blob: 5c5f709b43a7fa03b9c5c40ad01897c396278b8f [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
Austin Schuh05b70472020-01-01 17:11:17 -0800256// Class which handles reading the header and messages from the log file. This
257// handles any per-file state left before merging below.
258class MessageReader {
259 public:
260 MessageReader(std::string_view filename);
261
Austin Schuh6f3babe2020-01-26 20:34:50 -0800262 std::string_view filename() const { return span_reader_.filename(); }
263
Austin Schuh05b70472020-01-01 17:11:17 -0800264 // Returns the header from the log file.
265 const LogFileHeader *log_file_header() const {
Austin Schuh97789fc2020-08-01 14:42:45 -0700266 return &raw_log_file_header_.message();
267 }
268
269 // Returns the raw data of the header from the log file.
Austin Schuhadd6eb32020-11-09 21:24:26 -0800270 const SizePrefixedFlatbufferVector<LogFileHeader> &raw_log_file_header()
271 const {
Austin Schuh97789fc2020-08-01 14:42:45 -0700272 return raw_log_file_header_;
Austin Schuh05b70472020-01-01 17:11:17 -0800273 }
274
275 // Returns the minimum maount of data needed to queue up for sorting before
276 // ware guarenteed to not see data out of order.
277 std::chrono::nanoseconds max_out_of_order_duration() const {
278 return max_out_of_order_duration_;
279 }
280
Austin Schuhcde938c2020-02-02 17:30:07 -0800281 // Returns the newest timestamp read out of the log file.
Austin Schuh05b70472020-01-01 17:11:17 -0800282 monotonic_clock::time_point newest_timestamp() const {
283 return newest_timestamp_;
284 }
285
286 // Returns the next message if there is one.
Austin Schuhadd6eb32020-11-09 21:24:26 -0800287 std::optional<SizePrefixedFlatbufferVector<MessageHeader>> ReadMessage();
Austin Schuh05b70472020-01-01 17:11:17 -0800288
289 // The time at which we need to read another chunk from the logfile.
290 monotonic_clock::time_point queue_data_time() const {
291 return newest_timestamp() - max_out_of_order_duration();
292 }
293
294 private:
295 // Log chunk reader.
296 SpanReader span_reader_;
297
Austin Schuh97789fc2020-08-01 14:42:45 -0700298 // Vector holding the raw data for the log file header.
Austin Schuhadd6eb32020-11-09 21:24:26 -0800299 SizePrefixedFlatbufferVector<LogFileHeader> raw_log_file_header_;
Austin Schuh05b70472020-01-01 17:11:17 -0800300
301 // Minimum amount of data to queue up for sorting before we are guarenteed
302 // to not see data out of order.
303 std::chrono::nanoseconds max_out_of_order_duration_;
304
305 // Timestamp of the newest message in a channel queue.
306 monotonic_clock::time_point newest_timestamp_ = monotonic_clock::min_time;
307};
308
Austin Schuhc41603c2020-10-11 16:17:37 -0700309// A class to seamlessly read messages from a list of part files.
310class PartsMessageReader {
311 public:
312 PartsMessageReader(LogParts log_parts);
313
314 std::string_view filename() const { return message_reader_.filename(); }
315
Austin Schuhd2f96102020-12-01 20:27:29 -0800316 // Returns the LogParts that holds the filenames we are reading.
317 const LogParts &parts() const { return parts_; }
318
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800319 const LogFileHeader *log_file_header() const {
320 return message_reader_.log_file_header();
321 }
322
Austin Schuhc41603c2020-10-11 16:17:37 -0700323 // Returns the minimum amount of data needed to queue up for sorting before
324 // we are guarenteed to not see data out of order.
325 std::chrono::nanoseconds max_out_of_order_duration() const {
326 return message_reader_.max_out_of_order_duration();
327 }
328
329 // Returns the newest timestamp read out of the log file.
330 monotonic_clock::time_point newest_timestamp() const {
331 return newest_timestamp_;
332 }
333
334 // Returns the next message if there is one, or nullopt if we have reached the
335 // end of all the files.
336 // Note: reading the next message may change the max_out_of_order_duration().
