blob: 330c78ea68cb393a5aa777b080c05499dc77b2c4 [file] [log] [blame]
Austin Schuha36c8902019-12-30 18:07:15 -08001#include "aos/events/logging/logfile_utils.h"
2
3#include <fcntl.h>
Austin Schuha36c8902019-12-30 18:07:15 -08004#include <sys/stat.h>
5#include <sys/types.h>
6#include <sys/uio.h>
7
Brian Silvermanf51499a2020-09-21 12:49:08 -07008#include <algorithm>
9#include <climits>
Austin Schuha36c8902019-12-30 18:07:15 -080010
Austin Schuhe4fca832020-03-07 16:58:53 -080011#include "absl/strings/escaping.h"
Austin Schuh05b70472020-01-01 17:11:17 -080012#include "aos/configuration.h"
Austin Schuhfa895892020-01-07 20:07:41 -080013#include "aos/flatbuffer_merge.h"
Austin Schuh6f3babe2020-01-26 20:34:50 -080014#include "aos/util/file.h"
Austin Schuha36c8902019-12-30 18:07:15 -080015#include "flatbuffers/flatbuffers.h"
Austin Schuh05b70472020-01-01 17:11:17 -080016#include "gflags/gflags.h"
17#include "glog/logging.h"
Austin Schuha36c8902019-12-30 18:07:15 -080018
Brian Silvermanf59fe3f2020-09-22 21:04:09 -070019#if defined(__x86_64__)
20#define ENABLE_LZMA 1
21#elif defined(__aarch64__)
22#define ENABLE_LZMA 1
23#else
24#define ENABLE_LZMA 0
25#endif
26
27#if ENABLE_LZMA
28#include "aos/events/logging/lzma_encoder.h"
29#endif
30
Austin Schuh7fbf5a72020-09-21 16:28:13 -070031DEFINE_int32(flush_size, 128000,
Austin Schuha36c8902019-12-30 18:07:15 -080032 "Number of outstanding bytes to allow before flushing to disk.");
33
Brian Silvermanf51499a2020-09-21 12:49:08 -070034namespace aos::logger {
Austin Schuha36c8902019-12-30 18:07:15 -080035
Austin Schuh05b70472020-01-01 17:11:17 -080036namespace chrono = std::chrono;
37
Brian Silvermanf51499a2020-09-21 12:49:08 -070038DetachedBufferWriter::DetachedBufferWriter(
39 std::string_view filename, std::unique_ptr<DetachedBufferEncoder> encoder)
40 : filename_(filename), encoder_(std::move(encoder)) {
Brian Silvermana9f2ec92020-10-06 18:00:53 -070041 if (!util::MkdirPIfSpace(filename, 0777)) {
42 ran_out_of_space_ = true;
43 } else {
44 fd_ = open(std::string(filename).c_str(),
45 O_RDWR | O_CLOEXEC | O_CREAT | O_EXCL, 0774);
46 if (fd_ == -1 && errno == ENOSPC) {
47 ran_out_of_space_ = true;
48 } else {
49 PCHECK(fd_ != -1) << ": Failed to open " << filename << " for writing";
50 VLOG(1) << "Opened " << filename << " for writing";
51 }
52 }
Austin Schuha36c8902019-12-30 18:07:15 -080053}
54
55DetachedBufferWriter::~DetachedBufferWriter() {
Brian Silverman0465fcf2020-09-24 00:29:18 -070056 Close();
57 if (ran_out_of_space_) {
58 CHECK(acknowledge_ran_out_of_space_)
59 << ": Unacknowledged out of disk space, log file was not completed";
Brian Silvermanf51499a2020-09-21 12:49:08 -070060 }
Austin Schuh2f8fd752020-09-01 22:38:28 -070061}
62
Brian Silvermand90905f2020-09-23 14:42:56 -070063DetachedBufferWriter::DetachedBufferWriter(DetachedBufferWriter &&other) {
Austin Schuh2f8fd752020-09-01 22:38:28 -070064 *this = std::move(other);
65}
66
Brian Silverman87ac0402020-09-17 14:47:01 -070067// When other is destroyed "soon" (which it should be because we're getting an
68// rvalue reference to it), it will flush etc all the data we have queued up
69// (because that data will then be its data).
Austin Schuh2f8fd752020-09-01 22:38:28 -070070DetachedBufferWriter &DetachedBufferWriter::operator=(
71 DetachedBufferWriter &&other) {
Austin Schuh2f8fd752020-09-01 22:38:28 -070072 std::swap(filename_, other.filename_);
Brian Silvermanf51499a2020-09-21 12:49:08 -070073 std::swap(encoder_, other.encoder_);
Austin Schuh2f8fd752020-09-01 22:38:28 -070074 std::swap(fd_, other.fd_);
Brian Silverman0465fcf2020-09-24 00:29:18 -070075 std::swap(ran_out_of_space_, other.ran_out_of_space_);
76 std::swap(acknowledge_ran_out_of_space_, other.acknowledge_ran_out_of_space_);
Austin Schuh2f8fd752020-09-01 22:38:28 -070077 std::swap(iovec_, other.iovec_);
Brian Silvermanf51499a2020-09-21 12:49:08 -070078 std::swap(max_write_time_, other.max_write_time_);
79 std::swap(max_write_time_bytes_, other.max_write_time_bytes_);
80 std::swap(max_write_time_messages_, other.max_write_time_messages_);
81 std::swap(total_write_time_, other.total_write_time_);
82 std::swap(total_write_count_, other.total_write_count_);
83 std::swap(total_write_messages_, other.total_write_messages_);
84 std::swap(total_write_bytes_, other.total_write_bytes_);
Austin Schuh2f8fd752020-09-01 22:38:28 -070085 return *this;
Austin Schuha36c8902019-12-30 18:07:15 -080086}
87
Brian Silvermanf51499a2020-09-21 12:49:08 -070088void DetachedBufferWriter::QueueSpan(absl::Span<const uint8_t> span) {
Brian Silvermana9f2ec92020-10-06 18:00:53 -070089 if (ran_out_of_space_) {
90 // We don't want any later data to be written after space becomes
91 // available, so refuse to write anything more once we've dropped data
92 // because we ran out of space.
93 VLOG(1) << "Ignoring span: " << span.size();
94 return;
95 }
96
Brian Silvermanf51499a2020-09-21 12:49:08 -070097 if (encoder_->may_bypass() && span.size() > 4096u) {
98 // Over this threshold, we'll assume it's cheaper to add an extra
99 // syscall to write the data immediately instead of copying it to
100 // enqueue.
Austin Schuha36c8902019-12-30 18:07:15 -0800101
Brian Silvermanf51499a2020-09-21 12:49:08 -0700102 // First, flush everything.
103 while (encoder_->queue_size() > 0u) {
104 Flush();
105 }
Austin Schuhde031b72020-01-10 19:34:41 -0800106
Brian Silvermanf51499a2020-09-21 12:49:08 -0700107 // Then, write it directly.
108 const auto start = aos::monotonic_clock::now();
109 const ssize_t written = write(fd_, span.data(), span.size());
110 const auto end = aos::monotonic_clock::now();
Brian Silverman0465fcf2020-09-24 00:29:18 -0700111 HandleWriteReturn(written, span.size());
Brian Silvermanf51499a2020-09-21 12:49:08 -0700112 UpdateStatsForWrite(end - start, written, 1);
113 } else {
114 encoder_->Encode(CopySpanAsDetachedBuffer(span));
Austin Schuha36c8902019-12-30 18:07:15 -0800115 }
Brian Silvermanf51499a2020-09-21 12:49:08 -0700116
117 FlushAtThreshold();
Austin Schuha36c8902019-12-30 18:07:15 -0800118}
119
Brian Silverman0465fcf2020-09-24 00:29:18 -0700120void DetachedBufferWriter::Close() {
121 if (fd_ == -1) {
122 return;
123 }
124 encoder_->Finish();
125 while (encoder_->queue_size() > 0) {
126 Flush();
127 }
128 if (close(fd_) == -1) {
129 if (errno == ENOSPC) {
130 ran_out_of_space_ = true;
131 } else {
132 PLOG(ERROR) << "Closing log file failed";
133 }
134 }
135 fd_ = -1;
136 VLOG(1) << "Closed " << filename_;
137}
138
Austin Schuha36c8902019-12-30 18:07:15 -0800139void DetachedBufferWriter::Flush() {
Brian Silvermanf51499a2020-09-21 12:49:08 -0700140 const auto queue = encoder_->queue();
141 if (queue.empty()) {
Austin Schuha36c8902019-12-30 18:07:15 -0800142 return;
143 }
Brian Silverman0465fcf2020-09-24 00:29:18 -0700144 if (ran_out_of_space_) {
145 // We don't want any later data to be written after space becomes available,
146 // so refuse to write anything more once we've dropped data because we ran
147 // out of space.
148 VLOG(1) << "Ignoring queue: " << queue.size();
149 encoder_->Clear(queue.size());
150 return;
151 }
Brian Silvermanf51499a2020-09-21 12:49:08 -0700152
Austin Schuha36c8902019-12-30 18:07:15 -0800153 iovec_.clear();
Brian Silvermanf51499a2020-09-21 12:49:08 -0700154 const size_t iovec_size = std::min<size_t>(queue.size(), IOV_MAX);
155 iovec_.resize(iovec_size);
Austin Schuha36c8902019-12-30 18:07:15 -0800156 size_t counted_size = 0;
Brian Silvermanf51499a2020-09-21 12:49:08 -0700157 for (size_t i = 0; i < iovec_size; ++i) {
158 iovec_[i].iov_base = const_cast<uint8_t *>(queue[i].data());
159 iovec_[i].iov_len = queue[i].size();
160 counted_size += iovec_[i].iov_len;
Austin Schuha36c8902019-12-30 18:07:15 -0800161 }
Brian Silvermanf51499a2020-09-21 12:49:08 -0700162
163 const auto start = aos::monotonic_clock::now();
Austin Schuha36c8902019-12-30 18:07:15 -0800164 const ssize_t written = writev(fd_, iovec_.data(), iovec_.size());
Brian Silvermanf51499a2020-09-21 12:49:08 -0700165 const auto end = aos::monotonic_clock::now();
Brian Silverman0465fcf2020-09-24 00:29:18 -0700166 HandleWriteReturn(written, counted_size);
Brian Silvermanf51499a2020-09-21 12:49:08 -0700167
168 encoder_->Clear(iovec_size);
169
170 UpdateStatsForWrite(end - start, written, iovec_size);
171}
172
Brian Silverman0465fcf2020-09-24 00:29:18 -0700173void DetachedBufferWriter::HandleWriteReturn(ssize_t write_return,
174 size_t write_size) {
175 if (write_return == -1 && errno == ENOSPC) {
176 ran_out_of_space_ = true;
177 return;
178 }
179 PCHECK(write_return >= 0) << ": write failed";
180 if (write_return < static_cast<ssize_t>(write_size)) {
181 // Sometimes this happens instead of ENOSPC. On a real filesystem, this
182 // never seems to happen in any other case. If we ever want to log to a
183 // socket, this will happen more often. However, until we get there, we'll
184 // just assume it means we ran out of space.
