Nanopb: Basic concepts

The things outlined here are the underlying concepts of the nanopb design.

Proto files

All Protocol Buffers implementations use .proto files to describe the message format. The point of these files is to be a portable interface description language.

Compiling .proto files for nanopb

Nanopb comes with a Python script to generate .pb.c and .pb.h files from the .proto definition:

user@host:~$ nanopb/generator/nanopb_generator.py message.proto
Writing to message.pb.h and message.pb.c

Internally this script uses Google protoc to parse the input file. If you do not have it available, you may receive an error message. You can install either grpcio-tools Python package using pip, or the protoc compiler itself from protobuf-compiler distribution package. Generally the Python package is recommended, because nanopb requires protoc version 3.6 or newer to support all features, and some distributions come with an older version.

Modifying generator behaviour

Using generator options, you can set maximum sizes for fields in order to allocate them statically. The preferred way to do this is to create an .options file with the same name as your .proto file:

# Foo.proto
message Foo {
   required string name = 1;
}

# Foo.options
Foo.name max_size:16

For more information on this, see the Proto file options section in the reference manual.

Streams

Nanopb uses streams for accessing the data in encoded format. The stream abstraction is very lightweight, and consists of a structure (pb_ostream_t or pb_istream_t) which contains a pointer to a callback function.

There are a few generic rules for callback functions:

1. Return false on IO errors. The encoding or decoding process will abort immediately.
2. Use state to store your own data, such as a file descriptor.
3. bytes_written and bytes_left are updated by pb_write and pb_read.
4. Your callback may be used with substreams. In this case bytes_left, bytes_written and max_size have smaller values than the original stream. Don’t use these values to calculate pointers.
5. Always read or write the full requested length of data. For example, POSIX recv() needs the MSG_WAITALL parameter to accomplish this.

Output streams

struct _pb_ostream_t
{
   bool (*callback)(pb_ostream_t *stream, const uint8_t *buf, size_t count);
   void *state;
   size_t max_size;
   size_t bytes_written;
};

The callback for output stream may be NULL, in which case the stream simply counts the number of bytes written. In this case, max_size is ignored.

Otherwise, if bytes_written + bytes_to_be_written is larger than max_size, pb_write returns false before doing anything else. If you don't want to limit the size of the stream, pass SIZE_MAX.

Example 1:

This is the way to get the size of the message without storing it anywhere:

Person myperson = ...;
pb_ostream_t sizestream = {0};
pb_encode(&sizestream, Person_fields, &myperson);
printf("Encoded size is %d\n", sizestream.bytes_written);

Example 2:

Writing to stdout:

bool callback(pb_ostream_t `stream, const uint8_t `buf, size_t count)
{
   FILE *file = (FILE*) stream->state;
   return fwrite(buf, 1, count, file) == count;
}

pb_ostream_t stdoutstream = {&callback, stdout, SIZE_MAX, 0};

Input streams

For input streams, there is one extra rule:

6. You don’t need to know the length of the message in advance. After getting EOF error when reading, set bytes_left to 0 and return false. pb_decode() will detect this and if the EOF was in a proper position, it will return true.

Here is the structure:

struct _pb_istream_t
{
   bool (*callback)(pb_istream_t *stream, uint8_t *buf, size_t count);
   void *state;
   size_t bytes_left;
};

The callback must always be a function pointer. Bytes_left is an upper limit on the number of bytes that will be read. You can use SIZE_MAX if your callback handles EOF as described above.

Example:

This function binds an input stream to stdin:

bool callback(pb_istream_t *stream, uint8_t *buf, size_t count)
{
   FILE *file = (FILE*)stream->state;
   bool status;

   if (buf == NULL)
   {
       while (count-- && fgetc(file) != EOF);
       return count == 0;
   }

   status = (fread(buf, 1, count, file) == count);

   if (feof(file))
       stream->bytes_left = 0;

   return status;
}

pb_istream_t stdinstream = {&callback, stdin, SIZE_MAX};

Data types

Most Protocol Buffers datatypes have directly corresponding C datatypes, such as int32 is int32_t, float is float and bool is bool. However, the variable-length datatypes are more complex:

1. Strings, bytes and repeated fields of any type map to callback functions by default.
2. If there is a special option (nanopb).max_length or (nanopb).max_size specified in the .proto file, string maps to null-terminated char array and bytes map to a structure containing a char array and a size field.
3. If (nanopb).fixed_length is set to true and (nanopb).max_size is also set, then bytes map to an inline byte array of fixed size.
4. If there is a special option (nanopb).max_count specified on a repeated field, it maps to an array of whatever type is being repeated. Another field will be created for the actual number of entries stored.
5. If (nanopb).fixed_count is set to true and (nanopb).max_count is also set, the field for the actual number of entries will not by created as the count is always assumed to be max count.

