3 How to implement an alternative carrier for the Erlang distribution
This document describes how one can implement ones own carrier protocol for the Erlang distribution. The distribution is normally carried by the TCP/IP protocol. What's explained here is the method for replacing TCP/IP with another protocol.
The document is a step by step explanation of the uds_dist example application (seated in the kernel applications examples directory). The uds_dist application implements distribution over Unix domain sockets and is written for the Sun Solaris 2 operating environment. The mechanisms are however general and applies to any operating system Erlang runs on. The reason the C code is not made portable, is simply readability.
This document was written a long time ago. Most of it is still valid, but some things have changed since it was first written. Most notably the driver interface. There have been some updates to the documentation of the driver presented in this documentation, but more could be done and are planned for the future. The reader is encouraged to also read the erl_driver, and the driver_entry documentation.
3.1 Introduction
To implement a new carrier for the Erlang distribution, one must first make the protocol available to the Erlang machine, which involves writing an Erlang driver. There is no way one can use a port program, there has to be an Erlang driver. Erlang drivers can either be statically linked to the emulator, which can be an alternative when using the open source distribution of Erlang, or dynamically loaded into the Erlang machines address space, which is the only alternative if a precompiled version of Erlang is to be used.
Writing an Erlang driver is by no means easy. The driver is written as a couple of call-back functions called by the Erlang emulator when data is sent to the driver or the driver has any data available on a file descriptor. As the driver call-back routines execute in the main thread of the Erlang machine, the call-back functions can perform no blocking activity whatsoever. The call-backs should only set up file descriptors for waiting and/or read/write available data. All I/O has to be non blocking. Driver call-backs are however executed in sequence, why a global state can safely be updated within the routines.
When the driver is implemented, one would preferably write an Erlang interface for the driver to be able to test the functionality of the driver separately. This interface can then be used by the distribution module which will cover the details of the protocol from the net_kernel. The easiest path is to mimic the inet and inet_tcp interfaces, but a lot of functionality in those modules need not be implemented. In the example application, only a few of the usual interfaces are implemented, and they are much simplified.
When the protocol is available to Erlang through a driver and an Erlang interface module, a distribution module can be written. The distribution module is a module with well defined call-backs, much like a gen_server (there is no compiler support for checking the call-backs though). The details of finding other nodes (i.e. talking to epmd or something similar), creating a listen port (or similar), connecting to other nodes and performing the handshakes/cookie verification are all implemented by this module. There is however a utility module, dist_util, that will do most of the hard work of handling handshakes, cookies, timers and ticking. Using dist_util makes implementing a distribution module much easier and that's what we are doing in the example application.
The last step is to create boot scripts to make the protocol implementation available at boot time. The implementation can be debugged by starting the distribution when all of the system is running, but in a real system the distribution should start very early, why a boot-script and some command line parameters are necessary. This last step also implies that the Erlang code in the interface and distribution modules is written in such a way that it can be run in the startup phase. Most notably there can be no calls to the application module or to any modules not loaded at boot-time (i.e. only kernel, stdlib and the application itself can be used).
3.2 The driver
Although Erlang drivers in general may be beyond the scope of this document, a brief introduction seems to be in place.
Drivers in general
An Erlang driver is a native code module written in C (or assembler) which serves as an interface for some special operating system service. This is a general mechanism that is used throughout the Erlang emulator for all kinds of I/O. An Erlang driver can be dynamically linked (or loaded) to the Erlang emulator at runtime by using the erl_ddll Erlang module. Some of the drivers in OTP are however statically linked to the runtime system, but that's more an optimization than a necessity.
The driver data-types and the functions available to the driver writer are defined in the header file erl_driver.h (there is also an deprecated version called driver.h, don't use that one.) seated in Erlang's include directory (and in $ERL_TOP/erts/emulator/beam in the source code distribution). Refer to that file for function prototypes etc.
When writing a driver to make a communications protocol available to Erlang, one should know just about everything worth knowing about that particular protocol. All operation has to be non blocking and all possible situations should be accounted for in the driver. A non stable driver will affect and/or crash the whole Erlang runtime system, which is seldom what's wanted.
The emulator calls the driver in the following situations:
- When the driver is loaded. This call-back has to have a special name and will inform the emulator of what call-backs should be used by returning a pointer to a ErlDrvEntry struct, which should be properly filled in (see below).
- When a port to the driver is opened (by a open_port call from Erlang). This routine should set up internal data structures and return an opaque data entity of the type ErlDrvData, which is a data-type large enough to hold a pointer. The pointer returned by this function will be the first argument to all other call-backs concerning this particular port. It is usually called the port handle. The emulator only stores the handle and does never try to interpret it, why it can be virtually anything (well anything not larger than a pointer that is) and can point to anything if it is a pointer. Usually this pointer will refer to a structure holding information about the particular port, as i t does in our example.
