James Kuszmaul | 38735e8 | 2019-12-07 16:42:06 -0800 | [diff] [blame^] | 1 | #include "aos/events/logging/logger.h" |
Austin Schuh | e309d2a | 2019-11-29 13:25:21 -0800 | [diff] [blame] | 2 | |
| 3 | #include <fcntl.h> |
| 4 | #include <sys/stat.h> |
| 5 | #include <sys/types.h> |
| 6 | #include <sys/uio.h> |
| 7 | #include <vector> |
| 8 | |
| 9 | #include "absl/strings/string_view.h" |
| 10 | #include "absl/types/span.h" |
| 11 | #include "aos/events/event_loop.h" |
James Kuszmaul | 38735e8 | 2019-12-07 16:42:06 -0800 | [diff] [blame^] | 12 | #include "aos/events/logging/logger_generated.h" |
Austin Schuh | e309d2a | 2019-11-29 13:25:21 -0800 | [diff] [blame] | 13 | #include "aos/flatbuffer_merge.h" |
| 14 | #include "aos/time/time.h" |
| 15 | #include "flatbuffers/flatbuffers.h" |
| 16 | |
| 17 | DEFINE_int32(flush_size, 1000000, |
| 18 | "Number of outstanding bytes to allow before flushing to disk."); |
| 19 | |
| 20 | namespace aos { |
| 21 | namespace logger { |
| 22 | |
| 23 | namespace chrono = std::chrono; |
| 24 | |
| 25 | DetachedBufferWriter::DetachedBufferWriter(absl::string_view filename) |
| 26 | : fd_(open(std::string(filename).c_str(), |
| 27 | O_RDWR | O_CLOEXEC | O_CREAT | O_EXCL, 0774)) { |
| 28 | PCHECK(fd_ != -1) << ": Failed to open " << filename; |
| 29 | } |
| 30 | |
| 31 | DetachedBufferWriter::~DetachedBufferWriter() { |
| 32 | Flush(); |
| 33 | PLOG_IF(ERROR, close(fd_) == -1) << " Failed to close logfile"; |
| 34 | } |
| 35 | |
| 36 | void DetachedBufferWriter::QueueSizedFlatbuffer( |
| 37 | flatbuffers::FlatBufferBuilder *fbb) { |
| 38 | QueueSizedFlatbuffer(fbb->Release()); |
| 39 | } |
| 40 | |
| 41 | void DetachedBufferWriter::QueueSizedFlatbuffer( |
| 42 | flatbuffers::DetachedBuffer &&buffer) { |
| 43 | queued_size_ += buffer.size(); |
| 44 | queue_.emplace_back(std::move(buffer)); |
| 45 | |
| 46 | if (queued_size_ > static_cast<size_t>(FLAGS_flush_size)) { |
| 47 | Flush(); |
| 48 | } |
| 49 | } |
| 50 | |
| 51 | void DetachedBufferWriter::Flush() { |
| 52 | if (queue_.size() == 0u) { |
| 53 | return; |
| 54 | } |
| 55 | iovec_.clear(); |
| 56 | iovec_.reserve(queue_.size()); |
| 57 | size_t counted_size = 0; |
| 58 | for (size_t i = 0; i < queue_.size(); ++i) { |
| 59 | struct iovec n; |
| 60 | n.iov_base = queue_[i].data(); |
| 61 | n.iov_len = queue_[i].size(); |
| 62 | counted_size += n.iov_len; |
| 63 | iovec_.emplace_back(std::move(n)); |
| 64 | } |
| 65 | CHECK_EQ(counted_size, queued_size_); |
| 66 | const ssize_t written = writev(fd_, iovec_.data(), iovec_.size()); |
| 67 | |
| 68 | PCHECK(written == static_cast<ssize_t>(queued_size_)) |
| 69 | << ": Wrote " << written << " expected " << queued_size_; |
| 70 | |
| 71 | queued_size_ = 0; |
| 72 | queue_.clear(); |
| 73 | // TODO(austin): Handle partial writes in some way other than crashing... |
| 74 | } |
| 75 | |
| 76 | Logger::Logger(DetachedBufferWriter *writer, EventLoop *event_loop, |
| 77 | std::chrono::milliseconds polling_period) |
| 78 | : event_loop_(event_loop), |
| 79 | writer_(writer), |
| 80 | timer_handler_(event_loop_->AddTimer([this]() { DoLogData(); })), |
| 81 | polling_period_(polling_period) { |
| 82 | for (const Channel *channel : *event_loop_->configuration()->channels()) { |
| 83 | FetcherStruct fs; |
| 84 | fs.fetcher = event_loop->MakeRawFetcher(channel); |
| 85 | fs.written = false; |
| 86 | fetchers_.emplace_back(std::move(fs)); |
