Skip to main content

starnix_modules_perfetto_consumer/
lib.rs

1// Copyright 2023 The Fuchsia Authors. All rights reserved.
2// Use of this source code is governed by a BSD-style license that can be
3// found in the LICENSE file.
4
5#![recursion_limit = "256"]
6
7use anyhow::bail;
8use fuchsia_trace::{
9    BufferingMode, ProlongedContext, TraceState, category_enabled, trace_state, trace_string_ref_t,
10};
11use fuchsia_trace_observer::TraceObserver;
12use futures::{SinkExt, StreamExt};
13use fxt::blob::{BlobHeader, BlobType};
14use perfetto_protos::perfetto::protos::trace_config::buffer_config::FillPolicy;
15use perfetto_protos::perfetto::protos::trace_config::{BufferConfig, DataSource};
16use perfetto_protos::perfetto::protos::{
17    DataSourceConfig, DisableTracingRequest, EnableTracingRequest, FreeBuffersRequest,
18    FtraceConfig, ReadBuffersRequest, TraceConfig, ipc_frame,
19};
20use perfetto_trace_protos::perfetto::protos::frame_timeline_event::{
21    ActualDisplayFrameStart, ActualSurfaceFrameStart, Event, ExpectedDisplayFrameStart,
22    ExpectedSurfaceFrameStart,
23};
24use perfetto_trace_protos::perfetto::protos::ftrace_event::Event::Print;
25use perfetto_trace_protos::perfetto::protos::trace_packet;
26use starnix_core::security;
27use starnix_core::task::dynamic_thread_spawner::SpawnRequestBuilder;
28use starnix_core::task::tracing::TracePerformanceEventManager;
29use starnix_core::task::{CurrentTask, Kernel};
30use starnix_core::vfs::FsString;
31use starnix_logging::{
32    CATEGORY_ATRACE, NAME_PERFETTO_BLOB, log_debug, log_error, log_info, log_warn,
33};
34use starnix_perfetto_trace_decoder::{decode_read_buffers_response, decode_trace, encode_trace};
35
36use starnix_uapi::errors::Errno;
37
38mod atrace;
39
40const PERFETTO_BUFFER_SIZE_KB: u32 = 63488;
41
42/// State needed to act upon trace state changes.
43struct CallbackState {
44    /// The previously observed trace state.
45    prev_state: TraceState,
46    /// Path to the Perfetto consumer socket.
47    socket_path: FsString,
48    /// Connection to the consumer socket, if it has been initialized. This gets initialized the
49    /// first time it is needed.
50    connection: Option<perfetto::Consumer>,
51    /// Prolonged trace context to prevent the Fuchsia trace session from terminating while reading
52    /// data from Perfetto.
53    prolonged_context: Option<ProlongedContext>,
54    /// Partial trace packet returned from Perfetto but not yet written to Fuchsia.
55    packet_data: Vec<u8>,
56
57    event_manager: TracePerformanceEventManager,
58}
59
60impl CallbackState {
61    fn connection(
62        &mut self,
63        current_task: &CurrentTask,
64    ) -> Result<&mut perfetto::Consumer, anyhow::Error> {
65        match self.connection {
66            None => {
67                self.connection =
68                    Some(perfetto::Consumer::new(current_task, self.socket_path.as_ref())?);
69                Ok(self.connection.as_mut().unwrap())
70            }
71            Some(ref mut conn) => Ok(conn),
72        }
73    }
74
75    fn handle_stopped(&mut self) {
76        self.prolonged_context = None;
77        self.packet_data.clear();
78        self.event_manager.stop();
79        self.event_manager.clear();
80    }
81
82    fn on_state_change(
83        &mut self,
84        new_state: TraceState,
85        current_task: &CurrentTask,
86    ) -> Result<(), anyhow::Error> {
87        let prev_state = self.prev_state;
88        self.prev_state = new_state;
89        log_debug!(
90            "Perfetto consumer state change. new_state: {new_state:?}, prev_state: {prev_state:?}"
91        );
92        match new_state {
93            TraceState::Started => {
94                if prev_state != TraceState::Stopped {
95                    // This means something unexpected has caused the trace_engine to change
96                    // states faster than we're processing the trace observer events.
