blob: ccd8bd77d0a65d1da02a844e84b9918c4ae6c84c [file] [log] [blame]
Austin Schuha36c8902019-12-30 18:07:15 -08001#ifndef AOS_EVENTS_LOGGING_LOGFILE_UTILS_H_
2#define AOS_EVENTS_LOGGING_LOGFILE_UTILS_H_
3
4#include <sys/uio.h>
5
Austin Schuh97789fc2020-08-01 14:42:45 -07006#include <chrono>
Austin Schuh05b70472020-01-01 17:11:17 -08007#include <deque>
Austin Schuh97789fc2020-08-01 14:42:45 -07008#include <limits>
9#include <memory>
Austin Schuh05b70472020-01-01 17:11:17 -080010#include <optional>
Austin Schuhfa895892020-01-07 20:07:41 -080011#include <string>
Austin Schuha36c8902019-12-30 18:07:15 -080012#include <string_view>
Brian Silverman98360e22020-04-28 16:51:20 -070013#include <tuple>
Austin Schuh97789fc2020-08-01 14:42:45 -070014#include <utility>
Austin Schuha36c8902019-12-30 18:07:15 -080015#include <vector>
16
Austin Schuh4b5c22a2020-11-30 22:58:43 -080017#include "absl/container/btree_set.h"
Austin Schuh05b70472020-01-01 17:11:17 -080018#include "absl/types/span.h"
Brian Silvermanf51499a2020-09-21 12:49:08 -070019#include "aos/containers/resizeable_buffer.h"
Austin Schuha36c8902019-12-30 18:07:15 -080020#include "aos/events/event_loop.h"
Austin Schuh2dc8c7d2021-07-01 17:41:28 -070021#include "aos/events/logging/boot_timestamp.h"
Brian Silvermanf51499a2020-09-21 12:49:08 -070022#include "aos/events/logging/buffer_encoder.h"
Austin Schuhc41603c2020-10-11 16:17:37 -070023#include "aos/events/logging/logfile_sorting.h"
Austin Schuha36c8902019-12-30 18:07:15 -080024#include "aos/events/logging/logger_generated.h"
Brian Silvermanf51499a2020-09-21 12:49:08 -070025#include "aos/flatbuffers.h"
Austin Schuha36c8902019-12-30 18:07:15 -080026#include "flatbuffers/flatbuffers.h"
27
Brian Silvermanf51499a2020-09-21 12:49:08 -070028namespace aos::logger {
Austin Schuha36c8902019-12-30 18:07:15 -080029
30enum class LogType : uint8_t {
31 // The message originated on this node and should be logged here.
32 kLogMessage,
33 // The message originated on another node, but only the delivery times are
34 // logged here.
35 kLogDeliveryTimeOnly,
36 // The message originated on another node. Log it and the delivery times
37 // together. The message_gateway is responsible for logging any messages
38 // which didn't get delivered.
Austin Schuh6f3babe2020-01-26 20:34:50 -080039 kLogMessageAndDeliveryTime,
40 // The message originated on the other node and should be logged on this node.
41 kLogRemoteMessage
Austin Schuha36c8902019-12-30 18:07:15 -080042};
43
Austin Schuha36c8902019-12-30 18:07:15 -080044// This class manages efficiently writing a sequence of detached buffers to a
Brian Silvermanf51499a2020-09-21 12:49:08 -070045// file. It encodes them, queues them up, and batches the write operation.
Austin Schuha36c8902019-12-30 18:07:15 -080046class DetachedBufferWriter {
47 public:
Brian Silvermana9f2ec92020-10-06 18:00:53 -070048 // Marker struct for one of our constructor overloads.
49 struct already_out_of_space_t {};
50
Brian Silvermanf51499a2020-09-21 12:49:08 -070051 DetachedBufferWriter(std::string_view filename,
52 std::unique_ptr<DetachedBufferEncoder> encoder);
Brian Silvermana9f2ec92020-10-06 18:00:53 -070053 // Creates a dummy instance which won't even open a file. It will act as if
54 // opening the file ran out of space immediately.
55 DetachedBufferWriter(already_out_of_space_t) : ran_out_of_space_(true) {}
Austin Schuh2f8fd752020-09-01 22:38:28 -070056 DetachedBufferWriter(DetachedBufferWriter &&other);
57 DetachedBufferWriter(const DetachedBufferWriter &) = delete;
58
Austin Schuha36c8902019-12-30 18:07:15 -080059 ~DetachedBufferWriter();
60
Austin Schuh2f8fd752020-09-01 22:38:28 -070061 DetachedBufferWriter &operator=(DetachedBufferWriter &&other);
Brian Silverman98360e22020-04-28 16:51:20 -070062 DetachedBufferWriter &operator=(const DetachedBufferWriter &) = delete;
63
Austin Schuh6f3babe2020-01-26 20:34:50 -080064 std::string_view filename() const { return filename_; }
65
Brian Silvermana9f2ec92020-10-06 18:00:53 -070066 // This will be true until Close() is called, unless the file couldn't be
67 // created due to running out of space.
68 bool is_open() const { return fd_ != -1; }
69
Brian Silvermanf51499a2020-09-21 12:49:08 -070070 // Queues up a finished FlatBufferBuilder to be encoded and written.
71 //
72 // Triggers a flush if there's enough data queued up.
73 //
74 // Steals the detached buffer from it.