185 ran_out_of_space_ = true;
186 return;
187 }
188}
189
Brian Silvermanf51499a2020-09-21 12:49:08 -0700190void DetachedBufferWriter::UpdateStatsForWrite(
191 aos::monotonic_clock::duration duration, ssize_t written, int iovec_size) {
192 if (duration > max_write_time_) {
193 max_write_time_ = duration;
194 max_write_time_bytes_ = written;
195 max_write_time_messages_ = iovec_size;
196 }
197 total_write_time_ += duration;
198 ++total_write_count_;
199 total_write_messages_ += iovec_size;
200 total_write_bytes_ += written;
201}
202
203void DetachedBufferWriter::FlushAtThreshold() {
204 // Flush if we are at the max number of iovs per writev, because there's no
205 // point queueing up any more data in memory. Also flush once we have enough
206 // data queued up.
207 while (encoder_->queued_bytes() > static_cast<size_t>(FLAGS_flush_size) ||
208 encoder_->queue_size() >= IOV_MAX) {
209 Flush();
210 }
Austin Schuha36c8902019-12-30 18:07:15 -0800211}
212
213flatbuffers::Offset<MessageHeader> PackMessage(
214 flatbuffers::FlatBufferBuilder *fbb, const Context &context,
215 int channel_index, LogType log_type) {
216 flatbuffers::Offset<flatbuffers::Vector<uint8_t>> data_offset;
217
218 switch (log_type) {
219 case LogType::kLogMessage:
220 case LogType::kLogMessageAndDeliveryTime:
Austin Schuh6f3babe2020-01-26 20:34:50 -0800221 case LogType::kLogRemoteMessage:
Brian Silvermaneaa41d62020-07-08 19:47:35 -0700222 data_offset = fbb->CreateVector(
223 static_cast<const uint8_t *>(context.data), context.size);
Austin Schuha36c8902019-12-30 18:07:15 -0800224 break;
225
226 case LogType::kLogDeliveryTimeOnly:
227 break;
228 }
229
230 MessageHeader::Builder message_header_builder(*fbb);
231 message_header_builder.add_channel_index(channel_index);
Austin Schuh6f3babe2020-01-26 20:34:50 -0800232
233 switch (log_type) {
234 case LogType::kLogRemoteMessage:
235 message_header_builder.add_queue_index(context.remote_queue_index);
236 message_header_builder.add_monotonic_sent_time(
237 context.monotonic_remote_time.time_since_epoch().count());
238 message_header_builder.add_realtime_sent_time(
239 context.realtime_remote_time.time_since_epoch().count());
240 break;
241
242 case LogType::kLogMessage:
243 case LogType::kLogMessageAndDeliveryTime:
244 case LogType::kLogDeliveryTimeOnly:
245 message_header_builder.add_queue_index(context.queue_index);
246 message_header_builder.add_monotonic_sent_time(
247 context.monotonic_event_time.time_since_epoch().count());
248 message_header_builder.add_realtime_sent_time(
249 context.realtime_event_time.time_since_epoch().count());
250 break;
251 }
Austin Schuha36c8902019-12-30 18:07:15 -0800252
253 switch (log_type) {
254 case LogType::kLogMessage:
Austin Schuh6f3babe2020-01-26 20:34:50 -0800255 case LogType::kLogRemoteMessage:
Austin Schuha36c8902019-12-30 18:07:15 -0800256 message_header_builder.add_data(data_offset);
257 break;
258
259 case LogType::kLogMessageAndDeliveryTime:
260 message_header_builder.add_data(data_offset);
261 [[fallthrough]];
262
263 case LogType::kLogDeliveryTimeOnly:
264 message_header_builder.add_monotonic_remote_time(
265 context.monotonic_remote_time.time_since_epoch().count());
266 message_header_builder.add_realtime_remote_time(
267 context.realtime_remote_time.time_since_epoch().count());
268 message_header_builder.add_remote_queue_index(context.remote_queue_index);
269 break;
270 }
271
272 return message_header_builder.Finish();
273}
274
Brian Silvermanf51499a2020-09-21 12:49:08 -0700275SpanReader::SpanReader(std::string_view filename) : filename_(filename) {
Brian Silvermanf59fe3f2020-09-22 21:04:09 -0700276 static const std::string_view kXz = ".xz";
277 if (filename.substr(filename.size() - kXz.size()) == kXz) {
278#if ENABLE_LZMA
279 decoder_ = std::make_unique<LzmaDecoder>(filename);
280#else
281 LOG(FATAL) << "Reading xz-compressed files not supported on this platform";
282#endif
283 } else {
284 decoder_ = std::make_unique<DummyDecoder>(filename);
285 }
Austin Schuh05b70472020-01-01 17:11:17 -0800286}
287
288absl::Span<const uint8_t> SpanReader::ReadMessage() {
289 // Make sure we have enough for the size.
290 if (data_.size() - consumed_data_ < sizeof(flatbuffers::uoffset_t)) {
291 if (!ReadBlock()) {
292 return absl::Span<const uint8_t>();
293 }
294 }
295
296 // Now make sure we have enough for the message.
297 const size_t data_size =
298 flatbuffers::GetPrefixedSize(data_.data() + consumed_data_) +
299 sizeof(flatbuffers::uoffset_t);
Austin Schuhe4fca832020-03-07 16:58:53 -0800300 if (data_size == sizeof(flatbuffers::uoffset_t)) {
301 LOG(ERROR) << "Size of data is zero. Log file end is corrupted, skipping.";
302 LOG(ERROR) << " Rest of log file is "
303 << absl::BytesToHexString(std::string_view(
304 reinterpret_cast<const char *>(data_.data() +
305 consumed_data_),
306 data_.size() - consumed_data_));
307 return absl::Span<const uint8_t>();
308 }
Austin Schuh05b70472020-01-01 17:11:17 -0800309 while (data_.size() < consumed_data_ + data_size) {
310 if (!ReadBlock()) {
311 return absl::Span<const uint8_t>();
312 }
313 }
314
315 // And return it, consuming the data.
316 const uint8_t *data_ptr = data_.data() + consumed_data_;
317
318 consumed_data_ += data_size;
319
320 return absl::Span<const uint8_t>(data_ptr, data_size);
321}
322
Austin Schuh05b70472020-01-01 17:11:17 -0800323bool SpanReader::ReadBlock() {
Brian Silvermanf51499a2020-09-21 12:49:08 -0700324 // This is the amount of data we grab at a time. Doing larger chunks minimizes
325 // syscalls and helps decompressors batch things more efficiently.
Austin Schuh05b70472020-01-01 17:11:17 -0800326 constexpr size_t kReadSize = 256 * 1024;
327
328 // Strip off any unused data at the front.
329 if (consumed_data_ != 0) {
Brian Silvermanf51499a2020-09-21 12:49:08 -0700330 data_.erase_front(consumed_data_);
Austin Schuh05b70472020-01-01 17:11:17 -0800331 consumed_data_ = 0;
332 }
333
334 const size_t starting_size = data_.size();
335
336 // This should automatically grow the backing store. It won't shrink if we
337 // get a small chunk later. This reduces allocations when we want to append
338 // more data.
Brian Silvermanf51499a2020-09-21 12:49:08 -0700339 data_.resize(starting_size + kReadSize);
Austin Schuh05b70472020-01-01 17:11:17 -0800340
Brian Silvermanf51499a2020-09-21 12:49:08 -0700341 const size_t count =
342 decoder_->Read(data_.begin() + starting_size, data_.end());
343 data_.resize(starting_size + count);
Austin Schuh05b70472020-01-01 17:11:17 -0800344 if (count == 0) {
Austin Schuh05b70472020-01-01 17:11:17 -0800345 return false;
346 }
Austin Schuh05b70472020-01-01 17:11:17 -0800347
348 return true;
349}
350
Austin Schuhadd6eb32020-11-09 21:24:26 -0800351std::optional<SizePrefixedFlatbufferVector<LogFileHeader>> ReadHeader(
Austin Schuh3bd4c402020-11-06 18:19:06 -0800352 std::string_view filename) {
Austin Schuh6f3babe2020-01-26 20:34:50 -0800353 SpanReader span_reader(filename);
Austin Schuh6f3babe2020-01-26 20:34:50 -0800354 absl::Span<const uint8_t> config_data = span_reader.ReadMessage();
355
356 // Make sure something was read.
Austin Schuh3bd4c402020-11-06 18:19:06 -0800357 if (config_data == absl::Span<const uint8_t>()) {
358 return std::nullopt;
359 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800360
Austin Schuh5212cad2020-09-09 23:12:09 -0700361 // And copy the config so we have it forever, removing the size prefix.
Brian Silverman354697a2020-09-22 21:06:32 -0700362 ResizeableBuffer data;
Austin Schuhadd6eb32020-11-09 21:24:26 -0800363 data.resize(config_data.size());
364 memcpy(data.data(), config_data.begin(), data.size());
365 return SizePrefixedFlatbufferVector<LogFileHeader>(std::move(data));
Austin Schuh6f3babe2020-01-26 20:34:50 -0800366}
367
Austin Schuhadd6eb32020-11-09 21:24:26 -0800368std::optional<SizePrefixedFlatbufferVector<MessageHeader>> ReadNthMessage(
Austin Schuh3bd4c402020-11-06 18:19:06 -0800369 std::string_view filename, size_t n) {
Austin Schuh5212cad2020-09-09 23:12:09 -0700370 SpanReader span_reader(filename);
371 absl::Span<const uint8_t> data_span = span_reader.ReadMessage();
372 for (size_t i = 0; i < n + 1; ++i) {
373 data_span = span_reader.ReadMessage();
374
375 // Make sure something was read.
Austin Schuh3bd4c402020-11-06 18:19:06 -0800376 if (data_span == absl::Span<const uint8_t>()) {
377 return std::nullopt;
378 }
Austin Schuh5212cad2020-09-09 23:12:09 -0700379 }
380
Brian Silverman354697a2020-09-22 21:06:32 -0700381 // And copy the config so we have it forever, removing the size prefix.
382 ResizeableBuffer data;
Austin Schuhadd6eb32020-11-09 21:24:26 -0800383 data.resize(data_span.size());
384 memcpy(data.data(), data_span.begin(), data.size());
385 return SizePrefixedFlatbufferVector<MessageHeader>(std::move(data));
Austin Schuh5212cad2020-09-09 23:12:09 -0700386}
387
Austin Schuh05b70472020-01-01 17:11:17 -0800388MessageReader::MessageReader(std::string_view filename)
Austin Schuh97789fc2020-08-01 14:42:45 -0700389 : span_reader_(filename),
Austin Schuhadd6eb32020-11-09 21:24:26 -0800390 raw_log_file_header_(
391 SizePrefixedFlatbufferVector<LogFileHeader>::Empty()) {
Austin Schuh05b70472020-01-01 17:11:17 -0800392 // Make sure we have enough to read the size.
Austin Schuh97789fc2020-08-01 14:42:45 -0700393 absl::Span<const uint8_t> header_data = span_reader_.ReadMessage();
Austin Schuh05b70472020-01-01 17:11:17 -0800394
395 // Make sure something was read.
Austin Schuh97789fc2020-08-01 14:42:45 -0700396 CHECK(header_data != absl::Span<const uint8_t>())
397 << ": Failed to read header from: " << filename;
Austin Schuh05b70472020-01-01 17:11:17 -0800398
Austin Schuh97789fc2020-08-01 14:42:45 -0700399 // And copy the header data so we have it forever.