Austin Schuhadd6eb32020-11-09 21:24:26 -0800337 std::optional<SizePrefixedFlatbufferVector<MessageHeader>> ReadMessage();
Austin Schuhc41603c2020-10-11 16:17:37 -0700338
Austin Schuh48507722021-07-17 17:29:24 -0700339 // Returns the boot count for the requested node, or std::nullopt if we don't
340 // know.
341 std::optional<size_t> boot_count(size_t node_index) const {
342 CHECK_GE(node_index, 0u);
343 CHECK_LT(node_index, boot_counts_.size());
344 return boot_counts_[node_index];
345 }
346
Austin Schuhc41603c2020-10-11 16:17:37 -0700347 private:
348 // Opens the next log and updates message_reader_. Sets done_ if there is
349 // nothing more to do.
350 void NextLog();
Austin Schuh48507722021-07-17 17:29:24 -0700351 void ComputeBootCounts();
Austin Schuhc41603c2020-10-11 16:17:37 -0700352
353 const LogParts parts_;
354 size_t next_part_index_ = 1u;
355 bool done_ = false;
356 MessageReader message_reader_;
Brian Silvermanfee16972021-09-14 12:06:38 -0700357 // We instantiate the next one early, to allow implementations to prefetch.
358 // TODO(Brian): To get optimal performance when downloading, this needs more
359 // communication with the implementation to prioritize the next part and add
360 // more parallelism when it helps. Maybe some kind of a queue of parts in
361 // order, and the implementation gets to pull however many make sense off the
362 // front?
363 std::optional<MessageReader> next_message_reader_;
Austin Schuhc41603c2020-10-11 16:17:37 -0700364
Austin Schuh315b96b2020-12-11 21:21:12 -0800365 // True after we have seen a message after the start of the log. The
366 // guarentees on logging essentially are that all data from before the
367 // starting time of the log may be arbitrarily out of order, but once we get
368 // max_out_of_order_duration past the start, everything will remain within
369 // max_out_of_order_duration. We shouldn't see anything before the start
370 // after we've seen a message that is at least max_out_of_order_duration after
371 // the start.
372 bool after_start_ = false;
373
Austin Schuhc41603c2020-10-11 16:17:37 -0700374 monotonic_clock::time_point newest_timestamp_ = monotonic_clock::min_time;
Austin Schuh48507722021-07-17 17:29:24 -0700375
376 // Per node boot counts.
377 std::vector<std::optional<size_t>> boot_counts_;
Austin Schuhc41603c2020-10-11 16:17:37 -0700378};
379
Austin Schuh1be0ce42020-11-29 22:43:26 -0800380// Struct to hold a message as it gets sorted on a single node.
381struct Message {
382 // The channel.
383 uint32_t channel_index = 0xffffffff;
384 // The local queue index.
Austin Schuh58646e22021-08-23 23:51:46 -0700385 // TODO(austin): Technically the boot inside queue_index is redundant with
386 // timestamp. In practice, it is less error-prone to duplicate it. Maybe a
387 // function to return the combined struct?
388 BootQueueIndex queue_index;
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700389 // The local timestamp.
390 BootTimestamp timestamp;
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700391
Austin Schuh48507722021-07-17 17:29:24 -0700392 // Remote boot when this is a timestamp.
393 size_t monotonic_remote_boot = 0xffffff;
394
395 size_t monotonic_timestamp_boot = 0xffffff;
396
Austin Schuh1be0ce42020-11-29 22:43:26 -0800397 // The data (either a timestamp header, or a data header).
398 SizePrefixedFlatbufferVector<MessageHeader> data;
399
400 bool operator<(const Message &m2) const;
401 bool operator>=(const Message &m2) const;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800402 bool operator==(const Message &m2) const;
Austin Schuh1be0ce42020-11-29 22:43:26 -0800403};
404
405std::ostream &operator<<(std::ostream &os, const Message &m);
406
Austin Schuhd2f96102020-12-01 20:27:29 -0800407// Structure to hold a full message and all the timestamps, which may or may not
408// have been sent from a remote node. The remote_queue_index will be invalid if
409// this message is from the point of view of the node which sent it.
