Pipe/Filter Message Processing

Note

The system described below provides a message processing system with a straightforward API. However it makes many extra memory copies and allocations than would otherwise be required, and also tends to make applications using it somewhat opaque because it is not obvious what this or that Pipe& object actually does (type of operation, number of messages output (if any!), and so on), whereas using say a HashFunction or AEAD_Mode provides a much better idea in the code of what operation is occuring.

This filter interface is no longer used within the library itself (outside a few dusty corners) and will likely not see any further major development. However it will remain included because the API is often convenient and many applications use it.

Many common uses of cryptography involve processing one or more streams of data. Botan provides services that make setting up data flows through various operations, such as compression, encryption, and base64 encoding. Each of these operations is implemented in what are called filters in Botan. A set of filters are created and placed into a pipe, and information “flows” through the pipe until it reaches the end, where the output is collected for retrieval. If you’re familiar with the Unix shell environment, this design will sound quite familiar.

Here is an example that uses a pipe to base64 encode some strings:

Pipe pipe(new Base64_Encoder); // pipe owns the pointer
pipe.start_msg();
pipe.write("message 1");
pipe.end_msg(); // flushes buffers, increments message number

// process_msg(x) is start_msg() && write(x) && end_msg()
pipe.process_msg("message2");

std::string m1 = pipe.read_all_as_string(0); // "message1"
std::string m2 = pipe.read_all_as_string(1); // "message2"

Bytestreams in the pipe are grouped into messages; blocks of data that are processed in an identical fashion (ie, with the same sequence of filter operations). Messages are delimited by calls to start_msg and end_msg. Each message in a pipe has its own identifier, which currently is an integer that increments up from zero.

The Base64_Encoder was allocated using new; but where was it deallocated? When a filter object is passed to a Pipe, the pipe takes ownership of the object, and will deallocate it when it is no longer needed.

There are two different ways to make use of messages. One is to send several messages through a Pipe without changing the Pipe configuration, so you end up with a sequence of messages; one use of this would be to send a sequence of identically encrypted UDP packets, for example (note that the data need not be identical; it is just that each is encrypted, encoded, signed, etc in an identical fashion). Another is to change the filters that are used in the Pipe between each message, by adding or removing filters; functions that let you do this are documented in the Pipe API section.

Botan has about 40 filters that perform different operations on data. Here’s code that uses one of them to encrypt a string with AES:

AutoSeeded_RNG rng,
SymmetricKey key(rng, 16); // a random 128-bit key
InitializationVector iv(rng, 16); // a random 128-bit IV

// The algorithm we want is specified by a string
Pipe pipe(get_cipher("AES-128/CBC", key, iv, ENCRYPTION));

pipe.process_msg("secrets");
pipe.process_msg("more secrets");

secure_vector<uint8_t> c1 = pipe.read_all(0);

uint8_t c2[4096] = { 0 };
size_t got_out = pipe.read(c2, sizeof(c2), 1);
// use c2[0...got_out]

Note the use of AutoSeeded_RNG, which is a random number generator. If you want to, you can explicitly set up the random number generators and entropy sources you want to, however for 99% of cases AutoSeeded_RNG is preferable.

Pipe also has convenience methods for dealing with std::iostream. Here is an example of this, using the bzip2 compression filter:

std::ifstream in("data.bin", std::ios::binary)
std::ofstream out("data.bin.bz2", std::ios::binary)

Pipe pipe(new Compression_Filter("bzip2", 9));

pipe.start_msg();
in >> pipe;
pipe.end_msg();
out << pipe;

However there is a hitch to the code above; the complete contents of the compressed data will be held in memory until the entire message has been compressed, at which time the statement out << pipe is executed, and the data is freed as it is read from the pipe and written to the file. But if the file is very large, we might not have enough physical memory (or even enough virtual memory!) for that to be practical. So instead of storing the compressed data in the pipe for reading it out later, we divert it directly to the file:

std::ifstream in("data.bin", std::ios::binary)
std::ofstream out("data.bin.bz2", std::ios::binary)

