1 // Copyright (C) 2019 The Android Open Source Project
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 #include "common/debug.h"
16 #include "common/expected.h"
17 #include "perfetto/rx_producer.h"
18 
19 #include <android-base/file.h>
20 #include <android-base/properties.h>
21 #include <android-base/unique_fd.h>
22 
23 #include <iostream>
24 
25 #include <sched.h>
26 #include <sys/types.h>
27 #include <sys/stat.h>
28 #include <syscall.h>
29 #include <fcntl.h>
30 #include <unistd.h>
31 
32 // TODO: move to perfetto code
33 namespace perfetto {
34 namespace consumer {
35 
operator <<(std::ostream & os,State state)36 std::ostream& operator<<(std::ostream& os, State state) {
37   switch (state) {
38     case State::kTraceFailed:
39       os << "kTraceFailed";
40       break;
41     case State::kConnectionError:
42       os << "kConnectionError";
43       break;
44     case State::kSessionNotFound:
45       os << "kSessionNotFound";
46       break;
47     case State::kIdle:
48       os << "kIdle";
49       break;
50     case State::kConnecting:
51       os << "kConnecting";
52       break;
53     case State::kConfigured:
54       os << "kConfigured";
55       break;
56     case State::kTracing:
57       os << "kTracing";
58       break;
59     case State::kTraceEnded:
60       os << "kTraceEnded";
61       break;
62     default:
63       os << "(unknown)";  // did someone forget to update this code?
64       break;
65   }
66   return os;
67 }
68 
69 }  // namespace consumer
70 }  // namespace perfetto
71 
72 namespace iorap::perfetto {
73 
CreateComponent()74 PerfettoDependencies::Component PerfettoDependencies::CreateComponent() {
75   // TODO: read from config.
76   static const uint32_t kTraceDurationMs =
77       ::android::base::GetUintProperty("iorapd.perfetto.trace_duration_ms", /*default*/5000U);
78 
79   static const uint32_t kBufferSize =
80       ::android::base::GetUintProperty("iorapd.perfetto.buffer_size", /*default*/4096U);
81 
82   return fruit::createComponent()
83     .bind<PerfettoConsumer, PerfettoConsumerImpl>()
84     .registerProvider([]() /* -> TraceConfig */ {
85         return CreateConfig(kTraceDurationMs,
86                             /*deferred_start*/false,
87                             kBufferSize);
88     });
89 }
90 
CreateConfig(uint32_t duration_ms,bool deferred_start,uint32_t buffer_size)91 ::perfetto::protos::TraceConfig PerfettoDependencies::CreateConfig(uint32_t duration_ms,
92                                                                    bool deferred_start,
93                                                                    uint32_t buffer_size) {
94   ::perfetto::protos::TraceConfig trace_config;
95 
96   trace_config.set_duration_ms(duration_ms);
97   trace_config.add_buffers()->set_size_kb(buffer_size);
98   trace_config.set_deferred_start(deferred_start);
99 
100   auto* ds_config = trace_config.add_data_sources()->mutable_config();
101   ds_config->set_name("linux.ftrace");
102   ds_config->mutable_ftrace_config()->add_ftrace_events(
103       "mm_filemap_add_to_page_cache");
104   ds_config->mutable_ftrace_config()->add_ftrace_events(
105       "mm_filemap_delete_from_page_cache");
106   ds_config->set_target_buffer(0);
107 
108   return trace_config;
109 }
110 
111 // RAII-style wrapper around a perfetto handle that calls Destroy
112 // in a thread-safe manner.
113 struct PerfettoConsumerHandle {
114  private:
115   std::shared_ptr<PerfettoConsumer> consumer_;
116   PerfettoConsumer::Handle handle_;
117 
118  public:
119   // Takes over ownership of the 'handle'.
120   //
121   // Consumer must not be null.
PerfettoConsumerHandleiorap::perfetto::PerfettoConsumerHandle122   PerfettoConsumerHandle(std::shared_ptr<PerfettoConsumer> consumer,
123                          PerfettoConsumer::Handle handle)
124     : consumer_{std::move(consumer)},
125       handle_{std::move(handle)} {
126     DCHECK(consumer_ != nullptr);
127   }
128 
GetConsumeriorap::perfetto::PerfettoConsumerHandle129   std::shared_ptr<PerfettoConsumer> GetConsumer() const {
130     return consumer_;
131   }
132 
GetHandleiorap::perfetto::PerfettoConsumerHandle133   PerfettoConsumer::Handle GetHandle() const {
134     return handle_;
135   }
136 
~PerfettoConsumerHandleiorap::perfetto::PerfettoConsumerHandle137   ~PerfettoConsumerHandle() {
138     LOG(VERBOSE) << "PerfettoConsumerHandle::Destroy(" << handle_ << ")";
139     consumer_->Destroy(handle_);
140   }
141 
operator ==iorap::perfetto::PerfettoConsumerHandle142   bool operator==(const PerfettoConsumerHandle& other) const {
143     return handle_ == other.handle_ && consumer_ == other.consumer_;
144   }
145 
operator !=iorap::perfetto::PerfettoConsumerHandle146   bool operator!=(const PerfettoConsumerHandle& other) const {
147     return !(*this == other);
148   }
149 };
150 
151 
152 // Snapshot of a single perfetto OnStateChanged callback.
