paired with. |
| 322 | auto lock = std::lock_guard(m_notifyLock); |
| 323 | |
| 324 | auto expected = Generation; |
| 325 | if (Generation != 0 && m_notifiedGeneration.compare_exchange_strong(expected, 0) && m_hooks.OnVmStopping) |
| 326 | { |
| 327 | m_hooks.OnVmStopping(); |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | void WSLCSessionRuntime::BeginVmStopLockHeld() noexcept |
| 332 | { |
| 333 | // Reset before publishing the flag, so a lease cannot observe the pending stop while the event is |
| 334 | // still signaled from the previous teardown. |
| 335 | m_vmStopCompleteEvent.ResetEvent(); |
| 336 | m_vmStopPending.store(true); |
| 337 | } |
| 338 | |
| 339 | void WSLCSessionRuntime::EndVmStop() noexcept |
| 340 | { |
| 341 | m_vmStopPending.store(false); |
| 342 | m_vmStopCompleteEvent.SetEvent(); |
| 343 | } |
| 344 | |
| 345 | bool WSLCSessionRuntime::TryClaimExpectedStop() noexcept |
| 346 | { |
| 347 | auto expected = VmExitDisposition::Active; |
| 348 | return m_vmExitDisposition.compare_exchange_strong(expected, VmExitDisposition::StopRequested); |
| 349 | } |
| 350 | |
| 351 | bool WSLCSessionRuntime::TryClaimSpontaneousExit() noexcept |
| 352 | { |
| 353 | auto expected = VmExitDisposition::Active; |
| 354 | return m_vmExitDisposition.compare_exchange_strong(expected, VmExitDisposition::ExitClaimed); |
| 355 | } |
| 356 | |
| 357 | void WSLCSessionRuntime::StartVmLockHeld() |
| 358 | { |
| 359 | WI_ASSERT(m_vmState.load() != VmState::Running); |
| 360 | |
| 361 | WSL_LOG("WslcVmStarting", TraceLoggingValue(m_id, "SessionId")); |
| 362 | |
| 363 | m_vmState.store(VmState::Starting); |
| 364 | m_vmExitDisposition.store(VmExitDisposition::Active); |
| 365 | |
| 366 | // Identify this instance. Bumped under the runtime lock so a notification that had to drop the |
| 367 | // lock to fire can tell whether it still describes the VM it was raised for. The first VM is |
| 368 | // generation 1, so the 0 stored by TearDownVmLockHeld never collides with a live instance. |
| 369 | m_vmGeneration.fetch_add(1); |
| 370 | |
| 371 | // Tear back down if bring-up fails partway. The VM may have exited on its own during bring-up, |
| 372 | // so claim the stop first and only tear down if we win it (TryClaimExpectedStop()); otherwise |
| 373 | // OnVmExited() owns the teardown and we just release the lock to let its Terminate() finish. |
| 374 | auto startCleanup = wil::scope_exit_log(WI_DIAGNOSTICS_INFO, [&]() { |
| 375 | if (TryClaimExpectedStop()) |
| 376 | { |
| 377 | TearDownVmLockHeld(); |
| 378 | m_vmState.store(VmState::None); |
| 379 | } |
| 380 | else |
| 381 | { |
| 382 | WSL_LOG("WslcVmExitedDuringStart", TraceLoggingValue(m_id, "SessionId")); |
| 383 | } |
| 384 | }); |
| 385 | |
| 386 | // Create a fresh IO relay for this VM instance. The previous one (if any) was stopped |
| 387 | // during teardown and cannot be restarted. |
| 388 | m_ioRelay.emplace(); |
| 389 | |
| 390 | // Create the VM via the factory. Re-fetch the factory from the GIT so we call it through a |
| 391 | // proxy marshalled into this thread's apartment (see m_git). The VM produces crash events; |
| 392 | // the session multiplexes them out to any registered ICrashDumpCallback subscribers via |
| 393 | // OnCrashDumpWritten. |
| 394 | wil::com_ptr<IWSLCVirtualMachineFactory> vmFactory; |
| 395 | THROW_IF_FAILED(m_git->GetInterfaceFromGlobal(m_vmFactoryGitCookie, __uuidof(IWSLCVirtualMachineFactory), vmFactory.put_void())); |
| 396 | |
| 397 | wil::com_ptr<IWSLCVirtualMachine> vm; |
| 398 | THROW_IF_FAILED(vmFactory->CreateVirtualMachine(&vm)); |
| 399 | |
