aosp12/external/libwebm/webm_parser

WebM Parser

Introduction

This WebM parser is a C++11-based parser that aims to be a safe and complete parser for WebM. It supports all WebM elements (from the old deprecated ones to the newest ones like Colour), including recursive elements like ChapterAtom and SimpleTag. It supports incremental parsing; parsing may be stopped at any point and resumed later as needed. It also supports starting at an arbitrary WebM element, so parsing need not start from the beginning of the file.

The parser (WebmParser) works by being fed input data from a data source (an instance of Reader) that represents a WebM file. The parser will parse the WebM data into various data structures that represent the encoded WebM elements, and then call corresponding Callback event methods as the data structures are parsed.

Building

CMake support has been added to the root libwebm CMakeLists.txt file. Simply enable the ENABLE_WEBM_PARSER feature if using the interactive CMake builder, or alternatively pass the -DENABLE_WEBM_PARSER:BOOL=ON flag from the command line. By default, this parser is not enabled when building libwebm, so you must explicitly enable it.

Alternatively, the following illustrates the minimal commands necessary to compile the code into a static library without CMake:

c++ -Iinclude -I. -std=c++11 -c src/*.cc
ar rcs libwebm.a *.o

Using the parser

There are 3 basic components in the parser that are used: Reader, Callback, and WebmParser.

Reader

The Reader interface acts as a data source for the parser. You may subclass it and implement your own data source if you wish. Alternatively, use the FileReader, IstreamReader, or BufferReader if you wish to read from a FILE*, std::istream, or std::vector<std::uint8_t>, respectively.

The parser supports Reader implementations that do short reads. If Reader::Skip() or Reader::Read() do a partial read (returning Status::kOkPartial), the parser will call them again in an attempt to read more data. If no data is available, the Reader may return some other status (like Status::kWouldBlock) to indicate that no data is available. In this situation, the parser will stop parsing and return the status it received. Parsing may be resumed later when more data is available.

When the Reader has reached the end of the WebM document and no more data is available, it should return Status::kEndOfFile. This will cause parsing to stop. If the file ends at a valid location (that is, there aren't any elements that have specified a size that indicates the file ended prematurely), the parser will translate Status::kEndOfFile into Status::kOkCompleted and return it. If the file ends prematurely, the parser will return Status::kEndOfFile to indicate that.

Note that if the WebM file contains elements that have an unknown size (or a seek has been performed and the parser doesn't know the size of the root element(s)), and the parser is parsing them and hits end-of-file, the parser may still call Reader::Read()/Reader::Skip() multiple times (even though they've already reported Status::kEndOfFile) as nested parsers terminate parsing. Because of this, Reader::Read()/Reader::Skip() implementations should be able to handle being called multiple times after the file's end has been reached, and they should consistently return Status::kEndOfFile.

The three provided readers (FileReader, IstreamReader, and BufferReader) are blocking implementations (they won't return Status::kWouldBlock), so if you're using them the parser will run until it entirely consumes all their data (unless, of course, you request the parser to stop via Callback... see the next section).

Callback

As the parser progresses through the file, it builds objects (see webm/dom_types.h) that represent parsed data structures. The parser then notifies the Callback implementation as objects complete parsing. For some data structures (like frames or Void elements), the parser notifies the Callback and requests it to consume the data directly from the Reader (this is done for structures that can be large/frequent binary blobs in order to allow you to read the data directly into the object/type of your choice, rather than just reading them into a std::vector<std::uint8_t> and making you copy it into a different object if you wanted to work with something other than std::vector<std::uint8_t>).

The parser was designed to parse the data into objects that are small enough that the Callback can be quickly and frequently notified as soon as the object is ready, but large enough that the objects received by the Callback are still useful. Having Callback events for every tiny integer/float/string/etc. element would require too much assembly and work to be useful to most users, and pasing the file into a single DOM tree (or a small handful of large conglomerate structures) would unnecessarily delay video playback or consume too much memory on smaller devices.

The parser may call the following methods while nearly anywhere in the file:

• Callback::OnElementBegin(): This is called for every element that the parser encounters. This is primarily useful if you want to skip some elements or build a map of every element in the file.
• Callback::OnUnknownElement(): This is called when an element is either not a valid/recognized WebM element, or it is a WebM element but is improperly nested (e.g. an EBMLVersion element inside of a Segment element). The parser doesn't know how to handle the element; it could just skip it but instead defers to the Callback to decide how it should be handled. The default implementation just skips the element.
• Callback::OnVoid(): Void elements can appear anywhere in any master element. This method will be called to handle the Void element.