Brian Silverman354697a2020-09-22 21:06:32 -0700400 ResizeableBuffer header_data_copy;
Austin Schuhadd6eb32020-11-09 21:24:26 -0800401 header_data_copy.resize(header_data.size());
402 memcpy(header_data_copy.data(), header_data.begin(), header_data_copy.size());
Austin Schuh97789fc2020-08-01 14:42:45 -0700403 raw_log_file_header_ =
Austin Schuhadd6eb32020-11-09 21:24:26 -0800404 SizePrefixedFlatbufferVector<LogFileHeader>(std::move(header_data_copy));
Austin Schuh05b70472020-01-01 17:11:17 -0800405
Austin Schuhcde938c2020-02-02 17:30:07 -0800406 max_out_of_order_duration_ =
Austin Schuh2f8fd752020-09-01 22:38:28 -0700407 chrono::nanoseconds(log_file_header()->max_out_of_order_duration());
Austin Schuhcde938c2020-02-02 17:30:07 -0800408
409 VLOG(1) << "Opened " << filename << " as node "
410 << FlatbufferToJson(log_file_header()->node());
Austin Schuh05b70472020-01-01 17:11:17 -0800411}
412
Austin Schuhadd6eb32020-11-09 21:24:26 -0800413std::optional<SizePrefixedFlatbufferVector<MessageHeader>>
414MessageReader::ReadMessage() {
Austin Schuh05b70472020-01-01 17:11:17 -0800415 absl::Span<const uint8_t> msg_data = span_reader_.ReadMessage();
416 if (msg_data == absl::Span<const uint8_t>()) {
417 return std::nullopt;
418 }
419
Brian Silverman354697a2020-09-22 21:06:32 -0700420 ResizeableBuffer result_buffer;
Austin Schuhadd6eb32020-11-09 21:24:26 -0800421 result_buffer.resize(msg_data.size());
422 memcpy(result_buffer.data(), msg_data.begin(), result_buffer.size());
423 SizePrefixedFlatbufferVector<MessageHeader> result(std::move(result_buffer));
Austin Schuh05b70472020-01-01 17:11:17 -0800424
425 const monotonic_clock::time_point timestamp = monotonic_clock::time_point(
426 chrono::nanoseconds(result.message().monotonic_sent_time()));
427
428 newest_timestamp_ = std::max(newest_timestamp_, timestamp);
Austin Schuh8bd96322020-02-13 21:18:22 -0800429 VLOG(2) << "Read from " << filename() << " data " << FlatbufferToJson(result);
Austin Schuh6f3babe2020-01-26 20:34:50 -0800430 return std::move(result);
Austin Schuh05b70472020-01-01 17:11:17 -0800431}
432
Austin Schuhc41603c2020-10-11 16:17:37 -0700433PartsMessageReader::PartsMessageReader(LogParts log_parts)
434 : parts_(std::move(log_parts)), message_reader_(parts_.parts[0]) {}
435
Austin Schuhadd6eb32020-11-09 21:24:26 -0800436std::optional<SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuhc41603c2020-10-11 16:17:37 -0700437PartsMessageReader::ReadMessage() {
438 while (!done_) {
Austin Schuhadd6eb32020-11-09 21:24:26 -0800439 std::optional<SizePrefixedFlatbufferVector<MessageHeader>> message =
Austin Schuhc41603c2020-10-11 16:17:37 -0700440 message_reader_.ReadMessage();
441 if (message) {
442 newest_timestamp_ = message_reader_.newest_timestamp();
Austin Schuh32f68492020-11-08 21:45:51 -0800443 const monotonic_clock::time_point monotonic_sent_time(
444 chrono::nanoseconds(message->message().monotonic_sent_time()));
445 CHECK_GE(monotonic_sent_time,
446 newest_timestamp_ - max_out_of_order_duration());
Austin Schuhc41603c2020-10-11 16:17:37 -0700447 return message;
448 }
449 NextLog();
450 }
Austin Schuh32f68492020-11-08 21:45:51 -0800451 newest_timestamp_ = monotonic_clock::max_time;
Austin Schuhc41603c2020-10-11 16:17:37 -0700452 return std::nullopt;
453}
454
455void PartsMessageReader::NextLog() {
456 if (next_part_index_ == parts_.parts.size()) {
457 done_ = true;
458 return;
459 }
460 message_reader_ = MessageReader(parts_.parts[next_part_index_]);
461 ++next_part_index_;
462}
463
Austin Schuh6f3babe2020-01-26 20:34:50 -0800464SplitMessageReader::SplitMessageReader(
Austin Schuhfa895892020-01-07 20:07:41 -0800465 const std::vector<std::string> &filenames)
466 : filenames_(filenames),
Austin Schuhadd6eb32020-11-09 21:24:26 -0800467 log_file_header_(SizePrefixedFlatbufferVector<LogFileHeader>::Empty()) {
Austin Schuhfa895892020-01-07 20:07:41 -0800468 CHECK(NextLogFile()) << ": filenames is empty. Need files to read.";
469
Austin Schuh6f3babe2020-01-26 20:34:50 -0800470 // Grab any log file header. They should all match (and we will check as we
471 // open more of them).
Austin Schuh97789fc2020-08-01 14:42:45 -0700472 log_file_header_ = message_reader_->raw_log_file_header();
Austin Schuhfa895892020-01-07 20:07:41 -0800473
Austin Schuh2f8fd752020-09-01 22:38:28 -0700474 for (size_t i = 1; i < filenames_.size(); ++i) {
475 MessageReader message_reader(filenames_[i]);
476
477 const monotonic_clock::time_point new_monotonic_start_time(
478 chrono::nanoseconds(
479 message_reader.log_file_header()->monotonic_start_time()));
480 const realtime_clock::time_point new_realtime_start_time(
481 chrono::nanoseconds(
482 message_reader.log_file_header()->realtime_start_time()));
483
484 // There are 2 types of part files. Part files from before time estimation
485 // has started, and part files after. We don't declare a log file "started"
486 // until time estimation is up. And once a log file starts, it should never
487 // stop again, and should remain constant.
488 // To compare both types of headers, we mutate our saved copy of the header
489 // to match the next chunk by updating time if we detect a stopped ->
490 // started transition.
491 if (monotonic_start_time() == monotonic_clock::min_time) {
492 CHECK_EQ(realtime_start_time(), realtime_clock::min_time);
493 // We should only be missing the monotonic start time when logging data
Brian Silverman87ac0402020-09-17 14:47:01 -0700494 // for remote nodes. We don't have a good way to determine the remote
Austin Schuh2f8fd752020-09-01 22:38:28 -0700495 // realtime offset, so it shouldn't be filled out.
496 // TODO(austin): If we have a good way, feel free to fill it out. It
497 // probably won't be better than we could do in post though with the same
498 // data.
499 CHECK(!log_file_header_.mutable_message()->has_realtime_start_time());
500 if (new_monotonic_start_time != monotonic_clock::min_time) {
501 // If we finally found our start time, update the header. Do this once
502 // because it should never change again.
503 log_file_header_.mutable_message()->mutate_monotonic_start_time(
504 new_monotonic_start_time.time_since_epoch().count());
505 log_file_header_.mutable_message()->mutate_realtime_start_time(
506 new_realtime_start_time.time_since_epoch().count());
507 }
508 }
509
Austin Schuh64fab802020-09-09 22:47:47 -0700510 // We don't have a good way to set the realtime start time on remote nodes.
511 // Confirm it remains consistent.
512 CHECK_EQ(log_file_header_.mutable_message()->has_realtime_start_time(),
513 message_reader.log_file_header()->has_realtime_start_time());
514
515 // Parts index will *not* match unless we set them to match. We only want
516 // to accept the start time and parts mismatching, so set them.
517 log_file_header_.mutable_message()->mutate_parts_index(
518 message_reader.log_file_header()->parts_index());
519
Austin Schuh2f8fd752020-09-01 22:38:28 -0700520 // Now compare that the headers match.
Austin Schuh64fab802020-09-09 22:47:47 -0700521 if (!CompareFlatBuffer(message_reader.raw_log_file_header(),
522 log_file_header_)) {
Brian Silvermanae7c0332020-09-30 16:58:23 -0700523 if (message_reader.log_file_header()->has_log_event_uuid() &&
524 log_file_header_.message().has_log_event_uuid() &&
525 message_reader.log_file_header()->log_event_uuid()->string_view() !=
526 log_file_header_.message().log_event_uuid()->string_view()) {
Austin Schuh64fab802020-09-09 22:47:47 -0700527 LOG(FATAL) << "Logger UUIDs don't match between log file chunks "
528 << filenames_[0] << " and " << filenames_[i]
529 << ", this is not supported.";
530 }
531 if (message_reader.log_file_header()->has_parts_uuid() &&
532 log_file_header_.message().has_parts_uuid() &&
533 message_reader.log_file_header()->parts_uuid()->string_view() !=
534 log_file_header_.message().parts_uuid()->string_view()) {
535 LOG(FATAL) << "Parts UUIDs don't match between log file chunks "
536 << filenames_[0] << " and " << filenames_[i]
537 << ", this is not supported.";
538 }
539
540 LOG(FATAL) << "Header is different between log file chunks "
541 << filenames_[0] << " and " << filenames_[i]
542 << ", this is not supported.";
543 }
Austin Schuh2f8fd752020-09-01 22:38:28 -0700544 }
Austin Schuh64fab802020-09-09 22:47:47 -0700545 // Put the parts index back to the first log file chunk.
546 log_file_header_.mutable_message()->mutate_parts_index(
547 message_reader_->log_file_header()->parts_index());
Austin Schuh2f8fd752020-09-01 22:38:28 -0700548
Austin Schuh6f3babe2020-01-26 20:34:50 -0800549 // Setup per channel state.
Austin Schuh05b70472020-01-01 17:11:17 -0800550 channels_.resize(configuration()->channels()->size());
Austin Schuh6f3babe2020-01-26 20:34:50 -0800551 for (ChannelData &channel_data : channels_) {
552 channel_data.data.split_reader = this;
553 // Build up the timestamp list.
554 if (configuration::MultiNode(configuration())) {
555 channel_data.timestamps.resize(configuration()->nodes()->size());
556 for (MessageHeaderQueue &queue : channel_data.timestamps) {
557 queue.timestamps = true;
558 queue.split_reader = this;
559 }
560 }
561 }
Austin Schuh05b70472020-01-01 17:11:17 -0800562
Austin Schuh6f3babe2020-01-26 20:34:50 -0800563 // Build up channels_to_write_ as an optimization to make it fast to figure
564 // out which datastructure to place any new data from a channel on.
565 for (const Channel *channel : *configuration()->channels()) {
566 // This is the main case. We will only see data on this node.
567 if (configuration::ChannelIsSendableOnNode(channel, node())) {
568 channels_to_write_.emplace_back(
569 &channels_[channels_to_write_.size()].data);
570 } else
571 // If we can't send, but can receive, we should be able to see
572 // timestamps here.
573 if (configuration::ChannelIsReadableOnNode(channel, node())) {
574 channels_to_write_.emplace_back(
575 &(channels_[channels_to_write_.size()]
576 .timestamps[configuration::GetNodeIndex(configuration(),
577 node())]));
578 } else {
579 channels_to_write_.emplace_back(nullptr);
580 }
581 }
Austin Schuh05b70472020-01-01 17:11:17 -0800582}
583
Austin Schuh6f3babe2020-01-26 20:34:50 -0800584bool SplitMessageReader::NextLogFile() {
Austin Schuhfa895892020-01-07 20:07:41 -0800585 if (next_filename_index_ == filenames_.size()) {
586 return false;
587 }
588 message_reader_ =
589 std::make_unique<MessageReader>(filenames_[next_filename_index_]);
590
591 // We can't support the config diverging between two log file headers. See if
592 // they are the same.