| 400 | m_virtualMachine.emplace(vm.get(), m_settings, m_sessionTerminatingEvent.get(), WSLCVirtualMachine::TOnCrashDump(m_hooks.OnCrashDump)); |
| 401 | |
| 402 | // Publish only once the object is constructed. If Initialize() below throws, startCleanup tears |
| 403 | // the VM down and clears this again. |
| 404 | m_hasVm.store(true); |
| 405 | |
| 406 | m_virtualMachine->Initialize(); |
| 407 | |
| 408 | // Get an event from the service that is signaled when the VM exits. |
| 409 | m_vmExitedEvent.reset(); |
| 410 | THROW_IF_FAILED(vm->GetTerminationEvent(&m_vmExitedEvent)); |
| 411 | m_vmExited.store(false); |
| 412 | |
| 413 | if (m_hooks.BringUp) |
| 414 | { |
| 415 | m_hooks.BringUp(); |
| 416 | } |
| 417 | |
| 418 | // Monitor for unexpected VM exit. |
| 419 | m_ioRelay->AddHandle( |
| 420 | std::make_unique<windows::common::io::EventHandle>(m_vmExitedEvent.get(), std::bind(&WSLCSessionRuntime::OnVmExited, this))); |
| 421 | |
| 422 | if (m_hooks.RecoverState) |
| 423 | { |
| 424 | m_hooks.RecoverState(); |
| 425 | } |
| 426 | |
| 427 | m_vmState.store(VmState::Running); |
| 428 | startCleanup.release(); |
| 429 | |
| 430 | WSL_LOG("WslcVmStarted", TraceLoggingValue(m_id, "SessionId")); |
| 431 | } |
| 432 | |
| 433 | void WSLCSessionRuntime::InitializeDockerRuntime(const std::filesystem::path& storagePath) |
| 434 | { |
| 435 | // Wait for dockerd to be ready before starting the event tracker. |
| 436 | THROW_WIN32_IF_MSG( |
| 437 | ERROR_TIMEOUT, !m_dockerdReadyEvent.wait(m_settings->BootTimeoutMs), "Timed out waiting for dockerd to start"); |
| 438 | |
| 439 | [[maybe_unused]] auto [pid, ptyMaster, channel] = m_virtualMachine->Fork(WSLC_FORK::Thread); |
| 440 | |
| 441 | m_dockerClient.emplace(std::move(channel), m_virtualMachine->TerminatingEvent(), m_virtualMachine->VmId(), 10 * 1000); |
| 442 | |
| 443 | // (Re)bind the session-scoped event tracker to this VM's docker client and relay. Existing |
| 444 | // container subscriptions are preserved across restarts. |
| 445 | m_eventTracker->Connect(m_dockerClient.value(), *m_ioRelay); |
| 446 | |
| 447 | m_volumes.emplace(m_dockerClient.value(), m_virtualMachine.value(), m_eventTracker.value(), storagePath); |
| 448 | } |
| 449 | |
| 450 | void WSLCSessionRuntime::StopVmLockHeld() |
| 451 | { |
| 452 | if (m_vmState.load() != VmState::Running) |
| 453 | { |
| 454 | return; |
| 455 | } |
| 456 | |
| 457 | WSL_LOG("WslcVmIdleStop", TraceLoggingValue(m_id, "SessionId")); |
| 458 | |
| 459 | // N.B. The caller has claimed StopRequested (via TryClaimExpectedStop), so VM/dockerd/containerd |
| 460 | // exit callbacks firing from the relay thread during teardown are treated as expected, not as a |
| 461 | // crash. |
| 462 | m_vmState.store(VmState::Stopping); |
| 463 | |
| 464 | TearDownVmLockHeld(); |
| 465 | |
| 466 | m_vmState.store(VmState::None); |
| 467 | } |
| 468 | |
| 469 | void WSLCSessionRuntime::TearDownVmLockHeld(bool CaptureTerminationReason) |
| 470 | { |
| 471 | // The VM is committed to going away from here on, so the plugin's view of it ends here. Retiring |
| 472 | // the notified instance up front, before any step below can throw, also covers the case where the |
| 473 | // VM is torn down without an OnVmStopping ever being announced (e.g. bring-up failed). Written |
| 474 | // under the runtime lock; NotifyVmStarted/NotifyVmStopping read it under m_notifyLock without the |
| 475 | // runtime lock, so this store can land while one of them is inside a plugin handler. That is |
| 476 | // benign: both compare against a generation they captured earlier, so a store of 0 can only make |
