The Linux Telephony Kernel API
A year ago, Internet Telephony was a curiosity, and many people thought it would never work for real phone calls. Now, with services like Net2Phone, Deltathree.com and DialPad providing free or extremely low-priced phone calls delivered via the Internet, Voice over IP (VoIP) has reached near-mainstream status. While Linux clients for those services are not yet available, Linux is not being left behind. With the 2.2.14 kernel, Linux has taken a bold lead in the area of computer telephony integration: we have the first modern operating system with a defined kernel-layer application programming interface (API) for telephony support. To sweeten the pot, excellent quality open-source telephony software is already using this API. You can call around the world using Linux and the Internet—and the call is free!
This article will explain the basics of how the telephony device drivers are integrated into the kernel in a way suitable for creating a common API across vendors. Then we'll discuss the basic ideas behind the design and function of the API, and how data and event information are dealt with separately. Finally, we'll discuss how telephony events (like ringing or picking up a handset) are handled in a process called “asynchronous event notification”.
There are many people who've asked why we needed a new telephony support API for the kernel. Even Alan Cox had to be convinced! After all, most of the Internet Telephony software out there can use a sound card, and if that card supports full-duplex (simultaneous playing and recording) and the user has a decent microphone-speaker headset, the quality of the call can be adequate. Why add complexity and a new API?
The answer is quite simple: sound cards are not telephone devices! Sound cards can't generate dial tone, rings, wink, flash-hook or caller-ID—all of which are needed to operate a normal phone device properly. Yes, sound cards are functional for limited telephony use with a headset and microphone, but real telephony cards let you plug that nice cordless phone in and be free from the short cord to your computer...or plug into a sophisticated business phone system. Sound cards also have no chance of providing the ability to plug into the phone line provided by your local telephone company, therefore, losing the ability to do inbound and outbound call control, hop-on and hop-off toll bypass applications or use any of the other new generation telephony software applications being written today.
Also, for any real VoIP use over the Internet, it is necessary to compress the audio data. To be compatible with other VoIP applications and equipment you must support the commonly accepted compression codecs like G.723.1 or G.729a. Unfortunately, if you want to do this in software you have to license the codecs, and this can cost in the six-figure range (easily!) these libraries are most certainly not open source. Telephony cards (like the Quicknet cards mentioned below) have pre-licensed these codecs and provide them built into the hardware. This avoids the mess of paying per-copy royalties for the codecs (it's part of the hardware cost) and lets you develop code with your own choice of license, not one imposed on you by the codec developer. In other words, you can do open-source VoIP software and still use the advanced codecs! These codecs are most definitely not something you would find on a sound card.
Additionally, sound cards can't interface with phones or the phone system, and they can't support hardware-based audio compression codecs. Sound cards are simply different from phone cards.
To be fair, telephony cards are not sound cards either. Sound cards have audio capabilities that far exceed what telephones can do. For example, sound cards are stereo; phones are mono. Sound cards can sample and playback sounds at music frequencies (20Hz-20kHz); phones are limited to voice frequencies (300Hz-4kHz typically). Sound cards have advanced music capabilites and often support MIDI and extensive sound effects. Phone cards cannot do any of that. The hardware and functionality are very different. The device drivers and the API need to be different, too.
But wait, you say, “what about all the great software out there that can work with sound cards, like voice recognition and text-to-speech processing?” Wouldn't it be nice to be able to use that software on phone devices with a minimum of effort? You can. The API for Linux telephony was designed to work in a way that does not preclude the use of software originally intended for a sound card with a telephony device. Yes, some minor code changes will be needed, but it's not all that hard. I'll explain more about that towards the end of the article.
The push for this new API was spearheaded by Quicknet Technologies, Inc., which manufactures a line of phone cards. In November of 1999, Ed Okerson of Quicknet and myself (I was an employee of Quicknet at the time) proposed the precursor to today's API to Alan Cox. After several weeks of intense e-mail and a pile of code by Ed and Alan, the Linux telephony API was born.
So, how does it work? Let's get started.
At the operating system level, all devices are referred to by numbers. We look in /dev and see a collection of file names, but down deep Linux looks at devices based on a major and a minor number. Devices of a particular type all share a single major number; individual devices of that type must each have their own minor number. For example, if you do ls -al /dev/ttyS0 you'll see:
[email protected]:~/lj > ls -al /dev/ttyS0 crwxrwxr-x 1 root uucp 4, 64 Oct 27 06:23 /dev/ttyS0
Note that that the file permissions mask has the first character as a “c”, indicating that it's a character device. It's owned by root and is in the uucp group. The next two numbers are not file sizes, like you'd see on a normal file; they are the major and minor number. In this case ttyS0 has a major number of 4 and a minor number of 64.