- When an Erlang process sends data to the port. The data will arrive as a buffer of bytes, the interpretation is not defined, but is up to the implementor. This call-back returns nothing to the caller, answers are sent to the caller as messages (using a routine called driver_output available to all drivers). There is also a way to talk in a synchronous way to drivers, described below. There can be an additional call-back function for handling data that is fragmented (sent in a deep io-list). That interface will get the data in a form suitable for Unix writev rather than in a single buffer. There is no need for a distribution driver to implement such a call-back, so we wont.
- When a file descriptor is signaled for input. This call-back is called when the emulator detects input on a file descriptor which the driver has marked for monitoring by using the interface driver_select. The mechanism of driver select makes it possible to read non blocking from file descriptors by calling driver_select when reading is needed and then do the actual reading in this call-back (when reading is actually possible). The typical scenario is that driver_select is called when an Erlang process orders a read operation, and that this routine sends the answer when data is available on the file descriptor.
- When a file descriptor is signaled for output. This call-back is called in a similar way as the previous, but when writing to a file descriptor is possible. The usual scenario is that Erlang orders writing on a file descriptor and that the driver calls driver_select. When the descriptor is ready for output, this call-back is called an the driver can try to send the output. There may of course be queuing involved in such operations, and there are some convenient queue routines available to the driver writer to use in such situations.
- When a port is closed, either by an Erlang process or by the driver calling one of the driver_failure_XXX routines. This routine should clean up everything connected to one particular port. Note that when other call-backs call a driver_failure_XXX routine, this routine will be immediately called and the call-back routine issuing the error can make no more use of the data structures for the port, as this routine surely has freed all associated data and closed all file descriptors. If the queue utility available to driver writes is used, this routine will however not be called until the queue is empty.
- When an Erlang process calls erlang:port_control/3, which is a synchronous interface to drivers. The control interface is used to set driver options, change states of ports etc. We'll use this interface quite a lot in our example.
- When a timer expires. The driver can set timers with the function driver_set_timer. When such timers expire, a specific call-back function is called. We will not use timers in our example.
- When the whole driver is unloaded. Every resource allocated by the driver should be freed.
The distribution driver's data structures
The driver used for Erlang distribution should implement a reliable, order maintaining, variable length packet oriented protocol. All error correction, re-sending and such need to be implemented in the driver or by the underlying communications protocol. If the protocol is stream oriented (as is the case with both TCP/IP and our streamed Unix domain sockets), some mechanism for packaging is needed. We will use the simple method of having a header of four bytes containing the length of the package in a big endian 32 bit integer (as Unix domain sockets only can be used between processes on the same machine, we actually don't need to code the integer in some special endianess, but I'll do it anyway because in most situation you do need to do it. Unix domain sockets are reliable and order maintaining, so we don't need to implement resends and such in our driver.
Lets start writing our example Unix domain sockets driver by declaring prototypes and filling in a static ErlDrvEntry structure.
( 1) #include <stdio.h>
( 2) #include <stdlib.h>
( 3) #include <string.h>
( 4) #include <unistd.h>
( 5) #include <errno.h>
( 6) #include <sys/types.h>
( 7) #include <sys/stat.h>
( 8) #include <sys/socket.h>
( 9) #include <sys/un.h>
(10) #include <fcntl.h>
(11) #define HAVE_UIO_H
(12) #include "erl_driver.h"
(13) /*
(14) ** Interface routines
(15) */
(16) static ErlDrvData uds_start(ErlDrvPort port, char *buff);
(17) static void uds_stop(ErlDrvData handle);
(18) static void uds_command(ErlDrvData handle, char *buff, int bufflen);
(19) static void uds_input(ErlDrvData handle, ErlDrvEvent event);
(20) static void uds_output(ErlDrvData handle, ErlDrvEvent event);
(21) static void uds_finish(void);
(22) static int uds_control(ErlDrvData handle, unsigned int command,
(23) char* buf, int count, char** res, int res_size);
(24) /* The driver entry */
(25) static ErlDrvEntry uds_driver_entry = {
(26) NULL, /* init, N/A */
(27) uds_start, /* start, called when port is opened */
(28) uds_stop, /* stop, called when port is closed */
(29) uds_command, /* output, called when erlang has sent */
(30) uds_input, /* ready_input, called when input
(31) descriptor ready */
(32) uds_output, /* ready_output, called when output
(33) descriptor ready */
(34) "uds_drv", /* char *driver_name, the argument
(35) to open_port */
(36) uds_finish, /* finish, called when unloaded */
(37) NULL, /* void * that is not used (BC) */
(38) uds_control, /* control, port_control callback */
(39) NULL, /* timeout, called on timeouts */
(40) NULL, /* outputv, vector output interface */
(41) NULL, /* ready_async callback */
(42) NULL, /* flush callback */
(43) NULL, /* call callback */
(44) NULL, /* event callback */
(45) ERL_DRV_EXTENDED_MARKER, /* Extended driver interface marker */
(46) ERL_DRV_EXTENDED_MAJOR_VERSION, /* Major version number */
(47) ERL_DRV_EXTENDED_MINOR_VERSION, /* Minor version number */
(48) ERL_DRV_FLAG_SOFT_BUSY, /* Driver flags. Soft busy flag is
(49) required for distribution drivers */
(50) NULL, /* Reserved for internal use */
(51) NULL, /* process_exit callback */
(52) NULL /* stop_select callback */
(53) };On line 1 to 10 we have included the OS headers needed for our driver. As this driver is written for Solaris, we know that the header uio.h exists, why we can define the preprocessor variable HAVE_UIO_H before we include erl_driver.h at line 12. The definition of HAVE_UIO_H will make the I/O vectors used in Erlang's driver queues to correspond to the operating systems ditto, which is very convenient.