Pipe pipe(new Compression_Filter("bzip2", 9), new DataSink_Stream(out));

pipe.start_msg();
in >> pipe;
pipe.end_msg();

This is the first code we’ve seen so far that uses more than one filter in a pipe. The output of the compressor is sent to the DataSink_Stream. Anything written to a DataSink_Stream is written to a file; the filter produces no output. As soon as the compression algorithm finishes up a block of data, it will send it along to the sink filter, which will immediately write it to the stream; if you were to call pipe.read_all() after pipe.end_msg(), you’d get an empty vector out. This is particularly useful for cases where you are processing a large amount of data, as it means you don’t have to store everything in memory at once.

Here’s an example using two computational filters:

AutoSeeded_RNG rng,
SymmetricKey key(rng, 32);
InitializationVector iv(rng, 16);

Pipe encryptor(get_cipher("AES/CBC/PKCS7", key, iv, ENCRYPTION),
               new Base64_Encoder);

encryptor.start_msg();
file >> encryptor;
encryptor.end_msg(); // flush buffers, complete computations
std::cout << encryptor;

You can read from a pipe while you are still writing to it, which allows you to bound the amount of memory that is in use at any one time. A common idiom for this is:

pipe.start_msg();
std::vector<uint8_t> buffer(4096); // arbitrary size
while(infile.good())
   {
   infile.read((char*)&buffer[0], buffer.size());
   const size_t got_from_infile = infile.gcount();
   pipe.write(buffer, got_from_infile);

   if(infile.eof())
      pipe.end_msg();

   while(pipe.remaining() > 0)
      {
      const size_t buffered = pipe.read(buffer, buffer.size());
      outfile.write((const char*)&buffer[0], buffered);
      }
   }
if(infile.bad() || (infile.fail() && !infile.eof()))
   throw Some_Exception();

Fork

It is common that you might receive some data and want to perform more than one operation on it (ie, encrypt it with Serpent and calculate the SHA-256 hash of the plaintext at the same time). That’s where Fork comes in. Fork is a filter that takes input and passes it on to one or more filters that are attached to it. Fork changes the nature of the pipe system completely: instead of being a linked list, it becomes a tree or acyclic graph.

Each filter in the fork is given its own output buffer, and thus its own message. For example, if you had previously written two messages into a pipe, then you start a new one with a fork that has three paths of filter’s inside it, you add three new messages to the pipe. The data you put into the pipe is duplicated and sent into each set of filter and the eventual output is placed into a dedicated message slot in the pipe.

Messages in the pipe are allocated in a depth-first manner. This is only interesting if you are using more than one fork in a single pipe. As an example, consider the following:

Pipe pipe(new Fork(
             new Fork(
                new Base64_Encoder,
                new Fork(
                   NULL,
                   new Base64_Encoder
                   )
                ),
             new Hex_Encoder
             )
   );

In this case, message 0 will be the output of the first Base64_Encoder, message 1 will be a copy of the input (see below for how fork interprets NULL pointers), message 2 will be the output of the second Base64_Encoder, and message 3 will be the output of the Hex_Encoder. This results in message numbers being allocated in a top to bottom fashion, when looked at on the screen. However, note that there could be potential for bugs if this is not anticipated. For example, if your code is passed a filter, and you assume it is a “normal” one that only uses one message, your message offsets would be wrong, leading to some confusion during output.

If Fork’s first argument is a null pointer, but a later argument is not, then Fork will feed a copy of its input directly through. Here’s a case where that is useful:

// have std::string ciphertext, auth_code, key, iv, mac_key;

Pipe pipe(new Base64_Decoder,
          get_cipher("AES-128", key, iv, DECRYPTION),
          new Fork(
             0, // this message gets plaintext
             new MAC_Filter("HMAC(SHA-1)", mac_key)
          )
   );

pipe.process_msg(ciphertext);
std::string plaintext = pipe.read_all_as_string(0);
secure_vector<uint8_t> mac = pipe.read_all(1);

if(mac != auth_code)
   error();

Here we wanted to not only decrypt the message, but send the decrypted text through an additional computation, in order to compute the authentication code.

