| 1 | // Copyright (C) Microsoft Corporation. All rights reserved. |
| 2 | #include "precomp.h" |
| 3 | #include "p9io.h" |
| 4 | |
| 5 | #ifndef TEMP_FAILURE_RETRY |
| 6 | #define TEMP_FAILURE_RETRY(expression) \ |
| 7 | (__extension__({ \ |
| 8 | long int __result; \ |
| 9 | do \ |
| 10 | __result = (long int)(expression); \ |
| 11 | while (__result == -1L && errno == EINTR); \ |
| 12 | __result; \ |
| 13 | })) |
| 14 | #endif |
| 15 | |
| 16 | namespace p9fs { |
| 17 | |
| 18 | EpollWatcher g_Watcher; |
| 19 | |
| 20 | CoroutineIoIssuer::CoroutineIoIssuer(int fd) : m_FileDescriptor(fd) |
| 21 | { |
| 22 | } |
| 23 | |
| 24 | void CoroutineIoIssuer::Callback(sigval value) |
| 25 | { |
| 26 | const auto operation = static_cast<CoroutineIoOperation*>(value.sival_ptr); |
| 27 | auto error = aio_error(&operation->ControlBlock); |
| 28 | if (error == EINPROGRESS) |
| 29 | { |
| 30 | return; |
| 31 | } |
| 32 | auto bytesTransferred = aio_return(&operation->ControlBlock); |
| 33 | |
| 34 | operation->Result = {-error, error == 0 ? static_cast<size_t>(bytesTransferred) : 0}; |
| 35 | if (!operation->DoneOrCoroutine.exchange(true)) |
| 36 | { |
| 37 | return; |
| 38 | } |
| 39 | |
| 40 | // TODO: Can we use this thread to resume the coroutine? |
| 41 | g_Scheduler.Schedule(operation->Coroutine); |
| 42 | } |
| 43 | |
| 44 | bool CoroutineIoIssuer::PreIssue(CoroutineIoOperation& operation, CancelToken& token) |
| 45 | { |
| 46 | // Register the IO for cancellation. |
| 47 | operation.ControlBlock = {}; |
| 48 | operation.ControlBlock.aio_fildes = m_FileDescriptor; |
| 49 | operation.ControlBlock.aio_sigevent.sigev_notify = SIGEV_THREAD; |
| 50 | operation.ControlBlock.aio_sigevent.sigev_notify_function = Callback; |
| 51 | operation.ControlBlock.aio_sigevent.sigev_value.sival_ptr = &operation; |
| 52 | if (token.Register(operation)) |
| 53 | { |
| 54 | return true; |
| 55 | } |
| 56 | |
| 57 | // The operation has already been cancelled. Don't even issue the IO. |
| 58 | operation.Result = {-ECANCELED, 0}; |
| 59 | operation.DoneOrCoroutine = true; |
| 60 | return false; |
| 61 | } |
| 62 | |
| 63 | void CoroutineIoIssuer::IssueFailed(CancelToken& token) |
| 64 | { |
| 65 | // Unwind the work done in PreIssue. |
| 66 | token.Unregister(); |
| 67 | } |
| 68 | |
| 69 | void CoroutineIoIssuer::PostIssue(CoroutineIoOperation& operation, CancelToken& token, IoResult result) |
| 70 | { |
| 71 | if (result.Error != 0) |
| 72 | { |
| 73 | // The IO completed synchronously. |
| 74 | operation.Result = result; |
| 75 | operation.DoneOrCoroutine = true; |
| 76 | |
| 77 | WI_ASSERT(operation.Coroutine == nullptr); |
| 78 | } |
| 79 | else if (token.Cancelled()) |
| 80 | { |
| 81 | // The IO did not complete synchronously, but the operation has been |
| 82 | // cancelled. Depending on when the cancel occurred, the IO may not have |
| 83 | // been cancelled, so cancel it now. |
| 84 | aio_cancel(operation.ControlBlock.aio_fildes, &operation.ControlBlock); |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | // Register an epoll operation for either an in or an out event. |
| 89 | // Returns true if the operation must suspend to wait for the event, or false if the operation can |
| 90 | // resume immediately. |
| 91 | // N.B. There can be only one operation registered at a time for each event. |