| 477 | // them decline to notify, and by this point either the stop has already been announced (its CAS |
| 478 | // ran first) or the instance is gone and must not be announced at all. |
| 479 | m_notifiedGeneration.store(0); |
| 480 | |
| 481 | // Latch whether the guest is already dead before running session-state cleanup so container |
| 482 | // teardown (ReleaseRuntimeResources) can skip VM-dependent calls, e.g. volume unmounts, on a VM |
| 483 | // that has exited. A graceful stop reaches here with the VM still alive (is_signaled() false), so |
| 484 | // its mounts are unmounted through the live VM; StartVmLockHeld clears this for the next instance. |
| 485 | if (m_vmExitedEvent && m_vmExitedEvent.is_signaled()) |
| 486 | { |
| 487 | m_vmExited.store(true); |
| 488 | } |
| 489 | |
| 490 | if (m_hooks.TearDownSessionState) |
| 491 | { |
| 492 | m_hooks.TearDownSessionState(CaptureTerminationReason); |
| 493 | } |
| 494 | |
| 495 | m_volumes.reset(); |
| 496 | |
| 497 | // Stop the IO relay. |
| 498 | // This stops: |
| 499 | // - container state monitoring. |
| 500 | // - container init process relays |
| 501 | // - execs relays |
| 502 | // - container logs relays |
| 503 | if (m_ioRelay) |
| 504 | { |
| 505 | m_ioRelay->Stop(); |
| 506 | } |
| 507 | |
| 508 | { |
| 509 | std::lock_guard allocatedPortsLock(m_allocatedPortsLock); |
| 510 | m_allocatedPorts.clear(); |
| 511 | } |
| 512 | |
| 513 | // The session-scoped event tracker is intentionally not reset. Its stream handle dies with the IO |
| 514 | // relay above, and InitializeDockerRuntime re-binds it on the next start. |
| 515 | m_dockerClient.reset(); |
| 516 | |
| 517 | if (CaptureTerminationReason) |
| 518 | { |
| 519 | // Default: an explicit/graceful teardown is a shutdown (the VM is still alive and we are |
| 520 | // bringing it down). Overridden below if the VM exited on its own and recorded a cause. |
| 521 | m_lastTerminationReason = WSLCVirtualMachineTerminationReasonShutdown; |
| 522 | m_lastTerminationDetails.clear(); |
| 523 | } |
| 524 | |
| 525 | // Check if the VM has already exited (e.g., killed externally). |
| 526 | // If so, skip operations that require a live VM to avoid unnecessary waits. |
| 527 | // N.B. m_vmExitedEvent may be uninitialized if teardown runs before GetTerminationEvent() succeeds. |
| 528 | if (m_vmExitedEvent && m_vmExitedEvent.is_signaled()) |
| 529 | { |
| 530 | WSL_LOG("SkippingGracefulShutdown_VmDead", TraceLoggingValue(m_id, "SessionId")); |
| 531 | |
| 532 | // The VM exited on its own, so it recorded the cause. |
| 533 | if (CaptureTerminationReason && m_virtualMachine) |
| 534 | { |
| 535 | wil::unique_cotaskmem_string details; |
| 536 | LOG_IF_FAILED(m_virtualMachine->GetTerminationReason(&m_lastTerminationReason, &details)); |
| 537 | m_lastTerminationDetails = details ? details.get() : L""; |
| 538 | } |
| 539 | } |
| 540 | else if (m_virtualMachine) |
| 541 | { |
| 542 | m_virtualMachine->OnSessionTerminated(); |
| 543 | |
| 544 | // Stop dockerd first, then containerd (dockerd is a client of containerd). |
| 545 | // N.B. dockerd waits a couple seconds if there are any outstanding HTTP request sockets opened. |
| 546 | if (m_dockerdProcess.has_value()) |
| 547 | { |
| 548 | auto dockerdExitCode = |
| 549 | StopProcess(m_dockerdProcess.value(), wsl::windows::wslc::ProcessTerminateTimeoutMs, wsl::windows::wslc::ProcessKillTimeoutMs); |
| 550 | WSL_LOG("DockerdExit", TraceLoggingValue(dockerdExitCode, "code")); |
| 551 | } |
| 552 | |
| 553 | if (m_containerdProcess.has_value()) |
| 554 | { |
| 555 | auto containerdExitCode = |