593 if (next_filename_index_ != 0) {
Austin Schuh64fab802020-09-09 22:47:47 -0700594 // In order for the headers to identically compare, they need to have the
595 // same parts_index. Rewrite the saved header with the new parts_index,
596 // compare, and then restore.
597 const int32_t original_parts_index =
598 log_file_header_.message().parts_index();
599 log_file_header_.mutable_message()->mutate_parts_index(
600 message_reader_->log_file_header()->parts_index());
601
Austin Schuh97789fc2020-08-01 14:42:45 -0700602 CHECK(CompareFlatBuffer(message_reader_->raw_log_file_header(),
603 log_file_header_))
Austin Schuhfa895892020-01-07 20:07:41 -0800604 << ": Header is different between log file chunks "
605 << filenames_[next_filename_index_] << " and "
606 << filenames_[next_filename_index_ - 1] << ", this is not supported.";
Austin Schuh64fab802020-09-09 22:47:47 -0700607
608 log_file_header_.mutable_message()->mutate_parts_index(
609 original_parts_index);
Austin Schuhfa895892020-01-07 20:07:41 -0800610 }
611
612 ++next_filename_index_;
613 return true;
614}
615
Austin Schuh6f3babe2020-01-26 20:34:50 -0800616bool SplitMessageReader::QueueMessages(
Austin Schuhcde938c2020-02-02 17:30:07 -0800617 monotonic_clock::time_point last_dequeued_time) {
Austin Schuh6f3babe2020-01-26 20:34:50 -0800618 // TODO(austin): Once we are happy that everything works, read a 256kb chunk
619 // to reduce the need to re-heap down below.
Austin Schuhcde938c2020-02-02 17:30:07 -0800620
621 // Special case no more data. Otherwise we blow up on the CHECK statement
622 // confirming that we have enough data queued.
623 if (at_end_) {
624 return false;
625 }
626
627 // If this isn't the first time around, confirm that we had enough data queued
628 // to follow the contract.
629 if (time_to_queue_ != monotonic_clock::min_time) {
630 CHECK_LE(last_dequeued_time,
631 newest_timestamp() - max_out_of_order_duration())
632 << " node " << FlatbufferToJson(node()) << " on " << this;
633
634 // Bail if there is enough data already queued.
635 if (last_dequeued_time < time_to_queue_) {
Austin Schuhee711052020-08-24 16:06:09 -0700636 VLOG(1) << MaybeNodeName(target_node_) << "All up to date on " << this
637 << ", dequeued " << last_dequeued_time << " queue time "
638 << time_to_queue_;
Austin Schuhcde938c2020-02-02 17:30:07 -0800639 return true;
640 }
641 } else {
642 // Startup takes a special dance. We want to queue up until the start time,
643 // but we then want to find the next message to read. The conservative
644 // answer is to immediately trigger a second requeue to get things moving.
645 time_to_queue_ = monotonic_start_time();
Austin Schuheeba0292020-10-11 16:20:05 -0700646 CHECK_NE(time_to_queue_, monotonic_clock::min_time);
Austin Schuhcde938c2020-02-02 17:30:07 -0800647 QueueMessages(time_to_queue_);
648 }
649
650 // If we are asked to queue, queue for at least max_out_of_order_duration past
651 // the last known time in the log file (ie the newest timestep read). As long
652 // as we requeue exactly when time_to_queue_ is dequeued and go no further, we
653 // are safe. And since we pop in order, that works.
654 //
655 // Special case the start of the log file. There should be at most 1 message
656 // from each channel at the start of the log file. So always force the start
657 // of the log file to just be read.
658 time_to_queue_ = std::max(time_to_queue_, newest_timestamp());
Austin Schuhee711052020-08-24 16:06:09 -0700659 VLOG(1) << MaybeNodeName(target_node_) << "Queueing, going until "
660 << time_to_queue_ << " " << filename();
Austin Schuhcde938c2020-02-02 17:30:07 -0800661
662 bool was_emplaced = false;
Austin Schuh6f3babe2020-01-26 20:34:50 -0800663 while (true) {
Austin Schuhcde938c2020-02-02 17:30:07 -0800664 // Stop if we have enough.
Brian Silverman98360e22020-04-28 16:51:20 -0700665 if (newest_timestamp() > time_to_queue_ + max_out_of_order_duration() &&
Austin Schuhcde938c2020-02-02 17:30:07 -0800666 was_emplaced) {
Austin Schuhee711052020-08-24 16:06:09 -0700667 VLOG(1) << MaybeNodeName(target_node_) << "Done queueing on " << this
668 << ", queued to " << newest_timestamp() << " with requeue time "
669 << time_to_queue_;
Austin Schuh6f3babe2020-01-26 20:34:50 -0800670 return true;
671 }
Austin Schuh05b70472020-01-01 17:11:17 -0800672
Austin Schuhadd6eb32020-11-09 21:24:26 -0800673 if (std::optional<SizePrefixedFlatbufferVector<MessageHeader>> msg =
Austin Schuh6f3babe2020-01-26 20:34:50 -0800674 message_reader_->ReadMessage()) {
675 const MessageHeader &header = msg.value().message();
676
Austin Schuhcde938c2020-02-02 17:30:07 -0800677 const monotonic_clock::time_point timestamp = monotonic_clock::time_point(
678 chrono::nanoseconds(header.monotonic_sent_time()));
Austin Schuh6f3babe2020-01-26 20:34:50 -0800679
Austin Schuh0b5fd032020-03-28 17:36:49 -0700680 if (VLOG_IS_ON(2)) {
Brian Silvermand90905f2020-09-23 14:42:56 -0700681 LOG(INFO) << MaybeNodeName(target_node_) << "Queued " << this << " "
682 << filename() << " ttq: " << time_to_queue_ << " now "
Austin Schuhee711052020-08-24 16:06:09 -0700683 << newest_timestamp() << " start time "
684 << monotonic_start_time() << " " << FlatbufferToJson(&header);
Austin Schuh0b5fd032020-03-28 17:36:49 -0700685 } else if (VLOG_IS_ON(1)) {
Austin Schuhadd6eb32020-11-09 21:24:26 -0800686 SizePrefixedFlatbufferVector<MessageHeader> copy = msg.value();
Austin Schuh0b5fd032020-03-28 17:36:49 -0700687 copy.mutable_message()->clear_data();
Austin Schuhee711052020-08-24 16:06:09 -0700688 LOG(INFO) << MaybeNodeName(target_node_) << "Queued " << this << " "
689 << filename() << " ttq: " << time_to_queue_ << " now "
690 << newest_timestamp() << " start time "
691 << monotonic_start_time() << " " << FlatbufferToJson(copy);
Austin Schuh0b5fd032020-03-28 17:36:49 -0700692 }
Austin Schuhcde938c2020-02-02 17:30:07 -0800693
694 const int channel_index = header.channel_index();
695 was_emplaced = channels_to_write_[channel_index]->emplace_back(
696 std::move(msg.value()));
697 if (was_emplaced) {
698 newest_timestamp_ = std::max(newest_timestamp_, timestamp);
699 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800700 } else {
701 if (!NextLogFile()) {
Austin Schuhee711052020-08-24 16:06:09 -0700702 VLOG(1) << MaybeNodeName(target_node_) << "No more files, last was "
703 << filenames_.back();
Austin Schuhcde938c2020-02-02 17:30:07 -0800704 at_end_ = true;
Austin Schuh8bd96322020-02-13 21:18:22 -0800705 for (MessageHeaderQueue *queue : channels_to_write_) {
706 if (queue == nullptr || queue->timestamp_merger == nullptr) {
707 continue;
708 }
709 queue->timestamp_merger->NoticeAtEnd();
710 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800711 return false;
712 }
713 }
Austin Schuh05b70472020-01-01 17:11:17 -0800714 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800715}
716
717void SplitMessageReader::SetTimestampMerger(TimestampMerger *timestamp_merger,
718 int channel_index,
719 const Node *target_node) {
720 const Node *reinterpreted_target_node =
721 configuration::GetNodeOrDie(configuration(), target_node);
Austin Schuhee711052020-08-24 16:06:09 -0700722 target_node_ = reinterpreted_target_node;
723
Austin Schuh6f3babe2020-01-26 20:34:50 -0800724 const Channel *const channel =
725 configuration()->channels()->Get(channel_index);
726
Austin Schuhcde938c2020-02-02 17:30:07 -0800727 VLOG(1) << " Configuring merger " << this << " for channel " << channel_index
728 << " "
729 << configuration::CleanedChannelToString(
730 configuration()->channels()->Get(channel_index));
731
Austin Schuh6f3babe2020-01-26 20:34:50 -0800732 MessageHeaderQueue *message_header_queue = nullptr;
733
734 // Figure out if this log file is from our point of view, or the other node's
735 // point of view.
736 if (node() == reinterpreted_target_node) {
Austin Schuhcde938c2020-02-02 17:30:07 -0800737 VLOG(1) << " Replaying as logged node " << filename();
738
739 if (configuration::ChannelIsSendableOnNode(channel, node())) {
740 VLOG(1) << " Data on node";
741 message_header_queue = &(channels_[channel_index].data);
742 } else if (configuration::ChannelIsReadableOnNode(channel, node())) {
743 VLOG(1) << " Timestamps on node";
744 message_header_queue =
745 &(channels_[channel_index].timestamps[configuration::GetNodeIndex(
746 configuration(), node())]);
Austin Schuh6f3babe2020-01-26 20:34:50 -0800747 } else {
Austin Schuhcde938c2020-02-02 17:30:07 -0800748 VLOG(1) << " Dropping";
Austin Schuh6f3babe2020-01-26 20:34:50 -0800749 }
750 } else {
Austin Schuhcde938c2020-02-02 17:30:07 -0800751 VLOG(1) << " Replaying as other node " << filename();
Austin Schuh6f3babe2020-01-26 20:34:50 -0800752 // We are replaying from another node's point of view. The only interesting
Austin Schuhcde938c2020-02-02 17:30:07 -0800753 // data is data that is sent from our node and received on theirs.
754 if (configuration::ChannelIsReadableOnNode(channel,
755 reinterpreted_target_node) &&
756 configuration::ChannelIsSendableOnNode(channel, node())) {
757 VLOG(1) << " Readable on target node";
Austin Schuh6f3babe2020-01-26 20:34:50 -0800758 // Data from another node.
759 message_header_queue = &(channels_[channel_index].data);
760 } else {
Austin Schuhcde938c2020-02-02 17:30:07 -0800761 VLOG(1) << " Dropping";
Austin Schuh6f3babe2020-01-26 20:34:50 -0800762 // This is either not sendable on the other node, or is a timestamp and
763 // therefore not interesting.
764 }
765 }
766
767 // If we found one, write it down. This will be nullptr when there is nothing
768 // relevant on this channel on this node for the target node. In that case,
769 // we want to drop the message instead of queueing it.