Examples of .proto specifications vs. generated structure

Simple integer field:
.proto: int32 age = 1;
.pb.h: int32_t age;

String with unknown length:
.proto: string name = 1;
.pb.h: pb_callback_t name;

String with known maximum length:
.proto: string name = 1 [(nanopb).max_length = 40];
.pb.h: char name[41];

Repeated string with unknown count:
.proto: repeated string names = 1;
.pb.h: pb_callback_t names;

Repeated string with known maximum count and size:
.proto: repeated string names = 1 [(nanopb).max_length = 40, (nanopb).max_count = 5];
.pb.h: size_t names_count; char names[5][41];

Bytes field with known maximum size:
.proto: bytes data = 1 [(nanopb).max_size = 16];
.pb.h: PB_BYTES_ARRAY_T(16) data;, where the struct contains {pb_size_t size; pb_byte_t bytes[n];}

Bytes field with fixed length:
.proto: bytes data = 1 [(nanopb).max_size = 16, (nanopb).fixed_length = true];
.pb.h: pb_byte_t data[16];

Repeated integer array with known maximum size:
.proto: repeated int32 numbers = 1 [(nanopb).max_count = 5];
.pb.h: pb_size_t numbers_count; int32_t numbers[5];

Repeated integer array with fixed count:
.proto: repeated int32 numbers = 1 [(nanopb).max_count = 5, (nanopb).fixed_count = true];
.pb.h: int32_t numbers[5];

The maximum lengths are checked in runtime. If string/bytes/array exceeds the allocated length, pb_decode() will return false.

Note: For the bytes datatype, the field length checking may not be exact. The compiler may add some padding to the pb_bytes_t structure, and the nanopb runtime doesn’t know how much of the structure size is padding. Therefore it uses the whole length of the structure for storing data, which is not very smart but shouldn’t cause problems. In practise, this means that if you specify (nanopb).max_size=5 on a bytes field, you may be able to store 6 bytes there. For the string field type, the length limit is exact.

Note: The decoder only keeps track of one fixed_count repeated field at a time. Usually this it not an issue because all elements of a repeated field occur end-to-end. Interleaved array elements of several fixed_count repeated fields would be a valid protobuf message, but would get rejected by nanopb decoder with error "wrong size for fixed count field".

Field callbacks

The easiest way to handle repeated fields is to specify a maximum size for them, as shown in the previous section. However, sometimes you need to be able to handle arrays with unlimited length, possibly larger than available RAM memory.

For these cases, nanopb provides a callback interface. Nanopb core invokes the callback function when it gets to the specific field in the message. Your code can then handle the field in custom ways, for example decode the data piece-by-piece and store to filesystem.

The pb_callback_t structure contains a function pointer and a void pointer called arg you can use for passing data to the callback. If the function pointer is NULL, the field will be skipped. A pointer to the arg is passed to the function, so that it can modify it and retrieve the value.

The actual behavior of the callback function is different in encoding and decoding modes. In encoding mode, the callback is called once and should write out everything, including field tags. In decoding mode, the callback is called repeatedly for every data item.

To write more complex field callbacks, it is recommended to read the Google Protobuf Encoding Specification.

Encoding callbacks

bool (*encode)(pb_ostream_t *stream, const pb_field_iter_t *field, void * const *arg);

When encoding, the callback should write out complete fields, including the wire type and field number tag. It can write as many or as few fields as it likes. For example, if you want to write out an array as repeated field, you should do it all in a single call.

Usually you can use pb_encode_tag_for_field to encode the wire type and tag number of the field. However, if you want to encode a repeated field as a packed array, you must call pb_encode_tag instead to specify a wire type of PB_WT_STRING.

If the callback is used in a submessage, it will be called multiple times during a single call to pb_encode. In this case, it must produce the same amount of data every time. If the callback is directly in the main message, it is called only once.

This callback writes out a dynamically sized string:

bool write_string(pb_ostream_t *stream, const pb_field_iter_t *field, void * const *arg)
{
    char *str = get_string_from_somewhere();
    if (!pb_encode_tag_for_field(stream, field))
        return false;

    return pb_encode_string(stream, (uint8_t*)str, strlen(str));
}

Decoding callbacks

bool (*decode)(pb_istream_t *stream, const pb_field_iter_t *field, void **arg);

When decoding, the callback receives a length-limited substring that reads the contents of a single field. The field tag has already been read. For string and bytes, the length value has already been parsed, and is available at stream->bytes_left.