The different call-back functions are declared ("forward declarations") on line 16 to 23.
The driver structure is similar for statically linked in drivers and dynamically loaded. However some of the fields should be left empty (i.e. initialized to NULL) in the different types of drivers. The first field (the init function pointer) is always left blank in a dynamically loaded driver, which can be seen on line 26. The NULL on line 37 should always be there, the field is no longer used and is retained for backward compatibility. We use no timers in this driver, why no call-back for timers is needed. The outputv field (line 40) can be used to implement an interface similar to Unix writev for output. The Erlang runtime system could previously not use outputv for the distribution, but since erts version 5.7.2 it can. Since this driver was written before erts version 5.7.2 it does not use the outputv callback. Using the outputv callback is preferred since it reduces copying of data. (We will however use scatter/gather I/O internally in the driver).
As of erts version 5.5.3 the driver interface was extended with version control and the possibility to pass capability information. Capability flags are present at line 48. As of erts version 5.7.4 the ERL_DRV_FLAG_SOFT_BUSY flag is required for drivers that are to be used by the distribution. The soft busy flag implies that the driver is capable of handling calls to the output and outputv callbacks even though it has marked itself as busy. This has always been a requirement on drivers used by the distribution, but there have previously not been any capability information available about this. For more information see set_busy_port()).
This driver was written before the runtime system had SMP support. The driver will still function in the runtime system with SMP support, but performance will suffer from lock contention on the driver lock used for the driver. This can be alleviated by reviewing and perhaps rewriting the code so that each instance of the driver safely can execute in parallel. When instances safely can execute in parallel it is safe to enable instance specific locking on the driver. This is done by passing ERL_DRV_FLAG_USE_PORT_LOCKING as a driver flag. This is left as an exercise for the reader.
Our defined call-backs thus are:
- uds_start, which shall initiate data for a port. We wont create any actual sockets here, just initialize data structures.
- uds_stop, the function called when a port is closed.
- uds_command, which will handle messages from Erlang. The messages can either be plain data to be sent or more subtle instructions to the driver. We will use this function mostly for data pumping.
- uds_input, this is the call-back which is called when we have something to read from a socket.
- uds_output, this is the function called when we can write to a socket.
- uds_finish, which is called when the driver is unloaded. A distribution driver will actually (or hopefully) never be unloaded, but we include this for completeness. Being able to clean up after oneself is always a good thing.
- uds_control, the erlang:port_control/2 call-back, which will be used a lot in this implementation.
The ports implemented by this driver will operate in two major modes, which i will call the command and data modes. In command mode, only passive reading and writing (like gen_tcp:recv/gen_tcp:send) can be done, and this is the mode the port will be in during the distribution handshake. When the connection is up, the port will be switched to data mode and all data will be immediately read and passed further to the Erlang emulator. In data mode, no data arriving to the uds_command will be interpreted, but just packaged and sent out on the socket. The uds_control call-back will do the switching between those two modes.
While the net_kernel informs different subsystems that the connection is coming up, the port should accept data to send, but not receive any data, to avoid that data arrives from another node before every kernel subsystem is prepared to handle it. We have a third mode for this intermediate stage, lets call it the intermediate mode.