97                    log_error!(
98                        "Started received in {prev_state:?} state! Cleaning up then starting."
99                    );
100                    self.handle_stopped();
101                }
102                self.prolonged_context = ProlongedContext::acquire();
103                let connection = self.connection(current_task)?;
104                // A fixed set of data sources that may be of interest. As demand for other sources
105                // is found, add them here, and it may become worthwhile to allow this set to be
106                // configurable per trace session.
107                let mut data_sources = vec![
108                    DataSource {
109                        config: Some(DataSourceConfig {
110                            name: Some("track_event".to_string()),
111                            ..Default::default()
112                        }),
113                        ..Default::default()
114                    },
115                    DataSource {
116                        config: Some(DataSourceConfig {
117                            name: Some("android.surfaceflinger.frame".to_string()),
118                            target_buffer: Some(0),
119                            ..Default::default()
120                        }),
121                        ..Default::default()
122                    },
123                    DataSource {
124                        config: Some(DataSourceConfig {
125                            name: Some("android.surfaceflinger.frametimeline".to_string()),
126                            target_buffer: Some(0),
127                            ..Default::default()
128                        }),
129                        ..Default::default()
130                    },
131                ];
132                if category_enabled(CATEGORY_ATRACE) {
133                    data_sources.push(DataSource {
134                        config: Some(DataSourceConfig {
135                            name: Some("linux.ftrace".to_string()),
136                            ftrace_config: Some(FtraceConfig {
137                                ftrace_events: vec!["ftrace/print".to_string()],
138                                // Enable all supported atrace categories. This could be improved
139                                // in the future to be a subset that is configurable by each trace
140                                // session.
141                                atrace_categories: vec![
142                                    "am".to_string(),
143                                    "adb".to_string(),
144                                    "aidl".to_string(),
145                                    "dalvik".to_string(),
146                                    "audio".to_string(),
147                                    "binder_lock".to_string(),
148                                    "binder_driver".to_string(),
149                                    "bionic".to_string(),
150                                    "camera".to_string(),
151                                    "database".to_string(),
152                                    "gfx".to_string(),
153                                    "hal".to_string(),
154                                    "input".to_string(),
155                                    "network".to_string(),
156                                    "nnapi".to_string(),
157                                    "pm".to_string(),
158                                    "power".to_string(),
159                                    "rs".to_string(),
160                                    "res".to_string(),
161                                    "rro".to_string(),
162                                    "sched".to_string(),
163                                    "sm".to_string(),
164                                    "ss".to_string(),
165                                    "vibrator".to_string(),
166                                    "video".to_string(),
167                                    "view".to_string(),
168                                    "webview".to_string(),
169                                    "wm".to_string(),
170                                ],
171                                atrace_apps: vec!["*".to_string()],
172                                ..Default::default()
173                            }),
174                            ..Default::default()
175                        }),
176                        ..Default::default()
177                    });
178                }
179                connection.enable_tracing(
180                    current_task,
181                    EnableTracingRequest {
182                        trace_config: Some(TraceConfig {
183                            buffers: vec![BufferConfig {
184                                size_kb: Some(PERFETTO_BUFFER_SIZE_KB),
185                                fill_policy: Some(FillPolicy::Discard.into()),
186                                ..Default::default()
187                            }],
188                            data_sources,
189                            ..Default::default()
190                        }),
191                        attach_notification_only: None,
192                    },
193                )?;
194                // Once tracing has started, notify the event manager so it can start tracking processes.
195                self.event_manager.start(current_task.kernel());
196            }
197            TraceState::Stopping => {
198                if prev_state != TraceState::Started {
199                    // If we receive a stop request and we don't think we're actually tracing, our
200                    // local state likely desynced from the global trace state. Clean up our state
201                    // and ensure we're stopped so we re-synchronize.
202                    log_error!("Stopping received in {prev_state:?} state! Cleaning up.");
203                    self.handle_stopped();
204                    return Ok(());
205                }
206
207                // We want to hold the prolonged context to ensure the trace session doesn't
208                // exit out from under us, but we also want to ensure we drop the prolonged
209                // context if we bail for whatever reason below.