Austin Schuhbd06ae42021-03-31 22:48:21 -070075 void QueueSizedFlatbuffer(flatbuffers::FlatBufferBuilder *fbb,
76 aos::monotonic_clock::time_point now) {
77 QueueSizedFlatbuffer(fbb->Release(), now);
Brian Silvermanf51499a2020-09-21 12:49:08 -070078 }
79 // May steal the backing storage of buffer, or may leave it alone.
Austin Schuhbd06ae42021-03-31 22:48:21 -070080 void QueueSizedFlatbuffer(flatbuffers::DetachedBuffer &&buffer,
81 aos::monotonic_clock::time_point now) {
82 QueueSizedFlatbuffer(std::move(buffer));
83 FlushAtThreshold(now);
84 }
85 // Unconditionally queues the buffer.
Brian Silvermanf51499a2020-09-21 12:49:08 -070086 void QueueSizedFlatbuffer(flatbuffers::DetachedBuffer &&buffer) {
Brian Silvermana9f2ec92020-10-06 18:00:53 -070087 if (ran_out_of_space_) {
88 return;
89 }
Brian Silvermanf51499a2020-09-21 12:49:08 -070090 encoder_->Encode(std::move(buffer));
Brian Silvermanf51499a2020-09-21 12:49:08 -070091 }
Austin Schuha36c8902019-12-30 18:07:15 -080092
Brian Silvermanf51499a2020-09-21 12:49:08 -070093 // Queues up data in span. May copy or may write it to disk immediately.
94 void QueueSpan(absl::Span<const uint8_t> span);
Austin Schuha36c8902019-12-30 18:07:15 -080095
Brian Silverman0465fcf2020-09-24 00:29:18 -070096 // Indicates we got ENOSPC when trying to write. After this returns true, no
97 // further data is written.
98 bool ran_out_of_space() const { return ran_out_of_space_; }
99
100 // To avoid silently failing to write logfiles, you must call this before
101 // destruction if ran_out_of_space() is true and the situation has been
102 // handled.
103 void acknowledge_out_of_space() {
104 CHECK(ran_out_of_space_);
105 acknowledge_ran_out_of_space_ = true;
106 }
107
108 // Fully flushes and closes the underlying file now. No additional data may be
109 // enqueued after calling this.
110 //
111 // This will be performed in the destructor automatically.
112 //
113 // Note that this may set ran_out_of_space().
114 void Close();
115
Brian Silvermanf51499a2020-09-21 12:49:08 -0700116 // Returns the total number of bytes written and currently queued.
Austin Schuha426f1f2021-03-31 22:27:41 -0700117 size_t total_bytes() const {
118 if (!encoder_) {
119 return 0;
120 }
121 return encoder_->total_bytes();
122 }
Austin Schuha36c8902019-12-30 18:07:15 -0800123
Brian Silvermanf51499a2020-09-21 12:49:08 -0700124 // The maximum time for a single write call, or 0 if none have been performed.
125 std::chrono::nanoseconds max_write_time() const { return max_write_time_; }
126 // The number of bytes in the longest write call, or -1 if none have been
127 // performed.
128 int max_write_time_bytes() const { return max_write_time_bytes_; }
129 // The number of buffers in the longest write call, or -1 if none have been
130 // performed.
131 int max_write_time_messages() const { return max_write_time_messages_; }
132 // The total time spent in write calls.
133 std::chrono::nanoseconds total_write_time() const {
134 return total_write_time_;
135 }
136 // The total number of writes which have been performed.
137 int total_write_count() const { return total_write_count_; }
138 // The total number of messages which have been written.
139 int total_write_messages() const { return total_write_messages_; }
140 // The total number of bytes which have been written.
141 int total_write_bytes() const { return total_write_bytes_; }
142 void ResetStatistics() {
143 max_write_time_ = std::chrono::nanoseconds::zero();
144 max_write_time_bytes_ = -1;
145 max_write_time_messages_ = -1;
146 total_write_time_ = std::chrono::nanoseconds::zero();
147 total_write_count_ = 0;
148 total_write_messages_ = 0;
149 total_write_bytes_ = 0;
150 }
Brian Silverman98360e22020-04-28 16:51:20 -0700151
Austin Schuha36c8902019-12-30 18:07:15 -0800152 private:
Brian Silvermanf51499a2020-09-21 12:49:08 -0700153 // Performs a single writev call with as much of the data we have queued up as
154 // possible.
155 //
156 // This will normally take all of the data we have queued up, unless an
157 // encoder has spit out a big enough chunk all at once that we can't manage
158 // all of it.
159 void Flush();
160
Brian Silverman0465fcf2020-09-24 00:29:18 -0700161 // write_return is what write(2) or writev(2) returned. write_size is the
162 // number of bytes we expected it to write.
163 void HandleWriteReturn(ssize_t write_return, size_t write_size);
164
Brian Silvermanf51499a2020-09-21 12:49:08 -0700165 void UpdateStatsForWrite(aos::monotonic_clock::duration duration,
166 ssize_t written, int iovec_size);
167
168 // Flushes data if we've reached the threshold to do that as part of normal
Austin Schuhbd06ae42021-03-31 22:48:21 -0700169 // operation either due to the outstanding queued data, or because we have
170 // passed our flush period. now is the current time to save some CPU grabbing
171 // the current time. It just needs to be close.