| 92 | bool EpollDispatcher::Register(int event, CoroutineEpollOperation& operation) |
| 93 | { |
| 94 | std::scoped_lock<std::mutex> lock{m_lock}; |
| 95 | if (event == EPOLLIN) |
| 96 | { |
| 97 | FAIL_FAST_IF(m_inOperation != nullptr); |
| 98 | if (WI_IsFlagSet(m_currentEvents, EPOLLIN)) |
| 99 | { |
| 100 | WI_ClearFlag(m_currentEvents, EPOLLIN); |
| 101 | |
| 102 | // Since Register is called before the operation is registered for cancellation, |
| 103 | // it's not possible for this to return false. |
| 104 | operation.SetResult(0); |
| 105 | return false; |
| 106 | } |
| 107 | |
| 108 | m_inOperation = &operation; |
| 109 | } |
| 110 | else |
| 111 | { |
| 112 | FAIL_FAST_IF(event != EPOLLOUT); |
| 113 | FAIL_FAST_IF(m_outOperation != nullptr); |
| 114 | if (WI_IsFlagSet(m_currentEvents, EPOLLOUT)) |
| 115 | { |
| 116 | WI_ClearFlag(m_currentEvents, EPOLLOUT); |
| 117 | operation.SetResult(0); |
| 118 | return false; |
| 119 | } |
| 120 | |
| 121 | m_outOperation = &operation; |
| 122 | } |
| 123 | |
| 124 | return true; |
| 125 | } |
| 126 | |
| 127 | // Removes the handler for the specified event (EPOLLIN or EPOLLOUT). |
| 128 | void EpollDispatcher::Remove(int event) |
| 129 | { |
| 130 | std::scoped_lock<std::mutex> lock{m_lock}; |
| 131 | if (event == EPOLLIN) |
| 132 | { |
| 133 | m_inOperation = nullptr; |
| 134 | } |
| 135 | else |
| 136 | { |
| 137 | FAIL_FAST_IF(event != EPOLLOUT); |
| 138 | m_outOperation = nullptr; |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | // Notifies the dispatcher an event has occurred. |
| 143 | void EpollDispatcher::Notify(int events) |
| 144 | { |
| 145 | std::scoped_lock<std::mutex> lock{m_lock}; |
| 146 | |
| 147 | // Resume the out operation first, since that is responding to an existing message rather than |
| 148 | // reading the request for a new one. |
| 149 | if (WI_IsFlagSet(events, EPOLLOUT)) |
| 150 | { |
| 151 | if (m_outOperation != nullptr) |
| 152 | { |
| 153 | m_outOperation->Resume(0); |
| 154 | m_outOperation = nullptr; |
| 155 | } |
| 156 | else |
| 157 | { |
| 158 | // If no operation is registered, remember the event occurred for the next time one |
| 159 | // is registered. |
| 160 | WI_SetFlag(m_currentEvents, EPOLLOUT); |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | if (WI_IsFlagSet(events, EPOLLIN)) |
| 165 | { |
| 166 | if (m_inOperation != nullptr) |
| 167 | { |
| 168 | m_inOperation->Resume(0); |
| 169 | m_inOperation = nullptr; |
| 170 | } |
| 171 | else |
| 172 | { |
| 173 | // If no operation is registered, remember the event occurred for the next time one |
| 174 | // is registered. |
| 175 | WI_SetFlag(m_currentEvents, EPOLLIN); |
| 176 | } |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | void EpollWatcher::Run() |
| 181 | { |
| 182 | FAIL_FAST_IF(m_EpollFileDescriptor >= 0); |
| 183 | |
| 184 | m_EpollFileDescriptor = epoll_create1(EPOLL_CLOEXEC); |
| 185 | THROW_LAST_ERROR_IF(m_EpollFileDescriptor < 0); |
| 186 | |
| 187 | std::thread(WatchThread, this).detach(); |
| 188 | } |
| 189 | |
| 190 | void EpollWatcher::Add(int fd, int events, EpollDispatcher& dispatcher) |
| 191 | { |
| 192 | epoll_event event{}; |
| 193 | event.events = events; |
| 194 | event.data.ptr = &dispatcher; |
| 195 | THROW_LAST_ERROR_IF(epoll_ctl(m_EpollFileDescriptor, EPOLL_CTL_ADD, fd, &event) < 0); |
| 196 | } |
| 197 | |
| 198 | void EpollWatcher::Remove(int fd) |
