| 1 | // Copyright (C) Microsoft Corporation. All rights reserved. |
| 2 | |
| 3 | #include <filesystem> |
| 4 | #include <optional> |
| 5 | #include <iostream> |
| 6 | #include <linux/audit.h> /* Definition of AUDIT_* constants */ |
| 7 | #include <linux/sock_diag.h> |
| 8 | #include <linux/inet_diag.h> |
| 9 | #include <linux/net.h> |
| 10 | #include <sys/syscall.h> |
| 11 | #include <sys/xattr.h> |
| 12 | #include "common.h" // Needs to be included before sal.h before of __reserved macro |
| 13 | #include "NetlinkTransactionError.h" |
| 14 | #include "GnsPortTracker.h" |
| 15 | #include "lxinitshared.h" |
| 16 | #include "seccomp_defs.h" |
| 17 | |
| 18 | // TODO: Include <sys/pidfd.h> and remove this once musl provides it. |
| 19 | // <linux/pidfd.h> cannot be used because it conflicts with musl's <fcntl.h>. |
| 20 | #ifndef PIDFD_THREAD |
| 21 | #define PIDFD_THREAD O_EXCL |
| 22 | #endif |
| 23 | |
| 24 | constexpr size_t c_bind_timeout_seconds = 60; |
| 25 | constexpr auto c_sock_diag_refresh_delay = std::chrono::milliseconds(500); |
| 26 | constexpr auto c_sock_diag_poll_timeout = std::chrono::milliseconds(10); |
| 27 | constexpr auto c_bpf_poll_timeout = std::chrono::milliseconds(500); |
| 28 | |
| 29 | GnsPortTracker::GnsPortTracker( |
| 30 | std::shared_ptr<wsl::shared::SocketChannel> hvSocketChannel, |
| 31 | NetlinkChannel&& netlinkChannel, |
| 32 | std::shared_ptr<SecCompDispatcher> seccompDispatcher, |
| 33 | LX_MINI_INIT_NETWORKING_MODE networkingMode) : |
| 34 | m_hvSocketChannel(std::move(hvSocketChannel)), |
| 35 | m_channel(std::move(netlinkChannel)), |
| 36 | m_seccompDispatcher(seccompDispatcher), |
| 37 | m_networkingMode(networkingMode) |
| 38 | { |
| 39 | m_networkNamespace = std::filesystem::read_symlink("/proc/self/ns/net").string(); |
| 40 | GNS_LOG_INFO("GnsPortTracker initialized with networking mode ({})", static_cast<int>(m_networkingMode)); |
| 41 | } |
| 42 | |
| 43 | void GnsPortTracker::RunPortRefresh() |
| 44 | { |
| 45 | UtilSetThreadName("GnsPortTracker"); |
| 46 | |
| 47 | // The polling of bound sockets is done in a separate thread because |
| 48 | // sock_diag sometimes fails with EBUSY when a bind() is in progress. |
| 49 | // Doing this in a separate thread allows the main thread not to be delayed |
| 50 | // because of transient sock_diag failures |
| 51 | |
| 52 | for (;;) |
| 53 | { |
| 54 | // Netlink will sometimes return EBUSY. Don't fail for that |
| 55 | try |
| 56 | { |
| 57 | std::promise<void> resume; |
| 58 | auto result = PortRefreshResult{ListAllocatedPorts(), time(nullptr), std::bind(&std::promise<void>::set_value, &resume)}; |
| 59 | m_allocatedPortsRefresh.set_value(result); |
| 60 | |
| 61 | resume.get_future().wait(); |
| 62 | } |
| 63 | catch (const NetlinkTransactionError& e) |
| 64 | { |
| 65 | if (e.Error().value_or(0) != -EBUSY) |
| 66 | { |
| 67 | std::cerr << "Failed to refresh allocated ports, " << e.what() << std::endl; |
| 68 | } |
| 69 | } |
| 70 | |
| 71 | std::this_thread::sleep_for(c_sock_diag_refresh_delay); |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | int GnsPortTracker::ProcessSecCompNotification(seccomp_notif* notification) |
| 76 | { |
| 77 | seccomp_notif notificationCopy = *notification; |
| 78 | m_request.post(notificationCopy); |
| 79 | return m_reply.get(); |
| 80 | } |
| 81 | |
| 82 | void GnsPortTracker::Run() |
| 83 | { |
| 84 | // This method consumes seccomp notifications and allows / disallows port allocations |
| 85 | // depending on wsl core's response. |
| 86 | // After dealing with a notification it also looks at the bound ports list to check |
| 87 | // for port deallocation |
| 88 | |
| 89 | std::thread{std::bind(&GnsPortTracker::RunPortRefresh, this)}.detach(); |
| 90 | |
| 91 | auto future = std::make_optional(m_allocatedPortsRefresh.get_future()); |
| 92 | std::optional<PortRefreshResult> refreshResult; |
| 93 | |