172 void FlushAtThreshold(aos::monotonic_clock::time_point now);
Brian Silvermanf51499a2020-09-21 12:49:08 -0700173
Austin Schuh2f8fd752020-09-01 22:38:28 -0700174 std::string filename_;
Brian Silvermanf51499a2020-09-21 12:49:08 -0700175 std::unique_ptr<DetachedBufferEncoder> encoder_;
Austin Schuh6f3babe2020-01-26 20:34:50 -0800176
Austin Schuha36c8902019-12-30 18:07:15 -0800177 int fd_ = -1;
Brian Silverman0465fcf2020-09-24 00:29:18 -0700178 bool ran_out_of_space_ = false;
179 bool acknowledge_ran_out_of_space_ = false;
Austin Schuha36c8902019-12-30 18:07:15 -0800180
Austin Schuha36c8902019-12-30 18:07:15 -0800181 // List of iovecs to use with writev. This is a member variable to avoid
182 // churn.
183 std::vector<struct iovec> iovec_;
Brian Silvermanf51499a2020-09-21 12:49:08 -0700184
185 std::chrono::nanoseconds max_write_time_ = std::chrono::nanoseconds::zero();
186 int max_write_time_bytes_ = -1;
187 int max_write_time_messages_ = -1;
188 std::chrono::nanoseconds total_write_time_ = std::chrono::nanoseconds::zero();
189 int total_write_count_ = 0;
190 int total_write_messages_ = 0;
191 int total_write_bytes_ = 0;
Austin Schuhbd06ae42021-03-31 22:48:21 -0700192
193 aos::monotonic_clock::time_point last_flush_time_ =
194 aos::monotonic_clock::min_time;
Austin Schuha36c8902019-12-30 18:07:15 -0800195};
196
197// Packes a message pointed to by the context into a MessageHeader.
198flatbuffers::Offset<MessageHeader> PackMessage(
199 flatbuffers::FlatBufferBuilder *fbb, const Context &context,
200 int channel_index, LogType log_type);
201
Austin Schuh05b70472020-01-01 17:11:17 -0800202// Class to read chunks out of a log file.
203class SpanReader {
204 public:
205 SpanReader(std::string_view filename);
Austin Schuha36c8902019-12-30 18:07:15 -0800206
Austin Schuh6f3babe2020-01-26 20:34:50 -0800207 std::string_view filename() const { return filename_; }
208
Austin Schuhcf5f6442021-07-06 10:43:28 -0700209 // Returns a span with the data for the next message from the log file,
210 // including the size. The result is only guarenteed to be valid until
211 // ReadMessage() or PeekMessage() is called again.
Austin Schuh05b70472020-01-01 17:11:17 -0800212 absl::Span<const uint8_t> ReadMessage();
213
Austin Schuhcf5f6442021-07-06 10:43:28 -0700214 // Returns a span with the data for the next message without consuming it.
215 // Multiple calls to PeekMessage return the same data. ReadMessage or
216 // ConsumeMessage must be called to get the next message.
217 absl::Span<const uint8_t> PeekMessage();
218 // Consumes the message so the next call to ReadMessage or PeekMessage returns
219 // new data. This does not invalidate the data.
220 void ConsumeMessage();
221
Austin Schuh05b70472020-01-01 17:11:17 -0800222 private:
223 // TODO(austin): Optimization:
224 // Allocate the 256k blocks like we do today. But, refcount them with
225 // shared_ptr pointed to by the messageheader that is returned. This avoids
226 // the copy. Need to do more benchmarking.
Brian Silvermanf51499a2020-09-21 12:49:08 -0700227 // And (Brian): Consider just mmapping the file and handing out refcounted
228 // pointers into that too.
Austin Schuh05b70472020-01-01 17:11:17 -0800229
230 // Reads a chunk of data into data_. Returns false if no data was read.
231 bool ReadBlock();
232
Austin Schuhc41603c2020-10-11 16:17:37 -0700233 std::string filename_;
Austin Schuh6f3babe2020-01-26 20:34:50 -0800234
Brian Silvermanf51499a2020-09-21 12:49:08 -0700235 // File reader and data decoder.
236 std::unique_ptr<DataDecoder> decoder_;
Austin Schuh05b70472020-01-01 17:11:17 -0800237
Brian Silvermanf51499a2020-09-21 12:49:08 -0700238 // Vector to read into.
239 ResizeableBuffer data_;
Austin Schuh05b70472020-01-01 17:11:17 -0800240
241 // Amount of data consumed already in data_.
242 size_t consumed_data_ = 0;
Austin Schuh05b70472020-01-01 17:11:17 -0800243};
244
Brian Silvermanfee16972021-09-14 12:06:38 -0700245// Reads the last header from a log file. This handles any duplicate headers
246// that were written.
247std::optional<SizePrefixedFlatbufferVector<LogFileHeader>> ReadHeader(
248 SpanReader *span_reader);
249std::optional<SizePrefixedFlatbufferVector<LogFileHeader>> ReadHeader(
250 std::string_view filename);
251// Reads the Nth message from a log file, excluding the header. Note: this
252// doesn't handle duplicate headers.
253std::optional<SizePrefixedFlatbufferVector<MessageHeader>> ReadNthMessage(
254 std::string_view filename, size_t n);
255
Tyler Chatowb7c6eba2021-07-28 14:43:23 -0700256class UnpackedMessageHeader;
257
Austin Schuh05b70472020-01-01 17:11:17 -0800258// Class which handles reading the header and messages from the log file. This
259// handles any per-file state left before merging below.
