summaryrefslogtreecommitdiffstats
path: root/third_party/libwebrtc/g3doc/implementation_basics.md
blob: ae1f199b68ab2f71237c874239dc6690af3031f5 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
<!-- go/cmark -->
<!--* freshness: {owner: 'hta' reviewed: '2021-05-31'} *-->

# Basic concepts and primitives

## Time

Internally, time is represent using the [webrtc::Timestamp][1] class. This
represents
time with a resolution of one microsecond, using a 64-bit integer, and provides
converters to milliseconds or seconds as needed.

All timestamps need to be measured from the system monotonic time.

The epoch is not specified (because we can't always know if the system clock is
correct), but whenever an absolute epoch is needed, the Unix time
epoch (Jan 1, 1970 at 0:00 GMT) is used.

Conversion from/to other formats (for example milliseconds, NTP times,
timestamp strings) should happen as close to the interface requiring that
format as possible.

NOTE: There are parts of the codebase that don't use Timestamp, parts of the
codebase that use the NTP epoch, and parts of the codebase that don't use the
monotonic clock. They need to
be updated.

## Threads

All execution happens on a TaskQueue instance. How a TaskQueue is implemented
varies by platform, but they all have the [webrtc::TaskQueueBase][3] API.

This API offers primitives for posting tasks, with or without delay.

Some core parts use the [rtc::Thread][2], which is a subclass of TaskQueueBase.
This may contain a SocketServer for processing I/O, and is used for policing
certain calling pattern between a few core threads (the NetworkThread cannot
do Invoke on the Worker thread, for instance).

## Reserved class suffixes

C++ classes with names ending in the suffixes "Factory", "Builder" and "Manager" are supposed to behave
in certain well known ways.

For a particular class name Foo, the following classes, if they exist, should
behave as follows:

* FooFactory: Has a Create function that creates a Foo object and returns the
  object or an owning reference to it (for instance std::unique_ptr or
  rtc::scoped_refptr<Foo>). The Create function should NOT alter the factory
  state; ideally, it is marked const. Ownership of the returned object is only
  with the caller.

* FooBuilder: Has a Build function that returns ownership of a Foo object (as
  above). The Builder can only be used once, and resources given to the Builder
  before the Build function is called are either released or owned by the Foo
  object. The Create function may be reference-qualified (declared as ```Foo
  Build() &&```), which means it is invoked as ```std::move(builder).Build()```,
  and C++ will ensure that it is not used again.

* FooManager: Has a Create function that returns an rtc::scoped_refptr<Foo> (if
  shared ownership) or a Foo* (if the Manager retains sole ownership). If
  Create() cannot fail, consider returning a Foo&. The Manager is responsible
  for keeping track of the object; if the Create function returns a Foo*, the
  Foo object is guaranteed to be destroyed when the FooManager is destroyed.

If a Manager class manages multiple classes of objects, the Create functions
should be appropriately named (the FooAndBarManager would have CreateFoo() and
CreateBar() functions), and the class will have a suitable name for the group of
objects it is managing.

FooFactory is mainly useful for the case where preparation for producing Foo
objects is complex. If Foo can be created with just an argument list, consider
exposing its constructor instead; if Foo creation can fail, consider having
a free function called CreateFoo instead of a factory.

Note that classes with these names exist that do not follow these conventions.
When they are detected, they need to be marked with TODO statements and bugs
filed on them to get them into a conformant state.

## Synchronization primitives

### PostTask and thread-guarded variables

The preferred method for synchronization is to post tasks between threads,
and to let each thread take care of its own variables (lock-free programming).
All variables in
classes intended to be used with multiple threads should therefore be
annotated with RTC_GUARDED_BY(thread).

For classes used with only one thread, the recommended pattern is to let
them own a webrtc::SequenceChecker (conventionally named sequence_checker_)
and let all variables be RTC_GUARDED_BY(sequence_checker_).

Member variables marked const do not need to be guarded, since they never
change. (But note that they may point to objects that can change!)

When posting tasks with callbacks, it is the duty of the caller to check
that the object one is calling back into still exists when the callback
is made. A helper for this task is the [webrtc::ScopedTaskSafety][5]
flag, which can automatically drop callbacks in this situation, and
associated classes.

### Synchronization primitives to be used when needed

When it is absolutely necessary to let one thread wait for another thread
to do something, Thread::Invoke can be used. This function is DISCOURAGED,
since it leads to performance issues, but is currently still widespread.

When it is absolutely necessary to access one variable from multiple threads,
the webrtc::Mutex can be used. Such variables MUST be marked up with
RTC_GUARDED_BY(mutex), to allow static analysis that lessens the chance of
deadlocks or unintended consequences.

### Synchronization primitives that are being removed
The following non-exhaustive list of synchronization primitives are
in the (slow) process of being removed from the codebase.

* sigslot. Use [webrtc::CallbackList][4] instead, or, when there's only one
  signal consumer, a single std::function.
  
* AsyncInvoker.

* RecursiveCriticalSection. Try to use [webrtc::Mutex][6] instead, and don't recurse.

## Enum-To-String functions
If there is a need to convert an enum to a string representation, such as for
enums exposed at the Javascript API interface, the recommended way is to write
a function named AsString, declared "static constexpr" and returning an
absl::string_view. The declaration should be right after the enum declaration,
in the same scope; the implementation (which must be marked "inline") should
be at the end of the same header file.

If the enum is not defined within a class, the "static" keyword is not needed.

[1]: https://source.chromium.org/chromium/chromium/src/+/main:third_party/webrtc/api/units/timestamp.h;drc=b95d90b78a3491ef8e8aa0640dd521515ec881ca;l=29
[2]: https://source.chromium.org/chromium/chromium/src/+/main:third_party/webrtc/rtc_base/thread.h;drc=1107751b6f11c35259a1c5c8a0f716e227b7e3b4;l=194
[3]: https://source.chromium.org/chromium/chromium/src/+/main:third_party/webrtc/api/task_queue/task_queue_base.h;drc=1107751b6f11c35259a1c5c8a0f716e227b7e3b4;l=25
[4]: https://source.chromium.org/chromium/chromium/src/+/main:third_party/webrtc/rtc_base/callback_list.h;drc=54b91412de3f579a2d5ccdead6e04cc2cc5ca3a1;l=162
[5]: https://source.chromium.org/chromium/chromium/src/+/main:third_party/webrtc/rtc_base/task_utils/pending_task_safety_flag.h;drc=86ee89f73e4f4799b3ebcc0b5c65837c9601fe6d;l=117
[6]: https://source.chromium.org/chromium/chromium/src/+/main:third_party/webrtc/rtc_base/synchronization/mutex.h;drc=0d3c09a8fe5f12dfbc9f1bcd5790fda8830624ec;l=40