153 //
154 // Operate on the PerfettoConsumer to further change the state.
155 //
156 // The Handle is kept 'valid' until all references to the PerfettoConsumerHandle
157 // are dropped to 0. This ensures the Handle is not destroyed too early. All
158 // direct usages of 'Handle' must be scoped by the PerfettoConsumerHandle.
159 struct PerfettoStateChange {
160  public:
161   using State = ::perfetto::consumer::State;
162   using Handle = ::perfetto::consumer::Handle;
163 
164   State state;                                                           // Never invalid.
165   std::shared_ptr<PerfettoConsumerHandle> perfetto_consumer_and_handle;  // Never null.
166 
167   // Safety: Use only within scope of the PerfettoStateChange.
GetHandleiorap::perfetto::PerfettoStateChange168   Handle GetHandle() const {
169     // TODO: it would be even safer to wrap all the calls to the handle inside a class,
170     // instead of exposing this raw Handle.
171     return perfetto_consumer_and_handle->GetHandle();
172   }
173 
GetConsumeriorap::perfetto::PerfettoStateChange174   std::shared_ptr<PerfettoConsumer> GetConsumer() const {
175     return perfetto_consumer_and_handle->GetConsumer();
176   }
177 };
178 
operator <<(std::ostream & os,const PerfettoStateChange & state_change)179 std::ostream& operator<<(std::ostream& os, const PerfettoStateChange& state_change) {
180   os << "PerfettoStateChange{" << state_change.state << ","
181      << state_change.GetHandle() << ","
182      << state_change.GetConsumer().get() << "}";
183   return os;
184 }
185 
186 // Once created, this acts as a hot observable, emitting 'PerfettoStateChange' transition items.
187 // Only the 'state' will vary, the handle and perfetto_consumer are always the same value.
188 //
189 // Clients only need to handle the success states in #on_next, all failure states will go to
190 // #on_error.
191 //
192 // Upon reaching the appropriate terminal states, either #on_completed or #on_error is called.
193 // No future callbacks will then occur, so this object should be subsequently deleted.
194 //
195 // The Handle is destroyed automatically after the last item is emitted, so it must only be
196 // manipulated from the #on_next callbacks. Do not save the Handle and use it at other times.
197 class StateChangedSubject {
198  public:
199   using State = ::perfetto::consumer::State;
200   using Handle = ::perfetto::consumer::Handle;
201 
202   // Static members to solve use-after-free bug.
203   // The object is accessed from not only perfetto thread, but also iorap
204   // thread. Use this global map to manage it.
205   static std::mutex state_subject_mutex_;
206   static std::unordered_map<Handle, StateChangedSubject*> state_subject_map_;
207 
StateChangedSubject(const::perfetto::protos::TraceConfig & trace_config,rxcpp::subscriber<PerfettoStateChange> destination,std::shared_ptr<PerfettoConsumer> perfetto_consumer)208   StateChangedSubject(const ::perfetto::protos::TraceConfig& trace_config,
209                       rxcpp::subscriber<PerfettoStateChange> destination,
210                       std::shared_ptr<PerfettoConsumer> perfetto_consumer)
211     : deferred_start(trace_config.deferred_start()),
212       dest(std::move(destination)),
213       perfetto_consumer_(std::move(perfetto_consumer)) {
214     DCHECK(perfetto_consumer_ != nullptr);
215   }
216 
217  private:
218   struct StateChangedError : public std::runtime_error {
StateChangedErroriorap::perfetto::StateChangedSubject::StateChangedError219     explicit StateChangedError(const std::string& what_arg) : std::runtime_error(what_arg) {}
220   };
221 
222   std::shared_ptr<PerfettoConsumerHandle> handle_;  // non-null after bound_ == true.
223   std::atomic<bool> bound_{false};  // synchronize-with for BindHandle -> OnStateChanged.
224 
225   State last_state{State::kIdle};
226   bool deferred_start{false};
227 
228   rxcpp::subscriber<PerfettoStateChange> dest;
229   std::shared_ptr<PerfettoConsumer> perfetto_consumer_;  // This is never null.
230 
DcheckBadStateTransition(State state,bool fail_unless=false) const231   void DcheckBadStateTransition(State state, bool fail_unless = false) const {
232     DCHECK(fail_unless) << "Invalid state transition to " << state << " from " << last_state;
233   }
234 
DcheckValidStateTransition(State state)235   void DcheckValidStateTransition(State state) {
236     // State must not be out of range.
237     DCHECK_GE(state, State::kTraceFailed);
238     DCHECK_LE(state, State::kTraceEnded);
239 
240     // Internal state that should never leak out into public perfetto API:
241     DCHECK_NE(state, State::kIdle);
242     // These can only be returned by PollState:
243     DCHECK_NE(state, State::kSessionNotFound);
244 
245     // Validate state transitions as per the perfetto API contract.
246     // See the 'state diagram' in consumer_api.h
247     switch (last_state) {
248       case State::kTraceFailed:  // Final and unrecoverable.