Any filters that are attached to the pipe after the fork are implicitly attached onto the first branch created by the fork. For example, let’s say you created this pipe:

Pipe pipe(new Fork(new Hash_Filter("SHA-256"),
                   new Hash_Filter("SHA-512")),
          new Hex_Encoder);

And then called start_msg, inserted some data, then end_msg. Then pipe would contain two messages. The first one (message number 0) would contain the SHA-256 sum of the input in hex encoded form, and the other would contain the SHA-512 sum of the input in raw binary. In many situations you’ll want to perform a sequence of operations on multiple branches of the fork; in which case, use the filter described in Chain.

Chain

A Chain filter creates a chain of filters and encapsulates them inside a single filter (itself). This allows a sequence of filters to become a single filter, to be passed into or out of a function, or to a Fork constructor.

You can call Chain’s constructor with up to four Filter pointers (they will be added in order), or with an array of filter pointers and a size_t that tells Chain how many filters are in the array (again, they will be attached in order). Here’s the example from the last section, using chain instead of relying on the implicit passthrough the other version used:

Pipe pipe(new Fork(
              new Chain(new Hash_Filter("SHA-256"), new Hex_Encoder),
              new Hash_Filter("SHA-512")
              )
         );

Sources and Sinks

Data Sources

A DataSource is a simple abstraction for a thing that stores bytes. This type is used heavily in the areas of the API related to ASN.1 encoding/decoding. The following types are DataSource: Pipe, SecureQueue, and a couple of special purpose ones: DataSource_Memory and DataSource_Stream.

You can create a DataSource_Memory with an array of bytes and a length field. The object will make a copy of the data, so you don’t have to worry about keeping that memory allocated. This is mostly for internal use, but if it comes in handy, feel free to use it.

A DataSource_Stream is probably more useful than the memory based one. Its constructors take either a std::istream or a std::string. If it’s a stream, the data source will use the istream to satisfy read requests (this is particularly useful to use with std::cin). If the string version is used, it will attempt to open up a file with that name and read from it.

Data Sinks

A DataSink (in data_snk.h) is a Filter that takes arbitrary amounts of input, and produces no output. This means it’s doing something with the data outside the realm of what Filter/Pipe can handle, for example, writing it to a file (which is what the DataSink_Stream does). There is no need for DataSink``s that write to a ``std::string or memory buffer, because Pipe can handle that by itself.

Here’s a quick example of using a DataSink, which encrypts in.txt and sends the output to out.txt. There is no explicit output operation; the writing of out.txt is implicit:

DataSource_Stream in("in.txt");
Pipe pipe(get_cipher("AES-128/CTR-BE", key, iv),
          new DataSink_Stream("out.txt"));
pipe.process_msg(in);

A real advantage of this is that even if “in.txt” is large, only as much memory is needed for internal I/O buffers will be used.

The Pipe API

Initializing Pipe

By default, Pipe will do nothing at all; any input placed into the Pipe will be read back unchanged. Obviously, this has limited utility, and presumably you want to use one or more filters to somehow process the data. First, you can choose a set of filters to initialize the Pipe via the constructor. You can pass it either a set of up to four filter pointers, or a pre-defined array and a length:

Pipe pipe1(new Filter1(/*args*/), new Filter2(/*args*/),
           new Filter3(/*args*/), new Filter4(/*args*/));
Pipe pipe2(new Filter1(/*args*/), new Filter2(/*args*/));

Filter* filters[5] = {
  new Filter1(/*args*/), new Filter2(/*args*/), new Filter3(/*args*/),
  new Filter4(/*args*/), new Filter5(/*args*/) /* more if desired... */
};
Pipe pipe3(filters, 5);

This is by far the most common way to initialize a Pipe. However, occasionally a more flexible initialization strategy is necessary; this is supported by 4 member functions. These functions may only be used while the pipe in question is not in use; that is, either before calling start_msg, or after end_msg has been called (and no new calls to start_msg have been made yet).