| 87 | } |
| 88 | |
| 89 | // When things start, we want to log the header, then the most recent messages |
| 90 | // available on each fetcher to capture the previous state, then start |
| 91 | // polling. |
| 92 | event_loop_->OnRun([this, polling_period]() { |
| 93 | // Grab data from each channel right before we declare the log file started |
| 94 | // so we can capture the latest message on each channel. This lets us have |
| 95 | // non periodic messages with configuration that now get logged. |
| 96 | for (FetcherStruct &f : fetchers_) { |
| 97 | f.written = !f.fetcher->Fetch(); |
| 98 | } |
| 99 | |
| 100 | // We need to pick a point in time to declare the log file "started". This |
| 101 | // starts here. It needs to be after everything is fetched so that the |
| 102 | // fetchers are all pointed at the most recent message before the start |
| 103 | // time. |
| 104 | const monotonic_clock::time_point monotonic_now = |
| 105 | event_loop_->monotonic_now(); |
| 106 | const realtime_clock::time_point realtime_now = event_loop_->realtime_now(); |
| 107 | last_synchronized_time_ = monotonic_now; |
| 108 | |
| 109 | { |
| 110 | // Now write the header with this timestamp in it. |
| 111 | flatbuffers::FlatBufferBuilder fbb; |
| 112 | fbb.ForceDefaults(1); |
| 113 | |
| 114 | flatbuffers::Offset<aos::Configuration> configuration_offset = |
| 115 | CopyFlatBuffer(event_loop_->configuration(), &fbb); |
| 116 | |
| 117 | aos::logger::LogFileHeader::Builder log_file_header_builder(fbb); |
| 118 | |
| 119 | log_file_header_builder.add_configuration(configuration_offset); |
| 120 | // The worst case theoretical out of order is the polling period times 2. |
| 121 | // One message could get logged right after the boundary, but be for right |
| 122 | // before the next boundary. And the reverse could happen for another |
| 123 | // message. Report back 3x to be extra safe, and because the cost isn't |
| 124 | // huge on the read side. |
| 125 | log_file_header_builder.add_max_out_of_order_duration( |
| 126 | std::chrono::duration_cast<std::chrono::nanoseconds>(3 * |
| 127 | polling_period) |
| 128 | .count()); |
| 129 | |
| 130 | log_file_header_builder.add_monotonic_sent_time( |
| 131 | std::chrono::duration_cast<std::chrono::nanoseconds>( |
| 132 | monotonic_now.time_since_epoch()) |
| 133 | .count()); |
| 134 | log_file_header_builder.add_realtime_sent_time( |
| 135 | std::chrono::duration_cast<std::chrono::nanoseconds>( |
| 136 | realtime_now.time_since_epoch()) |
| 137 | .count()); |
| 138 | |
| 139 | fbb.FinishSizePrefixed(log_file_header_builder.Finish()); |
| 140 | writer_->QueueSizedFlatbuffer(&fbb); |
| 141 | } |
| 142 | |
| 143 | timer_handler_->Setup(event_loop_->monotonic_now() + polling_period, |
| 144 | polling_period); |
| 145 | }); |
| 146 | } |
| 147 | |
| 148 | void Logger::DoLogData() { |
| 149 | // We want to guarentee that messages aren't out of order by more than |
| 150 | // max_out_of_order_duration. To do this, we need sync points. Every write |
| 151 | // cycle should be a sync point. |
| 152 | const monotonic_clock::time_point monotonic_now = monotonic_clock::now(); |
| 153 | |
| 154 | do { |
| 155 | // Move the sync point up by at most polling_period. This forces one sync |
| 156 | // per iteration, even if it is small. |
| 157 | last_synchronized_time_ = |
| 158 | std::min(last_synchronized_time_ + polling_period_, monotonic_now); |