249         // b/122548195: this can transition to 'kConnectionError' if selinux is disabled.
250         if (state == State::kConnectionError) {
251           LOG(WARNING) << "b/122548195: kTraceFailed is non-terminal, ignoring.";
252           // This is a bit awkward: rxcpp will drop the #on_error calls if its more than once.
253           break;
254         }
255         DcheckBadStateTransition(state);
256         break;
257       case State::kConnectionError:  // Final and unrecoverable.
258         DcheckBadStateTransition(state);
259         break;
260       case State::kSessionNotFound:
261         DcheckBadStateTransition(state);
262         break;
263       case State::kIdle:
264         // OK: we initialized our own state to idle prior to the first callback.
265         break;
266       case State::kConnecting:
267         switch (state) {
268           case State::kConfigured:
269             // kConfigured, if |deferred_start| == true in the trace config.
270             DcheckBadStateTransition(state, deferred_start);
271             break;
272           case State::kTracing:
273             // kTracing, if |deferred_start| == false.
274             DcheckBadStateTransition(state, !deferred_start);
275             break;
276           case State::kConnectionError:
277             // An error state, e.g. if cannot reach the traced daemon.
278             break;
279           default:
280             // Unconditionally invalid state transitions from kConnecting to anything else.
281             DcheckBadStateTransition(state);
282         }
283         break;
284       case State::kConfigured:
285         DCHECK(deferred_start);
286         if (state != State::kTracing  // OK: this is documented.
287             && state != State::kTraceFailed) {  // Undocumented selinux failure.
288             // Undocumented, but it appears to go directly from Configured->TraceEnded
289             // it can also go to kTraceFailed if e.g. there's an selinux violation
290             // however this appears to be underdocumented.
291             // b/122607276 #2
292 
293           if (state != State::kTraceEnded) {  // b/122607276 #1
294             DcheckBadStateTransition(state);
295           }
296         }
297         break;
298       case State::kTracing:
299         switch (state) {
300           case State::kTraceEnded:
301             break;
302           case State::kTraceFailed:
303             break;
304           default:
305             DcheckBadStateTransition(state);
306         }
307         break;
308       case State::kTraceEnded:
309         // Cannot transition from terminal state to another state.
310         DcheckBadStateTransition(state);
311         break;
312 
313       // default: This list is exhaustive
314     }
315   }
316 
IsTerminalState() const317   constexpr bool IsTerminalState() const {
318     switch (last_state) {
319       case State::kTraceFailed:
320       case State::kConnectionError:
321       case State::kTraceEnded:
322         return true;
323       default:
324         return false;
325     }
326   }
327 
328   // Returns true for non-terminal states (i.e. this callback will be invoked again).
329   // Returns false otherwise.
OnStateChanged(Handle handle,State state)330   bool OnStateChanged(Handle handle, State state) {
331     using namespace ::perfetto::consumer;
332 
333     // Block until 'BoundHandle' is called by the other thread.
334     while (!bound_.load()) {}  // seq_cst acquire.
335 
336     std::shared_ptr<PerfettoConsumerHandle> handle_ptr = handle_;
337     DCHECK(handle_ptr != nullptr);
338 
339     DCHECK_EQ(handle_ptr->GetHandle(), handle);
340     DcheckValidStateTransition(state);
341 
342     switch (state) {
343       // Error states (terminal).
344       case State::kTraceFailed:
345         EmitError("kTraceFailed");
346         break;
347       case State::kConnectionError:
348         EmitError("kConnectionError");
349         break;
350 
351       // Regular transitions (non-terminal).
352       case State::kConnecting:
353       case State::kConfigured:
354       case State::kTracing:
355         EmitNext(state);
356         break;
357       // Regular transitions (terminal).
358       case State::kTraceEnded:  // XX: do we even need to emit the 'TraceEnded' state?
359         EmitNext(state);
360         dest.on_completed();
361         break;
362       default:
363         DcheckBadStateTransition(state);
364     }
365 
366     bool force_non_terminal = false;
367 
368     if (last_state == State::kConfigured &&  state == State::kConnectionError) {
369       // b/122548195: this can transition to 'kConnectionError' if selinux is disabled.
370       force_non_terminal = true;
371       // This function must 'return true' in this buggy case, otherwise we will
372       // call the destructor too early and subsequent callbacks will crash.
373     }
374 
375     // Remember the state to validate prior state transitions.
376     last_state = state;
377 
378     // The owner of this class should avoid leaking memory once we reach a terminal state.
379     return !IsTerminalState() || force_non_terminal;
380   }
381 
382  public:
383   // Thread safety: Called by main thread, terminates the rx stream.
384   // When this function is invoked, no calls to this class from other threads can occur.
OnCreateFailed()385   void OnCreateFailed() {
386     // returned when an invalid handle is passed to PollState().
387     last_state = State::kSessionNotFound;
388     EmitError("Create returned kInvalidHandle");
389   }
390 
391   // Thread safety: Called by main thread, this could be concurrent to
392   // 'CallbackOnStateChanged'.