410struct TimestampedMessage {
411 uint32_t channel_index = 0xffffffff;
412
Austin Schuh58646e22021-08-23 23:51:46 -0700413 BootQueueIndex queue_index;
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700414 BootTimestamp monotonic_event_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800415 realtime_clock::time_point realtime_event_time = realtime_clock::min_time;
416
Austin Schuh58646e22021-08-23 23:51:46 -0700417 BootQueueIndex remote_queue_index;
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700418 BootTimestamp monotonic_remote_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800419 realtime_clock::time_point realtime_remote_time = realtime_clock::min_time;
420
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700421 BootTimestamp monotonic_timestamp_time;
Austin Schuh8bf1e632021-01-02 22:41:04 -0800422
Austin Schuhd2f96102020-12-01 20:27:29 -0800423 SizePrefixedFlatbufferVector<MessageHeader> data;
424};
425
426std::ostream &operator<<(std::ostream &os, const TimestampedMessage &m);
427
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800428// Class to sort the resulting messages from a PartsMessageReader.
429class LogPartsSorter {
430 public:
431 LogPartsSorter(LogParts log_parts);
432
Austin Schuh0ca51f32020-12-25 21:51:45 -0800433 // Returns the parts that this is sorting messages from.
434 const LogParts &parts() const { return parts_message_reader_.parts(); }
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800435
Austin Schuhd2f96102020-12-01 20:27:29 -0800436 monotonic_clock::time_point monotonic_start_time() const {
Austin Schuh0ca51f32020-12-25 21:51:45 -0800437 return parts().monotonic_start_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800438 }
439 realtime_clock::time_point realtime_start_time() const {
Austin Schuh0ca51f32020-12-25 21:51:45 -0800440 return parts().realtime_start_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800441 }
442
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800443 // The time this data is sorted until.
444 monotonic_clock::time_point sorted_until() const { return sorted_until_; }
445
446 // Returns the next sorted message from the log file. It is safe to call
447 // std::move() on the result to move the data flatbuffer from it.
448 Message *Front();
449 // Pops the front message. This should only be called after a call to
450 // Front().
451 void PopFront();
452
453 // Returns a debug string representing the contents of this sorter.
454 std::string DebugString() const;
455
456 private:
457 // Log parts reader we are wrapping.
458 PartsMessageReader parts_message_reader_;
459 // Cache of the time we are sorted until.
460 aos::monotonic_clock::time_point sorted_until_ = monotonic_clock::min_time;
461
Austin Schuhb000de62020-12-03 22:00:40 -0800462 // Timestamp of the last message returned. Used to make sure nothing goes
463 // backwards.
464 monotonic_clock::time_point last_message_time_ = monotonic_clock::min_time;
465
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800466 // Set used for efficient sorting of messages. We can benchmark and evaluate
467 // other data structures if this proves to be the bottleneck.
468 absl::btree_set<Message> messages_;
Austin Schuh48507722021-07-17 17:29:24 -0700469
470 // Mapping from channel to source node.
471 // TODO(austin): Should we put this in Boots so it can be cached for everyone?
472 std::vector<size_t> source_node_index_;
Austin Schuh4b5c22a2020-11-30 22:58:43 -0800473};
474
Austin Schuh8f52ed52020-11-30 23:12:39 -0800475// Class to run merge sort on the messages from multiple LogPartsSorter
476// instances.
477class NodeMerger {
478 public:
Austin Schuhd2f96102020-12-01 20:27:29 -0800479 NodeMerger(std::vector<LogParts> parts);
480
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700481 // Copying and moving will mess up the internal raw pointers. Just don't do
482 // it.
483 NodeMerger(NodeMerger const &) = delete;
484 NodeMerger(NodeMerger &&) = delete;
485 void operator=(NodeMerger const &) = delete;
486 void operator=(NodeMerger &&) = delete;
487
Austin Schuhd2f96102020-12-01 20:27:29 -0800488 // Node index in the configuration of this node.
489 int node() const { return node_; }
Austin Schuh8f52ed52020-11-30 23:12:39 -0800490
Austin Schuh0ca51f32020-12-25 21:51:45 -0800491 // List of parts being sorted together.