| 94 | for (;;) |
| 95 | { |
| 96 | std::optional<BindCall> bindCall; |
| 97 | try |
| 98 | { |
| 99 | bindCall = ReadNextRequest(); |
| 100 | } |
| 101 | catch (const std::exception& e) |
| 102 | { |
| 103 | GNS_LOG_ERROR("Failed to read bind request, {}", e.what()); |
| 104 | } |
| 105 | |
| 106 | if (bindCall.has_value()) |
| 107 | { |
| 108 | int result = 0; |
| 109 | if (bindCall->Request.has_value()) |
| 110 | { |
| 111 | PortAllocation& allocationRequest = bindCall->Request.value(); |
| 112 | result = HandleRequest(allocationRequest); |
| 113 | if (result == 0) |
| 114 | { |
| 115 | TrackPort(allocationRequest); |
| 116 | GNS_LOG_INFO( |
| 117 | "Tracking bind call: family ({}) port ({}) protocol ({})", |
| 118 | allocationRequest.Family, |
| 119 | allocationRequest.Port, |
| 120 | allocationRequest.Protocol); |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | try |
| 125 | { |
| 126 | CompleteRequest(bindCall->CallId, result); |
| 127 | } |
| 128 | catch (const std::exception& e) |
| 129 | { |
| 130 | GNS_LOG_ERROR("Failed to complete bind request, {}", e.what()); |
| 131 | } |
| 132 | |
| 133 | if (bindCall->PortZeroBind.has_value()) |
| 134 | { |
| 135 | try |
| 136 | { |
| 137 | auto allocation = ResolvePortZeroBind(std::move(bindCall->PortZeroBind.value())); |
| 138 | if (allocation.has_value()) |
| 139 | { |
| 140 | const auto portResult = HandleRequest(allocation.value()); |
| 141 | if (portResult == 0) |
| 142 | { |
| 143 | TrackPort(std::move(allocation.value())); |
| 144 | } |
| 145 | else |
| 146 | { |
| 147 | GNS_LOG_ERROR( |
| 148 | "Failed to register resolved port-0 bind: family ({}) port ({}) protocol ({}), error {}", |
| 149 | allocation->Family, |
| 150 | allocation->Port, |
| 151 | allocation->Protocol, |
| 152 | portResult); |
| 153 | } |
| 154 | } |
| 155 | } |
| 156 | catch (const std::exception& e) |
| 157 | { |
| 158 | GNS_LOG_ERROR("Failed to resolve port-0 bind, {}", e.what()); |
| 159 | } |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | // If bindCall is empty, then the read() timed out. Look for any closed port |
| 164 | if (future.has_value() && future->wait_for(c_sock_diag_poll_timeout) == std::future_status::ready) |
| 165 | { |
| 166 | refreshResult.emplace(future->get()); |
| 167 | future.reset(); |
| 168 | m_allocatedPortsRefresh = {}; |
| 169 | |
| 170 | // If this loop's iteration had a bind call, it's possible that RefreshAllocatedPort |
| 171 | // was called before the bind called was processed. Make sure that the port list |
| 172 | // is up to date (If this is called, the next block will schedule another refresh) |
| 173 | if (!bindCall.has_value()) |
| 174 | { |
| 175 | OnRefreshAllocatedPorts(refreshResult->Ports, refreshResult->Timestamp); |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | // Only look at bound ports if there's something to deallocate to avoid wasting cycles |
| 180 | if (refreshResult.has_value()) |
| 181 | { |
| 182 | if (!m_allocatedPorts.empty()) |
| 183 | { |
| 184 | future = m_allocatedPortsRefresh.get_future(); |
| 185 | refreshResult->Resume(); // This will resume the sock_diag thread |
| 186 | refreshResult.reset(); |
| 187 | } |
| 188 | } |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | GnsPortTracker::ActivePorts GnsPortTracker::ListAllocatedPorts() |
| 193 | { |
| 194 | ActivePorts ports; |
| 195 | |
| 196 | inet_diag_req_v2 message{}; |
| 197 | message.sdiag_family = AF_INET; |
| 198 | message.sdiag_protocol = IPPROTO_TCP; |
| 199 | message.idiag_states = ~0; |
| 200 | |
| 201 | auto onMessage = [&](const NetlinkResponse& response) { |
| 202 | for (const auto& e : response.Messages<inet_diag_msg>(SOCK_DIAG_BY_FAMILY)) |
| 203 | { |
| 204 | const auto* payload = e.Payload(); |
| 205 | ports.PortProtocolPairs.emplace(ntohs(payload->id.idiag_sport), static_cast<int>(message.sdiag_protocol)); |
| 206 | |
| 207 | in6_addr address = {}; |