| 159 | size_t channel_index = 0; |
| 160 | // Write each channel to disk, one at a time. |
| 161 | for (FetcherStruct &f : fetchers_) { |
| 162 | while (true) { |
| 163 | if (f.fetcher.get() == nullptr) { |
| 164 | if (!f.fetcher->FetchNext()) { |
| 165 | VLOG(1) << "No new data on " |
| 166 | << FlatbufferToJson(f.fetcher->channel()); |
| 167 | break; |
| 168 | } else { |
| 169 | f.written = false; |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | if (f.written) { |
| 174 | if (!f.fetcher->FetchNext()) { |
| 175 | VLOG(1) << "No new data on " |
| 176 | << FlatbufferToJson(f.fetcher->channel()); |
| 177 | break; |
| 178 | } else { |
| 179 | f.written = false; |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | if (!f.written && f.fetcher->context().monotonic_sent_time < |
| 184 | last_synchronized_time_) { |
| 185 | // Write! |
| 186 | flatbuffers::FlatBufferBuilder fbb(f.fetcher->context().size + |
| 187 | max_header_size_); |
| 188 | fbb.ForceDefaults(1); |
| 189 | |
| 190 | flatbuffers::Offset<flatbuffers::Vector<uint8_t>> data_offset = |
| 191 | fbb.CreateVector( |
| 192 | static_cast<uint8_t *>(f.fetcher->context().data), |
| 193 | f.fetcher->context().size); |
| 194 | |
| 195 | VLOG(1) << "Writing data for channel " |
| 196 | << FlatbufferToJson(f.fetcher->channel()); |
| 197 | |
| 198 | MessageHeader::Builder message_header_builder(fbb); |
| 199 | message_header_builder.add_channel_index(channel_index); |
| 200 | message_header_builder.add_monotonic_sent_time( |
| 201 | f.fetcher->context() |
| 202 | .monotonic_sent_time.time_since_epoch() |
| 203 | .count()); |
| 204 | message_header_builder.add_realtime_sent_time( |
| 205 | f.fetcher->context() |
| 206 | .realtime_sent_time.time_since_epoch() |
| 207 | .count()); |
| 208 | |
| 209 | message_header_builder.add_queue_index( |
| 210 | f.fetcher->context().queue_index); |
| 211 | |
| 212 | message_header_builder.add_data(data_offset); |
| 213 | |
| 214 | fbb.FinishSizePrefixed(message_header_builder.Finish()); |
| 215 | max_header_size_ = std::max( |
| 216 | max_header_size_, fbb.GetSize() - f.fetcher->context().size); |
| 217 | writer_->QueueSizedFlatbuffer(&fbb); |
| 218 | |
| 219 | f.written = true; |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | ++channel_index; |
| 224 | } |
| 225 | |
| 226 | CHECK_EQ(channel_index, fetchers_.size()); |
| 227 | |
| 228 | // If we missed cycles, we could be pretty far behind. Spin until we are |
| 229 | // caught up. |
| 230 | } while (last_synchronized_time_ + polling_period_ < monotonic_now); |
| 231 | |
| 232 | writer_->Flush(); |
| 233 | } |
| 234 | |
| 235 | LogReader::LogReader(absl::string_view filename) |
| 236 | : fd_(open(std::string(filename).c_str(), O_RDONLY | O_CLOEXEC)) { |
| 237 | PCHECK(fd_ != -1) << ": Failed to open " << filename; |
| 238 | |
| 239 | // Make sure we have enough to read the size. |
| 240 | absl::Span<const uint8_t> config_data = ReadMessage(); |
| 241 | |
| 242 | // Make sure something was read. |
| 243 | CHECK(config_data != absl::Span<const uint8_t>()); |
| 244 | |
| 245 | // And copy the config so we have it forever. |
| 246 | configuration_ = std::vector<uint8_t>(config_data.begin(), config_data.end()); |
| 247 | |
| 248 | max_out_of_order_duration_ = std::chrono::nanoseconds( |
| 249 | flatbuffers::GetSizePrefixedRoot<LogFileHeader>(configuration_.data()) |