BindHandle(const std::shared_ptr<PerfettoConsumerHandle> & handle)393   void BindHandle(const std::shared_ptr<PerfettoConsumerHandle>& handle) {
394     handle_ = handle;
395 
396     // Unblock OnStateChanged.
397     bound_.store(true);  // seq_cst release.
398   }
399 
400 
401   // Called by libperfetto background thread (same one every time) and iorap
402   // thread.
CallbackOnStateChanged(Handle handle,State state,void * callback_arg)403   static void CallbackOnStateChanged(Handle handle, State state, void* callback_arg) {
404     LOG(VERBOSE) << "CallbackOnStateChanged(handle=" << handle << ",state=" << state
405                  << ",callback_arg=" << callback_arg << ")";
406 
407     // Validate OnStateChanged callback invariants, guaranteed by libperfetto.
408     DCHECK_NE(handle, ::perfetto::consumer::kInvalidHandle);
409 
410     // TODO: the memory ordering guarantees should be explicitly specified in consumer_api.h:
411     // This isn't specific enough:
412     // "The callback will be invoked on an internal thread and must not block."
413     // However looking at the implementation it posts onto a single-thread task runner,
414     // so this must be the case.
415 
416     // This current thread owns 'StateChangedSubject', no other threads must access it.
417     // Explicit synchronization is not necessary.
418 
419     {
420       std::lock_guard<std::mutex> guard(StateChangedSubject::state_subject_mutex_);
421       auto it = StateChangedSubject::state_subject_map_.find(handle);
422       // If the object is already deleted, do nothing.
423       if (it == StateChangedSubject::state_subject_map_.end()) {
424         return;
425       }
426 
427       StateChangedSubject* state_subject = it->second;
428       if (!state_subject->OnStateChanged(handle, state)) {
429         // Clean up the state tracker when we reach a terminal state.
430         // This means that no future callbacks will occur anymore.
431         StateChangedSubject::state_subject_map_.erase(it);
432         delete state_subject;
433       }
434     }
435   }
436 
437  private:
EmitError(const std::string & msg)438   void EmitError(const std::string& msg) {
439     // Sidenote: Exact error class does not matter, rxcpp only lets us access the error
440     // as a string (rxcpp::util::what).
441     //
442     // Either way, the recovery strategy is identical (log then try and restart).
443     dest.on_error(rxcpp::util::make_error_ptr(StateChangedError{msg}));
444   }
445 
EmitNext(State state)446   void EmitNext(State state) {
447     if (WOULD_LOG(VERBOSE) && !dest.is_subscribed()) {
448       // This is purely for logging: #on_next already filters out items after unsubscription.
449       LOG(VERBOSE) << "StateChangedSubject#EmitNext(" << state << ") - drop due to unsubscribe";
450     }
451 
452     auto handle_ptr = handle_;
453     DCHECK(handle_ptr != nullptr);
454 
455     // Non-null guarantee for the items emitted into this stream.
456     PerfettoStateChange state_change{state, handle_ptr};
457     dest.on_next(std::move(state_change));
458   }
459 
460   // TODO: inherit from rx subject and handle #unsubscribe explicitly, instead
461   // of just being subject-like?
462 };
463 
464 std::mutex StateChangedSubject::state_subject_mutex_;
465 std::unordered_map<::perfetto::consumer::Handle,
466     StateChangedSubject*> StateChangedSubject::state_subject_map_;
467 
468 // Note: The states will be emitted on a separate thread, so e.g. #as_blocking()
469 // needs to be used to avoid dropping everything on the floor.
470 //
471 // Important: The #on_error case must be handled explicitly by the observable,
472 // because the default behavior is to 'throw' which will cause an std::terminate with -fno-except.
473 static auto /*[observable<State>, shared_ptr<PerfettoConsumerHandle>]*/
CreatePerfettoStateStream(::perfetto::protos::TraceConfig perfetto_config,std::shared_ptr<PerfettoConsumer> perfetto_consumer)474     CreatePerfettoStateStream(::perfetto::protos::TraceConfig perfetto_config,
475                               std::shared_ptr<PerfettoConsumer> perfetto_consumer) {
476   auto obs = rxcpp::observable<>::create<PerfettoStateChange>(
477     [perfetto_config = std::move(perfetto_config), perfetto_consumer = std::move(perfetto_consumer)]
478         (rxcpp::subscriber<PerfettoStateChange> subscriber) {
479       std::unique_ptr<StateChangedSubject> state_subject{
480           new StateChangedSubject{perfetto_config, subscriber, perfetto_consumer}};
481 
482       // Perfetto API requires a pointer to a serialized protobuf, it doesn't accept
483       // the code-generated object.
484       std::string perfetto_config_str = perfetto_config.SerializeAsString();
485 
486       ::perfetto::consumer::Handle handle =
487           perfetto_consumer->Create(perfetto_config_str.data(),
488                                     perfetto_config_str.size(),
489                                     // executes on the same background thread repeatedly.
490                                     &StateChangedSubject::CallbackOnStateChanged,
491                                     // inter-thread-move
492                                     reinterpret_cast<void*>(state_subject.get()));
493       // perfetto::consumer::Create synchronizes-with OnStateChanged callback, this means
494       // we don't need to explicitly synchronize state_subject here so long as we don't access
495       // it on this thread again.