The Linux Telephony API assigns the major device number of 100 to phone-type devices. Your distribution of Linux has probably not created the devices for you, like it has for /dev/ttyS*, /dev/audio and other older, commonly accepted device mappings. If /dev/phone* devices are not present on your system, you'll need to make them before using the Linux Telephony API. You can fix this yourself quickly with the following command (as root):
mknod /dev/phone0 c 100 0
This creates a new device file called /dev/phone0. It's a character device with a major number of 100 and a minor number of 0. Refer to the mknod man page for more details on this command. You only actually need enough device files for your hardware, but most folks create devices 0 15 by default.
Note that there is presently an error in the file /usr/src/linux/Documentation/devices.txt. This file, which provides official assignments for all major numbers it currently states that Linux telephony is to use major number 159 and that 100 is deprecated. This is an acknowledged error and will be fixed in future documentation. The correct major number for Linux telephony (and the number used in the kernel) is 100.
Alan Cox developed the phonedev module based on a similar approach that he took for the Video4Linux Project. There will be many vendors who create products capable of being a phone device. Rather than having multiple telephony product vendors, each requiring their own major number, they can all use major number 100 and the commonly defined /dev/phoneN (where N is the device number).
All of these must present a common basic interface to user-space software, that is, they must all follow the same common API.al, though they may provide extensions particular to their product. Vendors will create their own device driver module that will implement this common API as an external interface by dealing with the inner details of their particular hardware.
The phonedev module solved the problem of mapping the minor device number to a specific vendor-type module at runtime. The source code is in the files/usr/src/linux/drivers/telephony/phonedev.c and the header files, phonedev.h and telephony.h, are located in /usr/include/linux. Here's how it works.
Listing 1. phone_device Structure
Every phone device must use a phone_device structure (see Listing 1). Likewise, every phone device must call two functions to interact with the phonedev module in order to register and unregister itself. These are defined as:
extern int phone_register_device (struct phone_device *, int unit); extern void phone_unregister_device (struct phone_device *);
At load time, the phonedev module sets itself up and waits to service other telephony devices. When a telephony driver is loaded (by modprobe or insmod, as discussed later) it calls the phone_register_device function. A simple explanation of this function is that it keeps a pointer within the phonedev module to the phone_device structure, searches for the first open minor number and assigns that to the phone device, and then increments a counter to track the fact that something is using the phonedev module (to prevent it being unloaded while in use).
The practical implications of this are simple: the first telephony modules loaded will be assigned the first available (lowest numbered) minor numbers. This is crucial to understand in cases where modules from different vendors need to coexist on the same system, and specific assignment of a particular card is desired to match a particular minor number (a specific /dev/phoneN device). In other words, if you have a device from XYZ company and a device from ABC company, and you want ABC's card to be /dev/phone0, you'll have to make sure that the driver for ABC gets loaded first.
All devices must provide at least the basic functions to interact with the device: open, read, write, close, etc. These are all part of the “file operations structure” (see linux/fs.h for details). Each device defines the functions appropriate for itself.
Anytime a program opens a /dev/phoneN device, it's actually calling the fopen function defined in the phonedev module's file operations structure. This function performs the following operations:
It grabs the minor device number from /dev/phoneN inode.
It builds a string of the form char-major-%d-%d using the major and minor numbers. In the case of minor number 0 (corresponding to /dev/phone0), this string would be char-major-100-0.
It uses this string in a call to request_module to ask that the module be loaded. This has the same effect as if the program modprobe was called (in fact, it actually starts a separate kernel thread and executes modprobe in it). This ensures that, if the device is a module instead of a part of the kernel, kmod has a chance to load the module before phonedev tries to use it.
It then calls the fopen function in the phone device module to perform the actual act of opening the individual device.
As you can see, the phonedev module has two basic purposes: to assign minor numbers to phone devices dynamically at load time and provide a clean way to autoload needed phone device modules at run time. It's clear, though, that a good understanding of modprobe and module dependency is needed to fully understand the telephony modules and their interactions.
If you are using kmod to automatically load kernel modules as you need them, then you will need to make sure that several aliases are defined in the /etc/modules.conf file.