| 556 | StopProcess(m_containerdProcess.value(), wsl::windows::wslc::ProcessTerminateTimeoutMs, wsl::windows::wslc::ProcessKillTimeoutMs); |
| 557 | WSL_LOG("ContainerdExit", TraceLoggingValue(containerdExitCode, "code")); |
| 558 | } |
| 559 | |
| 560 | // N.B. dockerd has exited by this point, so unmounting the VHD is safe since no container can be running. |
| 561 | if (m_storageMounted) |
| 562 | { |
| 563 | try |
| 564 | { |
| 565 | m_virtualMachine->Unmount(wsl::windows::wslc::ContainerdStorageMountPoint); |
| 566 | } |
| 567 | CATCH_LOG(); |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | m_dockerdProcess.reset(); |
| 572 | m_containerdProcess.reset(); |
| 573 | |
| 574 | // Retire the lock-free mirror before destroying the object so an unlocked reader never sees |
| 575 | // "a VM is present" while it is being torn down. |
| 576 | m_hasVm.store(false); |
| 577 | m_virtualMachine.reset(); |
| 578 | m_storageMounted = false; |
| 579 | |
| 580 | // Destroy the relay unless we're on its own thread (~IORelay joins the thread, which would |
| 581 | // deadlock). On unexpected-VM-exit path (runs on relay thread), leave it for ~WSLCSession. |
| 582 | if (!m_ioRelay || !m_ioRelay->IsRelayThread()) |
| 583 | { |
| 584 | m_ioRelay.reset(); |
| 585 | m_vmExitedEvent.reset(); |
| 586 | } |
| 587 | |
| 588 | // Delete the ephemeral swap VHD now that the VM is gone. |
| 589 | if (!m_swapVhdPath.empty()) |
| 590 | { |
| 591 | LOG_IF_WIN32_BOOL_FALSE(DeleteFileW(m_swapVhdPath.c_str())); |
| 592 | m_swapVhdPath.clear(); |
| 593 | } |
| 594 | } |
| 595 | |
| 596 | void WSLCSessionRuntime::OnIdleTimer() |
| 597 | try |
| 598 | { |
| 599 | // Idle teardown releases cross-process COM proxies (the VM and its VM-scoped state), so this |
| 600 | // threadpool callback must join the process MTA; otherwise those Release/calls fail with |
| 601 | // RPC_E_WRONG_THREAD. The function-try-block keeps this (and everything below) under CATCH_LOG: |
| 602 | // the threadpool callback that invokes us is noexcept, so an escaping throw would terminate. |
| 603 | const auto coInit = wil::CoInitializeEx(COINIT_MULTITHREADED); |
| 604 | |
| 605 | if (m_terminating->load() || !IdleTerminationEnabled()) |
| 606 | { |
| 607 | return; |
| 608 | } |
| 609 | |
| 610 | // Non-blocking acquire: a blocking exclusive would queue behind in-flight operations, and |
| 611 | // SRW locks favor waiting writers, stalling all new ops. If the lock is held, an operation |
| 612 | // is in flight; it holds an activity reference and will re-arm the timer (via the 1->0 |
| 613 | // transition) when it releases, so there is nothing to do here. |
| 614 | auto lock = m_lock.try_lock_exclusive(); |
| 615 | if (!lock) |
| 616 | { |
| 617 | return; |
| 618 | } |
| 619 | |
| 620 | // Re-check every teardown precondition under the lock. The activity count is the single |
| 621 | // source of truth for "the VM is needed"; a 0->1 transition since the timer fired (cancel |
| 622 | // raced the callback) is caught here. |
| 623 | if (m_terminating->load() || m_vmState.load() != VmState::Running || m_idleState->ActivityCount() != 0) |
| 624 | { |
| 625 | return; |
| 626 | } |
| 627 | |
| 628 | // Claim the stop. If we lose, OnVmExited() owns a spontaneous-exit teardown and is spinning for |
| 629 | // this lock, so release it and let that run instead of joining the relay ourselves. |
| 630 | if (!TryClaimExpectedStop()) |
| 631 | { |
| 632 | return; |
| 633 | } |
| 634 | |
| 635 | // Restore Active on completion (or early exit) so the next StartVmLockHeld starts clean; only |
| 636 | // clear our own claim. |