260class MessageReader {
261 public:
262 MessageReader(std::string_view filename);
263
Austin Schuh6f3babe2020-01-26 20:34:50 -0800264 std::string_view filename() const { return span_reader_.filename(); }
265
Austin Schuh05b70472020-01-01 17:11:17 -0800266 // Returns the header from the log file.
267 const LogFileHeader *log_file_header() const {
Austin Schuh97789fc2020-08-01 14:42:45 -0700268 return &raw_log_file_header_.message();
269 }
270
271 // Returns the raw data of the header from the log file.
Austin Schuhadd6eb32020-11-09 21:24:26 -0800272 const SizePrefixedFlatbufferVector<LogFileHeader> &raw_log_file_header()
273 const {
Austin Schuh97789fc2020-08-01 14:42:45 -0700274 return raw_log_file_header_;
Austin Schuh05b70472020-01-01 17:11:17 -0800275 }
276
277 // Returns the minimum maount of data needed to queue up for sorting before
278 // ware guarenteed to not see data out of order.
279 std::chrono::nanoseconds max_out_of_order_duration() const {
280 return max_out_of_order_duration_;
281 }
282
Austin Schuhcde938c2020-02-02 17:30:07 -0800283 // Returns the newest timestamp read out of the log file.
Austin Schuh05b70472020-01-01 17:11:17 -0800284 monotonic_clock::time_point newest_timestamp() const {
285 return newest_timestamp_;
286 }
287
288 // Returns the next message if there is one.
Tyler Chatowb7c6eba2021-07-28 14:43:23 -0700289 std::shared_ptr<UnpackedMessageHeader> ReadMessage();
Austin Schuh05b70472020-01-01 17:11:17 -0800290
291 // The time at which we need to read another chunk from the logfile.
292 monotonic_clock::time_point queue_data_time() const {
293 return newest_timestamp() - max_out_of_order_duration();
294 }
295
296 private:
297 // Log chunk reader.
298 SpanReader span_reader_;
299
Austin Schuh97789fc2020-08-01 14:42:45 -0700300 // Vector holding the raw data for the log file header.
Austin Schuhadd6eb32020-11-09 21:24:26 -0800301 SizePrefixedFlatbufferVector<LogFileHeader> raw_log_file_header_;
Austin Schuh05b70472020-01-01 17:11:17 -0800302
303 // Minimum amount of data to queue up for sorting before we are guarenteed
304 // to not see data out of order.
305 std::chrono::nanoseconds max_out_of_order_duration_;
306
307 // Timestamp of the newest message in a channel queue.
308 monotonic_clock::time_point newest_timestamp_ = monotonic_clock::min_time;
309};
310
Austin Schuhc41603c2020-10-11 16:17:37 -0700311// A class to seamlessly read messages from a list of part files.
312class PartsMessageReader {
313 public:
314 PartsMessageReader(LogParts log_parts);
315
316 std::string_view filename() const { return message_reader_.filename(); }
317
Austin Schuhd2f96102020-12-01 20:27:29 -0800318 // Returns the LogParts that holds the filenames we are reading.
319 const LogParts &parts() const { return parts_; }
320
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800321 const LogFileHeader *log_file_header() const {
322 return message_reader_.log_file_header();
323 }
324
Austin Schuhc41603c2020-10-11 16:17:37 -0700325 // Returns the minimum amount of data needed to queue up for sorting before
326 // we are guarenteed to not see data out of order.
327 std::chrono::nanoseconds max_out_of_order_duration() const {
328 return message_reader_.max_out_of_order_duration();
329 }
330
331 // Returns the newest timestamp read out of the log file.
332 monotonic_clock::time_point newest_timestamp() const {
333 return newest_timestamp_;
334 }
335
336 // Returns the next message if there is one, or nullopt if we have reached the
337 // end of all the files.
338 // Note: reading the next message may change the max_out_of_order_duration().
Tyler Chatowb7c6eba2021-07-28 14:43:23 -0700339 std::shared_ptr<UnpackedMessageHeader> ReadMessage();
Austin Schuhc41603c2020-10-11 16:17:37 -0700340
Austin Schuh48507722021-07-17 17:29:24 -0700341 // Returns the boot count for the requested node, or std::nullopt if we don't
342 // know.
343 std::optional<size_t> boot_count(size_t node_index) const {
344 CHECK_GE(node_index, 0u);
345 CHECK_LT(node_index, boot_counts_.size());
346 return boot_counts_[node_index];
347 }
348
Austin Schuhc41603c2020-10-11 16:17:37 -0700349 private:
350 // Opens the next log and updates message_reader_. Sets done_ if there is
351 // nothing more to do.
352 void NextLog();
Austin Schuh48507722021-07-17 17:29:24 -0700353 void ComputeBootCounts();
Austin Schuhc41603c2020-10-11 16:17:37 -0700354
355 const LogParts parts_;
356 size_t next_part_index_ = 1u;
357 bool done_ = false;
358 MessageReader message_reader_;
Brian Silvermanfee16972021-09-14 12:06:38 -0700359 // We instantiate the next one early, to allow implementations to prefetch.
360 // TODO(Brian): To get optimal performance when downloading, this needs more
361 // communication with the implementation to prioritize the next part and add
362 // more parallelism when it helps. Maybe some kind of a queue of parts in
363 // order, and the implementation gets to pull however many make sense off the
364 // front?