492 std::vector<const LogParts *> Parts() const;
493
494 const Configuration *configuration() const {
495 return parts_sorters_[0].parts().config.get();
Austin Schuhd2f96102020-12-01 20:27:29 -0800496 }
497
498 monotonic_clock::time_point monotonic_start_time() const {
499 return monotonic_start_time_;
500 }
501 realtime_clock::time_point realtime_start_time() const {
502 return realtime_start_time_;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800503 }
504
505 // The time this data is sorted until.
506 monotonic_clock::time_point sorted_until() const { return sorted_until_; }
507
508 // Returns the next sorted message from the set of log files. It is safe to
509 // call std::move() on the result to move the data flatbuffer from it.
510 Message *Front();
511 // Pops the front message. This should only be called after a call to
512 // Front().
513 void PopFront();
514
515 private:
516 // Unsorted list of all parts sorters.
Austin Schuhd2f96102020-12-01 20:27:29 -0800517 std::vector<LogPartsSorter> parts_sorters_;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800518 // Pointer to the parts sorter holding the current Front message if one
519 // exists, or nullptr if a new one needs to be found.
520 LogPartsSorter *current_ = nullptr;
521 // Cached sorted_until value.
522 aos::monotonic_clock::time_point sorted_until_ = monotonic_clock::min_time;
Austin Schuhd2f96102020-12-01 20:27:29 -0800523
524 // Cached node.
525 int node_;
526
Austin Schuhb000de62020-12-03 22:00:40 -0800527 // Timestamp of the last message returned. Used to make sure nothing goes
528 // backwards.
529 monotonic_clock::time_point last_message_time_ = monotonic_clock::min_time;
530
Austin Schuhd2f96102020-12-01 20:27:29 -0800531 realtime_clock::time_point realtime_start_time_ = realtime_clock::max_time;
532 monotonic_clock::time_point monotonic_start_time_ = monotonic_clock::max_time;
533};
534
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700535// Class to concatenate multiple boots worth of logs into a single per-node
536// stream.
537class BootMerger {
538 public:
539 BootMerger(std::vector<LogParts> file);
540
541 // Copying and moving will mess up the internal raw pointers. Just don't do
542 // it.
543 BootMerger(BootMerger const &) = delete;
544 BootMerger(BootMerger &&) = delete;
545 void operator=(BootMerger const &) = delete;
546 void operator=(BootMerger &&) = delete;
547
548 // Node index in the configuration of this node.
549 int node() const { return node_mergers_[0]->node(); }
550
551 // List of parts being sorted together.
552 std::vector<const LogParts *> Parts() const;
553
554 const Configuration *configuration() const {
555 return node_mergers_[0]->configuration();
556 }
557
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700558 monotonic_clock::time_point monotonic_start_time(size_t boot) const {
559 CHECK_LT(boot, node_mergers_.size());
560 return node_mergers_[boot]->monotonic_start_time();
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700561 }
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700562 realtime_clock::time_point realtime_start_time(size_t boot) const {
563 CHECK_LT(boot, node_mergers_.size());
564 return node_mergers_[boot]->realtime_start_time();
Austin Schuhf16ef6a2021-06-30 21:48:17 -0700565 }
566
567 bool started() const {
568 return node_mergers_[index_]->sorted_until() != monotonic_clock::min_time ||
569 index_ != 0;
570 }
571
572 // Returns the next sorted message from the set of log files. It is safe to
573 // call std::move() on the result to move the data flatbuffer from it.
574 Message *Front();
575 // Pops the front message. This should only be called after a call to
576 // Front().
577 void PopFront();
578
579 private:
580 int index_ = 0;
581
582 // TODO(austin): Sanjay points out this is pretty inefficient. Don't keep so
583 // many things open.
584 std::vector<std::unique_ptr<NodeMerger>> node_mergers_;
585};
586
Austin Schuhd2f96102020-12-01 20:27:29 -0800587// Class to match timestamps with the corresponding data from other nodes.
Austin Schuh79b30942021-01-24 22:32:21 -0800588//
589// This class also buffers data for the node it represents, and supports
590// notifying when new data is queued as well as queueing until a point in time.