770 if (message_header_queue != nullptr) {
771 message_header_queue->timestamp_merger = timestamp_merger;
772 }
773}
774
775std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -0800776 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -0800777SplitMessageReader::PopOldest(int channel_index) {
778 CHECK_GT(channels_[channel_index].data.size(), 0u);
Austin Schuhcde938c2020-02-02 17:30:07 -0800779 const std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
780 timestamp = channels_[channel_index].data.front_timestamp();
Austin Schuhadd6eb32020-11-09 21:24:26 -0800781 SizePrefixedFlatbufferVector<MessageHeader> front =
Austin Schuh6f3babe2020-01-26 20:34:50 -0800782 std::move(channels_[channel_index].data.front());
Austin Schuh2f8fd752020-09-01 22:38:28 -0700783 channels_[channel_index].data.PopFront();
Austin Schuhcde938c2020-02-02 17:30:07 -0800784
Austin Schuh2f8fd752020-09-01 22:38:28 -0700785 VLOG(1) << MaybeNodeName(target_node_) << "Popped Data " << this << " "
786 << std::get<0>(timestamp) << " for "
787 << configuration::StrippedChannelToString(
788 configuration()->channels()->Get(channel_index))
789 << " (" << channel_index << ")";
Austin Schuhcde938c2020-02-02 17:30:07 -0800790
791 QueueMessages(std::get<0>(timestamp));
Austin Schuh6f3babe2020-01-26 20:34:50 -0800792
793 return std::make_tuple(std::get<0>(timestamp), std::get<1>(timestamp),
794 std::move(front));
795}
796
797std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -0800798 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh2f8fd752020-09-01 22:38:28 -0700799SplitMessageReader::PopOldestTimestamp(int channel, int node_index) {
Austin Schuh6f3babe2020-01-26 20:34:50 -0800800 CHECK_GT(channels_[channel].timestamps[node_index].size(), 0u);
Austin Schuhcde938c2020-02-02 17:30:07 -0800801 const std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
802 timestamp = channels_[channel].timestamps[node_index].front_timestamp();
Austin Schuhadd6eb32020-11-09 21:24:26 -0800803 SizePrefixedFlatbufferVector<MessageHeader> front =
Austin Schuh6f3babe2020-01-26 20:34:50 -0800804 std::move(channels_[channel].timestamps[node_index].front());
Austin Schuh2f8fd752020-09-01 22:38:28 -0700805 channels_[channel].timestamps[node_index].PopFront();
Austin Schuhcde938c2020-02-02 17:30:07 -0800806
Austin Schuh2f8fd752020-09-01 22:38:28 -0700807 VLOG(1) << MaybeNodeName(target_node_) << "Popped timestamp " << this << " "
Austin Schuhee711052020-08-24 16:06:09 -0700808 << std::get<0>(timestamp) << " for "
809 << configuration::StrippedChannelToString(
810 configuration()->channels()->Get(channel))
Austin Schuh2f8fd752020-09-01 22:38:28 -0700811 << " on "
812 << configuration()->nodes()->Get(node_index)->name()->string_view()
813 << " (" << node_index << ")";
Austin Schuhcde938c2020-02-02 17:30:07 -0800814
815 QueueMessages(std::get<0>(timestamp));
Austin Schuh6f3babe2020-01-26 20:34:50 -0800816
817 return std::make_tuple(std::get<0>(timestamp), std::get<1>(timestamp),
818 std::move(front));
819}
820
Austin Schuhcde938c2020-02-02 17:30:07 -0800821bool SplitMessageReader::MessageHeaderQueue::emplace_back(
Austin Schuhadd6eb32020-11-09 21:24:26 -0800822 SizePrefixedFlatbufferVector<MessageHeader> &&msg) {
Austin Schuh6f3babe2020-01-26 20:34:50 -0800823 CHECK(split_reader != nullptr);
824
825 // If there is no timestamp merger for this queue, nobody is listening. Drop
826 // the message. This happens when a log file from another node is replayed,
827 // and the timestamp mergers down stream just don't care.
828 if (timestamp_merger == nullptr) {
Austin Schuhcde938c2020-02-02 17:30:07 -0800829 return false;
Austin Schuh6f3babe2020-01-26 20:34:50 -0800830 }
831
832 CHECK(timestamps != msg.message().has_data())
833 << ": Got timestamps and data mixed up on a node. "
834 << FlatbufferToJson(msg);
835
836 data_.emplace_back(std::move(msg));
837
838 if (data_.size() == 1u) {
839 // Yup, new data. Notify.
840 if (timestamps) {
841 timestamp_merger->UpdateTimestamp(split_reader, front_timestamp());
842 } else {
843 timestamp_merger->Update(split_reader, front_timestamp());
844 }
845 }
Austin Schuhcde938c2020-02-02 17:30:07 -0800846
847 return true;
Austin Schuh6f3babe2020-01-26 20:34:50 -0800848}
849
Austin Schuh2f8fd752020-09-01 22:38:28 -0700850void SplitMessageReader::MessageHeaderQueue::PopFront() {
Austin Schuh6f3babe2020-01-26 20:34:50 -0800851 data_.pop_front();
852 if (data_.size() != 0u) {
853 // Yup, new data.
854 if (timestamps) {
855 timestamp_merger->UpdateTimestamp(split_reader, front_timestamp());
856 } else {
857 timestamp_merger->Update(split_reader, front_timestamp());
858 }
Austin Schuh2f8fd752020-09-01 22:38:28 -0700859 } else {
860 // Poke anyways to update the heap.
861 if (timestamps) {
862 timestamp_merger->UpdateTimestamp(
863 nullptr, std::make_tuple(monotonic_clock::min_time, 0, nullptr));
864 } else {
865 timestamp_merger->Update(
866 nullptr, std::make_tuple(monotonic_clock::min_time, 0, nullptr));
867 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800868 }
Austin Schuh05b70472020-01-01 17:11:17 -0800869}
870
871namespace {
872
Austin Schuh6f3babe2020-01-26 20:34:50 -0800873bool SplitMessageReaderHeapCompare(
874 const std::tuple<monotonic_clock::time_point, uint32_t,
875 SplitMessageReader *>
876 first,
877 const std::tuple<monotonic_clock::time_point, uint32_t,
878 SplitMessageReader *>
879 second) {
880 if (std::get<0>(first) > std::get<0>(second)) {
881 return true;
882 } else if (std::get<0>(first) == std::get<0>(second)) {
883 if (std::get<1>(first) > std::get<1>(second)) {
884 return true;
885 } else if (std::get<1>(first) == std::get<1>(second)) {
886 return std::get<2>(first) > std::get<2>(second);
887 } else {
888 return false;
889 }
890 } else {
891 return false;
892 }
893}
894
Austin Schuh05b70472020-01-01 17:11:17 -0800895bool ChannelHeapCompare(
896 const std::pair<monotonic_clock::time_point, int> first,
897 const std::pair<monotonic_clock::time_point, int> second) {
898 if (first.first > second.first) {
899 return true;
900 } else if (first.first == second.first) {
901 return first.second > second.second;
902 } else {
903 return false;
904 }
905}
906
907} // namespace
908
Austin Schuh6f3babe2020-01-26 20:34:50 -0800909TimestampMerger::TimestampMerger(
910 const Configuration *configuration,
911 std::vector<SplitMessageReader *> split_message_readers, int channel_index,
912 const Node *target_node, ChannelMerger *channel_merger)
913 : configuration_(configuration),
914 split_message_readers_(std::move(split_message_readers)),
915 channel_index_(channel_index),
916 node_index_(configuration::MultiNode(configuration)
917 ? configuration::GetNodeIndex(configuration, target_node)
918 : -1),
919 channel_merger_(channel_merger) {
920 // Tell the readers we care so they know who to notify.
Austin Schuhcde938c2020-02-02 17:30:07 -0800921 VLOG(1) << "Configuring channel " << channel_index << " target node "
922 << FlatbufferToJson(target_node);
Austin Schuh6f3babe2020-01-26 20:34:50 -0800923 for (SplitMessageReader *reader : split_message_readers_) {
924 reader->SetTimestampMerger(this, channel_index, target_node);
925 }
926
927 // And then determine if we need to track timestamps.
928 const Channel *channel = configuration->channels()->Get(channel_index);
929 if (!configuration::ChannelIsSendableOnNode(channel, target_node) &&
930 configuration::ChannelIsReadableOnNode(channel, target_node)) {
931 has_timestamps_ = true;
932 }
933}
934
935void TimestampMerger::PushMessageHeap(
Austin Schuhcde938c2020-02-02 17:30:07 -0800936 std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
937 timestamp,
Austin Schuh6f3babe2020-01-26 20:34:50 -0800938 SplitMessageReader *split_message_reader) {
Austin Schuh2f8fd752020-09-01 22:38:28 -0700939 if (split_message_reader != nullptr) {
940 DCHECK(std::find_if(message_heap_.begin(), message_heap_.end(),
941 [split_message_reader](
942 const std::tuple<monotonic_clock::time_point,
943 uint32_t, SplitMessageReader *>
944 x) {
945 return std::get<2>(x) == split_message_reader;
946 }) == message_heap_.end())
947 << ": Pushing message when it is already in the heap.";
Austin Schuh6f3babe2020-01-26 20:34:50 -0800948
Austin Schuh2f8fd752020-09-01 22:38:28 -0700949 message_heap_.push_back(std::make_tuple(
950 std::get<0>(timestamp), std::get<1>(timestamp), split_message_reader));
Austin Schuh6f3babe2020-01-26 20:34:50 -0800951
Austin Schuh2f8fd752020-09-01 22:38:28 -0700952 std::push_heap(message_heap_.begin(), message_heap_.end(),
953 &SplitMessageReaderHeapCompare);
954 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800955
956 // If we are just a data merger, don't wait for timestamps.
957 if (!has_timestamps_) {
Austin Schuh2f8fd752020-09-01 22:38:28 -0700958 if (!message_heap_.empty()) {
959 channel_merger_->Update(std::get<0>(message_heap_[0]), channel_index_);
960 pushed_ = true;
961 } else {
962 // Remove ourselves if we are empty.
963 channel_merger_->Update(monotonic_clock::min_time, channel_index_);
964 }
Austin Schuh6f3babe2020-01-26 20:34:50 -0800965 }
966}
967
Austin Schuhcde938c2020-02-02 17:30:07 -0800968std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
969TimestampMerger::oldest_message() const {
970 CHECK_GT(message_heap_.size(), 0u);
971 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
972 oldest_message_reader = message_heap_.front();
973 return std::get<2>(oldest_message_reader)->oldest_message(channel_index_);
974}
975
976std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
977TimestampMerger::oldest_timestamp() const {
978 CHECK_GT(timestamp_heap_.size(), 0u);
979 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
980 oldest_message_reader = timestamp_heap_.front();
981 return std::get<2>(oldest_message_reader)
982 ->oldest_message(channel_index_, node_index_);
983}
984
Austin Schuh6f3babe2020-01-26 20:34:50 -0800985void TimestampMerger::PushTimestampHeap(
Austin Schuhcde938c2020-02-02 17:30:07 -0800986 std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
987 timestamp,
Austin Schuh6f3babe2020-01-26 20:34:50 -0800988 SplitMessageReader *split_message_reader) {
Austin Schuh2f8fd752020-09-01 22:38:28 -0700989 if (split_message_reader != nullptr) {
990 DCHECK(std::find_if(timestamp_heap_.begin(), timestamp_heap_.end(),
991 [split_message_reader](
992 const std::tuple<monotonic_clock::time_point,
993 uint32_t, SplitMessageReader *>
994 x) {
995 return std::get<2>(x) == split_message_reader;
996 }) == timestamp_heap_.end())
997 << ": Pushing timestamp when it is already in the heap.";
Austin Schuh6f3babe2020-01-26 20:34:50 -0800998
Austin Schuh2f8fd752020-09-01 22:38:28 -0700999 timestamp_heap_.push_back(std::make_tuple(
1000 std::get<0>(timestamp), std::get<1>(timestamp), split_message_reader));
Austin Schuh6f3babe2020-01-26 20:34:50 -08001001
Austin Schuh2f8fd752020-09-01 22:38:28 -07001002 std::push_heap(timestamp_heap_.begin(), timestamp_heap_.end(),
1003 SplitMessageReaderHeapCompare);
1004 }
Austin Schuh6f3babe2020-01-26 20:34:50 -08001005
1006 // If we are a timestamp merger, don't wait for data. Missing data will be
1007 // caught at read time.