| 250 | ->max_out_of_order_duration()); |
| 251 | |
| 252 | channels_.resize(configuration()->channels()->size()); |
| 253 | |
| 254 | QueueMessages(); |
| 255 | } |
| 256 | |
| 257 | bool LogReader::ReadBlock() { |
| 258 | if (end_of_file_) { |
| 259 | return false; |
| 260 | } |
| 261 | |
| 262 | // Appends 256k. This is enough that the read call is efficient. We don't |
| 263 | // want to spend too much time reading small chunks because the syscalls for |
| 264 | // that will be expensive. |
| 265 | constexpr size_t kReadSize = 256 * 1024; |
| 266 | |
| 267 | // Strip off any unused data at the front. |
| 268 | if (consumed_data_ != 0) { |
| 269 | data_.erase(data_.begin(), data_.begin() + consumed_data_); |
| 270 | consumed_data_ = 0; |
| 271 | } |
| 272 | |
| 273 | const size_t starting_size = data_.size(); |
| 274 | |
| 275 | // This should automatically grow the backing store. It won't shrink if we |
| 276 | // get a small chunk later. This reduces allocations when we want to append |
| 277 | // more data. |
| 278 | data_.resize(data_.size() + kReadSize); |
| 279 | |
| 280 | ssize_t count = read(fd_, &data_[starting_size], kReadSize); |
| 281 | data_.resize(starting_size + std::max(count, static_cast<ssize_t>(0))); |
| 282 | if (count == 0) { |
| 283 | end_of_file_ = true; |
| 284 | return false; |
| 285 | } |
| 286 | PCHECK(count > 0); |
| 287 | |
| 288 | return true; |
| 289 | } |
| 290 | |
| 291 | bool LogReader::MessageAvailable() { |
| 292 | // Are we big enough to read the size? |
| 293 | if (data_.size() - consumed_data_ < sizeof(flatbuffers::uoffset_t)) { |
| 294 | return false; |
| 295 | } |
| 296 | |
| 297 | // Then, are we big enough to read the full message? |
| 298 | const size_t data_size = |
| 299 | flatbuffers::GetPrefixedSize(data_.data() + consumed_data_) + |
| 300 | sizeof(flatbuffers::uoffset_t); |
| 301 | if (data_.size() < consumed_data_ + data_size) { |
| 302 | return false; |
| 303 | } |
| 304 | |
| 305 | return true; |
| 306 | } |
| 307 | |
| 308 | absl::Span<const uint8_t> LogReader::ReadMessage() { |
| 309 | // Make sure we have enough for the size. |
| 310 | if (data_.size() - consumed_data_ < sizeof(flatbuffers::uoffset_t)) { |
| 311 | if (!ReadBlock()) { |
| 312 | return absl::Span<const uint8_t>(); |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | // Now make sure we have enough for the message. |
| 317 | const size_t data_size = |
| 318 | flatbuffers::GetPrefixedSize(data_.data() + consumed_data_) + |
| 319 | sizeof(flatbuffers::uoffset_t); |
| 320 | while (data_.size() < consumed_data_ + data_size) { |
| 321 | if (!ReadBlock()) { |
| 322 | return absl::Span<const uint8_t>(); |
| 323 | } |
| 324 | } |
| 325 | |
| 326 | // And return it, consuming the data. |
| 327 | const uint8_t *data_ptr = data_.data() + consumed_data_; |
| 328 | |
| 329 | consumed_data_ += data_size; |
| 330 | |
| 331 | return absl::Span<const uint8_t>(data_ptr, data_size); |
| 332 | } |
| 333 | |
| 334 | void LogReader::QueueMessages() { |
| 335 | while (true) { |
| 336 | // Don't queue if we have enough data already. |
| 337 | // When a log file starts, there should be a message from each channel. |
| 338 | // Those messages might be very old. Make sure to read a chunk past the |
| 339 | // starting time. |
| 340 | if (channel_heap_.size() > 0 && |
| 341 | newest_timestamp_ > |
| 342 | std::max(oldest_message().first, monotonic_start_time()) + |