| 208 | if (payload->idiag_family == AF_INET6) |
| 209 | { |
| 210 | static_assert(sizeof(address.s6_addr32) == 16); |
| 211 | static_assert(sizeof(address.s6_addr32) == sizeof(payload->id.idiag_src)); |
| 212 | memcpy(address.s6_addr32, payload->id.idiag_src, sizeof(address.s6_addr32)); |
| 213 | } |
| 214 | else |
| 215 | { |
| 216 | address.s6_addr32[0] = payload->id.idiag_src[0]; |
| 217 | } |
| 218 | |
| 219 | ports.FullAllocations.emplace( |
| 220 | ntohs(payload->id.idiag_sport), static_cast<int>(payload->idiag_family), static_cast<int>(message.sdiag_protocol), address); |
| 221 | } |
| 222 | }; |
| 223 | |
| 224 | { |
| 225 | auto transaction = m_channel.CreateTransaction(message, SOCK_DIAG_BY_FAMILY, NLM_F_DUMP); |
| 226 | transaction.Execute(onMessage); |
| 227 | } |
| 228 | |
| 229 | message.sdiag_family = AF_INET6; |
| 230 | { |
| 231 | auto transaction = m_channel.CreateTransaction(message, SOCK_DIAG_BY_FAMILY, NLM_F_DUMP); |
| 232 | transaction.Execute(onMessage); |
| 233 | } |
| 234 | |
| 235 | message.sdiag_protocol = IPPROTO_UDP; |
| 236 | message.sdiag_family = AF_INET; |
| 237 | { |
| 238 | auto transaction = m_channel.CreateTransaction(message, SOCK_DIAG_BY_FAMILY, NLM_F_DUMP); |
| 239 | transaction.Execute(onMessage); |
| 240 | } |
| 241 | |
| 242 | message.sdiag_family = AF_INET6; |
| 243 | { |
| 244 | auto transaction = m_channel.CreateTransaction(message, SOCK_DIAG_BY_FAMILY, NLM_F_DUMP); |
| 245 | transaction.Execute(onMessage); |
| 246 | } |
| 247 | |
| 248 | return ports; |
| 249 | } |
| 250 | |
| 251 | void GnsPortTracker::OnRefreshAllocatedPorts(const ActivePorts& Ports, time_t Timestamp) |
| 252 | { |
| 253 | // Because there's no way to get notified when the bind() call actually completes, it' possible |
| 254 | // that this method is called before the bind() completion and so the port allocation may not be visible yet. |
| 255 | // To avoid deallocating ports that simply haven't been done allocating yet, m_allocatedPorts stores a timeout |
| 256 | // that prevents deallocating the port unless: |
| 257 | // |
| 258 | // - The port has been seen to be allocated (if so, then the timeout is empty) |
| 259 | // - The timeout has expired |
| 260 | |
| 261 | for (auto it = m_allocatedPorts.begin(); it != m_allocatedPorts.end();) |
| 262 | { |
| 263 | bool portStillActive = false; |
| 264 | if (IsMirroredMode()) |
| 265 | { |
| 266 | // In mirrored mode, match by port+protocol only. Sockets can use a source port even though an explicit |
| 267 | // bind() was not made for that port. As long as there is a socket using the source port, we should not |
| 268 | // deallocate it yet. |
| 269 | // |
| 270 | // For example, a listening socket on port X and a connection socket accepted from that listening socket |
| 271 | // will both use port X. Even if the listening socket is closed the connection socket may still be using |
| 272 | // the same port. |
| 273 | portStillActive = Ports.PortProtocolPairs.contains({it->first.Port, it->first.Protocol}); |
| 274 | } |
| 275 | else |
| 276 | { |
| 277 | // In other modes, match by full allocation (port+family+protocol+address). |
| 278 | portStillActive = Ports.FullAllocations.contains(it->first); |
| 279 | } |
| 280 | |
| 281 | if (!portStillActive) |
| 282 | { |
| 283 | if (!it->second.has_value() || it->second.value() < Timestamp) |
| 284 | { |
| 285 | auto result = RequestPort(it->first, false); |
| 286 | if (result != 0) |
| 287 | { |
| 288 | std::cerr << "GnsPortTracker: Failed to deallocate port " << it->first << ", " << result << std::endl; |
| 289 | } |
| 290 | |
| 291 | GNS_LOG_INFO( |
| 292 | "No longer tracking bind call: family ({}) port ({}) protocol ({})", |
| 293 | it->first.Family, |
| 294 | it->first.Port, |
| 295 | it->first.Protocol); |
| 296 | |
| 297 | it = m_allocatedPorts.erase(it); |
| 298 | continue; |
| 299 | } |
| 300 | } |