void Pipe::prepend(Filter *filter)

Calling prepend will put the passed filter first in the list of transformations. For example, if you prepend a filter implementing encryption, and the pipe already had a filter that hex encoded the input, then the next message processed would be first encrypted, and then hex encoded.

void Pipe::append(Filter *filter)

Like prepend, but places the filter at the end of the message flow. This doesn’t always do what you expect if there is a fork.

void Pipe::pop()

Removes the first filter in the flow.

void Pipe::reset()

Removes all the filters that the pipe currently holds - it is reset to an empty/no-op state. Any data that is being retained by the pipe is retained after a reset, and reset does not affect message numbers (discussed later).

Giving Data to a Pipe

Input to a Pipe is delimited into messages, which can be read from independently (ie, you can read 5 bytes from one message, and then all of another message, without either read affecting any other messages).

void Pipe::start_msg()

Starts a new message; if a message was already running, an exception is thrown. After this function returns, you can call write.

void Pipe::write(const uint8_t *input, size_t length)
void Pipe::write(const std::vector<uint8_t> &input)
void Pipe::write(const std::string &input)
void Pipe::write(DataSource &input)
void Pipe::write(uint8_t input)

All versions of write write the input into the filter sequence. If a message is not currently active, an exception is thrown.

void Pipe::end_msg()

End the currently active message

Sometimes, you may want to do only a single write per message. In this case, you can use the process_msg series of functions, which start a message, write their argument into the pipe, and then end the message. In this case you would not make any explicit calls to start_msg/end_msg.

Pipes can also be used with the >> operator, and will accept a std::istream, or on Unix systems with the fd_unix module, a Unix file descriptor. In either case, the entire contents of the file will be read into the pipe.

Getting Output from a Pipe

Retrieving the processed data from a pipe is a bit more complicated, for various reasons. The pipe will separate each message into a separate buffer, and you have to retrieve data from each message independently. Each of the reader functions has a final parameter that specifies what message to read from. If this parameter is set to Pipe::DEFAULT_MESSAGE, it will read the current default message (DEFAULT_MESSAGE is also the default value of this parameter).

Functions in Pipe related to reading include:

size_t Pipe::read(uint8_t *out, size_t len)

Reads up to len bytes into out, and returns the number of bytes actually read.

size_t Pipe::peek(uint8_t *out, size_t len)

Acts exactly like read, except the data is not actually read; the next read will return the same data.

secure_vector<uint8_t> Pipe::read_all()

Reads the entire message into a buffer and returns it

std::string Pipe::read_all_as_string()

Like read_all, but it returns the data as a std::string. No encoding is done; if the message contains raw binary, so will the string.

size_t Pipe::remaining()

Returns how many bytes are left in the message

Pipe::message_id Pipe::default_msg()

Returns the current default message number

Pipe::message_id Pipe::message_count()

Returns the total number of messages currently in the pipe

Pipe::set_default_msg(Pipe::message_id msgno)

Sets the default message number (which must be a valid message number for that pipe). The ability to set the default message number is particularly important in the case of using the file output operations (<< with a std::ostream or Unix file descriptor), because there is no way to specify the message explicitly when using the output operator.

Pipe I/O for Unix File Descriptors

This is a minor feature, but it comes in handy sometimes. In all installations of the library, Botan’s Pipe object overloads the << and >> operators for C++ iostream objects, which is usually more than sufficient for doing I/O.

However, there are cases where the iostream hierarchy does not map well to local ‘file types’, so there is also the ability to do I/O directly with Unix file descriptors. This is most useful when you want to read from or write to something like a TCP or Unix-domain socket, or a pipe, since for simple file access it’s usually easier to just use C++’s file streams.

If BOTAN_EXT_PIPE_UNIXFD_IO is defined, then you can use the overloaded I/O operators with Unix file descriptors. For an example of this, check out the hash_fd example, included in the Botan distribution.

Filter Catalog

This section documents most of the useful filters included in the library.