210                let _local_prolonged_context = std::mem::replace(&mut self.prolonged_context, None);
211                let start_time = std::time::Instant::now();
212
213                let connection = self.connection(current_task)?;
214                let disable_request =
215                    connection.disable_tracing(current_task, DisableTracingRequest {})?;
216                loop {
217                    let frame = connection.next_frame_blocking(current_task)?;
218                    if frame.request_id == Some(disable_request) {
219                        break;
220                    } else {
221                        log_error!(
222                            "Ignoring frame while looking for DisableTracingRequest: {frame:?}"
223                        );
224                    }
225                }
226
227                let read_buffers_request =
228                    connection.read_buffers(current_task, ReadBuffersRequest {})?;
229
230                let blob_name_ref = {
231                    let Some(context) = fuchsia_trace::Context::acquire() else {
232                        bail!("Tracing stopped despite holding prolonged context");
233                    };
234                    context.register_string_literal(NAME_PERFETTO_BLOB)
235                };
236
237                // IPC responses may be spread across multiple frames, so loop until we get a
238                // message that indicates it is the last one. Additionally, if there are
239                // unrelated messages on the socket (e.g. leftover from a previous trace
240                // session), the loop will read past and ignore them.
241                loop {
242                    let frame = self.connection(current_task)?.next_frame_blocking(current_task)?;
243                    if frame.request_id != Some(read_buffers_request) {
244                        continue;
245                    } else {
246                        log_debug!(
247                            "perfetto_consumer ignoring frame while looking for ReadBuffersRequest {read_buffers_request}: {frame:?}"
248                        );
249                    }
250                    if let Some(ipc_frame::Msg::MsgInvokeMethodReply(reply)) = &frame.msg {
251                        if let Ok(response) = decode_read_buffers_response(
252                            reply.reply_proto.as_deref().unwrap_or(&[]),
253                        ) {
254                            for slice in &response.slices {
255                                if let Some(data) = &slice.data {
256                                    self.packet_data.extend(data);
257                                }
258                                if slice.last_slice_for_packet.unwrap_or(false) {
259                                    let mut blob_data = Vec::new();
260                                    // Packet field number = 1, length delimited type = 2.
261                                    blob_data.push(1 << 3 | 2);
262                                    // Push a varint encoded length.
263                                    // See https://protobuf.dev/programming-guides/encoding/
264                                    const HIGH_BIT: u8 = 0x80;
265                                    const LOW_SEVEN_BITS: usize = 0x7F;
266                                    let mut value = self.packet_data.len();
267                                    while value >= HIGH_BIT as usize {
268                                        blob_data.push((value & LOW_SEVEN_BITS) as u8 | HIGH_BIT);
269                                        value >>= 7;
270                                    }
271                                    blob_data.push(value as u8);
272                                    // `append` moves all data out of the passed Vec, so
273                                    // s.packet_data will be empty after this call.
274                                    blob_data.append(&mut self.packet_data);
275
276                                    // At this point blob_data is a full Perfetto Trace protobuf.
277                                    // Parse the data and replace the linux pids with their
278                                    // corresponding koid.
279                                    let rewritten =
280                                        self.rewrite_pids(&blob_data).unwrap_or(blob_data);
281
282                                    // Ignore a failure to write the packet here. We don't
283                                    // return immediately because we want to allow the
284                                    // remaining records to be recorded as dropped.
285                                    //
286                                    // Once we fill a buffer in oneshot mode, we expect to drop
287                                    // the remaining packets here.
288                                    //
289                                    // Rather than logging here, allow the trace system to
290                                    // aggregate the number of records dropped and we can query
291                                    // the trace system later to determine if we dropped
292                                    // records when it's more efficient to do so.
293                                    let _ = self.forward_packet(blob_name_ref, rewritten);
294                                }
295                            }
296                        } else {
297                            log_error!("perfetto_consumer cannot decode protobuf from {reply:?}");
298                        }
299                        if reply.has_more != Some(true) {
300                            break;
301                        }
302                    } else {
303                        log_error!(
304                            "perfetto_consumer ignoring non-MsgInvokeMethodReply message: {frame:?}"
305                        );
306                    }
307                }
308                // The response to a free buffers request does not have anything meaningful,
309                // so we don't need to worry about tracking the request id to match to the
310                // response.