| 343 | max_out_of_order_duration_) { |
| 344 | break; |
| 345 | } |
| 346 | |
| 347 | absl::Span<const uint8_t> msg_data = ReadMessage(); |
| 348 | if (msg_data == absl::Span<const uint8_t>()) { |
| 349 | break; |
| 350 | } |
| 351 | |
| 352 | FlatbufferVector<MessageHeader> msg(std::vector<uint8_t>( |
| 353 | msg_data.begin() + sizeof(flatbuffers::uoffset_t), msg_data.end())); |
| 354 | |
| 355 | EmplaceDataBack(std::move(msg)); |
| 356 | } |
| 357 | |
| 358 | queue_data_time_ = newest_timestamp_ - max_out_of_order_duration_; |
| 359 | } |
| 360 | |
| 361 | const Configuration *LogReader::configuration() { |
| 362 | return flatbuffers::GetSizePrefixedRoot<LogFileHeader>(configuration_.data()) |
| 363 | ->configuration(); |
| 364 | } |
| 365 | |
| 366 | monotonic_clock::time_point LogReader::monotonic_start_time() { |
| 367 | return monotonic_clock::time_point(std::chrono::nanoseconds( |
| 368 | flatbuffers::GetSizePrefixedRoot<LogFileHeader>(configuration_.data()) |
| 369 | ->monotonic_sent_time())); |
| 370 | } |
| 371 | |
| 372 | realtime_clock::time_point LogReader::realtime_start_time() { |
| 373 | return realtime_clock::time_point(std::chrono::nanoseconds( |
| 374 | flatbuffers::GetSizePrefixedRoot<LogFileHeader>(configuration_.data()) |
| 375 | ->realtime_sent_time())); |
| 376 | } |
| 377 | |
| 378 | void LogReader::Register(EventLoop *event_loop) { |
| 379 | event_loop_ = event_loop; |
| 380 | |
Austin Schuh | 39788ff | 2019-12-01 18:22:57 -0800 | [diff] [blame] | 381 | // Otherwise we replay the timing report and try to resend it... |
| 382 | event_loop_->SkipTimingReport(); |
| 383 | |
Austin Schuh | e309d2a | 2019-11-29 13:25:21 -0800 | [diff] [blame] | 384 | for (size_t i = 0; i < channels_.size(); ++i) { |
| 385 | CHECK_EQ(configuration()->channels()->Get(i)->name(), |
| 386 | event_loop_->configuration()->channels()->Get(i)->name()); |
| 387 | CHECK_EQ(configuration()->channels()->Get(i)->type(), |
| 388 | event_loop_->configuration()->channels()->Get(i)->type()); |
| 389 | |
| 390 | channels_[i].raw_sender = event_loop_->MakeRawSender( |
| 391 | event_loop_->configuration()->channels()->Get(i)); |
| 392 | } |
| 393 | |
| 394 | timer_handler_ = event_loop_->AddTimer([this]() { |
| 395 | std::pair<monotonic_clock::time_point, int> oldest_channel_index = |
| 396 | PopOldestChannel(); |
| 397 | const monotonic_clock::time_point monotonic_now = |
Austin Schuh | 39788ff | 2019-12-01 18:22:57 -0800 | [diff] [blame] | 398 | event_loop_->context().monotonic_sent_time; |
Austin Schuh | e309d2a | 2019-11-29 13:25:21 -0800 | [diff] [blame] | 399 | CHECK(monotonic_now == oldest_channel_index.first) |
| 400 | << ": Now " << monotonic_now.time_since_epoch().count() |
| 401 | << " trying to send " |
| 402 | << oldest_channel_index.first.time_since_epoch().count(); |
| 403 | |
| 404 | struct LogReader::ChannelData &channel = |
| 405 | channels_[oldest_channel_index.second]; |
| 406 | |
| 407 | FlatbufferVector<MessageHeader> front = std::move(channel.front()); |
| 408 | |
| 409 | CHECK(front.message().data() != nullptr); |
| 410 | if (oldest_channel_index.first > monotonic_start_time()) { |
| 411 | channel.raw_sender->Send(front.message().data()->Data(), |
| 412 | front.message().data()->size()); |
| 413 | } else { |
| 414 | LOG(WARNING) << "Not sending data from before the start of the log file. " |