496       LOG(DEBUG) << "Create Perfetto handle " << handle;
497 
498       if (handle == ::perfetto::consumer::kInvalidHandle) {
499         LOG(ERROR) << "Failed to create Perfetto handle";
500         // No callbacks will occur, so our thread still owns the state subject.
501         state_subject->OnCreateFailed();
502         return;
503       }
504 
505       {
506         std::lock_guard<std::mutex> guard(StateChangedSubject::state_subject_mutex_);
507         StateChangedSubject::state_subject_map_[handle] = state_subject.get();
508       }
509 
510       std::shared_ptr<PerfettoConsumerHandle> safe_handle{
511           new PerfettoConsumerHandle{perfetto_consumer, handle}};
512 
513       // Share ownership of the Handle with the StateSubject.
514       // This way we defer calling 'Destroy' until the callback reaches a terminal state
515       // *and* all users of the stream are done with the handle.
516       state_subject->BindHandle(safe_handle);
517 
518       // state_subject ownership is taken over by OnStateChanged.
519       // It will also be touched in a separate thread, so we must never access it here again.
520       state_subject.release();
521 
522       // 'subscriber#add' is actually a call to register an on_unsubscribe listener.
523       subscriber.add([safe_handle]() {
524         LOG(VERBOSE) << "PerfettoStateChange#unsubscribe";
525 
526         // Release our ref-count to the handle.
527         // safe_handle.reset();  // This happens implicitly.
528 
529         // TODO: I think this won't handle the case where we need to shut down early.
530         // Need to use the explicit kShutdown for that?
531       });
532 
533       // TODO: this would be an excellent place to shuffle the perfetto config protobuf
534       // into a global debug state for dumpsys.
535     });
536 
537   return obs;
538 }
539 
540 template <typename T>
WriteFullyToFile(const std::string & path,bool follow_symlinks) const541 bool BinaryWireProtobuf<T>::WriteFullyToFile(const std::string& path,
542                                              bool follow_symlinks) const {
543   // TODO: it would be great if android::base had a string_view overload to avoid copying
544   // data into an std::string.
545 
546   // u  g  o
547   // rw-rw----
548   //
549   // Protobufs can be read/written but not executed.
550   static constexpr const mode_t kMode = S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP;
551 
552   int flags =
553       O_WRONLY | O_CREAT | O_TRUNC | O_CLOEXEC | O_BINARY | (follow_symlinks ? 0 : O_NOFOLLOW);
554   android::base::unique_fd fd(TEMP_FAILURE_RETRY(open(path.c_str(), flags, kMode)));
555 
556   if (fd == -1) {
557     PLOG(ERROR) << "BinaryWireProtobuf::WriteFullyToFile open failed";
558     return false;
559   }
560 
561   if (!::android::base::WriteFully(fd, data_.data(), size())) {
562     PLOG(ERROR) << "BinaryWireProtobuf::WriteFullyToFile write failed";
563     return CleanUpAfterFailedWrite(path);
564   }
565 
566   return true;
567 }
568 
569 template <typename T>
CleanUpAfterFailedWrite(const std::string & path)570 bool BinaryWireProtobuf<T>::CleanUpAfterFailedWrite(const std::string& path) {
571   // Something went wrong. Let's not leave a corrupt file lying around.
572   int saved_errno = errno;
573   unlink(path.c_str());
574   errno = saved_errno;
575   return false;
576 }
577 
578 template <typename T>
WriteStringToFd(int fd) const579 bool BinaryWireProtobuf<T>::WriteStringToFd(int fd) const {
580   const char* p = reinterpret_cast<const char*>(data_.data());
581   size_t left = size();
582   while (left > 0) {
583     ssize_t n = TEMP_FAILURE_RETRY(write(fd, p, left));
584     if (n == -1) {
585       return false;
586     }
587     p += n;
588     left -= n;
589   }
590   return true;
591 }
592 
593 template <typename T>
ReadFullyFromFile(const std::string & path,bool follow_symlinks)594 std::optional<BinaryWireProtobuf<T>> BinaryWireProtobuf<T>::ReadFullyFromFile(
595     const std::string& path,
596     bool follow_symlinks) {
597   std::vector<std::byte> data;
598 
599   int flags = O_RDONLY | O_CLOEXEC | O_BINARY | (follow_symlinks ? 0 : O_NOFOLLOW);
600   android::base::unique_fd fd(TEMP_FAILURE_RETRY(open(path.c_str(), flags)));
601   if (fd == -1) {
602     return std::nullopt;
603   }
604 
605   if (ReadFdToString(fd.get(), /*out*/&data)) {
606     return BinaryWireProtobuf<T>{std::move(data)};
607   } else {
608     return std::nullopt;
609   }
610 }
611 
612 template <typename T>
operator ==(const BinaryWireProtobuf<T> & other) const613 bool BinaryWireProtobuf<T>::operator==(const BinaryWireProtobuf<T>& other) const {
614   if (data_.size() != other.data_.size()) {
615     return false;
616   }
617   return std::equal(data_.begin(), data_.end(), other.data_.begin());
618 }
619 
620 template <typename T>
ReadFdToString(int fd,std::vector<std::byte> * content)621 bool BinaryWireProtobuf<T>::ReadFdToString(int fd, /*out*/std::vector<std::byte>* content) {
622   DCHECK(content != nullptr);
623 
624   content->clear();
625 
626   struct stat sb;
627   if (fstat(fd, /*out*/&sb) != -1 && sb.st_size > 0) {
628     content->reserve(sb.st_size);
629   }
630 
631   char buf[BUFSIZ];
632   auto it = content->begin();
633   ssize_t n;
634   while ((n = TEMP_FAILURE_RETRY(read(fd, &buf[0], sizeof(buf)))) > 0) {
635     content->insert(it,
636                     reinterpret_cast<std::byte*>(&buf[0]),
637                     reinterpret_cast<std::byte*>(&buf[n]));
638 
639     std::advance(/*inout*/it, static_cast<size_t>(n));
640 
641     static_assert(sizeof(char) == sizeof(std::byte), "sanity check for reinterpret cast");
642   }
643   return (n == 0) ? true : false;
644 }
645 
646 // explicit template instantiation.