| 301 | else |
| 302 | { |
| 303 | it->second.reset(); // The port is known to be allocated, remove the timeout |
| 304 | } |
| 305 | |
| 306 | ++it; |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | int GnsPortTracker::RequestPort(const PortAllocation& Port, bool allocate) |
| 311 | { |
| 312 | LX_GNS_PORT_ALLOCATION_REQUEST request{}; |
| 313 | request.Header.MessageType = LxGnsMessagePortMappingRequest; |
| 314 | request.Header.MessageSize = sizeof(request); |
| 315 | request.Af = Port.Family; |
| 316 | request.Protocol = Port.Protocol; |
| 317 | request.Port = Port.Port; |
| 318 | request.Allocate = allocate; |
| 319 | static_assert(sizeof(request.Address32) == 16); |
| 320 | static_assert(sizeof(request.Address32) == sizeof(Port.Address.s6_addr32)); |
| 321 | memcpy(request.Address32, Port.Address.s6_addr32, sizeof(request.Address32)); |
| 322 | |
| 323 | const auto& response = m_hvSocketChannel->Transaction(request); |
| 324 | |
| 325 | return response.Result; |
| 326 | } |
| 327 | |
| 328 | int GnsPortTracker::HandleRequest(const PortAllocation& Port) |
| 329 | { |
| 330 | // If the port is already allocated, let the call go through and the kernel will |
| 331 | // decide if bind() should succeed or not |
| 332 | // Note: Returning 0 will also cause the port's timeout to be updated |
| 333 | |
| 334 | if (m_allocatedPorts.contains(Port)) |
| 335 | { |
| 336 | GNS_LOG_INFO("Request for a port that's already reserved (family {}, port {}, protocol {})", Port.Family, Port.Port, Port.Protocol); |
| 337 | return 0; |
| 338 | } |
| 339 | |
| 340 | // Ask the host for this port otherwise |
| 341 | const auto error = RequestPort(Port, true); |
| 342 | GNS_LOG_INFO( |
| 343 | "Requested the host for port allocation on port (family {}, port {}, protocol {}) - returned {}", Port.Family, Port.Port, Port.Protocol, error); |
| 344 | return error; |
| 345 | } |
| 346 | |
| 347 | std::optional<GnsPortTracker::BindCall> GnsPortTracker::ReadNextRequest() |
| 348 | { |
| 349 | // Read the call information |
| 350 | auto request_value = m_request.try_get(c_bpf_poll_timeout); |
| 351 | if (!request_value.has_value()) |
| 352 | { |
| 353 | return {}; |
| 354 | } |
| 355 | |
| 356 | auto callInfo = request_value.value(); |
| 357 | |
| 358 | // This logic needs to be defensive because the calling process is blocked until |
| 359 | // CompleteRequest() is called, so if the call information can't be processed because |
| 360 | // the caller has done something wrong (bad pointer, fd, or protocol), just let it go through |
| 361 | // and the kernel will fail it |
| 362 | |
| 363 | try |
| 364 | { |
| 365 | return GetCallInfo(callInfo.id, callInfo.pid, callInfo.data.arch, callInfo.data.nr, gsl::make_span(callInfo.data.args)); |
| 366 | } |
| 367 | catch (const std::exception& e) |
| 368 | { |
| 369 | GNS_LOG_ERROR("Failed to read bind() call info with ID {} for pid {}, {}", callInfo.id, callInfo.pid, e.what()); |
| 370 | return {{{}, {}, callInfo.id}}; |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | std::optional<GnsPortTracker::BindCall> GnsPortTracker::GetCallInfo( |
| 375 | uint64_t CallId, pid_t Pid, int Arch, int SysCallNumber, const gsl::span<unsigned long long>& Arguments) |
| 376 | { |
| 377 | auto ParseSocket = [&](int Socket, size_t AddressPtr, size_t AddressLength) -> std::optional<BindCall> { |
| 378 | if (AddressLength < sizeof(sockaddr) || AddressLength > sizeof(sockaddr_storage)) |
| 379 | { |
| 380 | return {{{}, {}, CallId}}; // Invalid sockaddr. Let it go through. |
| 381 | } |
| 382 | |
| 383 | auto networkNamespace = std::filesystem::read_symlink(std::format("/proc/{}/ns/net", Pid)).string(); |
| 384 | if (networkNamespace != m_networkNamespace) |
| 385 | { |
| 386 | GNS_LOG_INFO("Skipping bind() call for pid {} in network namespace {}", Pid, networkNamespace.c_str()); |