365 std::optional<MessageReader> next_message_reader_;
Austin Schuhc41603c2020-10-11 16:17:37 -0700366
Austin Schuh315b96b2020-12-11 21:21:12 -0800367 // True after we have seen a message after the start of the log. The
368 // guarentees on logging essentially are that all data from before the
369 // starting time of the log may be arbitrarily out of order, but once we get
370 // max_out_of_order_duration past the start, everything will remain within
371 // max_out_of_order_duration. We shouldn't see anything before the start
372 // after we've seen a message that is at least max_out_of_order_duration after
373 // the start.
374 bool after_start_ = false;
375
Austin Schuhc41603c2020-10-11 16:17:37 -0700376 monotonic_clock::time_point newest_timestamp_ = monotonic_clock::min_time;
Austin Schuh48507722021-07-17 17:29:24 -0700377
378 // Per node boot counts.
379 std::vector<std::optional<size_t>> boot_counts_;
Austin Schuhc41603c2020-10-11 16:17:37 -0700380};
381
Tyler Chatowb7c6eba2021-07-28 14:43:23 -0700382// Stores MessageHeader as a flat header and inline, aligned block of data.
383class UnpackedMessageHeader {
384 public:
385 UnpackedMessageHeader(const UnpackedMessageHeader &) = delete;
386 UnpackedMessageHeader &operator=(const UnpackedMessageHeader &) = delete;
387
388 // The channel.
389 uint32_t channel_index = 0xffffffff;
390
391 monotonic_clock::time_point monotonic_sent_time;
392 realtime_clock::time_point realtime_sent_time;
393
394 // The local queue index.
395 uint32_t queue_index = 0xffffffff;
396
397 std::optional<std::chrono::nanoseconds> monotonic_remote_time;
398
399 std::optional<realtime_clock::time_point> realtime_remote_time;
400 std::optional<uint32_t> remote_queue_index;
401
402 // This field is defaulted in the flatbuffer, so we need to store both the
403 // possibly defaulted value and whether it is defaulted.
404 monotonic_clock::time_point monotonic_timestamp_time;
405 bool has_monotonic_timestamp_time;
406
407 static std::shared_ptr<UnpackedMessageHeader> MakeMessage(
408 const MessageHeader &message);
409
410 // Note: we are storing a span here because we need something to put in the
411 // SharedSpan pointer that RawSender takes. We are using the aliasing
412 // constructor of shared_ptr to avoid the allocation, and it needs a nice
413 // pointer to track.
414 absl::Span<const uint8_t> span;
415
416 char actual_data[];
417
418 private:
419 ~UnpackedMessageHeader() {}
420
421 static void DestroyAndFree(UnpackedMessageHeader *p) {
422 p->~UnpackedMessageHeader();
423 free(p);
424 }
425};
426
427std::ostream &operator<<(std::ostream &os,
428 const UnpackedMessageHeader &message);
429
Austin Schuh1be0ce42020-11-29 22:43:26 -0800430// Struct to hold a message as it gets sorted on a single node.
431struct Message {
432 // The channel.
433 uint32_t channel_index = 0xffffffff;
434 // The local queue index.
Austin Schuh58646e22021-08-23 23:51:46 -0700435 // TODO(austin): Technically the boot inside queue_index is redundant with
436 // timestamp. In practice, it is less error-prone to duplicate it. Maybe a
437 // function to return the combined struct?
438 BootQueueIndex queue_index;
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700439 // The local timestamp.
440 BootTimestamp timestamp;
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700441
Austin Schuh48507722021-07-17 17:29:24 -0700442 // Remote boot when this is a timestamp.
443 size_t monotonic_remote_boot = 0xffffff;
444
445 size_t monotonic_timestamp_boot = 0xffffff;
446
Tyler Chatowb7c6eba2021-07-28 14:43:23 -0700447 std::shared_ptr<UnpackedMessageHeader> data;
Austin Schuh1be0ce42020-11-29 22:43:26 -0800448
449 bool operator<(const Message &m2) const;
450 bool operator>=(const Message &m2) const;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800451 bool operator==(const Message &m2) const;
Austin Schuh1be0ce42020-11-29 22:43:26 -0800452};
453
454std::ostream &operator<<(std::ostream &os, const Message &m);
455
Austin Schuhd2f96102020-12-01 20:27:29 -0800456// Structure to hold a full message and all the timestamps, which may or may not
457// have been sent from a remote node. The remote_queue_index will be invalid if
458// this message is from the point of view of the node which sent it.
459struct TimestampedMessage {
460 uint32_t channel_index = 0xffffffff;
461
Austin Schuh58646e22021-08-23 23:51:46 -0700462 BootQueueIndex queue_index;
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700463 BootTimestamp monotonic_event_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800464 realtime_clock::time_point realtime_event_time = realtime_clock::min_time;
465
Austin Schuh58646e22021-08-23 23:51:46 -0700466 BootQueueIndex remote_queue_index;
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700467 BootTimestamp monotonic_remote_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800468 realtime_clock::time_point realtime_remote_time = realtime_clock::min_time;
469
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700470 BootTimestamp monotonic_timestamp_time;
Austin Schuh8bf1e632021-01-02 22:41:04 -0800471
Tyler Chatowb7c6eba2021-07-28 14:43:23 -0700472 std::shared_ptr<UnpackedMessageHeader> data;
Austin Schuhd2f96102020-12-01 20:27:29 -0800473};
474
475std::ostream &operator<<(std::ostream &os, const TimestampedMessage &m);
476
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800477// Class to sort the resulting messages from a PartsMessageReader.