The parser may call the following methods in the proper nesting order, as shown in the list. A *Begin() method will always be matched up with its corresponding *End() method (unless a seek has been performed). The parser will only call the methods in the proper nesting order as specified in the WebM DOM. For example, Callback::OnEbml() will never be called in between Callback::OnSegmentBegin()/Callback::OnSegmentEnd() (since the EBML element is not a child of the Segment element), and Callback::OnTrackEntry() will only ever be called in between Callback::OnSegmentBegin()/Callback::OnSegmentEnd() (since the TrackEntry element is a (grand-)child of the Segment element and must be contained by a Segment element). Callback::OnFrame() is listed twice because it will be called to handle frames contained in both SimpleBlock and Block elements.

• Callback::OnEbml()
• Callback::OnSegmentBegin()
  • Callback::OnSeek()
  • Callback::OnInfo()
  • Callback::OnClusterBegin()
    • Callback::OnSimpleBlockBegin()
      • Callback::OnFrame()
    • Callback::OnSimpleBlockEnd()
    • Callback::OnBlockGroupBegin()
      • Callback::OnBlockBegin()
        • Callback::OnFrame()
      • Callback::OnBlockEnd()
    • Callback::OnBlockGroupEnd()
  • Callback::OnClusterEnd()
  • Callback::OnTrackEntry()
  • Callback::OnCuePoint()
  • Callback::OnEditionEntry()
  • Callback::OnTag()
• Callback::OnSegmentEnd()

Only Callback::OnFrame() (and no other Callback methods) will be called in between Callback::OnSimpleBlockBegin()/Callback::OnSimpleBlockEnd() or Callback::OnBlockBegin()/Callback::OnBlockEnd(), since the SimpleBlock and Block elements are not master elements only contain frames.

Note that seeking into the middle of the file may cause the parser to skip some *Begin() methods. For example, if a seek is performed to a SimpleBlock element, Callback::OnSegmentBegin() and Callback::OnClusterBegin() will not be called. In this situation, the full sequence of callback events would be (assuming the file ended after the SimpleBlock): Callback::OnSimpleBlockBegin(), Callback::OnFrame() (for every frame in the SimpleBlock), Callback::OnSimpleBlockEnd(), Callback::OnClusterEnd(), and Callback::OnSegmentEnd(). Since the Cluster and Segment elements were skipped, the Cluster DOM object may have some members marked as absent, and the *End() events for the Cluster and Segment elements will have metadata with unknown header position, header length, and body size (see kUnknownHeaderSize, kUnknownElementSize, and kUnknownElementPosition).

When a Callback method has completed, it should return Status::kOkCompleted to allow parsing to continue. If you would like parsing to stop, return any other status code (except Status::kEndOfFile, since that's treated somewhat specially and is intended for Readers to use), which the parser will return. If you return a non-parsing-error status code (.e.g. Status::kOkPartial, Status::kWouldBlock, etc. or your own status code with a value > 0), parsing may be resumed again. When parsing is resumed, the parser will call the same callback method again (and once again, you may return Status::kOkCompleted to let parsing continue or some other value to stop parsing).

You may subclass the Callback element and override methods which you are interested in receiving events for. By default, methods taking an Action parameter will set it to Action::kRead so the entire file is parsed. The Callback::OnFrame() method will just skip over the frame bytes by default.

WebmParser

The actual parsing work is done with WebmParser. Simply construct a WebmParser and call WebmParser::Feed() (providing it a Callback and Reader instance) to parse a file. It will return Status::kOkCompleted when the entire file has been successfully parsed. WebmParser::Feed() doesn't store any internal references to the Callback or Reader.

If you wish to start parsing from the middle of a file, call WebmParser::DidSeek() before calling WebmParser::Feed() to prepare the parser to receive data starting at an arbitrary point in the file. When seeking, you should seek to the beginning of a WebM element; seeking to a location that is not the start of a WebM element (e.g. seeking to a frame, rather than its containing SimpleBlock/Block element) will cause parsing to fail. Calling WebmParser::DidSeek() will reset the state of the parser and clear any internal errors, so a WebmParser instance may be reused (even if it has previously failed to parse a file).