| 387 | return {{{}, {}, CallId}}; // Different network namespace. Let it go through. |
| 388 | } |
| 389 | |
| 390 | auto processMemory = m_seccompDispatcher->ReadProcessMemory(CallId, Pid, AddressPtr, AddressLength); |
| 391 | if (!processMemory.has_value()) |
| 392 | { |
| 393 | throw RuntimeErrorWithSourceLocation("Failed to read process memory"); |
| 394 | } |
| 395 | |
| 396 | sockaddr& address = *reinterpret_cast<sockaddr*>(processMemory->data()); |
| 397 | |
| 398 | if ((address.sa_family != AF_INET && address.sa_family != AF_INET6) || |
| 399 | (address.sa_family == AF_INET6 && AddressLength < sizeof(sockaddr_in6))) |
| 400 | { |
| 401 | return {{{}, {}, CallId}}; // This is a non IP call, or invalid sockaddr_in6. Let it go through |
| 402 | } |
| 403 | |
| 404 | // Read the port. The port *happens* to be in the same spot in memory for both sockaddr_in |
| 405 | // and sockaddr_in6. To avoid a second memory read, we take advantage of this fact to fetch |
| 406 | // the port from the currently read memory, regardless of the address family. |
| 407 | static_assert(sizeof(sockaddr_in) <= sizeof(sockaddr)); |
| 408 | |
| 409 | const auto* inAddr = reinterpret_cast<sockaddr_in*>(&address); |
| 410 | in_port_t port = ntohs(inAddr->sin_port); |
| 411 | if (port == 0) |
| 412 | { |
| 413 | // Port 0 means the kernel will assign an ephemeral port. We can't know |
| 414 | // the port until after the bind() completes, so duplicate the socket fd |
| 415 | // now (while the process is still stopped by seccomp) and defer the |
| 416 | // getsockname() lookup to after CompleteRequest() unblocks it. |
| 417 | try |
| 418 | { |
| 419 | const int protocol = GetSocketProtocol(Pid, Socket); |
| 420 | auto dupFd = DuplicateSocketFd(Pid, Socket); |
| 421 | if (!dupFd) |
| 422 | { |
| 423 | return {{{}, {}, CallId}}; |
| 424 | } |
| 425 | if (!m_seccompDispatcher->ValidateCookie(CallId)) |
| 426 | { |
| 427 | return {{{}, {}, CallId}}; |
| 428 | } |
| 429 | return {{{}, DeferredPortLookup{Pid, std::move(dupFd), protocol}, CallId}}; |
| 430 | } |
| 431 | catch (const std::exception&) |
| 432 | { |
| 433 | return {{{}, {}, CallId}}; // Can't determine protocol, just let it through |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | in6_addr storedAddress = {}; |
| 438 | |
| 439 | if (address.sa_family == AF_INET) |
| 440 | { |
| 441 | storedAddress.s6_addr32[0] = inAddr->sin_addr.s_addr; |
| 442 | } |
| 443 | else |
| 444 | { |
| 445 | const auto* inAddr6 = reinterpret_cast<sockaddr_in6*>(&address); |
| 446 | memcpy(storedAddress.s6_addr32, inAddr6->sin6_addr.s6_addr32, sizeof(storedAddress.s6_addr32)); |
| 447 | } |
| 448 | |
| 449 | // It's possible that the calling process lied and passed a sockaddr that |
| 450 | // doesn't match the underlying socket family or a bad fd. If that's the case, |
| 451 | // then GetSocketProtocol() will throw |
| 452 | const int protocol = GetSocketProtocol(Pid, Socket); |
| 453 | |
| 454 | // As GetSocketProtocol interacts with /proc/<pid>, to avoid TOCTOU races we need to |
| 455 | // verify that call is still valid (call id is the same thing as cookie) |
| 456 | if (!m_seccompDispatcher->ValidateCookie(CallId)) |
| 457 | { |
| 458 | throw RuntimeErrorWithSourceLocation(std::format("Invalid call id {}", CallId)); |
| 459 | } |
| 460 | |
| 461 | return {{{PortAllocation(port, address.sa_family, protocol, storedAddress)}, {}, CallId}}; |
| 462 | }; |
| 463 | |
| 464 | // listen() can trigger an implicit autobind (assigning an ephemeral port) on a socket that |
| 465 | // was never explicitly bind()'d. There's no sockaddr to inspect here (listen() only takes a |
| 466 | // socket fd and a backlog), and we can't tell in advance whether the socket is already bound, |
| 467 | // so duplicate the fd and check getsockname() immediately: if it's already bound (the common |
| 468 | // bind()+listen() case), resolve the port synchronously here. Otherwise defer resolution to |