Austin Schuhd2f96102020-12-01 20:27:29 -0800591class TimestampMapper {
592 public:
593 TimestampMapper(std::vector<LogParts> file);
594
595 // Copying and moving will mess up the internal raw pointers. Just don't do
596 // it.
597 TimestampMapper(TimestampMapper const &) = delete;
598 TimestampMapper(TimestampMapper &&) = delete;
599 void operator=(TimestampMapper const &) = delete;
600 void operator=(TimestampMapper &&) = delete;
601
602 // TODO(austin): It would be super helpful to provide a way to queue up to
603 // time X without matching timestamps, and to then be able to pull the
604 // timestamps out of this queue. This lets us bootstrap time estimation
605 // without exploding memory usage worst case.
606
Austin Schuh0ca51f32020-12-25 21:51:45 -0800607 const Configuration *configuration() const { return configuration_.get(); }
Austin Schuhd2f96102020-12-01 20:27:29 -0800608
609 // Returns which node this is sorting for.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700610 size_t node() const { return boot_merger_.node(); }
Austin Schuhd2f96102020-12-01 20:27:29 -0800611
612 // The start time of this log.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700613 monotonic_clock::time_point monotonic_start_time(size_t boot) const {
614 return boot_merger_.monotonic_start_time(boot);
Austin Schuhd2f96102020-12-01 20:27:29 -0800615 }
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700616 realtime_clock::time_point realtime_start_time(size_t boot) const {
617 return boot_merger_.realtime_start_time(boot);
Austin Schuhd2f96102020-12-01 20:27:29 -0800618 }
619
620 // Uses timestamp_mapper as the peer for its node. Only one mapper may be set
621 // for each node. Peers are used to look up the data for timestamps on this
622 // node.
623 void AddPeer(TimestampMapper *timestamp_mapper);
624
Austin Schuh24bf4972021-06-29 22:09:08 -0700625 // Returns true if anything has been queued up.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700626 bool started() const { return boot_merger_.started(); }
Austin Schuhd2f96102020-12-01 20:27:29 -0800627
628 // Returns the next message for this node.
629 TimestampedMessage *Front();
630 // Pops the next message. Front must be called first.
631 void PopFront();
632
633 // Returns debug information about this node.
634 std::string DebugString() const;
635
Austin Schuh79b30942021-01-24 22:32:21 -0800636 // Queues data the provided time.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700637 void QueueUntil(BootTimestamp queue_time);
Austin Schuhe639ea12021-01-25 13:00:22 -0800638 // Queues until we have time_estimation_buffer of data in the queue.
639 void QueueFor(std::chrono::nanoseconds time_estimation_buffer);
Austin Schuh79b30942021-01-24 22:32:21 -0800640
Austin Schuh06601222021-01-26 17:02:50 -0800641 // Queues until the condition is met.
642 template <typename T>
643 void QueueUntilCondition(T fn) {
644 while (true) {
645 if (fn()) {
646 break;
647 }
648 if (!QueueMatched()) {
649 break;
650 }
651 }
652 }
653
Austin Schuh79b30942021-01-24 22:32:21 -0800654 // Sets a callback to be called whenever a full message is queued.
655 void set_timestamp_callback(std::function<void(TimestampedMessage *)> fn) {
656 timestamp_callback_ = fn;
657 }
658
Austin Schuhd2f96102020-12-01 20:27:29 -0800659 private:
660 // The state for a remote node. This holds the data that needs to be matched
661 // with the remote node's timestamps.
662 struct NodeData {
663 // True if we should save data here. This should be true if any of the
664 // bools in delivered below are true.
665 bool any_delivered = false;
666
Austin Schuh36c00932021-07-19 18:13:21 -0700667 // True if we have a peer and therefore should be saving data for it.
668 bool save_for_peer = false;
669
Austin Schuhd2f96102020-12-01 20:27:29 -0800670 // Peer pointer. This node is only to be considered if a peer is set.
671 TimestampMapper *peer = nullptr;
672
673 struct ChannelData {
674 // Deque per channel. This contains the data from the outside
675 // TimestampMapper node which is relevant for the node this NodeData
676 // points to.