311                let _free_buffers_request_id = self
312                    .connection(current_task)?
313                    .free_buffers(current_task, FreeBuffersRequest { buffer_ids: vec![0] })?;
314                let elapsed = start_time.elapsed().as_millis();
315                log_info!(
316                    "Perfetto frames copied, dropping prolonged trace context. Processing took {elapsed} ms"
317                );
318            }
319            TraceState::Stopped => {
320                self.handle_stopped();
321            }
322        }
323        Ok(())
324    }
325
326    // Forward `data` to the trace buffer by wrapping it in fxt blob records with the name
327    // `blob_name_ref`..
328    fn forward_packet(&self, blob_name_ref: trace_string_ref_t, data: Vec<u8>) -> Option<usize> {
329        // The blob data may be larger than what we can fit in a single record. If so, split it up
330        // over multiple chunks.
331        let mut bytes_written = 0;
332        let mut data_to_write = &data[..];
333
334        // We want to break the data into chunks:
335        // - Bigger chunks means less per-write overheader
336        // - Bigger chunks means less overhead due to blob meta
337        //
338        // However, too big and the blobs won't fit nicely into the trace buffer.
339        // The trace buffer is minimum 1MiB in size, so writing 4k at a time seems like a
340        // reasonable place to start that is both reasonably large and not going to leave a ton of
341        // space at the end of the trace buffer.
342        let max_chunk_size = 4096;
343        while !data_to_write.is_empty() {
344            let chunk_size = data_to_write.len().min(max_chunk_size);
345            let chunk = &data_to_write[..chunk_size];
346            self.forward_blob(blob_name_ref, &chunk)?;
347            data_to_write = &data_to_write[chunk_size..];
348            bytes_written += chunk_size;
349        }
350        Some(bytes_written)
351    }
352
353    // Given a blob name, wrap the data in an fxt perfetto blob and write it to the trace buffer.
354    fn forward_blob(&self, blob_name_ref: trace_string_ref_t, blob_data: &[u8]) -> Option<usize> {
355        let mut header = BlobHeader::empty();
356        header.set_name_ref(blob_name_ref.encoded_value);
357        header.set_payload_len(blob_data.len() as u16);
358        header.set_blob_format_type(BlobType::Perfetto.into());
359
360        let record_bytes = fxt::fxt_builder::FxtBuilder::new(header).atom(blob_data).build();
361        assert!(record_bytes.len() % std::mem::size_of::<u64>() == 0);
362        let num_words = record_bytes.len() / std::mem::size_of::<u64>();
363        let record_data = record_bytes.as_ptr();
364        #[allow(
365            clippy::undocumented_unsafe_blocks,
366            reason = "Force documented unsafe blocks in Starnix"
367        )]
368        let record_words =
369            unsafe { std::slice::from_raw_parts(record_data.cast::<u64>(), num_words) };
370
371        while let Some(context) = fuchsia_trace::Context::acquire() {
372            if let Some(bytes) = context.copy_record(record_words) {
373                return Some(bytes);
374            }
375            if context.buffering_mode() != BufferingMode::Streaming {
376                // If we're not in streaming mode, there will never be room for this record. Drop
377                // it.
378                return None;
379            }
380            // We're writing records pretty quick here, we're just forwarding data from
381            // perfetto with no breaks. trace_manager might not be able to keep up if it's also
382            // servicing other trace-providers. We want to back off we if find that we run out
383            // of space.
384            //
385            // We drop the context to decrement the refcount on the trace session. This allows
386            // trace-engine to switch the buffers if needed and drain out the buffers so that
387            // when we wake, there will hopefully be room.
388            //
389            // TODO(b/304532640)
390            drop(context);
391            std::thread::sleep(std::time::Duration::from_millis(100));
392        }
393        None
394    }
395
396    fn rewrite_pids(&mut self, protobuf_blob: &Vec<u8>) -> anyhow::Result<Vec<u8>> {
397        let mut proto = decode_trace(protobuf_blob.as_slice())?;
398        for p in &mut proto.packet {
399            if let Some(ref mut data) = p.data {
400                match data {
401                    trace_packet::Data::FrameTimelineEvent(frame_timeline_event) => {
402                        if let Some(evt) = &mut frame_timeline_event.event {
403                            // Update the linux pid to the Fuchsia pid. Each event has its own
404                            // match arm since the variant data is of a different type for each event.