| 469 | // ResolvePortZeroBind(), the same as a bind(port=0) call, since the port isn't assigned until |
| 470 | // the listen() syscall (which performs the implicit autobind) actually completes in-kernel. |
| 471 | auto ParseListen = [&](int Socket) -> std::optional<BindCall> { |
| 472 | try |
| 473 | { |
| 474 | auto networkNamespace = std::filesystem::read_symlink(std::format("/proc/{}/ns/net", Pid)).string(); |
| 475 | if (networkNamespace != m_networkNamespace) |
| 476 | { |
| 477 | GNS_LOG_INFO("Skipping listen() call for pid {} in network namespace {}", Pid, networkNamespace.c_str()); |
| 478 | return {{{}, {}, CallId}}; // Different network namespace. Let it go through. |
| 479 | } |
| 480 | |
| 481 | const int protocol = GetSocketProtocol(Pid, Socket); |
| 482 | auto dupFd = DuplicateSocketFd(Pid, Socket); |
| 483 | if (!dupFd) |
| 484 | { |
| 485 | return {{{}, {}, CallId}}; |
| 486 | } |
| 487 | if (!m_seccompDispatcher->ValidateCookie(CallId)) |
| 488 | { |
| 489 | return {{{}, {}, CallId}}; |
| 490 | } |
| 491 | |
| 492 | // If the socket was already explicitly bind()'d - the common case of a normal |
| 493 | // bind() followed by listen() - its port is already known right now, before the |
| 494 | // listen() syscall even runs, so resolve it immediately instead of deferring |
| 495 | // through the post-completion polling path. That path is reserved for the case |
| 496 | // where listen() itself is what triggers the kernel's implicit autobind (i.e. no |
| 497 | // prior bind() call), which can only be observed after listen() has completed. |
| 498 | sockaddr_storage storage{}; |
| 499 | socklen_t addressLength = sizeof(storage); |
| 500 | if (getsockname(dupFd.get(), reinterpret_cast<sockaddr*>(&storage), &addressLength) == 0) |
| 501 | { |
| 502 | in_port_t port = 0; |
| 503 | in6_addr address = {}; |
| 504 | if (storage.ss_family == AF_INET) |
| 505 | { |
| 506 | const auto* sin = reinterpret_cast<const sockaddr_in*>(&storage); |
| 507 | port = ntohs(sin->sin_port); |
| 508 | address.s6_addr32[0] = sin->sin_addr.s_addr; |
| 509 | } |
| 510 | else if (storage.ss_family == AF_INET6) |
| 511 | { |
| 512 | const auto* sin6 = reinterpret_cast<const sockaddr_in6*>(&storage); |
| 513 | port = ntohs(sin6->sin6_port); |
| 514 | memcpy(address.s6_addr32, sin6->sin6_addr.s6_addr32, sizeof(address.s6_addr32)); |
| 515 | } |
| 516 | |
| 517 | if (port != 0) |
| 518 | { |
| 519 | return {{{PortAllocation(port, static_cast<int>(storage.ss_family), protocol, address)}, {}, CallId}}; |
| 520 | } |
| 521 | } |
| 522 | |
| 523 | return {{{}, DeferredPortLookup{Pid, std::move(dupFd), protocol}, CallId}}; |
| 524 | } |
| 525 | catch (const std::exception&) |
| 526 | { |
| 527 | return {{{}, {}, CallId}}; // Not an IP socket (or can't determine its protocol), just let it through |
| 528 | } |
| 529 | }; |
| 530 | |
| 531 | #ifdef __x86_64__ |
| 532 | if (Arch & __AUDIT_ARCH_64BIT) |
| 533 | { |
| 534 | if (SysCallNumber == __NR_listen) |
| 535 | { |
| 536 | return ParseListen(Arguments[0]); |
| 537 | } |
| 538 | |
| 539 | return ParseSocket(Arguments[0], Arguments[1], Arguments[2]); |
| 540 | } |
| 541 | // Note: 32bit on x86_64 uses the __NR_socketcall with the first argument |
| 542 | // set to SYS_BIND/SYS_LISTEN to make bind()/listen() system calls and the |
| 543 | // second argument is a pointer to a block of memory containing the original arguments. |
| 544 | else |
| 545 | { |
| 546 | if (Arguments[0] == SYS_LISTEN) |
| 547 | { |
| 548 | // Grab the first parameter (the socket fd). |
| 549 | auto processMemory = m_seccompDispatcher->ReadProcessMemory(CallId, Pid, Arguments[1], sizeof(uint32_t)); |
| 550 | if (!processMemory.has_value()) |
| 551 | { |
| 552 | throw RuntimeErrorWithSourceLocation("Failed to read process memory"); |
| 553 | } |
| 554 | |