The callback will be called multiple times for repeated fields. For packed fields, you can either read multiple values until the stream ends, or leave it to pb_decode to call your function over and over until all values have been read.

This callback reads multiple integers and prints them:

bool read_ints(pb_istream_t *stream, const pb_field_iter_t *field, void **arg)
{
    while (stream->bytes_left)
    {
        uint64_t value;
        if (!pb_decode_varint(stream, &value))
            return false;
        printf("%lld\n", value);
    }
    return true;
}

Function name bound callbacks

bool MyMessage_callback(pb_istream_t *istream, pb_ostream_t *ostream, const pb_field_iter_t *field);

Storing function pointer in pb_callback_t fields inside the message requires extra storage space and is often cumbersome. As an alternative, the generator options callback_function and callback_datatype can be used to bind a callback function based on its name.

Typically this feature is used by setting callback_datatype to e.g. void\* or even a struct type used to store encoded or decoded data. The generator will automatically set callback_function to MessageName_callback and produce a prototype for it in generated .pb.h. By implementing this function in your own code, you will receive callbacks for fields without having to separately set function pointers.

If you want to use function name bound callbacks for some fields and pb_callback_t for other fields, you can call pb_default_field_callback from the message-level callback. It will then read a function pointer from pb_callback_t and call it.

Message descriptor

For using the pb_encode() and pb_decode() functions, you need a description of all the fields contained in a message. This description is usually autogenerated from .proto file.

For example this submessage in the Person.proto file:

message Person {
    message PhoneNumber {
        required string number = 1 [(nanopb).max_size = 40];
        optional PhoneType type = 2 [default = HOME];
    }
}

This in turn generates a macro list in the .pb.h file:

#define Person_PhoneNumber_FIELDLIST(X, a) \
X(a, STATIC,   REQUIRED, STRING,   number,            1) \
X(a, STATIC,   OPTIONAL, UENUM,    type,              2)

Inside the .pb.c file there is a macro call to PB_BIND:

PB_BIND(Person_PhoneNumber, Person_PhoneNumber, AUTO)

These macros will in combination generate pb_msgdesc_t structure and associated lists:

const uint32_t Person_PhoneNumber_field_info[] = { ... };
const pb_msgdesc_t * const Person_PhoneNumber_submsg_info[] = { ... };
const pb_msgdesc_t Person_PhoneNumber_msg = {
  2,
  Person_PhoneNumber_field_info,
  Person_PhoneNumber_submsg_info,
  Person_PhoneNumber_DEFAULT,
  NULL,
};

The encoding and decoding functions take a pointer to this structure and use it to process each field in the message.

Oneof

Protocol Buffers supports oneof sections, where only one of the fields contained within can be present. Here is an example of oneof usage:

message MsgType1 {
    required int32 value = 1;
}

message MsgType2 {
    required bool value = 1;
}

message MsgType3 {
    required int32 value1 = 1;
    required int32 value2 = 2;
} 

message MyMessage {
    required uint32 uid = 1;
    required uint32 pid = 2;
    required uint32 utime = 3;

    oneof payload {
        MsgType1 msg1 = 4;
        MsgType2 msg2 = 5;
        MsgType3 msg3 = 6;
    }
}

Nanopb will generate payload as a C union and add an additional field which_payload:

typedef struct _MyMessage {
  uint32_t uid;
  uint32_t pid;
  uint32_t utime;
  pb_size_t which_payload;
  union {
      MsgType1 msg1;
      MsgType2 msg2;
      MsgType3 msg3;
  } payload;
} MyMessage;

which_payload indicates which of the oneof fields is actually set. The user is expected to set the field manually using the correct field tag:

MyMessage msg = MyMessage_init_zero;
msg.payload.msg2.value = true;
msg.which_payload = MyMessage_msg2_tag;

Notice that neither which_payload field nor the unused fields in payload will consume any space in the resulting encoded message.

When a field inside oneof contains pb_callback_t fields, the callback values cannot be set before decoding. This is because the different fields share the same storage space in C union. Instead either function name bound callbacks or a separate message level callback can be used. See tests/oneof_callback for an example on this.

Extension fields

Protocol Buffers supports a concept of extension fields, which are additional fields to a message, but defined outside the actual message. The definition can even be in a completely separate .proto file.

The base message is declared as extensible by keyword extensions in the .proto file:

message MyMessage {
    .. fields ..
    extensions 100 to 199;
}

For each extensible message, nanopb_generator.py declares an additional callback field called extensions. The field and associated datatype pb_extension_t forms a linked list of handlers. When an unknown field is encountered, the decoder calls each handler in turn until either one of them handles the field, or the list is exhausted.