1008 if (has_timestamps_) {
Austin Schuh2f8fd752020-09-01 22:38:28 -07001009 if (!timestamp_heap_.empty()) {
1010 channel_merger_->Update(std::get<0>(timestamp_heap_[0]), channel_index_);
1011 pushed_ = true;
1012 } else {
1013 // Remove ourselves if we are empty.
1014 channel_merger_->Update(monotonic_clock::min_time, channel_index_);
1015 }
Austin Schuh6f3babe2020-01-26 20:34:50 -08001016 }
1017}
1018
1019std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001020 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001021TimestampMerger::PopMessageHeap() {
1022 // Pop the oldest message reader pointer off the heap.
1023 CHECK_GT(message_heap_.size(), 0u);
1024 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
1025 oldest_message_reader = message_heap_.front();
1026
1027 std::pop_heap(message_heap_.begin(), message_heap_.end(),
1028 &SplitMessageReaderHeapCompare);
1029 message_heap_.pop_back();
1030
1031 // Pop the oldest message. This re-pushes any messages from the reader to the
1032 // message heap.
1033 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001034 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001035 oldest_message =
1036 std::get<2>(oldest_message_reader)->PopOldest(channel_index_);
1037
1038 // Confirm that the time and queue_index we have recorded matches.
1039 CHECK_EQ(std::get<0>(oldest_message), std::get<0>(oldest_message_reader));
1040 CHECK_EQ(std::get<1>(oldest_message), std::get<1>(oldest_message_reader));
1041
1042 // Now, keep reading until we have found all duplicates.
Brian Silverman8a32ce62020-08-12 12:02:38 -07001043 while (!message_heap_.empty()) {
Austin Schuh6f3babe2020-01-26 20:34:50 -08001044 // See if it is a duplicate.
1045 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
1046 next_oldest_message_reader = message_heap_.front();
1047
Austin Schuhcde938c2020-02-02 17:30:07 -08001048 std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
1049 next_oldest_message_time = std::get<2>(next_oldest_message_reader)
1050 ->oldest_message(channel_index_);
Austin Schuh6f3babe2020-01-26 20:34:50 -08001051
1052 if (std::get<0>(next_oldest_message_time) == std::get<0>(oldest_message) &&
1053 std::get<1>(next_oldest_message_time) == std::get<1>(oldest_message)) {
1054 // Pop the message reader pointer.
1055 std::pop_heap(message_heap_.begin(), message_heap_.end(),
1056 &SplitMessageReaderHeapCompare);
1057 message_heap_.pop_back();
1058
1059 // Pop the next oldest message. This re-pushes any messages from the
1060 // reader.
1061 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001062 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001063 next_oldest_message = std::get<2>(next_oldest_message_reader)
1064 ->PopOldest(channel_index_);
1065
1066 // And make sure the message matches in it's entirety.
1067 CHECK(std::get<2>(oldest_message).span() ==
1068 std::get<2>(next_oldest_message).span())
1069 << ": Data at the same timestamp doesn't match.";
1070 } else {
1071 break;
1072 }
1073 }
1074
1075 return oldest_message;
1076}
1077
1078std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001079 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001080TimestampMerger::PopTimestampHeap() {
1081 // Pop the oldest message reader pointer off the heap.
1082 CHECK_GT(timestamp_heap_.size(), 0u);
1083
1084 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
1085 oldest_timestamp_reader = timestamp_heap_.front();
1086
1087 std::pop_heap(timestamp_heap_.begin(), timestamp_heap_.end(),
1088 &SplitMessageReaderHeapCompare);
1089 timestamp_heap_.pop_back();
1090
1091 CHECK(node_index_ != -1) << ": Timestamps in a single node environment";
1092
1093 // Pop the oldest message. This re-pushes any timestamps from the reader to
1094 // the timestamp heap.
1095 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001096 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001097 oldest_timestamp = std::get<2>(oldest_timestamp_reader)
Austin Schuh2f8fd752020-09-01 22:38:28 -07001098 ->PopOldestTimestamp(channel_index_, node_index_);
Austin Schuh6f3babe2020-01-26 20:34:50 -08001099
1100 // Confirm that the time we have recorded matches.
1101 CHECK_EQ(std::get<0>(oldest_timestamp), std::get<0>(oldest_timestamp_reader));
1102 CHECK_EQ(std::get<1>(oldest_timestamp), std::get<1>(oldest_timestamp_reader));
1103
Austin Schuh2f8fd752020-09-01 22:38:28 -07001104 // Now, keep reading until we have found all duplicates.
1105 while (!timestamp_heap_.empty()) {
1106 // See if it is a duplicate.
1107 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
1108 next_oldest_timestamp_reader = timestamp_heap_.front();
Austin Schuh6f3babe2020-01-26 20:34:50 -08001109
Austin Schuh2f8fd752020-09-01 22:38:28 -07001110 std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
1111 next_oldest_timestamp_time =
1112 std::get<2>(next_oldest_timestamp_reader)
1113 ->oldest_message(channel_index_, node_index_);
Austin Schuh6f3babe2020-01-26 20:34:50 -08001114
Austin Schuh2f8fd752020-09-01 22:38:28 -07001115 if (std::get<0>(next_oldest_timestamp_time) ==
1116 std::get<0>(oldest_timestamp) &&
1117 std::get<1>(next_oldest_timestamp_time) ==
1118 std::get<1>(oldest_timestamp)) {
1119 // Pop the timestamp reader pointer.
1120 std::pop_heap(timestamp_heap_.begin(), timestamp_heap_.end(),
1121 &SplitMessageReaderHeapCompare);
1122 timestamp_heap_.pop_back();
1123
1124 // Pop the next oldest timestamp. This re-pushes any messages from the
1125 // reader.
1126 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001127 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh2f8fd752020-09-01 22:38:28 -07001128 next_oldest_timestamp =
1129 std::get<2>(next_oldest_timestamp_reader)
1130 ->PopOldestTimestamp(channel_index_, node_index_);
1131
1132 // And make sure the contents matches in it's entirety.
1133 CHECK(std::get<2>(oldest_timestamp).span() ==
1134 std::get<2>(next_oldest_timestamp).span())
1135 << ": Data at the same timestamp doesn't match, "
1136 << aos::FlatbufferToJson(std::get<2>(oldest_timestamp)) << " vs "
1137 << aos::FlatbufferToJson(std::get<2>(next_oldest_timestamp)) << " "
1138 << absl::BytesToHexString(std::string_view(
1139 reinterpret_cast<const char *>(
1140 std::get<2>(oldest_timestamp).span().data()),
1141 std::get<2>(oldest_timestamp).span().size()))
1142 << " vs "
1143 << absl::BytesToHexString(std::string_view(
1144 reinterpret_cast<const char *>(
1145 std::get<2>(next_oldest_timestamp).span().data()),
1146 std::get<2>(next_oldest_timestamp).span().size()));
1147
1148 } else {
1149 break;
1150 }
Austin Schuh8bd96322020-02-13 21:18:22 -08001151 }
1152
Austin Schuh2f8fd752020-09-01 22:38:28 -07001153 return oldest_timestamp;
Austin Schuh8bd96322020-02-13 21:18:22 -08001154}
1155
Austin Schuhadd6eb32020-11-09 21:24:26 -08001156std::tuple<TimestampMerger::DeliveryTimestamp,
1157 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001158TimestampMerger::PopOldest() {
1159 if (has_timestamps_) {
Austin Schuh2f8fd752020-09-01 22:38:28 -07001160 VLOG(1) << "Looking for matching timestamp for "
1161 << configuration::StrippedChannelToString(
1162 configuration_->channels()->Get(channel_index_))
1163 << " (" << channel_index_ << ") "
1164 << " at " << std::get<0>(oldest_timestamp());
1165
Austin Schuh8bd96322020-02-13 21:18:22 -08001166 // Read the timestamps.
Austin Schuh6f3babe2020-01-26 20:34:50 -08001167 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001168 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001169 oldest_timestamp = PopTimestampHeap();
1170
1171 TimestampMerger::DeliveryTimestamp timestamp;
1172 timestamp.monotonic_event_time =
1173 monotonic_clock::time_point(chrono::nanoseconds(
1174 std::get<2>(oldest_timestamp).message().monotonic_sent_time()));
1175 timestamp.realtime_event_time =
1176 realtime_clock::time_point(chrono::nanoseconds(
1177 std::get<2>(oldest_timestamp).message().realtime_sent_time()));
Austin Schuh8d7e0bb2020-10-02 17:57:00 -07001178 timestamp.queue_index =
1179 std::get<2>(oldest_timestamp).message().queue_index();
Austin Schuh6f3babe2020-01-26 20:34:50 -08001180
1181 // Consistency check.
1182 CHECK_EQ(timestamp.monotonic_event_time, std::get<0>(oldest_timestamp));
1183 CHECK_EQ(std::get<2>(oldest_timestamp).message().queue_index(),
1184 std::get<1>(oldest_timestamp));
1185
1186 monotonic_clock::time_point remote_timestamp_monotonic_time(
1187 chrono::nanoseconds(
1188 std::get<2>(oldest_timestamp).message().monotonic_remote_time()));
1189
Austin Schuh8bd96322020-02-13 21:18:22 -08001190 // See if we have any data. If not, pass the problem up the chain.
Brian Silverman8a32ce62020-08-12 12:02:38 -07001191 if (message_heap_.empty()) {
Austin Schuhee711052020-08-24 16:06:09 -07001192 LOG(WARNING) << MaybeNodeName(configuration_->nodes()->Get(node_index_))
1193 << "No data to match timestamp on "
1194 << configuration::CleanedChannelToString(
1195 configuration_->channels()->Get(channel_index_))
1196 << " (" << channel_index_ << ")";
Austin Schuh8bd96322020-02-13 21:18:22 -08001197 return std::make_tuple(timestamp,
1198 std::move(std::get<2>(oldest_timestamp)));
1199 }
1200
Austin Schuh6f3babe2020-01-26 20:34:50 -08001201 while (true) {
Austin Schuhcde938c2020-02-02 17:30:07 -08001202 {
1203 // Ok, now try grabbing data until we find one which matches.
1204 std::tuple<monotonic_clock::time_point, uint32_t, const MessageHeader *>
1205 oldest_message_ref = oldest_message();
1206
1207 // Time at which the message was sent (this message is written from the
1208 // sending node's perspective.