| 199 | { |
| 200 | THROW_LAST_ERROR_IF(epoll_ctl(m_EpollFileDescriptor, EPOLL_CTL_DEL, fd, nullptr) < 0); |
| 201 | } |
| 202 | |
| 203 | void EpollWatcher::WatchThread(EpollWatcher* watcher) |
| 204 | { |
| 205 | for (;;) |
| 206 | { |
| 207 | epoll_event events[10]; |
| 208 | int result = TEMP_FAILURE_RETRY(epoll_wait(watcher->m_EpollFileDescriptor, events, 10, -1)); |
| 209 | THROW_LAST_ERROR_IF(result < 0); |
| 210 | |
| 211 | for (int i = 0; i < result; ++i) |
| 212 | { |
| 213 | if (events[i].data.ptr != nullptr) |
| 214 | { |
| 215 | const auto dispatcher = static_cast<EpollDispatcher*>(events[i].data.ptr); |
| 216 | dispatcher->Notify(events[i].events); |
| 217 | } |
| 218 | } |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | Task<size_t> RecvAsync(CoroutineEpollIssuer& socket, gsl::span<gsl::byte> buffer, CancelToken& token) |
| 223 | { |
| 224 | CoroutineEpollOperation operation; |
| 225 | auto result = |
| 226 | co_await socket.Issue<ssize_t>(operation, token, EPOLLIN, [&](int fd) { return recv(fd, buffer.data(), buffer.size(), 0); }); |
| 227 | |
| 228 | if (result < 0) |
| 229 | { |
| 230 | THROW_ERRNO(-result); |
| 231 | } |
| 232 | |
| 233 | co_return static_cast<size_t>(result); |
| 234 | } |
| 235 | |
| 236 | Task<size_t> SendAsync(CoroutineEpollIssuer& socket, gsl::span<const gsl::byte> buffer, CancelToken& token) |
| 237 | { |
| 238 | CoroutineEpollOperation operation; |
| 239 | auto result = co_await socket.Issue<ssize_t>( |
| 240 | operation, token, EPOLLOUT, [&](int fd) { return send(fd, buffer.data(), buffer.size(), 0); }); |
| 241 | |
| 242 | if (result < 0) |
| 243 | { |
| 244 | THROW_ERRNO(-result); |
| 245 | } |
| 246 | |
| 247 | co_return static_cast<size_t>(result); |
| 248 | } |
| 249 | |
| 250 | Task<int> AcceptAsync(CoroutineEpollIssuer& listen, CancelToken& token) |
| 251 | { |
| 252 | CoroutineEpollOperation operation; |
| 253 | auto result = co_await listen.Issue<int>( |
| 254 | operation, token, EPOLLIN, [&](int fd) { return accept4(fd, nullptr, nullptr, SOCK_NONBLOCK | SOCK_CLOEXEC); }); |
| 255 | |
| 256 | if (result < 0) |
| 257 | { |
| 258 | THROW_ERRNO(-result); |
| 259 | } |
| 260 | |
| 261 | co_return result; |
| 262 | } |
| 263 | |
| 264 | Task<IoResult> ReadAsync(CoroutineIoIssuer& file, std::uint64_t offset, gsl::span<gsl::byte> buffer, CancelToken& token) |
| 265 | { |
| 266 | CoroutineIoOperation operation; |
| 267 | co_return co_await file.Issue(operation, token, [&](aiocb& cb) -> IoResult { |
| 268 | cb.aio_buf = buffer.data(); |
| 269 | cb.aio_nbytes = buffer.size(); |
| 270 | cb.aio_offset = offset; |
| 271 | if (aio_read(&cb) < 0) |
| 272 | { |
| 273 | return {-errno, 0}; |
| 274 | } |
| 275 | |
| 276 | return {}; |
| 277 | }); |
| 278 | } |
| 279 | |
| 280 | Task<IoResult> WriteAsync(CoroutineIoIssuer& file, std::uint64_t offset, gsl::span<const gsl::byte> buffer, CancelToken& token) |
| 281 | { |
| 282 | CoroutineIoOperation operation; |
| 283 | co_return co_await file.Issue(operation, token, [&](aiocb& cb) -> IoResult { |
| 284 | cb.aio_buf = (volatile void*)buffer.data(); |
| 285 | cb.aio_nbytes = buffer.size(); |
| 286 | cb.aio_offset = offset; |
| 287 | if (aio_write(&cb) < 0) |
| 288 | { |
| 289 | return {-errno, 0}; |
| 290 | } |
| 291 | |
| 292 | return {}; |
| 293 | }); |
| 294 | } |
| 295 | |
| 296 | } // namespace p9fs |