405                            match evt {
406                                Event::ExpectedDisplayFrameStart(ExpectedDisplayFrameStart {
407                                    pid,
408                                    ..
409                                })
410                                | Event::ActualDisplayFrameStart(ActualDisplayFrameStart {
411                                    pid,
412                                    ..
413                                })
414                                | Event::ExpectedSurfaceFrameStart(ExpectedSurfaceFrameStart {
415                                    pid,
416                                    ..
417                                })
418                                | Event::ActualSurfaceFrameStart(ActualSurfaceFrameStart {
419                                    pid,
420                                    ..
421                                }) => {
422                                    pid.as_mut().map(|pid| {
423                                        *pid = self.map_to_koid_val(*pid);
424                                    });
425                                }
426                                Event::FrameEnd(_frame_end) => {}
427                            }
428                        }
429                    }
430                    trace_packet::Data::FtraceEvents(ftrace_bundle) => {
431                        for evt in &mut ftrace_bundle.event {
432                            if let Some(ref mut pid) = evt.pid {
433                                *pid = self.map_thread_to_koid_val(*pid as i32) as u32;
434                            }
435                            if let Some(ref mut event_data) = evt.event {
436                                match event_data {
437                                    Print(print) => {
438                                        if let Some(ref mut data) = print.buf {
439                                            *data = self.map_print_event(data)
440                                        }
441                                    }
442                                    _ => (),
443                                }
444                            }
445                        }
446                    }
447                    // No need to process other data; we only fixup data that references the pid.
448                    _ => (),
449                }
450            }
451        }
452        Ok(encode_trace(&proto))
453    }
454
455    fn map_print_event(&mut self, data: &String) -> String {
456        if let Some(mut event) = atrace::ATraceEvent::parse(&data) {
457            match event {
458                atrace::ATraceEvent::Begin { ref mut pid, .. }
459                | atrace::ATraceEvent::End { ref mut pid }
460                | atrace::ATraceEvent::Instant { ref mut pid, .. }
461                | atrace::ATraceEvent::AsyncBegin { ref mut pid, .. }
462                | atrace::ATraceEvent::AsyncEnd { ref mut pid, .. }
463                | atrace::ATraceEvent::Counter { ref mut pid, .. }
464                | atrace::ATraceEvent::AsyncTrackBegin { ref mut pid, .. }
465                | atrace::ATraceEvent::AsyncTrackEnd { ref mut pid, .. }
466                | atrace::ATraceEvent::Track { ref mut pid, .. } => {
467                    *pid = self.map_to_koid_val(*pid as i32) as u64
468                }
469            }
470            event.data()
471        } else {
472            data.to_string()
473        }
474    }
475
476    fn map_thread_to_koid_val(&mut self, pid: i32) -> i32 {
477        if pid == 0 {
478            return 0;
479        }
480        self.event_manager
481            .map_tid_to_koid(pid)
482            .map(|k| {
483                // Truncate the koid down to 32 bits in order to match the perfetto data schema. This is
484                // usually not an issue except for artificial koids which have the 2^63 bit set, such as
485                // virtual threads. This is consistent with the perfetto data importer code:
486                // https://github.com/google/perfetto/blob/c343c8a77c6e665c679e5c1ec845ac6dde0fc685/src/trace_processor/importers/fuchsia/fuchsia_trace_tokenizer.cc#L490
487                k.raw_koid() as i32
488            })
489            .unwrap_or(pid)
490    }
491
492    fn map_to_koid_val(&mut self, pid: i32) -> i32 {
493        if pid == 0 {
494            return 0;
495        }
496        self.event_manager
497            .map_pid_to_koid(pid)
498            .map(|k| {
499                // Truncate the koid down to 32 bits in order to match the perfetto data schema. This is
500                // usually not an issue except for artificial koids which have the 2^63 bit set, such as
501                // virtual threads. This is consistent with the perfetto data importer code:
502                // https://github.com/google/perfetto/blob/c343c8a77c6e665c679e5c1ec845ac6dde0fc685/src/trace_processor/importers/fuchsia/fuchsia_trace_tokenizer.cc#L490
503                k.raw_koid() as i32
504            })
505            .unwrap_or(pid)
506    }
507}
508
509pub fn start_perfetto_consumer_thread(kernel: &Kernel, socket_path: FsString) -> Result<(), Errno> {
510    let (mut tx, mut rx) = futures::channel::mpsc::channel::<TraceState>(32);
511
512    // Listens for trace state changes and sends them to Perfetto consumer thread.