Keyed Filters

A few sections ago, it was mentioned that Pipe can process multiple messages, treating each of them the same. Well, that was a bit of a lie. There are some algorithms (in particular, block ciphers not in ECB mode, and all stream ciphers) that change their state as data is put through them.

Naturally, you might well want to reset the keys or (in the case of block cipher modes) IVs used by such filters, so multiple messages can be processed using completely different keys, or new IVs, or new keys and IVs, or whatever. And in fact, even for a MAC or an ECB block cipher, you might well want to change the key used from message to message.

Enter Keyed_Filter, which acts as an abstract interface for any filter that is uses keys: block cipher modes, stream ciphers, MACs, and so on. It has two functions, set_key and set_iv. Calling set_key will set (or reset) the key used by the algorithm. Setting the IV only makes sense in certain algorithms – a call to set_iv on an object that doesn’t support IVs will cause an exception. You must call set_key before calling set_iv.

Here’s a example:

Keyed_Filter *aes, *hmac;
Pipe pipe(new Base64_Decoder,
          // Note the assignments to the cast and hmac variables
          aes = get_cipher("AES-128/CBC", aes_key, iv),
          new Fork(
             0, // Read the section 'Fork' to understand this
             new Chain(
                hmac = new MAC_Filter("HMAC(SHA-1)", mac_key, 12),
                new Base64_Encoder
                )
             )
   );
pipe.start_msg();
// use pipe for a while, decrypt some stuff, derive new keys and IVs
pipe.end_msg();

aes->set_key(aes_key2);
aes->set_iv(iv2);
hmac->set_key(mac_key2);

pipe.start_msg();
// use pipe for some other things
pipe.end_msg();

There are some requirements to using Keyed_Filter that you must follow. If you call set_key or set_iv on a filter that is owned by a Pipe, you must do so while the Pipe is “unlocked”. This refers to the times when no messages are being processed by Pipe – either before Pipe’s start_msg is called, or after end_msg is called (and no new call to start_msg has happened yet). Doing otherwise will result in undefined behavior, probably silently getting invalid output.

And remember: if you’re resetting both values, reset the key first.

Cipher Filters

Getting a hold of a Filter implementing a cipher is very easy. Make sure you’re including the header lookup.h, and then call get_cipher. You will pass the return value directly into a Pipe. There are a couple different functions which do varying levels of initialization:

Keyed_Filter *get_cipher(std::string cipher_spec, SymmetricKey key, InitializationVector iv, Cipher_Dir dir)
Keyed_Filter *get_cipher(std::string cipher_spec, SymmetricKey key, Cipher_Dir dir)

The version that doesn’t take an IV is useful for things that don’t use them, like block ciphers in ECB mode, or most stream ciphers. If you specify a cipher spec that does want a IV, and you use the version that doesn’t take one, an exception will be thrown. The dir argument can be either ENCRYPTION or DECRYPTION.

The cipher_spec is a string that specifies what cipher is to be used. The general syntax for “cipher_spec” is “STREAM_CIPHER”, “BLOCK_CIPHER/MODE”, or “BLOCK_CIPHER/MODE/PADDING”. In the case of stream ciphers, no mode is necessary, so just the name is sufficient. A block cipher requires a mode of some sort, which can be “ECB”, “CBC”, “CFB(n)”, “OFB”, “CTR-BE”, or “EAX(n)”. The argument to CFB mode is how many bits of feedback should be used. If you just use “CFB” with no argument, it will default to using a feedback equal to the block size of the cipher. EAX mode also takes an optional bit argument, which tells EAX how large a tag size to use~–~generally this is the size of the block size of the cipher, which is the default if you don’t specify any argument.

In the case of the ECB and CBC modes, a padding method can also be specified. If it is not supplied, ECB defaults to not padding, and CBC defaults to using PKCS #5/#7 compatible padding. The padding methods currently available are “NoPadding”, “PKCS7”, “OneAndZeros”, and “CTS”. CTS padding is currently only available for CBC mode, but the others can also be used in ECB mode.