478class LogPartsSorter {
479 public:
480 LogPartsSorter(LogParts log_parts);
481
Austin Schuh0ca51f32020-12-25 21:51:45 -0800482 // Returns the parts that this is sorting messages from.
483 const LogParts &parts() const { return parts_message_reader_.parts(); }
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800484
Austin Schuhd2f96102020-12-01 20:27:29 -0800485 monotonic_clock::time_point monotonic_start_time() const {
Austin Schuh0ca51f32020-12-25 21:51:45 -0800486 return parts().monotonic_start_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800487 }
488 realtime_clock::time_point realtime_start_time() const {
Austin Schuh0ca51f32020-12-25 21:51:45 -0800489 return parts().realtime_start_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800490 }
491
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800492 // The time this data is sorted until.
493 monotonic_clock::time_point sorted_until() const { return sorted_until_; }
494
495 // Returns the next sorted message from the log file. It is safe to call
496 // std::move() on the result to move the data flatbuffer from it.
497 Message *Front();
498 // Pops the front message. This should only be called after a call to
499 // Front().
500 void PopFront();
501
502 // Returns a debug string representing the contents of this sorter.
503 std::string DebugString() const;
504
505 private:
506 // Log parts reader we are wrapping.
507 PartsMessageReader parts_message_reader_;
508 // Cache of the time we are sorted until.
509 aos::monotonic_clock::time_point sorted_until_ = monotonic_clock::min_time;
510
Austin Schuhb000de62020-12-03 22:00:40 -0800511 // Timestamp of the last message returned. Used to make sure nothing goes
512 // backwards.
513 monotonic_clock::time_point last_message_time_ = monotonic_clock::min_time;
514
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800515 // Set used for efficient sorting of messages. We can benchmark and evaluate
516 // other data structures if this proves to be the bottleneck.
517 absl::btree_set<Message> messages_;
Austin Schuh48507722021-07-17 17:29:24 -0700518
519 // Mapping from channel to source node.
520 // TODO(austin): Should we put this in Boots so it can be cached for everyone?
521 std::vector<size_t> source_node_index_;
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800522};
523
Austin Schuh8f52ed52020-11-30 23:12:39 -0800524// Class to run merge sort on the messages from multiple LogPartsSorter
525// instances.
526class NodeMerger {
527 public:
Austin Schuhd2f96102020-12-01 20:27:29 -0800528 NodeMerger(std::vector<LogParts> parts);
529
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700530 // Copying and moving will mess up the internal raw pointers. Just don't do
531 // it.
532 NodeMerger(NodeMerger const &) = delete;
533 NodeMerger(NodeMerger &&) = delete;
534 void operator=(NodeMerger const &) = delete;
535 void operator=(NodeMerger &&) = delete;
536
Austin Schuhd2f96102020-12-01 20:27:29 -0800537 // Node index in the configuration of this node.
538 int node() const { return node_; }
Austin Schuh8f52ed52020-11-30 23:12:39 -0800539
Austin Schuh0ca51f32020-12-25 21:51:45 -0800540 // List of parts being sorted together.
541 std::vector<const LogParts *> Parts() const;
542
543 const Configuration *configuration() const {
544 return parts_sorters_[0].parts().config.get();
Austin Schuhd2f96102020-12-01 20:27:29 -0800545 }
546
547 monotonic_clock::time_point monotonic_start_time() const {
548 return monotonic_start_time_;
549 }
550 realtime_clock::time_point realtime_start_time() const {
551 return realtime_start_time_;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800552 }
Austin Schuh5dd22842021-11-17 16:09:39 -0800553 monotonic_clock::time_point monotonic_oldest_time() const {
554 return monotonic_oldest_time_;
555 }
Austin Schuh8f52ed52020-11-30 23:12:39 -0800556
557 // The time this data is sorted until.
558 monotonic_clock::time_point sorted_until() const { return sorted_until_; }
559
560 // Returns the next sorted message from the set of log files. It is safe to
561 // call std::move() on the result to move the data flatbuffer from it.
562 Message *Front();
563 // Pops the front message. This should only be called after a call to
564 // Front().
565 void PopFront();
566
567 private:
568 // Unsorted list of all parts sorters.
Austin Schuhd2f96102020-12-01 20:27:29 -0800569 std::vector<LogPartsSorter> parts_sorters_;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800570 // Pointer to the parts sorter holding the current Front message if one
571 // exists, or nullptr if a new one needs to be found.
572 LogPartsSorter *current_ = nullptr;
573 // Cached sorted_until value.
574 aos::monotonic_clock::time_point sorted_until_ = monotonic_clock::min_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800575
576 // Cached node.
577 int node_;
578
Austin Schuhb000de62020-12-03 22:00:40 -0800579 // Timestamp of the last message returned. Used to make sure nothing goes
580 // backwards.
581 monotonic_clock::time_point last_message_time_ = monotonic_clock::min_time;
582
Austin Schuhd2f96102020-12-01 20:27:29 -0800583 realtime_clock::time_point realtime_start_time_ = realtime_clock::max_time;
584 monotonic_clock::time_point monotonic_start_time_ = monotonic_clock::max_time;
Austin Schuh5dd22842021-11-17 16:09:39 -0800585 monotonic_clock::time_point monotonic_oldest_time_ = monotonic_clock::max_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800586};
587
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700588// Class to concatenate multiple boots worth of logs into a single per-node
589// stream.