1209 monotonic_clock::time_point remote_monotonic_time(chrono::nanoseconds(
1210 std::get<2>(oldest_message_ref)->monotonic_sent_time()));
1211
1212 if (remote_monotonic_time < remote_timestamp_monotonic_time) {
Austin Schuhee711052020-08-24 16:06:09 -07001213 LOG(WARNING) << configuration_->nodes()
1214 ->Get(node_index_)
1215 ->name()
1216 ->string_view()
1217 << " Undelivered message, skipping. Remote time is "
1218 << remote_monotonic_time << " timestamp is "
1219 << remote_timestamp_monotonic_time << " on channel "
1220 << configuration::StrippedChannelToString(
1221 configuration_->channels()->Get(channel_index_))
1222 << " (" << channel_index_ << ")";
Austin Schuhcde938c2020-02-02 17:30:07 -08001223 PopMessageHeap();
1224 continue;
1225 } else if (remote_monotonic_time > remote_timestamp_monotonic_time) {
Austin Schuhee711052020-08-24 16:06:09 -07001226 LOG(WARNING) << configuration_->nodes()
1227 ->Get(node_index_)
1228 ->name()
1229 ->string_view()
1230 << " Data not found. Remote time should be "
1231 << remote_timestamp_monotonic_time
1232 << ", message time is " << remote_monotonic_time
1233 << " on channel "
1234 << configuration::StrippedChannelToString(
1235 configuration_->channels()->Get(channel_index_))
Austin Schuh2f8fd752020-09-01 22:38:28 -07001236 << " (" << channel_index_ << ")"
1237 << (VLOG_IS_ON(1) ? DebugString() : "");
Austin Schuhcde938c2020-02-02 17:30:07 -08001238 return std::make_tuple(timestamp,
1239 std::move(std::get<2>(oldest_timestamp)));
1240 }
1241
1242 timestamp.monotonic_remote_time = remote_monotonic_time;
1243 }
1244
Austin Schuh2f8fd752020-09-01 22:38:28 -07001245 VLOG(1) << "Found matching data "
1246 << configuration::StrippedChannelToString(
1247 configuration_->channels()->Get(channel_index_))
1248 << " (" << channel_index_ << ")";
Austin Schuh6f3babe2020-01-26 20:34:50 -08001249 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001250 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001251 oldest_message = PopMessageHeap();
1252
Austin Schuh6f3babe2020-01-26 20:34:50 -08001253 timestamp.realtime_remote_time =
1254 realtime_clock::time_point(chrono::nanoseconds(
1255 std::get<2>(oldest_message).message().realtime_sent_time()));
1256 timestamp.remote_queue_index =
1257 std::get<2>(oldest_message).message().queue_index();
1258
Austin Schuhcde938c2020-02-02 17:30:07 -08001259 CHECK_EQ(timestamp.monotonic_remote_time,
1260 remote_timestamp_monotonic_time);
1261
1262 CHECK_EQ(timestamp.remote_queue_index,
1263 std::get<2>(oldest_timestamp).message().remote_queue_index())
1264 << ": " << FlatbufferToJson(&std::get<2>(oldest_timestamp).message())
1265 << " data "
1266 << FlatbufferToJson(&std::get<2>(oldest_message).message());
Austin Schuh6f3babe2020-01-26 20:34:50 -08001267
Austin Schuh30dd5c52020-08-01 14:43:44 -07001268 return std::make_tuple(timestamp, std::move(std::get<2>(oldest_message)));
Austin Schuh6f3babe2020-01-26 20:34:50 -08001269 }
1270 } else {
1271 std::tuple<monotonic_clock::time_point, uint32_t,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001272 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001273 oldest_message = PopMessageHeap();
1274
1275 TimestampMerger::DeliveryTimestamp timestamp;
1276 timestamp.monotonic_event_time =
1277 monotonic_clock::time_point(chrono::nanoseconds(
1278 std::get<2>(oldest_message).message().monotonic_sent_time()));
1279 timestamp.realtime_event_time =
1280 realtime_clock::time_point(chrono::nanoseconds(
1281 std::get<2>(oldest_message).message().realtime_sent_time()));
Austin Schuh8d7e0bb2020-10-02 17:57:00 -07001282 timestamp.queue_index = std::get<2>(oldest_message).message().queue_index();
Austin Schuh6f3babe2020-01-26 20:34:50 -08001283 timestamp.remote_queue_index = 0xffffffff;
1284
1285 CHECK_EQ(std::get<0>(oldest_message), timestamp.monotonic_event_time);
1286 CHECK_EQ(std::get<1>(oldest_message),
1287 std::get<2>(oldest_message).message().queue_index());
1288
Austin Schuh30dd5c52020-08-01 14:43:44 -07001289 return std::make_tuple(timestamp, std::move(std::get<2>(oldest_message)));
Austin Schuh6f3babe2020-01-26 20:34:50 -08001290 }
1291}
1292
Austin Schuh8bd96322020-02-13 21:18:22 -08001293void TimestampMerger::NoticeAtEnd() { channel_merger_->NoticeAtEnd(); }
1294
Austin Schuh6f3babe2020-01-26 20:34:50 -08001295namespace {
1296std::vector<std::unique_ptr<SplitMessageReader>> MakeSplitMessageReaders(
1297 const std::vector<std::vector<std::string>> &filenames) {
1298 CHECK_GT(filenames.size(), 0u);
1299 // Build up all the SplitMessageReaders.
1300 std::vector<std::unique_ptr<SplitMessageReader>> result;
1301 for (const std::vector<std::string> &filenames : filenames) {
1302 result.emplace_back(std::make_unique<SplitMessageReader>(filenames));
1303 }
1304 return result;
1305}
1306} // namespace
1307
1308ChannelMerger::ChannelMerger(
1309 const std::vector<std::vector<std::string>> &filenames)
1310 : split_message_readers_(MakeSplitMessageReaders(filenames)),
Austin Schuh97789fc2020-08-01 14:42:45 -07001311 log_file_header_(split_message_readers_[0]->raw_log_file_header()) {
Austin Schuh6f3babe2020-01-26 20:34:50 -08001312 // Now, confirm that the configuration matches for each and pick a start time.
1313 // Also return the list of possible nodes.
1314 for (const std::unique_ptr<SplitMessageReader> &reader :
1315 split_message_readers_) {
1316 CHECK(CompareFlatBuffer(log_file_header_.message().configuration(),
1317 reader->log_file_header()->configuration()))
1318 << ": Replaying log files with different configurations isn't "
1319 "supported";
1320 }
1321
1322 nodes_ = configuration::GetNodes(configuration());
1323}
1324
1325bool ChannelMerger::SetNode(const Node *target_node) {
1326 std::vector<SplitMessageReader *> split_message_readers;
1327 for (const std::unique_ptr<SplitMessageReader> &reader :
1328 split_message_readers_) {
1329 split_message_readers.emplace_back(reader.get());
1330 }
1331
1332 // Go find a log_file_header for this node.
1333 {
1334 bool found_node = false;
1335
1336 for (const std::unique_ptr<SplitMessageReader> &reader :
1337 split_message_readers_) {
James Kuszmaulfc273dc2020-05-09 17:56:19 -07001338 // In order to identify which logfile(s) map to the target node, do a
1339 // logical comparison of the nodes, by confirming that we are either in a
1340 // single-node setup (where the nodes will both be nullptr) or that the
1341 // node names match (but the other node fields--e.g., hostname lists--may
1342 // not).
1343 const bool both_null =
1344 reader->node() == nullptr && target_node == nullptr;
1345 const bool both_have_name =
1346 (reader->node() != nullptr) && (target_node != nullptr) &&
1347 (reader->node()->has_name() && target_node->has_name());
1348 const bool node_names_identical =
Brian Silvermand90905f2020-09-23 14:42:56 -07001349 both_have_name && (reader->node()->name()->string_view() ==
1350 target_node->name()->string_view());
James Kuszmaulfc273dc2020-05-09 17:56:19 -07001351 if (both_null || node_names_identical) {
Austin Schuh6f3babe2020-01-26 20:34:50 -08001352 if (!found_node) {
1353 found_node = true;
Austin Schuhadd6eb32020-11-09 21:24:26 -08001354 log_file_header_ = reader->raw_log_file_header();
Austin Schuhcde938c2020-02-02 17:30:07 -08001355 VLOG(1) << "Found log file " << reader->filename() << " with node "
1356 << FlatbufferToJson(reader->node()) << " start_time "
1357 << monotonic_start_time();
Austin Schuh6f3babe2020-01-26 20:34:50 -08001358 } else {
Austin Schuh2f8fd752020-09-01 22:38:28 -07001359 // Find the earliest start time. That way, if we get a full log file
1360 // directly from the node, and a partial later, we start with the
1361 // full. Update our header to match that.
1362 const monotonic_clock::time_point new_monotonic_start_time(
1363 chrono::nanoseconds(
1364 reader->log_file_header()->monotonic_start_time()));
1365 const realtime_clock::time_point new_realtime_start_time(
1366 chrono::nanoseconds(
1367 reader->log_file_header()->realtime_start_time()));
1368
1369 if (monotonic_start_time() == monotonic_clock::min_time ||
1370 (new_monotonic_start_time != monotonic_clock::min_time &&
1371 new_monotonic_start_time < monotonic_start_time())) {
1372 log_file_header_.mutable_message()->mutate_monotonic_start_time(
1373 new_monotonic_start_time.time_since_epoch().count());
1374 log_file_header_.mutable_message()->mutate_realtime_start_time(
1375 new_realtime_start_time.time_since_epoch().count());
1376 VLOG(1) << "Updated log file " << reader->filename()
1377 << " with node " << FlatbufferToJson(reader->node())
1378 << " start_time " << new_monotonic_start_time;
1379 }
Austin Schuh6f3babe2020-01-26 20:34:50 -08001380 }
1381 }
1382 }
1383
1384 if (!found_node) {
1385 LOG(WARNING) << "Failed to find log file for node "
1386 << FlatbufferToJson(target_node);
1387 return false;
1388 }
1389 }
1390
1391 // Build up all the timestamp mergers. This connects up all the
1392 // SplitMessageReaders.
1393 timestamp_mergers_.reserve(configuration()->channels()->size());
1394 for (size_t channel_index = 0;
1395 channel_index < configuration()->channels()->size(); ++channel_index) {
1396 timestamp_mergers_.emplace_back(
1397 configuration(), split_message_readers, channel_index,
1398 configuration::GetNode(configuration(), target_node), this);
1399 }
1400
1401 // And prime everything.
Austin Schuh6f3babe2020-01-26 20:34:50 -08001402 for (std::unique_ptr<SplitMessageReader> &split_message_reader :
1403 split_message_readers_) {
Austin Schuhcde938c2020-02-02 17:30:07 -08001404 split_message_reader->QueueMessages(
1405 split_message_reader->monotonic_start_time());
Austin Schuh6f3babe2020-01-26 20:34:50 -08001406 }
1407
1408 node_ = configuration::GetNodeOrDie(configuration(), target_node);
1409 return true;
1410}
1411
Austin Schuh858c9f32020-08-31 16:56:12 -07001412monotonic_clock::time_point ChannelMerger::OldestMessageTime() const {
Brian Silverman8a32ce62020-08-12 12:02:38 -07001413 if (channel_heap_.empty()) {
Austin Schuh6f3babe2020-01-26 20:34:50 -08001414 return monotonic_clock::max_time;
1415 }
1416 return channel_heap_.front().first;
1417}
1418
1419void ChannelMerger::PushChannelHeap(monotonic_clock::time_point timestamp,
1420 int channel_index) {
1421 // Pop and recreate the heap if it has already been pushed. And since we are
1422 // pushing again, we don't need to clear pushed.
1423 if (timestamp_mergers_[channel_index].pushed()) {
Brian Silverman8a32ce62020-08-12 12:02:38 -07001424 const auto channel_iterator = std::find_if(
Austin Schuh6f3babe2020-01-26 20:34:50 -08001425 channel_heap_.begin(), channel_heap_.end(),
1426 [channel_index](const std::pair<monotonic_clock::time_point, int> x) {
1427 return x.second == channel_index;
Brian Silverman8a32ce62020-08-12 12:02:38 -07001428 });
1429 DCHECK(channel_iterator != channel_heap_.end());
1430 if (std::get<0>(*channel_iterator) == timestamp) {
1431 // It's already in the heap, in the correct spot, so nothing
1432 // more for us to do here.