| 637 | auto dispositionCleanup = wil::scope_exit([this]() { |
| 638 | auto stopRequested = VmExitDisposition::StopRequested; |
| 639 | m_vmExitDisposition.compare_exchange_strong(stopRequested, VmExitDisposition::Active); |
| 640 | }); |
| 641 | |
| 642 | // Final activity re-check under the lock, and the last point at which the stop can still be called |
| 643 | // off: a VmLease that bumped the count since the check above is now blocked on the shared lock we |
| 644 | // hold. Back out here -- nothing has been announced yet -- rather than stopping a VM that is about |
| 645 | // to be used again. |
| 646 | if (m_idleState->ActivityCount() != 0) |
| 647 | { |
| 648 | return; |
| 649 | } |
| 650 | |
| 651 | // Past this point the VM is going away, whatever happens. Publishing the pending stop under the |
| 652 | // lock is what makes the announcement below true: an ordinary lease that races in while the lock |
| 653 | // is dropped no longer gets this VM, it waits for the teardown and is served by the next one. The |
| 654 | // exception is a lease taken by the OnVmStopping handler itself, which must be served or it would |
| 655 | // deadlock against the very callback this teardown is waiting on. Work that handler leaves running |
| 656 | // dies with the VM -- which is exactly what it was just told would happen. |
| 657 | // |
| 658 | // N.B. endStop is declared after 'lock' so it runs *before* the lock is released (reverse |
| 659 | // declaration order): waiters are released only once the teardown is complete and published. |
| 660 | BeginVmStopLockHeld(); |
| 661 | auto endStop = wil::scope_exit([this]() { EndVmStop(); }); |
| 662 | |
| 663 | // Fire OnVmStopping with m_lock dropped so a plugin handler may take a VM lease without |
| 664 | // deadlocking, and while the VM is still running so the handler can still use it. |
| 665 | // |
| 666 | // The notification cannot be allowed to abort the teardown: the stop is already published and its |
| 667 | // generation retired, so bailing out here would leave waiters to be served by the VM they were |
| 668 | // promised would die, and its eventual teardown would be silent. |
| 669 | const auto generation = m_vmGeneration.load(); |
| 670 | |
| 671 | lock.reset(); |
| 672 | try |
| 673 | { |
| 674 | NotifyVmStopping(generation); |
| 675 | } |
| 676 | CATCH_LOG(); |
| 677 | lock = m_lock.lock_exclusive(); |
| 678 | |
| 679 | // Unconditional: the stop was announced, so it happens. Reacquiring the exclusive lock first |
| 680 | // drains every in-flight operation, so nothing is cut off mid-call. StopVmLockHeld no-ops if a |
| 681 | // concurrent Terminate already tore the VM down, and copes with a VM that crashed while the lock |
| 682 | // was dropped -- OnVmExited() declined that exit (we hold the expected-stop claim) and its exit |
| 683 | // handle is one-shot, so this is the only teardown left to run. |
| 684 | StopVmLockHeld(); |
| 685 | } |
| 686 | CATCH_LOG(); |
| 687 | |
| 688 | bool WSLCSessionRuntime::TriggerIdleTerminationForTest() |
| 689 | { |
| 690 | // Mirror OnIdleTimer's MTA context on a dedicated thread: the incoming RPC thread is an STA, so |
| 691 | // both re-initializing MTA on it and running teardown inline would fail (RPC_E_CHANGED_MODE / |
| 692 | // RPC_E_WRONG_THREAD when releasing the VM's cross-process COM proxies). |
| 693 | bool wasAlreadyIdle = false; |
| 694 | std::exception_ptr error; |
| 695 | |
| 696 | std::thread worker([&]() { |
| 697 | try |
| 698 | { |
| 699 | const auto coInit = wil::CoInitializeEx(COINIT_MULTITHREADED); |
| 700 | |