677 std::deque<Message> messages;
678 // Bool tracking per channel if a message is delivered to the node this
679 // NodeData represents.
680 bool delivered = false;
681 };
682
683 // Vector with per channel data.
684 std::vector<ChannelData> channels;
685 };
686
687 // Returns (and forgets about) the data for the provided timestamp message
688 // showing when it was delivered to this node.
689 Message MatchingMessageFor(const Message &message);
690
691 // Queues up a single message into our message queue, and any nodes that this
692 // message is delivered to. Returns true if one was available, false
693 // otherwise.
694 bool Queue();
695
Austin Schuh79b30942021-01-24 22:32:21 -0800696 // Queues up a single matched message into our matched message queue. Returns
697 // true if one was queued, and false otherwise.
698 bool QueueMatched();
699
Austin Schuhd2f96102020-12-01 20:27:29 -0800700 // Queues up data until we have at least one message >= to time t.
701 // Useful for triggering a remote node to read enough data to have the
702 // timestamp you care about available.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700703 void QueueUnmatchedUntil(BootTimestamp t);
Austin Schuhd2f96102020-12-01 20:27:29 -0800704
Austin Schuh79b30942021-01-24 22:32:21 -0800705 // Queues m into matched_messages_.
706 void QueueMessage(Message *m);
Austin Schuhd2f96102020-12-01 20:27:29 -0800707
Austin Schuh58646e22021-08-23 23:51:46 -0700708 // Returns the name of the node this class is sorting for.
709 std::string_view node_name() const {
710 return configuration_->has_nodes() ? configuration_->nodes()
711 ->Get(boot_merger_.node())
712 ->name()
713 ->string_view()
714 : "(single node)";
715 }
716
Austin Schuhd2f96102020-12-01 20:27:29 -0800717 // The node merger to source messages from.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700718 BootMerger boot_merger_;
Austin Schuh0ca51f32020-12-25 21:51:45 -0800719
720 std::shared_ptr<const Configuration> configuration_;
721
Austin Schuhd2f96102020-12-01 20:27:29 -0800722 // The buffer of messages for this node. These are not matched with any
723 // remote data.
724 std::deque<Message> messages_;
725 // The node index for the source node for each channel.
726 std::vector<size_t> source_node_;
727
728 // Vector per node. Not all nodes will have anything.
729 std::vector<NodeData> nodes_data_;
730
731 // Latest message to return.
Austin Schuh79b30942021-01-24 22:32:21 -0800732 std::deque<TimestampedMessage> matched_messages_;
Austin Schuhd2f96102020-12-01 20:27:29 -0800733
Austin Schuh79b30942021-01-24 22:32:21 -0800734 // Tracks the state of the first message in matched_messages_. Do we need to
735 // update it, is it valid, or should we return nullptr?
Austin Schuhd2f96102020-12-01 20:27:29 -0800736 enum class FirstMessage {
737 kNeedsUpdate,
738 kInMessage,
739 kNullptr,
740 };
741 FirstMessage first_message_ = FirstMessage::kNeedsUpdate;
742
743 // Timestamp of the last message returned. Used to make sure nothing goes
744 // backwards.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700745 BootTimestamp last_message_time_ = BootTimestamp::min_time();
Austin Schuhd2f96102020-12-01 20:27:29 -0800746 // Time this node is queued up until. Used for caching.
Austin Schuh2dc8c7d2021-07-01 17:41:28 -0700747 BootTimestamp queued_until_ = BootTimestamp::min_time();
Austin Schuh79b30942021-01-24 22:32:21 -0800748
749 std::function<void(TimestampedMessage *)> timestamp_callback_;
Austin Schuh8f52ed52020-11-30 23:12:39 -0800750};
751
Austin Schuhee711052020-08-24 16:06:09 -0700752// Returns the node name with a trailing space, or an empty string if we are on
753// a single node.
754std::string MaybeNodeName(const Node *);
755
Brian Silvermanf51499a2020-09-21 12:49:08 -0700756} // namespace aos::logger
Austin Schuha36c8902019-12-30 18:07:15 -0800757
758#endif // AOS_EVENTS_LOGGING_LOGFILE_UTILS_H_