1433 return;
1434 }
1435 channel_heap_.erase(channel_iterator);
Austin Schuh6f3babe2020-01-26 20:34:50 -08001436 std::make_heap(channel_heap_.begin(), channel_heap_.end(),
1437 ChannelHeapCompare);
1438 }
1439
Austin Schuh2f8fd752020-09-01 22:38:28 -07001440 if (timestamp == monotonic_clock::min_time) {
1441 timestamp_mergers_[channel_index].set_pushed(false);
1442 return;
1443 }
1444
Austin Schuh05b70472020-01-01 17:11:17 -08001445 channel_heap_.push_back(std::make_pair(timestamp, channel_index));
1446
1447 // The default sort puts the newest message first. Use a custom comparator to
1448 // put the oldest message first.
1449 std::push_heap(channel_heap_.begin(), channel_heap_.end(),
1450 ChannelHeapCompare);
1451}
1452
Austin Schuh2f8fd752020-09-01 22:38:28 -07001453void ChannelMerger::VerifyHeaps() {
Austin Schuh661a8d82020-09-13 17:25:56 -07001454 std::vector<std::pair<monotonic_clock::time_point, int>> channel_heap =
1455 channel_heap_;
1456 std::make_heap(channel_heap.begin(), channel_heap.end(), &ChannelHeapCompare);
Austin Schuh2f8fd752020-09-01 22:38:28 -07001457
Austin Schuh661a8d82020-09-13 17:25:56 -07001458 for (size_t i = 0; i < channel_heap_.size(); ++i) {
1459 CHECK(channel_heap_[i] == channel_heap[i]) << ": Heaps diverged...";
1460 CHECK_EQ(
1461 std::get<0>(channel_heap[i]),
1462 timestamp_mergers_[std::get<1>(channel_heap[i])].channel_merger_time());
Austin Schuh2f8fd752020-09-01 22:38:28 -07001463 }
1464}
1465
Austin Schuh6f3babe2020-01-26 20:34:50 -08001466std::tuple<TimestampMerger::DeliveryTimestamp, int,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001467 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001468ChannelMerger::PopOldest() {
Austin Schuh8bd96322020-02-13 21:18:22 -08001469 CHECK_GT(channel_heap_.size(), 0u);
Austin Schuh05b70472020-01-01 17:11:17 -08001470 std::pair<monotonic_clock::time_point, int> oldest_channel_data =
1471 channel_heap_.front();
Austin Schuh6f3babe2020-01-26 20:34:50 -08001472 int channel_index = oldest_channel_data.second;
Austin Schuh05b70472020-01-01 17:11:17 -08001473 std::pop_heap(channel_heap_.begin(), channel_heap_.end(),
1474 &ChannelHeapCompare);
1475 channel_heap_.pop_back();
Austin Schuh8bd96322020-02-13 21:18:22 -08001476
Austin Schuh6f3babe2020-01-26 20:34:50 -08001477 timestamp_mergers_[channel_index].set_pushed(false);
Austin Schuh05b70472020-01-01 17:11:17 -08001478
Austin Schuh6f3babe2020-01-26 20:34:50 -08001479 TimestampMerger *merger = &timestamp_mergers_[channel_index];
Austin Schuh05b70472020-01-01 17:11:17 -08001480
Austin Schuhcde938c2020-02-02 17:30:07 -08001481 // Merger handles any queueing needed from here.
Austin Schuh6f3babe2020-01-26 20:34:50 -08001482 std::tuple<TimestampMerger::DeliveryTimestamp,
Austin Schuhadd6eb32020-11-09 21:24:26 -08001483 SizePrefixedFlatbufferVector<MessageHeader>>
Austin Schuh6f3babe2020-01-26 20:34:50 -08001484 message = merger->PopOldest();
Brian Silverman8a32ce62020-08-12 12:02:38 -07001485 DCHECK_EQ(std::get<0>(message).monotonic_event_time,
1486 oldest_channel_data.first)
1487 << ": channel_heap_ was corrupted for " << channel_index << ": "
1488 << DebugString();
Austin Schuh05b70472020-01-01 17:11:17 -08001489
Austin Schuh2f8fd752020-09-01 22:38:28 -07001490 CHECK_GE(std::get<0>(message).monotonic_event_time, last_popped_time_)
1491 << ": " << MaybeNodeName(log_file_header()->node())
1492 << "Messages came off the queue out of order. " << DebugString();
1493 last_popped_time_ = std::get<0>(message).monotonic_event_time;
1494
1495 VLOG(1) << "Popped " << last_popped_time_ << " "
1496 << configuration::StrippedChannelToString(
1497 configuration()->channels()->Get(channel_index))
1498 << " (" << channel_index << ")";
1499
Austin Schuh6f3babe2020-01-26 20:34:50 -08001500 return std::make_tuple(std::get<0>(message), channel_index,
1501 std::move(std::get<1>(message)));
1502}
1503
Austin Schuhcde938c2020-02-02 17:30:07 -08001504std::string SplitMessageReader::MessageHeaderQueue::DebugString() const {
1505 std::stringstream ss;
1506 for (size_t i = 0; i < data_.size(); ++i) {
Austin Schuh2f8fd752020-09-01 22:38:28 -07001507 if (i < 5 || i + 5 > data_.size()) {
1508 if (timestamps) {
1509 ss << " msg: ";
1510 } else {
1511 ss << " timestamp: ";
1512 }
1513 ss << monotonic_clock::time_point(
1514 chrono::nanoseconds(data_[i].message().monotonic_sent_time()))
Austin Schuhcde938c2020-02-02 17:30:07 -08001515 << " ("
Austin Schuh2f8fd752020-09-01 22:38:28 -07001516 << realtime_clock::time_point(
1517 chrono::nanoseconds(data_[i].message().realtime_sent_time()))
1518 << ") " << data_[i].message().queue_index();
1519 if (timestamps) {
1520 ss << " <- remote "
1521 << monotonic_clock::time_point(chrono::nanoseconds(
1522 data_[i].message().monotonic_remote_time()))
1523 << " ("
1524 << realtime_clock::time_point(chrono::nanoseconds(
1525 data_[i].message().realtime_remote_time()))
1526 << ")";
1527 }
1528 ss << "\n";
1529 } else if (i == 5) {
1530 ss << " ...\n";
Austin Schuh6f3babe2020-01-26 20:34:50 -08001531 }
Austin Schuhcde938c2020-02-02 17:30:07 -08001532 }
Austin Schuh6f3babe2020-01-26 20:34:50 -08001533
Austin Schuhcde938c2020-02-02 17:30:07 -08001534 return ss.str();
1535}
Austin Schuh6f3babe2020-01-26 20:34:50 -08001536
Austin Schuhcde938c2020-02-02 17:30:07 -08001537std::string SplitMessageReader::DebugString(int channel) const {
1538 std::stringstream ss;
1539 ss << "[\n";
1540 ss << channels_[channel].data.DebugString();
1541 ss << " ]";
1542 return ss.str();
1543}
Austin Schuh6f3babe2020-01-26 20:34:50 -08001544
Austin Schuhcde938c2020-02-02 17:30:07 -08001545std::string SplitMessageReader::DebugString(int channel, int node_index) const {
1546 std::stringstream ss;
1547 ss << "[\n";
1548 ss << channels_[channel].timestamps[node_index].DebugString();
1549 ss << " ]";
1550 return ss.str();
1551}
1552
1553std::string TimestampMerger::DebugString() const {
1554 std::stringstream ss;
1555
1556 if (timestamp_heap_.size() > 0) {
1557 ss << " timestamp_heap {\n";
1558 std::vector<
1559 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>>
1560 timestamp_heap = timestamp_heap_;
1561 while (timestamp_heap.size() > 0u) {
1562 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
1563 oldest_timestamp_reader = timestamp_heap.front();
1564
1565 ss << " " << std::get<2>(oldest_timestamp_reader) << " "
1566 << std::get<0>(oldest_timestamp_reader) << " queue_index ("
1567 << std::get<1>(oldest_timestamp_reader) << ") ttq "
1568 << std::get<2>(oldest_timestamp_reader)->time_to_queue() << " "
1569 << std::get<2>(oldest_timestamp_reader)->filename() << " -> "
1570 << std::get<2>(oldest_timestamp_reader)
1571 ->DebugString(channel_index_, node_index_)
1572 << "\n";
1573
1574 std::pop_heap(timestamp_heap.begin(), timestamp_heap.end(),
1575 &SplitMessageReaderHeapCompare);
1576 timestamp_heap.pop_back();
1577 }
1578 ss << " }\n";
1579 }
1580
1581 ss << " message_heap {\n";
1582 {
1583 std::vector<
1584 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>>
1585 message_heap = message_heap_;
Brian Silverman8a32ce62020-08-12 12:02:38 -07001586 while (!message_heap.empty()) {
Austin Schuhcde938c2020-02-02 17:30:07 -08001587 std::tuple<monotonic_clock::time_point, uint32_t, SplitMessageReader *>
1588 oldest_message_reader = message_heap.front();
1589
1590 ss << " " << std::get<2>(oldest_message_reader) << " "
1591 << std::get<0>(oldest_message_reader) << " queue_index ("
1592 << std::get<1>(oldest_message_reader) << ") ttq "
1593 << std::get<2>(oldest_message_reader)->time_to_queue() << " "
1594 << std::get<2>(oldest_message_reader)->filename() << " -> "
1595 << std::get<2>(oldest_message_reader)->DebugString(channel_index_)
1596 << "\n";
1597
1598 std::pop_heap(message_heap.begin(), message_heap.end(),
1599 &SplitMessageReaderHeapCompare);
1600 message_heap.pop_back();
Austin Schuh6f3babe2020-01-26 20:34:50 -08001601 }
Austin Schuh05b70472020-01-01 17:11:17 -08001602 }
Austin Schuhcde938c2020-02-02 17:30:07 -08001603 ss << " }";
1604
1605 return ss.str();
1606}
1607
1608std::string ChannelMerger::DebugString() const {
1609 std::stringstream ss;
1610 ss << "start_time " << realtime_start_time() << " " << monotonic_start_time()
1611 << "\n";
1612 ss << "channel_heap {\n";
1613 std::vector<std::pair<monotonic_clock::time_point, int>> channel_heap =
1614 channel_heap_;
Brian Silverman8a32ce62020-08-12 12:02:38 -07001615 while (!channel_heap.empty()) {
Austin Schuhcde938c2020-02-02 17:30:07 -08001616 std::tuple<monotonic_clock::time_point, int> channel = channel_heap.front();
1617 ss << " " << std::get<0>(channel) << " (" << std::get<1>(channel) << ") "
1618 << configuration::CleanedChannelToString(
1619 configuration()->channels()->Get(std::get<1>(channel)))
1620 << "\n";
1621
1622 ss << timestamp_mergers_[std::get<1>(channel)].DebugString() << "\n";
1623
1624 std::pop_heap(channel_heap.begin(), channel_heap.end(),
1625 &ChannelHeapCompare);
1626 channel_heap.pop_back();
1627 }
1628 ss << "}";
1629
1630 return ss.str();
Austin Schuh05b70472020-01-01 17:11:17 -08001631}
1632
Austin Schuhee711052020-08-24 16:06:09 -07001633std::string MaybeNodeName(const Node *node) {
1634 if (node != nullptr) {
1635 return node->name()->str() + " ";
1636 }
1637 return "";
1638}
1639
Brian Silvermanf51499a2020-09-21 12:49:08 -07001640} // namespace aos::logger