513    // Unlike the perfetto thread, it won't block, so we can spawn it on the main async executor.
514    kernel.kthreads.spawn_future(
515        move || async move {
516            let observer = TraceObserver::new();
517            while let Ok(state) = observer.on_state_changed().await {
518                if let Err(e) = tx.send(state).await {
519                    log_error!("perfetto-trace-observer failed to send trace state change: {:?}. Receiver dropped.", e);
520                    return;
521                }
522            }
523        },
524        "perfetto-trace-observer",
525    );
526
527    // Perfetto consumer task: reads state changes from the channel and handles them.
528    // This task can block, so we spawn it on a dedicated thread to not block the observer or the
529    // main async executor.
530    let worker_closure = async move |current_task: &CurrentTask| {
531        let mut callback_state = CallbackState {
532            prev_state: TraceState::Stopped,
533            socket_path,
534            connection: None,
535            prolonged_context: None,
536            packet_data: Vec::new(),
537            event_manager: TracePerformanceEventManager::new(),
538        };
539
540        fn handle_state_change(
541            callback_state: &mut CallbackState,
542            current_task: &&CurrentTask,
543            state: TraceState,
544        ) -> Result<(), anyhow::Error> {
545            let current_task = current_task;
546            // TODO: https://fxbug.dev/457381697 - Revise how this kernel-internal work is security-
547            // checked.
548            let creds = security::creds_start_internal_operation(current_task);
549            current_task
550                .override_creds(creds, || callback_state.on_state_change(state, current_task))
551        }
552
553        // Check for tracing already started before we began observing.
554        // This happens when tracing is started on boot.
555        let mut state = trace_state();
556        if trace_state() == TraceState::Started {
557            const RETRY_DELAY: std::time::Duration = std::time::Duration::from_secs(5);
558            // When we do boot tracing, it is possible (even likely), that starnix has started but
559            // perfetto may not be ready to be connected to yet.
560            // In that case poll until it has started.
561            loop {
562                match handle_state_change(&mut callback_state, &current_task, state) {
563                    Ok(_) => break, // Success, exit loop.
564                    Err(e) => {
565                        if let Some(errno) = e.downcast_ref::<Errno>() {
566                            if errno == &starnix_uapi::errors::ENOENT
567                                || errno == &starnix_uapi::errors::ECONNREFUSED
568                            {
569                                log_warn!(
570                                    "perfetto_consumer initial start tracing failed because perfetto socket connection not established: {e:?} retrying in 5 seconds..."
571                                );
572                                std::thread::sleep(RETRY_DELAY);
573                                callback_state.prev_state = TraceState::Stopped;
574                                callback_state.connection = None;
575                                callback_state.prolonged_context = None;
576                                state = trace_state();
577                                continue; // Retry
578                            }
579                        }
580                        // For any other error, log and exit loop.
581                        log_error!(
582                            "perfetto_consumer initial start tracing failed with error: {e:?}"
583                        );
584                        break;
585                    }
586                }
587            }
588        }
589
590        while let Some(state) = rx.next().await {
591            handle_state_change(&mut callback_state, &current_task, state).unwrap_or_else(|e| {
592                log_error!("perfetto_consumer state change callback error: {:?}", e);
593            })
594        }
595    };
596    let worker_req = SpawnRequestBuilder::new()
597        .with_debug_name("perfetto-consumer")
598        .with_async_closure(worker_closure)
599        .build();
600    kernel.kthreads.spawner().spawn_from_request(worker_req);
601
602    Ok(())
603}