590class BootMerger {
591 public:
592 BootMerger(std::vector<LogParts> file);
593
594 // Copying and moving will mess up the internal raw pointers. Just don't do
595 // it.
596 BootMerger(BootMerger const &) = delete;
597 BootMerger(BootMerger &&) = delete;
598 void operator=(BootMerger const &) = delete;
599 void operator=(BootMerger &&) = delete;
600
601 // Node index in the configuration of this node.
602 int node() const { return node_mergers_[0]->node(); }
603
604 // List of parts being sorted together.
605 std::vector<const LogParts *> Parts() const;
606
607 const Configuration *configuration() const {
608 return node_mergers_[0]->configuration();
609 }
610
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700611 monotonic_clock::time_point monotonic_start_time(size_t boot) const {
612 CHECK_LT(boot, node_mergers_.size());
613 return node_mergers_[boot]->monotonic_start_time();
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700614 }
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700615 realtime_clock::time_point realtime_start_time(size_t boot) const {
616 CHECK_LT(boot, node_mergers_.size());
617 return node_mergers_[boot]->realtime_start_time();
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700618 }
Austin Schuh5dd22842021-11-17 16:09:39 -0800619 monotonic_clock::time_point monotonic_oldest_time(size_t boot) const {
620 CHECK_LT(boot, node_mergers_.size());
621 return node_mergers_[boot]->monotonic_oldest_time();
622 }
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700623
624 bool started() const {
625 return node_mergers_[index_]->sorted_until() != monotonic_clock::min_time ||
626 index_ != 0;
627 }
628
629 // Returns the next sorted message from the set of log files. It is safe to
630 // call std::move() on the result to move the data flatbuffer from it.
631 Message *Front();
632 // Pops the front message. This should only be called after a call to
633 // Front().
634 void PopFront();
635
636 private:
637 int index_ = 0;
638
639 // TODO(austin): Sanjay points out this is pretty inefficient. Don't keep so
640 // many things open.
641 std::vector<std::unique_ptr<NodeMerger>> node_mergers_;
642};
643
Austin Schuhd2f96102020-12-01 20:27:29 -0800644// Class to match timestamps with the corresponding data from other nodes.
Austin Schuh79b30942021-01-24 22:32:21 -0800645//
646// This class also buffers data for the node it represents, and supports
647// notifying when new data is queued as well as queueing until a point in time.
Austin Schuhd2f96102020-12-01 20:27:29 -0800648class TimestampMapper {
649 public:
650 TimestampMapper(std::vector<LogParts> file);
651
652 // Copying and moving will mess up the internal raw pointers. Just don't do
653 // it.
654 TimestampMapper(TimestampMapper const &) = delete;
655 TimestampMapper(TimestampMapper &&) = delete;
656 void operator=(TimestampMapper const &) = delete;
657 void operator=(TimestampMapper &&) = delete;
658
659 // TODO(austin): It would be super helpful to provide a way to queue up to
660 // time X without matching timestamps, and to then be able to pull the
661 // timestamps out of this queue. This lets us bootstrap time estimation
662 // without exploding memory usage worst case.
663
Austin Schuh0ca51f32020-12-25 21:51:45 -0800664 const Configuration *configuration() const { return configuration_.get(); }
Austin Schuhd2f96102020-12-01 20:27:29 -0800665
666 // Returns which node this is sorting for.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700667 size_t node() const { return boot_merger_.node(); }
Austin Schuhd2f96102020-12-01 20:27:29 -0800668
669 // The start time of this log.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700670 monotonic_clock::time_point monotonic_start_time(size_t boot) const {
671 return boot_merger_.monotonic_start_time(boot);
Austin Schuhd2f96102020-12-01 20:27:29 -0800672 }
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700673 realtime_clock::time_point realtime_start_time(size_t boot) const {
674 return boot_merger_.realtime_start_time(boot);
Austin Schuhd2f96102020-12-01 20:27:29 -0800675 }
Austin Schuh5dd22842021-11-17 16:09:39 -0800676 // Returns the oldest timestamp on a message on this boot.
677 monotonic_clock::time_point monotonic_oldest_time(size_t boot) const {
678 return boot_merger_.monotonic_oldest_time(boot);
679 }
Austin Schuhd2f96102020-12-01 20:27:29 -0800680
681 // Uses timestamp_mapper as the peer for its node. Only one mapper may be set
682 // for each node. Peers are used to look up the data for timestamps on this
683 // node.
684 void AddPeer(TimestampMapper *timestamp_mapper);
685
Austin Schuh24bf4972021-06-29 22:09:08 -0700686 // Returns true if anything has been queued up.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700687 bool started() const { return boot_merger_.started(); }
Austin Schuhd2f96102020-12-01 20:27:29 -0800688
689 // Returns the next message for this node.
690 TimestampedMessage *Front();
691 // Pops the next message. Front must be called first.
692 void PopFront();
693
694 // Returns debug information about this node.
695 std::string DebugString() const;
696
Austin Schuh79b30942021-01-24 22:32:21 -0800697 // Queues data the provided time.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700698 void QueueUntil(BootTimestamp queue_time);
Austin Schuhe639ea12021-01-25 13:00:22 -0800699 // Queues until we have time_estimation_buffer of data in the queue.
700 void QueueFor(std::chrono::nanoseconds time_estimation_buffer);
Austin Schuh79b30942021-01-24 22:32:21 -0800701
Austin Schuh06601222021-01-26 17:02:50 -0800702 // Queues until the condition is met.