647 template struct BinaryWireProtobuf<::google::protobuf::MessageLite>;
648 // TODO: refactor this not to need the template instantiation.
649 
650 // Copy of the 2.6.18 kernel header (linux/ioprio.h)
651 
652 #define IOPRIO_WHO_PROCESS (1)
653 #define IOPRIO_CLASS_IDLE (3)
654 
655 #define IOPRIO_BITS		(16)
656 #define IOPRIO_CLASS_SHIFT	(13)
657 #define IOPRIO_PRIO_MASK	((1UL << IOPRIO_CLASS_SHIFT) - 1)
658 
659 #define IOPRIO_PRIO_CLASS(mask)	((mask) >> IOPRIO_CLASS_SHIFT)
660 #define IOPRIO_PRIO_DATA(mask)	((mask) & IOPRIO_PRIO_MASK)
661 #define IOPRIO_PRIO_VALUE(class, data)	(((class) << IOPRIO_CLASS_SHIFT) | data)
662 
ioprio_get(int which,int who)663 static int ioprio_get(int which, int who) {
664   return syscall(SYS_ioprio_get, which, who);
665 }
666 
ioprio_set(int which,int who,int ioprio)667 static int ioprio_set(int which, int who, int ioprio) {
668   return syscall(SYS_ioprio_set, which, who, ioprio);
669 }
670 
671 // An rx Coordination, which will cause a new thread to spawn for each new Worker.
672 //
673 // Idle-class priority is set for the CPU and IO priorities on the new thread.
ObserveOnNewIoThread()674 rxcpp::observe_on_one_worker ObserveOnNewIoThread() {
675   // IO thread factory for idle-priority threads.
676   // Both the CPU scheduler and the IO scheduler are set to idle.
677   //
678   // Use this when needing to schedule disk access from a normal-priority thread onto a
679   // very low priority thread, but not so low that we need to use a BackgroundJobScheduler.
680   struct io_thread_factory {
681     std::thread operator()(std::function<void()> start) const {
682       return std::thread{
683         [start=std::move(start)]() {
684           // Set IO priority to idle.
685           do {
686             int value = ioprio_get(IOPRIO_WHO_PROCESS, /*pid*/0);
687             if (value == -1) {
688               PLOG(ERROR) << "io_thread_factory failed ioprio_get";
689               break;  // Can't set the ioprio, we don't know what data to use.
690             }
691 
692             int data = IOPRIO_PRIO_DATA(value); // priority level
693             // This appears to be '4' in practice. We may want to raise to
694             // be the highest-priority within the idle class.
695 
696             // idle scheduling class. only access disk when nobody else needs disk.
697             int res = ioprio_set(IOPRIO_WHO_PROCESS,
698                                  /*pid*/0,
699                                  IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, data));
700             if (res < 0) {
701               PLOG(ERROR) << "io_thread_factory failed ioprio_set";
702               break;
703             }
704 
705             // Changing the IO priority only has any effect with cfq scheduler:
706             // $> cat /sys/block/sda/queue/scheduler
707             LOG(VERBOSE) << "ioprio_set(WHO_PROCESS, class=IDLE, data=" << data << ")";
708           } while (false);
709 
710           // Set CPU priority to idle.
711           do {
712             struct sched_param param{};
713             param.sched_priority = 0;  // Required to be statically 0 when used with SCHED_IDLE.
714 
715             if (sched_setscheduler(/*pid*/0,  // current thread,
716                                    SCHED_IDLE,
717                                    /*in*/&param) != 0) {
718               PLOG(ERROR) << "io_thread_factory failed sched_setscheduler";
719               break;
720             }
721 
722             LOG(VERBOSE) << "sched_setscheduler(self, IDLE)";
723           } while (false);
724 
725           // XX: if changing the scheduling is too aggressive (i.e. it causes starvation),
726           // we may want to stick with the default class and change the nice (priority) levels
727           // to the minimum.
728 
729           // TODO: future work, maybe use cgroups configuration file instead?
730 
731           // Call the rxcpp-supplied code.