First of all, the system needs to know that char-major-100 is the phonedev module. Add this line to define:
alias char-major-100 phonedev
Now the system will need to know what actual phone device modules to associate with specific minor numbers. As we learned above, this may not guarantee that the phonedev module actually assigned that minor number to the phone device at module load time, but we'll cover that in more detail below. In the following examples we'll use the telephony cards from Quicknet Technologies Inc. (www.quicknet.net). These cards have Linux drivers in the kernel and are relatively inexpensive compared to most telephony cards. Quicknet's device driver is ixj.o and the module name is ixj. This driver is used for all of their telephony card products (it's smart enough to handle the ISA, PCI or PCMCIA flavors as well as knowing which cards have which kinds of phone interface circuits). To define that Quicknet's driver is associated with /dev/phone0, add this line to /etc/modules.conf:
alias char-major-100-0 ixjAs you'll recall from the discussion of phonedev's fopen function above, phonedev will build a string of the form char-major-%d-%d and fill in the parameters with 100 (major number) and the requested minor number. In our example, attempting to open /dev/phone0 would result in phonedev attempting to load char-major-100-0. That device is unknown to the kernel module loader. The alias command above maps that string to the module name ixj. When we try to open /dev/phone0 the phondev module will autoload the ixj module for us, and then call the fopen function defined in the ixj module to open the device (assuming your kernel was built to support the kernel module loader with CONFIG_KMOD=y).
A fundamental premise of the Linux telephony API is that the only thing read-from and written-to phone devices using read and write is actual audio data. This is separate from—and handled quite differently from—event type information.
What, specifically, is audio data, and how is it different from event data? Audio data is the result of the analog-to-digital (A/D) conversion (and possibly data compression) on the audio signal present at the microphone of the telephony device. The offhook signal resulting from picking up the phone handset is an event. The tone that results from the user pressing a digit on the phone handset is an event even though that action may also generate audio. An incoming ringing signal is called an event. In short, all non-microphone inputs are events. All microphone input (and the corresponding speaker output) is audio data. This audio data is read and written from the phone device using the standard read and write system calls.
Most phone devices will provide audio data compression in the device. In fact, for successful Internet telephony applications, some form of audio data compression is required. These compression techniques are called “audio codecs” (or codecs for short), and there are a set of commonly implemented and interoperable codecs. The Linux telephony API includes defined constants for most of these common codecs; however, a particular phone device may or may not support them all. Control of what codec is in use is handled by an ioctl system call.
One large difference in how Linux telephony API differs from what used to work with sound cards is that Linux telephony API is “frame-oriented” while sound cards are “byte oriented”. A device that is frame-oriented reads a discrete frame of data corresponding to a unit of time. This is done because all audio codecs operate on a period of audio data at a time (usually 10, 20 or 30 milliseconds' worth of data). Because use of a compressed codec is the norm for network telephony applications, this is the normal and expected mode of operation for phone devices. Sound cards do not have this restriction and are free to read and write a variable number of data bytes on any give call. The API defines “frame sizes” for each codec, and raw uncompressed audio data from the device is one of the codec choices. For example, the LINEAR16 codec (uncompressed 16-bit sound samples) has a default frame size of 240 bytes—corresponding to 30 milliseconds of data at the default 8000Hz sample rate. Every read operation from the device will result in 240 bytes of data or nothing. Of course, you can change this behavior with ioctl calls to adjust the size of a frame for uncompressed codecs.
Commands for the phone device to perform certain actions are not written to the phone device using a write call—only audio data is written to the device. To control the phone device, a set of ioctl functions are defined to handle the basic phone activities. This basic set of ioctl functions are defined in /usr/include/linux/telephony.h. Vendors who wish to extend that basic set of capabilities may do so, but those functions are limited to their own device driver and are outside the scope of the common Linux telephony API.
An example will best illustrate this potential. Using the telephony API with a Quicknet Internet PhoneJACK card and a phone plugged into the card (called an FXS port or POTS port by telephony folks), Listing 2 shows a brief program to ring the phone.
You'll notice that the device is opened, defaulting to /dev/phone0 if the user does not provide a specific device name on the command line. The maximum number of rings is set with an ioctl call using the PHONE_MAXRINGS constant. The phone is then instructed to ring using the PHONE_RING ioctl. This example program is a simplified version of the LGPL module ring.c found in Quicknet's Software Developer Kit (SDK). It's too simple to do anything more than illustrate the technique and demonstrate the use of ioctls to control a phone device, but real-world programs need not be much more complicated, the API is fairly simple. All the Linux telephony API ioctl constants are defined in the header file /etc/include/linux/telephony.h.
It's this basic fact that allows software written for sound cards to be adapted easily for use with phone devices; like sound cards, the only data written using the read and write system calls is audio data. In addition, the low-level constants used in the defined telephony ioctl calls were designed to avoid conflict with existing sound card ioctls. Porting an application that expects a sound card to use a phone card instead will primarily involve handling the errors returned from the sound card ioctl calls your code makes. It should be possible (though perhaps not easy) to write a wrapper that opens the telephony device and spawns a child process (inheriting the open file descriptor to the phone device) that runs software expecting a sound card interface. While it's not totally transparent, it's possible and should not actually be that difficult.