703 template <typename T>
704 void QueueUntilCondition(T fn) {
705 while (true) {
706 if (fn()) {
707 break;
708 }
709 if (!QueueMatched()) {
710 break;
711 }
712 }
713 }
714
Austin Schuh79b30942021-01-24 22:32:21 -0800715 // Sets a callback to be called whenever a full message is queued.
716 void set_timestamp_callback(std::function<void(TimestampedMessage *)> fn) {
717 timestamp_callback_ = fn;
718 }
719
Austin Schuhd2f96102020-12-01 20:27:29 -0800720 private:
721 // The state for a remote node. This holds the data that needs to be matched
722 // with the remote node's timestamps.
723 struct NodeData {
724 // True if we should save data here. This should be true if any of the
725 // bools in delivered below are true.
726 bool any_delivered = false;
727
Austin Schuh36c00932021-07-19 18:13:21 -0700728 // True if we have a peer and therefore should be saving data for it.
729 bool save_for_peer = false;
730
Austin Schuhd2f96102020-12-01 20:27:29 -0800731 // Peer pointer. This node is only to be considered if a peer is set.
732 TimestampMapper *peer = nullptr;
733
734 struct ChannelData {
735 // Deque per channel. This contains the data from the outside
736 // TimestampMapper node which is relevant for the node this NodeData
737 // points to.
738 std::deque<Message> messages;
739 // Bool tracking per channel if a message is delivered to the node this
740 // NodeData represents.
741 bool delivered = false;
742 };
743
744 // Vector with per channel data.
745 std::vector<ChannelData> channels;
746 };
747
748 // Returns (and forgets about) the data for the provided timestamp message
749 // showing when it was delivered to this node.
750 Message MatchingMessageFor(const Message &message);
751
752 // Queues up a single message into our message queue, and any nodes that this
753 // message is delivered to. Returns true if one was available, false
754 // otherwise.
755 bool Queue();
756
Austin Schuh79b30942021-01-24 22:32:21 -0800757 // Queues up a single matched message into our matched message queue. Returns
758 // true if one was queued, and false otherwise.
759 bool QueueMatched();
760
Austin Schuhd2f96102020-12-01 20:27:29 -0800761 // Queues up data until we have at least one message >= to time t.
762 // Useful for triggering a remote node to read enough data to have the
763 // timestamp you care about available.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700764 void QueueUnmatchedUntil(BootTimestamp t);
Austin Schuhd2f96102020-12-01 20:27:29 -0800765
Austin Schuh79b30942021-01-24 22:32:21 -0800766 // Queues m into matched_messages_.
767 void QueueMessage(Message *m);
Austin Schuhd2f96102020-12-01 20:27:29 -0800768
Austin Schuh58646e22021-08-23 23:51:46 -0700769 // Returns the name of the node this class is sorting for.
770 std::string_view node_name() const {
771 return configuration_->has_nodes() ? configuration_->nodes()
772 ->Get(boot_merger_.node())
773 ->name()
774 ->string_view()
775 : "(single node)";
776 }
777
Austin Schuhd2f96102020-12-01 20:27:29 -0800778 // The node merger to source messages from.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700779 BootMerger boot_merger_;
Austin Schuh0ca51f32020-12-25 21:51:45 -0800780
781 std::shared_ptr<const Configuration> configuration_;
782
Austin Schuhd2f96102020-12-01 20:27:29 -0800783 // The buffer of messages for this node. These are not matched with any
784 // remote data.
785 std::deque<Message> messages_;
786 // The node index for the source node for each channel.
787 std::vector<size_t> source_node_;
788
789 // Vector per node. Not all nodes will have anything.
790 std::vector<NodeData> nodes_data_;
791
792 // Latest message to return.
Austin Schuh79b30942021-01-24 22:32:21 -0800793 std::deque<TimestampedMessage> matched_messages_;
Austin Schuhd2f96102020-12-01 20:27:29 -0800794
Austin Schuh79b30942021-01-24 22:32:21 -0800795 // Tracks the state of the first message in matched_messages_. Do we need to
796 // update it, is it valid, or should we return nullptr?
Austin Schuhd2f96102020-12-01 20:27:29 -0800797 enum class FirstMessage {
798 kNeedsUpdate,
799 kInMessage,
800 kNullptr,
801 };
802 FirstMessage first_message_ = FirstMessage::kNeedsUpdate;
803
804 // Timestamp of the last message returned. Used to make sure nothing goes
805 // backwards.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700806 BootTimestamp last_message_time_ = BootTimestamp::min_time();
Austin Schuhd2f96102020-12-01 20:27:29 -0800807 // Time this node is queued up until. Used for caching.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700808 BootTimestamp queued_until_ = BootTimestamp::min_time();
Austin Schuh79b30942021-01-24 22:32:21 -0800809
810 std::function<void(TimestampedMessage *)> timestamp_callback_;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800811};
812
Austin Schuhee711052020-08-24 16:06:09 -0700813// Returns the node name with a trailing space, or an empty string if we are on
814// a single node.
815std::string MaybeNodeName(const Node *);
816
Brian Silvermanf51499a2020-09-21 12:49:08 -0700817} // namespace aos::logger
Austin Schuha36c8902019-12-30 18:07:15 -0800818
819#endif // AOS_EVENTS_LOGGING_LOGFILE_UTILS_H_