Building your program

The following program is a small program that completely parses a file from stdin:

#include <webm/callback.h>
#include <webm/file_reader.h>
#include <webm/webm_parser.h>

int main() {
  webm::Callback callback;
  webm::FileReader reader(std::freopen(nullptr, "rb", stdin));
  webm::WebmParser parser;
  parser.Feed(&callback, &reader);
}

It completely parses the input file, but we need to make a new class that derives from Callback if we want to receive any parsing events. So if we change it to:

#include <iomanip>
#include <iostream>

#include <webm/callback.h>
#include <webm/file_reader.h>
#include <webm/status.h>
#include <webm/webm_parser.h>

class MyCallback : public webm::Callback {
 public:
  webm::Status OnElementBegin(const webm::ElementMetadata& metadata,
                              webm::Action* action) override {
    std::cout << "Element ID = 0x"
              << std::hex << static_cast<std::uint32_t>(metadata.id);
    std::cout << std::dec;  // Reset to decimal mode.
    std::cout << " at position ";
    if (metadata.position == webm::kUnknownElementPosition) {
      // The position will only be unknown if we've done a seek. But since we
      // aren't seeking in this demo, this will never be the case. However, this
      // if-statement is included for completeness.
      std::cout << "<unknown>";
    } else {
      std::cout << metadata.position;
    }
    std::cout << " with header size ";
    if (metadata.header_size == webm::kUnknownHeaderSize) {
      // The header size will only be unknown if we've done a seek. But since we
      // aren't seeking in this demo, this will never be the case. However, this
      // if-statement is included for completeness.
      std::cout << "<unknown>";
    } else {
      std::cout << metadata.header_size;
    }
    std::cout << " and body size ";
    if (metadata.size == webm::kUnknownElementSize) {
      // WebM master elements may have an unknown size, though this is rare.
      std::cout << "<unknown>";
    } else {
      std::cout << metadata.size;
    }
    std::cout << '\n';

    *action = webm::Action::kRead;
    return webm::Status(webm::Status::kOkCompleted);
  }
};

int main() {
  MyCallback callback;
  webm::FileReader reader(std::freopen(nullptr, "rb", stdin));
  webm::WebmParser parser;
  webm::Status status = parser.Feed(&callback, &reader);
  if (status.completed_ok()) {
    std::cout << "Parsing successfully completed\n";
  } else {
    std::cout << "Parsing failed with status code: " << status.code << '\n';
  }
}

This will output information about every element in the entire file: it's ID, position, header size, and body size. The status of the parse is also checked and reported.

For a more complete example, see demo/demo.cc, which parses an entire file and prints out all of its information. That example overrides every Callback method to show exactly what information is available while parsing and how to access it. The example is verbose, but that's primarily due to pretty-printing and string formatting operations.

When compiling your program, add the include directory to your compiler's header search paths and link to the compiled library. Be sure your compiler has C++11 mode enabled (-std=c++11 in clang++ or g++).

Testing

Unit tests are located in the tests directory. Google Test and Google Mock are used as testing frameworks. Building and running the tests will be supported in the upcoming CMake scripts, but they can currently be built and run by manually compiling them (and linking to Google Test and Google Mock).

Fuzzing

The parser has been fuzzed with AFL and libFuzzer. If you wish to fuzz the parser with AFL or libFuzzer but don't want to write an executable that exercises the parsing API, you may use fuzzing/webm_fuzzer.cc.

When compiling for fuzzing, define the macro WEBM_FUZZER_BYTE_ELEMENT_SIZE_LIMIT to be some integer in order to limit the maximum size of ASCII/UTF-8/binary elements. It's too easy for the fuzzer to generate elements that claim to have a ridiculously massive size, which will cause allocations to fail or the program to allocate too much memory. AFL will terminate the process if it allocates too much memory (by default, 50 MB), and the [Address Sanitizer doesn't throw std::bad_alloc when an allocation fails] (https://github.com/google/sanitizers/issues/295). Defining WEBM_FUZZER_BYTE_ELEMENT_SIZE_LIMIT to a low number (say, 1024) will cause the ASCII/UTF-8/binary element parsers to return Status::kNotEnoughMemory if the element's size exceeds WEBM_FUZZER_BYTE_ELEMENT_SIZE_LIMIT, which will avoid false positives when fuzzing. The parser expects std::string and std::vector to throw std::bad_alloc when an allocation fails, which doesn't necessarily happen due to the fuzzers' limitations.

You may also define the macro WEBM_FUZZER_SEEK_FIRST to have fuzzing/webm_fuzzer.cc call WebmParser::DidSeek() before doing any parsing. This will test the seeking code paths.