Events that occur on the telephony side of the device need to be communicated to the user-space software running the phone. Old and crude methods for this required the software to poll the device continuously for status and changes. The Linux telephony API avoids that, of course, and provides two different techniques, both generically called “asynchronous event notification”. The first method uses signals for indication, and the second uses the “exception bit” in the file descriptor set for exceptions. I'll cover both techniques in order.
The use of signals for event notification requires three steps: first, prepare and declare a signal handler function for the SIGIO signal; second, set the process ID (PID) for the running process to receive the signal; third, enable the generation of the signal on the open files descriptor using the fcntl system call. An excellent description of these steps can be found in Chapter 12 of Advanced Programming in the UNIX Environment by W. Richard Stevens (a worthwhile book to have in any case). Again, a short example will probably clear this up. Assuming you have an open file descriptor ixj1 that is an open phone device, you can enable asynchronous event notification with signals with the following code snip:
signal(SIGIO, &getdata); fcntl(ixj1, F_SETOWN, getpid()); oflags1 = fcntl(ixj1, F_GETFL); fcntl(ixj1, F_SETFL, oflags1 | FASYNC);
and a related signal handling function (getdata in the code above) to process the data. When you get the signal, you still do not know what kind of event occurred—only that an event did occur. Your program would then have to make an ioctl call to the phone device to ask what kind of event was detected (more on this below). In addition, if your program has more than one open phone device file descriptor, you will not know which one generated the signal. And, signals can be complicated to deal with and can be unreliable in a multi-threaded program and, so, are avoided by some developers. These factors limit the effectiveness of this method, leading to the more useful case of using the “exception bit” in the file descriptor set for exceptions.
It's common for programs to set read and write exception sets and then use the select( ) system call to wait for a file descriptor to be readable or writable. A less well-known aspect of the select( ) call is the “exception set”. The Linux telephony API uses this exception set to signal a process that a telephony event has occurred. Listing 3 presents a simple example using the file descriptor ixj1.
Listing 3. Using an Exception Set
This extremely simple (and not really useful) example is purely to illustrate the technique of using the exception set and select( ) to detect events. The phone device driver will set the appropriate bit in the exception set if an event has occurred, causing select( ) to return. The user can screen the read, write and exception descriptor sets to determine if its own file descriptor was marked by the device driver as ready for that kind of operation. If data is ready to be read, the statement FD_ISSET(ixj1,&rfds) will return TRUE; FD_ISSET(ixj1,&wfds) statement returns TRUE if the device is ready to be written to. And, if a telephony event has occurred, the FD_ISSET(ixj1,&efds) will return TRUE. So, how do you detect what specific event occurred?
The API provides a special PHONE_EXCEPTION ioctl call and an associated telephony_exception structure to decode the return value. This call will set bits in the structure that indicate which of the telephony events occurred (there may have been several). In the above example “hookstate” bit is examined in the statement if(ixje.bits.hookstate), and if that bit is set it indicates a change of status. An ioctl call is then made to determine if the phone is on or off hook. Real-world code would have used a large select or an extended if-else-if ladder to examine the contents of ixje.bits after the PHONE_EXCEPTION ioctl call. A detailed explanation of how to use this technique is beyond the scope of this article, but please refer to the /usr/include/linux/telephony.h file for details of what kinds of events can be detected.
There are many open-source programs available now that use this API. However, the most well-known and widely used program is ohphone, a console application using the open-source OpenH323 library. Ohphone is part of the OpenH323 Project (www.openh323.org) and is in daily use by thousands of people to make free, high-quality phone calls over the Internet. Ohphone not only fully supports the Linux telephony API, but it is also compatible with other H.323-based products like Microsoft NetMeeting<+H>tm<+H> and Cisco voice-enabled routers. A more detailed discussion of this fine software is too much for this article, but you're encouraged to look at its web site to catch the latest news. The company that developed the OpenH323 library was recently acquired by Quicknet Technologies, Inc. as part of Quicknet's efforts to ensure continued major development effort of this open-source project. With such full commercial backing and a commitment to open source, I expect the OpenH323 Project software to become even better in the near future.
The Linux telephony API provides a common and consistent interface for developing telephony software on Linux. While there is currently only one vendor (Quicknet Technologies, Inc.) with fully compliant drivers for this API, several others are working toward compliant drivers. The API is lean, well designed, will not conflict with the existing API for sound cards and provides the ability to support multiple vendors behind the same interface. There's sure to be some exciting new telephony software developed for Linux in the coming year.
Greg Herlein has been an avid Linux developer since 1994. His company, Herlein Engineering, currently offers Linux/UNIX consulting, especially in the areas of telephony software development. He lives and works in San Francisco, California. He can be reached at [email protected]