| 555 | const uint32_t* CopiedArguments = reinterpret_cast<uint32_t*>(processMemory->data()); |
| 556 | return ParseListen(CopiedArguments[0]); |
| 557 | } |
| 558 | |
| 559 | if (Arguments[0] != SYS_BIND) |
| 560 | { |
| 561 | return {{{}, {}, CallId}}; // Not a bind or listen call, just let the call go through |
| 562 | } |
| 563 | // Grab the first 3 parameters |
| 564 | auto processMemory = m_seccompDispatcher->ReadProcessMemory(CallId, Pid, Arguments[1], sizeof(uint32_t) * 3); |
| 565 | if (!processMemory.has_value()) |
| 566 | { |
| 567 | throw RuntimeErrorWithSourceLocation("Failed to read process memory"); |
| 568 | } |
| 569 | |
| 570 | uint32_t* CopiedArguments = reinterpret_cast<uint32_t*>(processMemory->data()); |
| 571 | return ParseSocket(CopiedArguments[0], CopiedArguments[1], CopiedArguments[2]); |
| 572 | } |
| 573 | #else |
| 574 | // Both native 64-bit listen() (trapped via __NR_listen, e.g. on aarch64) and 32-bit ARM |
| 575 | // compat listen() (trapped via the hardcoded ARMV7_NR_listen syscall number) land here, so |
| 576 | // both syscall numbers must be checked. |
| 577 | if (SysCallNumber == __NR_listen || SysCallNumber == ARMV7_NR_listen) |
| 578 | { |
| 579 | return ParseListen(Arguments[0]); |
| 580 | } |
| 581 | |
| 582 | return ParseSocket(Arguments[0], Arguments[1], Arguments[2]); |
| 583 | #endif |
| 584 | } |
| 585 | |
| 586 | void GnsPortTracker::CompleteRequest(uint64_t id, int result) |
| 587 | { |
| 588 | m_reply.post(result); |
| 589 | } |
| 590 | |
| 591 | int GnsPortTracker::GetSocketProtocol(int pid, int fd) |
| 592 | { |
| 593 | const auto path = std::format("/proc/{}/fd/{}", pid, fd); |
| 594 | |
| 595 | // Because there's a race between the time where the buffer size is determined |
| 596 | // and the actual getxattr() call, retry until the buffer size is big enough |
| 597 | std::string protocol; |
| 598 | int result = -1; |
| 599 | do |
| 600 | { |
| 601 | int bufferSize = Syscall(getxattr, path.c_str(), "system.sockprotoname", nullptr, 0); |
| 602 | protocol.resize(std::max(0, bufferSize - 1)); |
| 603 | |
| 604 | result = getxattr(path.c_str(), "system.sockprotoname", protocol.data(), bufferSize); |
| 605 | } while (result < 0 && errno == ERANGE); |
| 606 | |
| 607 | if (result < 0) |
| 608 | { |
| 609 | throw RuntimeErrorWithSourceLocation(std::format("Failed to read protocol for socket: {}, {}", path, errno)); |
| 610 | } |
| 611 | |
| 612 | // In case the size of the attribute shrunk between the two getxattr calls |
| 613 | protocol.resize(std::max(0, result - 1)); |
| 614 | |
| 615 | if (protocol == "TCP" || protocol == "TCPv6") |
| 616 | { |
| 617 | return IPPROTO_TCP; |
| 618 | } |
| 619 | else if (protocol == "UDP" || protocol == "UDPv6") |
| 620 | { |
| 621 | return IPPROTO_UDP; |
| 622 | } |
| 623 | |
| 624 | throw RuntimeErrorWithSourceLocation(std::format("Unexpected IP socket protocol: {}", protocol)); |
| 625 | } |
| 626 | |
| 627 | wil::unique_fd GnsPortTracker::DuplicateSocketFd(pid_t Pid, int SocketFd) |
| 628 | { |
| 629 | // Duplicate the socket fd from the target process into our address space. |
| 630 | // We cannot use open("/proc/pid/fd/N") for sockets because the symlink target |
| 631 | // (socket:[inode]) is not a valid filesystem path. Use pidfd_getfd() instead. |
| 632 | // PIDFD_THREAD requires kernel >= 6.9. Fallback to process only if not supported. |
| 633 | int pidFdResult = static_cast<int>(syscall(SYS_pidfd_open, Pid, PIDFD_THREAD)); |
| 634 | if (pidFdResult < 0 && errno == EINVAL) |
| 635 | { |
| 636 | pidFdResult = static_cast<int>(syscall(SYS_pidfd_open, Pid, 0u)); |
| 637 | } |
| 638 | |
| 639 | wil::unique_fd pidFd(pidFdResult); |
| 640 | if (!pidFd) |
| 641 | { |
| 642 | GNS_LOG_INFO("Port-0 bind: pidfd_open failed for pid {} (errno {})", Pid, errno); |
| 643 | return {}; |
| 644 | } |
| 645 | |
| 646 | wil::unique_fd dupFd(static_cast<int>(syscall(SYS_pidfd_getfd, pidFd.get(), SocketFd, 0u))); |