732           start();
733         }
734       };
735     }
736   };
737 
738   static rxcpp::schedulers::scheduler thread_scheduler =
739       rxcpp::schedulers::make_new_thread(io_thread_factory{});
740 
741   static rxcpp::observe_on_one_worker observe_on_io_thread{thread_scheduler};
742 
743   return observe_on_io_thread;
744 }
745 
746 static auto/*observable<PerfettoTraceProto>*/
CreatePerfettoStream(rxcpp::observable<PerfettoStreamCommand> input,std::shared_ptr<PerfettoConsumer> perfetto_consumer,const::perfetto::protos::TraceConfig & trace_config)747     CreatePerfettoStream(rxcpp::observable<PerfettoStreamCommand> input,
748                          std::shared_ptr<PerfettoConsumer> perfetto_consumer,
749                          const ::perfetto::protos::TraceConfig& trace_config) {
750         // XX: should I also take a scheduler for input here???
751 
752   auto /*observable<PerfettoStateChange>*/ perfetto_states =
753     CreatePerfettoStateStream(trace_config, perfetto_consumer);
754 
755   using State = ::perfetto::consumer::State;
756 
757   auto/*coordinator*/ serialize_coordinator = rxcpp::observe_on_new_thread();
758   // Rx note:
759   // The optimal thing to do would be to have a lock/unlock for an entire subset of a chain.
760   // This would avoid creating new threads, and could also be used to intentionally block
761   // the regular C-callback perfetto thread.
762   //
763   // It seems possible to create a coordinator to lock a single operator in a chain, but this
764   // appears to be unsound. In particular, it doesn't even make life any simpler below because
765   // it would only apply the synchronization to 'zip' but not 'flat_map' which is unsound.
766   //
767   // There is also the built-in 'serialize_new_thread' which seems to create a new thread but
768   // then never actually uses it, that seems unfortunate and wasteful.
769   //
770   // Instead, do the simple thing which is create a new thread and always queue on there.
771   // Execution an action on that worker is itself unsynchronized, but this doesn't matter since
772   // the worker is only backed by 1 thread (no 2 schedulables can be executed concurrently
773   // on the 'observe_new_thread' worker).
774   return input
775     .tap([](PerfettoStreamCommand command) {
776            LOG(VERBOSE) << "CreatePerfettoStreamCommand#tap(command=" << command << ")";
777          })
778     // Input A, thread tA. Input B, thread tB. Continue execution with (A,B) on thread tC.
779     .zip(serialize_coordinator,  // rest of chain is also executed on the same thread.
780          perfetto_states)
781     // Note: zip terminates when either of the streams complete.
782     .flat_map(
783          [](std::tuple<PerfettoStreamCommand, PerfettoStateChange> p) {
784            auto& [command, state_change] = p;
785            LOG(VERBOSE) << "CreatePerfettoStream#combine("
786                         << command << "," << state_change << ")";
787            if (command == PerfettoStreamCommand::kShutdown) {
788              // Perfetto: Always safe to call ::perfetto::consumer::Destroy
789              // at any time.
790              //
791              // XX: How do we clean up the StateChangedSubject without racing
792              // against the callback? It strikes me that we may need a 'kDestroyed'
793              // state that perfetto can transition to from kConfigured.
794              LOG(VERBOSE) << "Call Perfetto_Consumer->Destroy";
795              state_change.GetConsumer()->Destroy(state_change.GetHandle());
796 
797              // XX: Do we even have any guarantees about not getting more callbacks?
798              // We could just say 'there can still be spurious output after Shutdown'
799              // and just ignore it (e.g. Shutdown and immediately unsubscribe).
800            } else if (command == PerfettoStreamCommand::kStartTracing
801                           && state_change.state == State::kConfigured) {
802              LOG(VERBOSE) << "Call Perfetto_Consumer->StartTracing";
803              state_change.GetConsumer()->StartTracing(state_change.GetHandle());
804            } else if (command == PerfettoStreamCommand::kStopTracing &&
805                           state_change.state == State::kTraceEnded) {
806              // TODO: if perfetto actually had a 'StopTracing' we could call that here.
807              // right now we just pretend it exists, but rely on the config timer instead.
808              ::perfetto::consumer::TraceBuffer trace_buffer =
809                  state_change.GetConsumer()->ReadTrace(state_change.GetHandle());
810 
811              LOG(VERBOSE) << "Perfetto Trace ended"
812                           << ", addr=" << reinterpret_cast<void*>(trace_buffer.begin)
813                           << ",size= " << trace_buffer.size;
814 
815              PerfettoTraceProto wire_proto{trace_buffer.begin, trace_buffer.size};
816              return rxcpp::observable<>::just(std::move(wire_proto)).as_dynamic();
817            }
818            return rxcpp::observable<>::empty<PerfettoTraceProto>().as_dynamic();
819          }
820     );
821 }
822 
operator <<(std::ostream & os,PerfettoStreamCommand c)823 std::ostream& operator<<(std::ostream& os, PerfettoStreamCommand c) {
824   switch (c) {
825     case PerfettoStreamCommand::kStartTracing:
826       os << "kStartTracing";
827       break;
828     case PerfettoStreamCommand::kStopTracing:
829       os << "kStopTracing";
830       break;
831     case PerfettoStreamCommand::kShutdown:
832       os << "kShutdown";
833       break;
834     default:
835       os << "(unknown)";
836       break;
837   }
838   return os;
839 }
840 
RxProducerFactory(PerfettoDependencies::Injector & injector)841 RxProducerFactory::RxProducerFactory(PerfettoDependencies::Injector& injector)
842   : injector_(injector) {
843 }
844 
845 // TODO: (fruit) maybe this could be streamlined further by avoiding this boilerplate?