| 415 | << oldest_channel_index.first.time_since_epoch().count() |
| 416 | << " start " |
| 417 | << monotonic_start_time().time_since_epoch().count() << " " |
| 418 | << FlatbufferToJson(front); |
| 419 | } |
| 420 | channel.data.pop_front(); |
| 421 | |
| 422 | // Re-push it and update the oldest timestamp. |
| 423 | if (channel.data.size() != 0) { |
| 424 | const monotonic_clock::time_point timestamp = monotonic_clock::time_point( |
| 425 | chrono::nanoseconds(channel.front().message().monotonic_sent_time())); |
| 426 | PushChannelHeap(timestamp, oldest_channel_index.second); |
| 427 | channel.oldest_timestamp = timestamp; |
| 428 | } else { |
| 429 | channel.oldest_timestamp = monotonic_clock::min_time; |
| 430 | } |
| 431 | |
| 432 | if (monotonic_now > queue_data_time_) { |
| 433 | QueueMessages(); |
| 434 | } |
| 435 | |
| 436 | if (channel_heap_.size() != 0) { |
| 437 | timer_handler_->Setup(oldest_message().first); |
| 438 | } |
| 439 | }); |
| 440 | |
| 441 | if (channel_heap_.size() > 0u) { |
| 442 | event_loop_->OnRun( |
| 443 | [this]() { timer_handler_->Setup(oldest_message().first); }); |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | void LogReader::Deregister() { |
| 448 | for (size_t i = 0; i < channels_.size(); ++i) { |
| 449 | channels_[i].raw_sender.reset(); |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | void LogReader::EmplaceDataBack(FlatbufferVector<MessageHeader> &&new_data) { |
| 454 | const monotonic_clock::time_point timestamp = monotonic_clock::time_point( |
| 455 | chrono::nanoseconds(new_data.message().monotonic_sent_time())); |
| 456 | const size_t channel_index = new_data.message().channel_index(); |
| 457 | CHECK_LT(channel_index, channels_.size()); |
| 458 | newest_timestamp_ = std::max(newest_timestamp_, timestamp); |
| 459 | if (channels_[channel_index].data.size() == 0) { |
| 460 | channels_[channel_index].oldest_timestamp = timestamp; |
| 461 | PushChannelHeap(timestamp, channel_index); |
| 462 | } |
| 463 | channels_[channel_index].data.emplace_back(std::move(new_data)); |
| 464 | } |
| 465 | |
| 466 | namespace { |
| 467 | |
| 468 | bool ChannelHeapCompare( |
| 469 | const std::pair<monotonic_clock::time_point, int> first, |
| 470 | const std::pair<monotonic_clock::time_point, int> second) { |
| 471 | if (first.first > second.first) { |
| 472 | return true; |
| 473 | } else if (first.first == second.first) { |
| 474 | return first.second > second.second; |
| 475 | } else { |
| 476 | return false; |
| 477 | } |
| 478 | } |
| 479 | |
| 480 | } // namespace |
| 481 | |
| 482 | void LogReader::PushChannelHeap(monotonic_clock::time_point timestamp, |
| 483 | int channel_index) { |
| 484 | channel_heap_.push_back(std::make_pair(timestamp, channel_index)); |
| 485 | |
| 486 | // The default sort puts the newest message first. Use a custom comparator to |
| 487 | // put the oldest message first. |
| 488 | std::push_heap(channel_heap_.begin(), channel_heap_.end(), |
| 489 | ChannelHeapCompare); |
| 490 | } |
| 491 | |
| 492 | std::pair<monotonic_clock::time_point, int> LogReader::PopOldestChannel() { |
| 493 | std::pair<monotonic_clock::time_point, int> result = channel_heap_.front(); |
| 494 | std::pop_heap(channel_heap_.begin(), channel_heap_.end(), |
| 495 | &ChannelHeapCompare); |
| 496 | channel_heap_.pop_back(); |
| 497 | return result; |
| 498 | } |
| 499 | |
| 500 | } // namespace logger |
| 501 | } // namespace aos |