| 647 | if (!dupFd) |
| 648 | { |
| 649 | GNS_LOG_INFO("Port-0 bind: pidfd_getfd failed for pid {} fd {} (errno {})", Pid, SocketFd, errno); |
| 650 | } |
| 651 | |
| 652 | return dupFd; |
| 653 | } |
| 654 | |
| 655 | void GnsPortTracker::TrackPort(PortAllocation allocation) |
| 656 | try |
| 657 | { |
| 658 | // Use insert_or_assign so the deallocation timeout is refreshed if the same |
| 659 | // port key is already present (emplace would silently keep the old entry). |
| 660 | m_allocatedPorts.insert_or_assign(std::move(allocation), std::make_optional(time(nullptr) + c_bind_timeout_seconds)); |
| 661 | } |
| 662 | catch (const std::exception& e) |
| 663 | { |
| 664 | GNS_LOG_ERROR("Failed to track port allocation, {}", e.what()); |
| 665 | } |
| 666 | |
| 667 | std::optional<GnsPortTracker::PortAllocation> GnsPortTracker::ResolvePortZeroBind(DeferredPortLookup lookup) |
| 668 | { |
| 669 | // This resolves the port for both an explicit bind(port=0) and an implicit |
| 670 | // autobind triggered by listen(), since neither can be known until after the |
| 671 | // syscall has actually completed in-kernel. |
| 672 | // |
| 673 | // The socket fd was already duplicated (via pidfd_getfd) while the target process |
| 674 | // was stopped by seccomp, so it remains valid even if the process has closed or |
| 675 | // reused the original fd number. |
| 676 | |
| 677 | // The bind() syscall has been completed (CompleteRequest() already unblocked the |
| 678 | // caller). Poll getsockname() briefly until the kernel assigns a port. |
| 679 | constexpr int maxRetries = 25; |
| 680 | constexpr auto retryDelay = std::chrono::milliseconds(10); |
| 681 | |
| 682 | in_port_t port = 0; |
| 683 | in6_addr address = {}; |
| 684 | int resolvedFamily = 0; |
| 685 | |
| 686 | for (int attempt = 0; attempt < maxRetries; ++attempt) |
| 687 | { |
| 688 | if (attempt > 0) |
| 689 | { |
| 690 | std::this_thread::sleep_for(retryDelay); |
| 691 | } |
| 692 | |
| 693 | sockaddr_storage storage{}; |
| 694 | socklen_t addrLen = sizeof(storage); |
| 695 | if (getsockname(lookup.DuplicatedSocketFd.get(), reinterpret_cast<sockaddr*>(&storage), &addrLen) != 0) |
| 696 | { |
| 697 | GNS_LOG_ERROR("Port-0 bind: getsockname failed for pid {} (errno {})", lookup.Pid, errno); |
| 698 | return {}; |
| 699 | } |
| 700 | |
| 701 | resolvedFamily = static_cast<int>(storage.ss_family); |
| 702 | |
| 703 | if (storage.ss_family == AF_INET) |
| 704 | { |
| 705 | const auto* sin = reinterpret_cast<const sockaddr_in*>(&storage); |
| 706 | port = ntohs(sin->sin_port); |
| 707 | address.s6_addr32[0] = sin->sin_addr.s_addr; |
| 708 | } |
| 709 | else if (storage.ss_family == AF_INET6) |
| 710 | { |
| 711 | const auto* sin6 = reinterpret_cast<const sockaddr_in6*>(&storage); |
| 712 | port = ntohs(sin6->sin6_port); |
| 713 | memcpy(address.s6_addr32, sin6->sin6_addr.s6_addr32, sizeof(address.s6_addr32)); |
| 714 | } |
| 715 | else |
| 716 | { |
| 717 | GNS_LOG_ERROR("Port-0 bind: unexpected address family ({}) for pid {}", resolvedFamily, lookup.Pid); |
| 718 | return {}; |
| 719 | } |
| 720 | |
| 721 | if (port != 0) |
| 722 | { |
| 723 | break; |
| 724 | } |
| 725 | } |
| 726 | |
| 727 | if (port == 0) |
| 728 | { |
| 729 | GNS_LOG_ERROR("Port-0 bind: kernel did not assign a port for pid {} after retries", lookup.Pid); |
| 730 | return {}; |
| 731 | } |
| 732 | |
| 733 | GNS_LOG_INFO( |
| 734 | "Port-0 bind resolved: family ({}) port ({}) protocol ({}) for pid {}", resolvedFamily, port, lookup.Protocol, lookup.Pid); |
| 735 | return PortAllocation(port, resolvedFamily, lookup.Protocol, address); |
| 736 | } |
| 737 | |
| 738 | std::ostream& operator<<(std::ostream& out, const GnsPortTracker::PortAllocation& entry) |
| 739 | { |
| 740 | return out << "Port=" << entry.Port << ", Family=" << entry.Family << ", Protocol=" << entry.Protocol; |
| 741 | } |