CreateTraceStream(rxcpp::observable<PerfettoStreamCommand> commands)846 rxcpp::observable<PerfettoTraceProto> RxProducerFactory::CreateTraceStream(
847     rxcpp::observable<PerfettoStreamCommand> commands) {
848   std::shared_ptr<PerfettoConsumer> perfetto_consumer =
849       injector_.get<std::shared_ptr<PerfettoConsumer>>();
850   const ::perfetto::protos::TraceConfig& trace_config =
851       injector_.get<::perfetto::protos::TraceConfig>();
852 
853   DCHECK(perfetto_consumer != nullptr);
854   DCHECK(reinterpret_cast<volatile const void*>(&trace_config) != nullptr);
855 
856   return CreatePerfettoStream(commands,
857                               perfetto_consumer,
858                               trace_config);
859 }
860 
861 // For testing/debugging only.
862 //
863 // Saves protobuf results in file name specified by 'arg_output_proto'.
CollectPerfettoTraceBufferImmediately(RxProducerFactory & producer_factory,const std::string & arg_output_proto)864 void CollectPerfettoTraceBufferImmediately(
865     RxProducerFactory& producer_factory,
866     const std::string& arg_output_proto) {
867   LOG(VERBOSE) << "CollectPerfettoTraceBufferImmediately";
868 
869   std::shared_ptr<PerfettoConsumer> perfetto_consumer =
870       producer_factory.injector_.get<std::shared_ptr<PerfettoConsumer>>();
871   const ::perfetto::protos::TraceConfig& trace_config =
872       producer_factory.injector_.get<const ::perfetto::protos::TraceConfig&>();
873 
874   auto /*observable<PerfettoStateChange>*/ perfetto_states =
875     CreatePerfettoStateStream(trace_config, perfetto_consumer);
876 
877   perfetto_states
878     .as_blocking()  // Wait for observable to terminate with on_completed or on_error.
879     .subscribe(/*on_next*/[&](auto state_change) {
880        LOG(VERBOSE) << "Perfetto post-processed State change: " << state_change;
881 
882        using State = ::perfetto::consumer::State;
883        switch (state_change.state) {
884          case State::kConnecting:
885            LOG(VERBOSE) << "Perfetto Tracing is Connecting";
886            // Transitional state. No-op.
887            break;
888          case State::kConfigured:
889            state_change.GetConsumer()->StartTracing(state_change.GetHandle());
890            break;
891          case State::kTracing:
892            LOG(VERBOSE) << "Perfetto Tracing started";
893            // Transitional state. No-op.
894            break;
895          case State::kTraceEnded: {
896            ::perfetto::consumer::TraceBuffer trace_buffer =
897              state_change.GetConsumer()->ReadTrace(state_change.GetHandle());
898 
899            LOG(VERBOSE) << "Perfetto Trace ended"
900                         << ", addr=" << reinterpret_cast<void*>(trace_buffer.begin)
901                         << ",size= " << trace_buffer.size;
902 
903            if (!arg_output_proto.empty()) {
904              std::string trace_buffer_str;
905              trace_buffer_str.resize(trace_buffer.size);
906              std::copy(trace_buffer.begin,
907                        trace_buffer.begin + trace_buffer.size,
908                        trace_buffer_str.data());
909              if (!android::base::WriteStringToFile(trace_buffer_str, arg_output_proto)) {
910                LOG(ERROR) << "Failed to save TraceBuffer to " << arg_output_proto;
911              } else {
912                LOG(INFO) << "TraceBuffer saved to file: " << arg_output_proto;
913                LOG(INFO);
914                LOG(INFO) << "To print this in a human readable form, execute these commands:";
915                LOG(INFO) << "$> adb pull '" << arg_output_proto << "'";
916                LOG(INFO) << "$> trace_to_text systrace <filename.pb>";
917              }
918            }
919 
920            // TODO: something more useful with this TraceBuffer, such as saving it to a file
921            // and printing the output.
922            break;
923          }
924          default:
925            // No other states are possible, because they go to #on_error or cause a dcheck.
926            DCHECK(false) << "Invalid state: " << state_change;
927        }
928 
929        //INTENTIONAL_COMPILER_ERROR_HERE // lets make sure this code actually does a trace.
930 
931      }, /*on_error*/[](rxcpp::util::error_ptr err) {
932        LOG(ERROR) << "Perfetto post-processed state change failed: " << rxcpp::util::what(err);
933      }, /*on_completed*/[]() {
934        LOG(VERBOSE) << "Perfetto post-processed State #on_completed";
935      });
936 }
937 
938 
939 }  // namespace iorap::perfetto
940