| 701 | auto lock = m_lock.lock_exclusive(); |
| 702 | |
| 703 | if (m_terminating->load() || m_vmState.load() != VmState::Running) |
| 704 | { |
| 705 | wasAlreadyIdle = true; |
| 706 | return; |
| 707 | } |
| 708 | |
| 709 | // tmpfs sessions have no persistent storage to recover from, so tearing a running VM down |
| 710 | // loses all state. Check this after the VM state so a never-started session still reports |
| 711 | // that it was already idle. |
| 712 | if (!IdleTerminationEnabled()) |
| 713 | { |
| 714 | return; |
| 715 | } |
| 716 | |
| 717 | // Match the production idle timer: an active container or operation keeps the VM alive. |
| 718 | if (m_idleState->ActivityCount() != 0) |
| 719 | { |
| 720 | return; |
| 721 | } |
| 722 | |
| 723 | if (!TryClaimExpectedStop()) |
| 724 | { |
| 725 | wasAlreadyIdle = true; |
| 726 | return; |
| 727 | } |
| 728 | |
| 729 | auto dispositionCleanup = wil::scope_exit([this]() { |
| 730 | auto stopRequested = VmExitDisposition::StopRequested; |
| 731 | m_vmExitDisposition.compare_exchange_strong(stopRequested, VmExitDisposition::Active); |
| 732 | }); |
| 733 | |
| 734 | // Mirror OnIdleTimer's final re-check under the lock: a lease taken while the stop claim |
| 735 | // was being made is now blocked on the lock we hold, and must back the teardown out here |
| 736 | // rather than announce a stop production code would have suppressed. |
| 737 | if (m_idleState->ActivityCount() != 0) |
| 738 | { |
| 739 | return; |
| 740 | } |
| 741 | |
| 742 | // Commit to the stop before announcing it, then fire OnVmStopping without holding m_lock |
| 743 | // and tear down unconditionally. See OnIdleTimer for the full rationale. |
| 744 | BeginVmStopLockHeld(); |
| 745 | auto endStop = wil::scope_exit([this]() { EndVmStop(); }); |
| 746 | |
| 747 | const auto generation = m_vmGeneration.load(); |
| 748 | |
| 749 | lock.reset(); |
| 750 | try |
| 751 | { |
| 752 | NotifyVmStopping(generation); |
| 753 | } |
| 754 | CATCH_LOG(); |
| 755 | lock = m_lock.lock_exclusive(); |
| 756 | |
| 757 | StopVmLockHeld(); |
| 758 | } |
| 759 | catch (...) |
| 760 | { |
| 761 | error = std::current_exception(); |
| 762 | } |
| 763 | }); |
| 764 | |
| 765 | worker.join(); |
| 766 | |
| 767 | if (error) |
| 768 | { |
| 769 | std::rethrow_exception(error); |
| 770 | } |
| 771 | |
| 772 | return wasAlreadyIdle; |
| 773 | } |
| 774 | |
| 775 | WSLCSessionRuntime::VmLease WSLCSessionRuntime::AcquireVmLease(VmLeasePolicy Policy) |
| 776 | { |
| 777 | return VmLease(*this, Policy); |
| 778 | } |
| 779 | |
| 780 | WSLCSessionRuntime::LockedRuntime WSLCSessionRuntime::Acquire(VmLeasePolicy Policy) |
| 781 | { |
| 782 | return LockedRuntime(*this, Policy); |
| 783 | } |
| 784 | |
| 785 | WSLCSessionRuntime::VmLease::VmLease(WSLCSessionRuntime& Runtime, VmLeasePolicy Policy) : m_runtime(&Runtime) |
| 786 | { |
| 787 | // Record an in-flight operation before bringing the VM up so idle teardown cannot tear it down |
| 788 | // between EnsureVmRunning() and acquiring the shared lock. AddActivity cancels any pending idle |
| 789 | // timer. |
| 790 | m_runtime->m_idleState->AddActivity(); |
| 791 | |
| 792 | auto countCleanup = wil::scope_exit([this]() { |
| 793 | m_runtime->m_idleState->ReleaseActivity(); |
| 794 | m_runtime = nullptr; |
| 795 | }); |
| 796 | |
| 797 | // Activity increment may race with idle teardown. Retry until we hold the lock with VM running. |
| 798 | for (;;) |
| 799 | { |
| 800 | if (Policy == VmLeasePolicy::Acquire) |
| 801 | { |
| 802 | m_runtime->EnsureVmRunning(); |
| 803 | } |
| 804 | |
| 805 | m_lock = m_runtime->m_lock.lock_shared(); |