The actual extensions are declared using the extend keyword in the .proto, and are in the global namespace:

extend MyMessage {
    optional int32 myextension = 100;
}

For each extension, nanopb_generator.py creates a constant of type pb_extension_type_t. To link together the base message and the extension, you have to:

1. Allocate storage for your field, matching the datatype in the .proto. For example, for a int32 field, you need a int32_t variable to store the value.
2. Create a pb_extension_t constant, with pointers to your variable and to the generated pb_extension_type_t.
3. Set the message.extensions pointer to point to the pb_extension_t.

An example of this is available in tests/test_encode_extensions.c and tests/test_decode_extensions.c.

Default values

Protobuf has two syntax variants, proto2 and proto3. Of these proto2 has user definable default values that can be given in .proto file:

message MyMessage {
    optional bytes foo = 1 [default = "ABC\x01\x02\x03"];
    optional string bar = 2 [default = "åäö"];
}

Nanopb will generate both static and runtime initialization for the default values. In myproto.pb.h there will be a #define MyMessage_init_default {...} that can be used to initialize whole message into default values:

MyMessage msg = MyMessage_init_default;

In addition to this, pb_decode() will initialize message fields to defaults at runtime. If this is not desired, pb_decode_ex() can be used instead.

Message framing

Protocol Buffers does not specify a method of framing the messages for transmission. This is something that must be provided by the library user, as there is no one-size-fits-all solution. Typical needs for a framing format are to:

1. Encode the message length.
2. Encode the message type.
3. Perform any synchronization and error checking that may be needed depending on application.

For example UDP packets already fulfill all the requirements, and TCP streams typically only need a way to identify the message length and type. Lower level interfaces such as serial ports may need a more robust frame format, such as HDLC (high-level data link control).

Nanopb provides a few helpers to facilitate implementing framing formats:

1. Functions pb_encode_ex and pb_decode_ex prefix the message data with a varint-encoded length.
2. Union messages and oneofs are supported in order to implement top-level container messages.
3. Message IDs can be specified using the (nanopb_msgopt).msgid option and can then be accessed from the header.

Return values and error handling

Most functions in nanopb return bool: true means success, false means failure. There is also support for error messages for debugging purposes: the error messages go in stream->errmsg.

The error messages help in guessing what is the underlying cause of the error. The most common error conditions are:

1. Invalid protocol buffers binary message.
2. Mismatch between binary message and .proto message type.
3. Unterminated message (incorrect message length).
4. Exceeding the max_size or bytes_left of a stream.
5. Exceeding the max_size/max_count of a string or array field
6. IO errors in your own stream callbacks.
7. Errors that happen in your callback functions.
8. Running out of memory, i.e. stack overflow.
9. Invalid field descriptors (would usually mean a bug in the generator).

Static assertions

Nanopb code uses static assertions to check size of structures at the compile time. The PB_STATIC_ASSERT macro is defined in pb.h. If ISO C11 standard is available, the C standard _Static_assert keyword is used, otherwise a negative sized array definition trick is used.

Common reasons for static assertion errors are:

1. FIELDINFO_DOES_NOT_FIT_width2 with width1 or width2: Message that is larger than 256 bytes, but nanopb generator does not detect it for some reason. Often resolved by giving all .proto files as argument to nanopb_generator.py at the same time, to ensure submessage definitions are found. Alternatively (nanopb).descriptorsize = DS_4 option can be given manually.

2. FIELDINFO_DOES_NOT_FIT_width4 with width4: Message that is larger than 64 kilobytes. There will be a better error message for this in a future nanopb version, but currently it asserts here. The compile time option PB_FIELD_32BIT should be specified either on C compiler command line or by editing pb.h. This will increase the sizes of integer types used internally in nanopb code.

3. DOUBLE_MUST_BE_8_BYTES: Some platforms, most notably AVR, do not support the 64-bit double type, only 32-bit float. The compile time option PB_CONVERT_DOUBLE_FLOAT can be defined to convert between the types automatically. The conversion results in small rounding errors and takes unnecessary space in transmission, so changing the .proto to use float type is often better.

4. INT64_T_WRONG_SIZE: The stdint.h system header is incorrect for the C compiler being used. This can result from erroneous compiler include path. If the compiler actually does not support 64-bit types, the compile time option PB_WITHOUT_64BIT can be used.

5. variably modified array size: The compiler used has problems resolving the array-based static assert at compile time. Try setting the compiler to C11 standard mode if possible. If static assertions cannot be made to work on the compiler used, the compile-time option PB_NO_STATIC_ASSERT can be specified to turn them off.