Some example “cipher_spec arguments are: “AES-128/CBC”, “Blowfish/CTR-BE”, “Serpent/XTS”, and “AES-256/EAX”.

“CTR-BE” refers to counter mode where the counter is incremented as if it were a big-endian encoded integer. This is compatible with most other implementations, but it is possible some will use the incompatible little endian convention. This version would be denoted as “CTR-LE” if it were supported.

“EAX” is a new cipher mode designed by Wagner, Rogaway, and Bellare. It is an authenticated cipher mode (that is, no separate authentication is needed), has provable security, and is free from patent entanglements. It runs about half as fast as most of the other cipher modes (like CBC, OFB, or CTR), which is not bad considering you don’t need to use an authentication code.

Hashes and MACs

Hash functions and MACs don’t need anything special when it comes to filters. Both just take their input and produce no output until end_msg is called, at which time they complete the hash or MAC and send that as output.

These filters take a string naming the type to be used. If for some reason you name something that doesn’t exist, an exception will be thrown.

Hash_Filter::Hash_Filter(std::string hash, size_t outlen = 0)

This constructor creates a filter that hashes its input with hash. When end_msg is called on the owning pipe, the hash is completed and the digest is sent on to the next filter in the pipeline. The parameter outlen specifies how many bytes of the hash output will be passed along to the next filter when end_msg is called. By default, it will pass the entire hash.

Examples of names for Hash_Filter are “SHA-1” and “Whirlpool”.

MAC_Filter::MAC_Filter(std::string mac, SymmetricKey key, size_t outlen = 0)

This constructor takes a name for a mac, such as “HMAC(SHA-1)” or “CMAC(AES-128)”, along with a key to use. The optional outlen works the same as in Hash_Filter.

Encoders

Often you want your data to be in some form of text (for sending over channels that aren’t 8-bit clean, printing it, etc). The filters Hex_Encoder and Base64_Encoder will convert arbitrary binary data into hex or base64 formats. Not surprisingly, you can use Hex_Decoder and Base64_Decoder to convert it back into its original form.

Both of the encoders can take a few options about how the data should be formatted (all of which have defaults). The first is a bool which says if the encoder should insert line breaks. This defaults to false. Line breaks don’t matter either way to the decoder, but it makes the output a bit more appealing to the human eye, and a few transport mechanisms (notably some email systems) limit the maximum line length.

The second encoder option is an integer specifying how long such lines will be (obviously this will be ignored if line-breaking isn’t being used). The default tends to be in the range of 60-80 characters, but is not specified. If you want a specific value, set it. Otherwise the default should be fine.

Lastly, Hex_Encoder takes an argument of type Case, which can be Uppercase or Lowercase (default is Uppercase). This specifies what case the characters A-F should be output as. The base64 encoder has no such option, because it uses both upper and lower case letters for its output.

You can find the declarations for these types in hex_filt.h and b64_filt.h.

Writing New Filters

The system of filters and pipes was designed in an attempt to make it as simple as possible to write new filter types. There are four functions that need to be implemented by a class deriving from Filter:

void Filter::write(const uint8_t *input, size_t length)

This function is what is called when a filter receives input for it to process. The filter is not required to process the data right away; many filters buffer their input before producing any output. A filter will usually have write called many times during its lifetime.

void Filter::send(uint8_t *output, size_t length)

Eventually, a filter will want to produce some output to send along to the next filter in the pipeline. It does so by calling send with whatever it wants to send along to the next filter. There is also a version of send taking a single byte argument, as a convenience.

void Filter::start_msg()

Implementing this function is optional. Implement it if your filter would like to do some processing or setup at the start of each message, such as allocating a data structure.

void Filter::end_msg()

Implementing this function is optional. It is called when it has been requested that filters finish up their computations. The filter should finish up with whatever computation it is working on (for example, a compressing filter would flush the compressor and send the final block), and empty any buffers in preparation for processing a fresh new set of input.

Additionally, if necessary, filters can define a constructor that takes any needed arguments, and a destructor to deal with deallocating memory, closing files, etc.