| 806 | |
| 807 | if (m_runtime->m_vmState.load() == VmState::Running) |
| 808 | { |
| 809 | // An announced stop always happens, so a VM with one pending is unusable even though it is |
| 810 | // still Running: wait for the teardown, then retry, which brings up a fresh VM. ExistingOnly |
| 811 | // callers are exempt and are served by the stopping VM -- see VmLeasePolicy. |
| 812 | if (Policy == VmLeasePolicy::ExistingOnly || !m_runtime->m_vmStopPending.load()) |
| 813 | { |
| 814 | break; |
| 815 | } |
| 816 | |
| 817 | // Release the shared lock before waiting. The teardown must be able to take the exclusive |
| 818 | // lock, and it is the only thing that will ever signal us. |
| 819 | m_lock.reset(); |
| 820 | |
| 821 | // The wait is bounded only so that a handler wedging the teardown is visible in traces; |
| 822 | // there is no correct way to proceed without a VM, so the retry is unconditional. |
| 823 | while (!m_runtime->m_vmStopCompleteEvent.wait(c_vmStopWaitLogIntervalMs)) |
| 824 | { |
| 825 | WSL_LOG( |
| 826 | "WslcVmLeaseWaitingForStop", |
| 827 | TraceLoggingLevel(WINEVENT_LEVEL_WARNING), |
| 828 | TraceLoggingValue(m_runtime->m_id, "SessionId")); |
| 829 | } |
| 830 | |
| 831 | continue; |
| 832 | } |
| 833 | |
| 834 | // ExistingOnly never starts a VM, so retrying could only spin. Reject the caller instead. |
| 835 | THROW_HR_IF(WSLC_E_VM_NOT_RUNNING, Policy == VmLeasePolicy::ExistingOnly); |
| 836 | |
| 837 | WSL_LOG( |
| 838 | "WslcVmLeaseRetry", |
| 839 | TraceLoggingValue(m_runtime->m_id, "SessionId"), |
| 840 | TraceLoggingValue(static_cast<uint32_t>(m_runtime->m_vmState.load()), "VmState")); |
| 841 | m_lock.reset(); |
| 842 | } |
| 843 | |
| 844 | countCleanup.release(); |
| 845 | } |
| 846 | |
| 847 | WSLCSessionRuntime::VmLease::VmLease(VmLease&& Other) noexcept : |
| 848 | m_runtime(std::exchange(Other.m_runtime, nullptr)), m_lock(std::move(Other.m_lock)) |
| 849 | { |
| 850 | } |
| 851 | |
| 852 | WSLCSessionRuntime::VmLease& WSLCSessionRuntime::VmLease::operator=(VmLease&& Other) noexcept |
| 853 | { |
| 854 | if (this != &Other) |
| 855 | { |
| 856 | if (m_runtime != nullptr) |
| 857 | { |
| 858 | // Release the shared lock before the activity reference so that, if this was the last |
| 859 | // activity, idle teardown can immediately take the exclusive lock. |
| 860 | m_lock.reset(); |
| 861 | m_runtime->m_idleState->ReleaseActivity(); |
| 862 | } |
| 863 | |
| 864 | m_runtime = std::exchange(Other.m_runtime, nullptr); |
| 865 | m_lock = std::move(Other.m_lock); |
| 866 | } |
| 867 | |
| 868 | return *this; |
| 869 | } |
| 870 | |
| 871 | WSLCSessionRuntime::VmLease::~VmLease() |
| 872 | { |
| 873 | if (m_runtime != nullptr) |
| 874 | { |
| 875 | // Release the shared lock before the activity reference so that, if this was the last |
| 876 | // activity, idle teardown can immediately take the exclusive lock. ReleaseActivity arms the |
| 877 | // idle timer on the 1->0 transition. |
| 878 | m_lock.reset(); |
| 879 | m_runtime->m_idleState->ReleaseActivity(); |
| 880 | } |
| 881 | } |
| 882 | |
| 883 | WSLCSessionRuntime::LockedRuntime::LockedRuntime(WSLCSessionRuntime& Runtime, VmLeasePolicy Policy) : |
| 884 | m_runtime(&Runtime), m_lease(Runtime.AcquireVmLease(Policy)) |
| 885 | { |
| 886 | } |
| 887 | |
| 888 | WSLCVirtualMachine& WSLCSessionRuntime::LockedRuntime::Vm() |
| 889 | { |
| 890 | return m_runtime->Vm(); |
| 891 | } |
| 892 | |
| 893 | IORelay* WSLCSessionRuntime::LockedRuntime::Relay() |
| 894 | { |
| 895 | return m_runtime->Relay(); |
| 896 | } |
| 897 | |
| 898 | DockerHTTPClient& WSLCSessionRuntime::LockedRuntime::Docker() |
| 899 | { |
| 900 | return m_runtime->Docker(); |
| 901 | } |
| 902 | |
| 903 | void WSLCSessionRuntime::OnVmExited() |
| 904 | { |
| 905 | // A spontaneous exit we must permanently terminate, unless an expected stop already claimed it, |
| 906 | // in which case the exit was wanted and we decline. |
| 907 | if (!TryClaimSpontaneousExit()) |
| 908 | { |
| 909 | WSL_LOG("WslcVmExitedDuringStop", TraceLoggingValue(m_id, "SessionId")); |
| 910 | return; |
| 911 | } |
| 912 | |
| 913 | WSL_LOG( |
| 914 | "VmExited", |
| 915 | TraceLoggingLevel(WINEVENT_LEVEL_WARNING), |
| 916 | TraceLoggingValue(m_id, "SessionId"), |
| 917 | TraceLoggingValue(m_displayName.c_str(), "Name"), |
| 918 | TraceLoggingValue(!m_sessionTerminatingEvent.is_signaled(), "Unexpected")); |
| 919 | |
| 920 | if (m_hooks.OnSpontaneousExit) |
| 921 | { |
| 922 | m_hooks.OnSpontaneousExit(); |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | void WSLCSessionRuntime::Shutdown( |
| 927 | wil::rwlock_release_exclusive_scope_exit& runtimeLock, WSLCVirtualMachineTerminationReason& terminationReason, std::wstring& terminationDetails) |
| 928 | { |
| 929 | if (!m_initialized) |
| 930 | { |
| 931 | return; |
| 932 | } |
| 933 | |
| 934 | // runtimeLock is an exclusive hold on m_lock, guaranteeing no operation is running. |
| 935 | WI_VERIFY(runtimeLock); |
| 936 | |
| 937 | // Permanently disable idle teardown and drain any in-flight timer callback on every exit path, |
| 938 | // including an exception escaping the teardown below. The timer callback captures this runtime, |
| 939 | // but IdleState outlives it (activity tokens hold shared_ptr copies), so skipping the disarm |
| 940 | // would leave a late 1->0 transition able to re-arm a timer that references freed memory. |
| 941 | // |
| 942 | // The exclusive lock is released first: a timer callback already blocked acquiring it must be |
| 943 | // able to obtain it, observe m_terminating, and return, otherwise the drain below deadlocks. |
| 944 | auto disarmIdleState = wil::scope_exit([&]() { |
| 945 | runtimeLock.reset(); |
| 946 | m_idleState->Disarm(); |
| 947 | }); |
| 948 | |
| 949 | // Notify with m_lock dropped, then re-lock for the teardown. The handler may call back into the |
| 950 | // session (e.g. WSLCCreateProcess) which takes a VM lease and this lock; firing under it would |
| 951 | // deadlock. m_terminating is set, so any such reentrant lease fails at EnsureVmRunning's gate |
| 952 | // rather than restarting the VM, and the reacquire can't block on it. A throwing handler must not |
| 953 | // skip the teardown below -- the session is terminating either way. |
| 954 | runtimeLock.reset(); |
| 955 | try |
| 956 | { |
| 957 | NotifyVmStopping(m_vmGeneration.load()); |
| 958 | } |
| 959 | CATCH_LOG(); |
| 960 | runtimeLock = m_lock.lock_exclusive(); |
| 961 | |
| 962 | // Tear down the VM (if running) and all VM-scoped state, capturing the termination reason; the |
| 963 | // reacquired exclusive hold on m_lock satisfies TearDownVmLockHeld's precondition. |
| 964 | TearDownVmLockHeld(/* CaptureTerminationReason */ true); |
| 965 | |
| 966 | m_vmState.store(VmState::None); |
| 967 | |
| 968 | terminationReason = m_lastTerminationReason; |
| 969 | terminationDetails = m_lastTerminationDetails; |
| 970 | |
| 971 | // Signal completion last so any observer of the terminated event sees a fully torn-down |
| 972 | // session and a populated termination reason. |
| 973 | m_sessionTerminatedEvent.SetEvent(); |
| 974 | } |
| 975 | |
| 976 | } // namespace wsl::windows::service::wslc |