Lets define an enum for the different types of ports we have:
( 1) typedef enum {
( 2) portTypeUnknown, /* An uninitialized port */
( 3) portTypeListener, /* A listening port/socket */
( 4) portTypeAcceptor, /* An intermediate stage when accepting
( 5) on a listen port */
( 6) portTypeConnector, /* An intermediate stage when connecting */
( 7) portTypeCommand, /* A connected open port in command mode */
( 8) portTypeIntermediate, /* A connected open port in special
( 9) half active mode */
(10) portTypeData /* A connectec open port in data mode */
(11) } PortType; Lets look at the different types:
- portTypeUnknown - The type a port has when it's opened, but not actually bound to any file descriptor.
- portTypeListener - A port that is connected to a listen socket. This port will not do especially much, there will be no data pumping done on this socket, but there will be read data available when one is trying to do an accept on the port.
- portTypeAcceptor - This is a port that is to represent the result of an accept operation. It is created when one wants to accept from a listen socket, and it will be converted to a portTypeCommand when the accept succeeds.
- portTypeConnector - Very similar to portTypeAcceptor, an intermediate stage between the request for a connect operation and that the socket is really connected to an accepting ditto in the other end. As soon as the sockets are connected, the port will switch type to portTypeCommand.
- portTypeCommand - A connected socket (or accepted socket if you want) that is in the command mode mentioned earlier.
- portTypeIntermediate - The intermediate stage for a connected socket. There should be no processing of input for this socket.
- portTypeData - The mode where data is pumped through the port and the uds_command routine will regard every call as a call where sending is wanted. In this mode all input available will be read and sent to Erlang as soon as it arrives on the socket, much like in the active mode of a gen_tcp socket.
Now lets look at the state we'll need for our ports. One can note that not all fields are used for all types of ports and that one could save some space by using unions, but that would clutter the code with multiple indirections, so i simply use one struct for all types of ports, for readability.
( 1) typedef unsigned char Byte;
( 2) typedef unsigned int Word;
( 3) typedef struct uds_data {
( 4) int fd; /* File descriptor */
( 5) ErlDrvPort port; /* The port identifier */
( 6) int lockfd; /* The file descriptor for a lock file in
( 7) case of listen sockets */
( 8) Byte creation; /* The creation serial derived from the
( 9) lockfile */
(10) PortType type; /* Type of port */
(11) char *name; /* Short name of socket for unlink */
(12) Word sent; /* Bytes sent */
(13) Word received; /* Bytes received */
(14) struct uds_data *partner; /* The partner in an accept/listen pair */
(15) struct uds_data *next; /* Next structure in list */
(16) /* The input buffer and its data */
(17) int buffer_size; /* The allocated size of the input buffer */
(18) int buffer_pos; /* Current position in input buffer */
(19) int header_pos; /* Where the current header is in the
(20) input buffer */
(21) Byte *buffer; /* The actual input buffer */
(22) } UdsData; This structure is used for all types of ports although some fields are useless for some types. The least memory consuming solution would be to arrange this structure as a union of structures, but the multiple indirections in the code to access a field in such a structure will clutter the code to much for an example.
Let's look at the fields in our structure:
- fd - The file descriptor of the socket associated with the port.
- port - The port identifier for the port which this structure corresponds to. It is needed for most driver_XXX calls from the driver back to the emulator.
-
lockfd - If the socket is a listen socket, we use a separate (regular) file for two purposes:
- We want a locking mechanism that gives no race conditions, so that we can be sure of if another Erlang node uses the listen socket name we require or if the file is only left there from a previous (crashed) session.
-
We store the creation serial number in the file. The creation is a number that should change between different instances of different Erlang emulators with the same name, so that process identifiers from one emulator won't be valid when sent to a new emulator with the same distribution name. The creation can be between 0 and 3 (two bits) and is stored in every process identifier sent to another node.
In a system with TCP based distribution, this data is kept in the Erlang port mapper daemon (epmd), which is contacted when a distributed node starts. The lock-file and a convention for the UDS listen socket's name will remove the need for epmd when using this distribution module. UDS is always restricted to one host, why avoiding a port mapper is easy.
- creation - The creation number for a listen socket, which is calculated as (the value found in the lock-file + 1) rem 4. This creation value is also written back into the lock-file, so that the next invocation of the emulator will found our value in the file.
- type - The current type/state of the port, which can be one of the values declared above.
- name - The name of the socket file (the path prefix removed), which allows for deletion (unlink) when the socket is closed.
- sent - How many bytes that have been sent over the socket. This may wrap, but that's no problem for the distribution, as the only thing that interests the Erlang distribution is if this value has changed (the Erlang net_kernel ticker uses this value by calling the driver to fetch it, which is done through the erlang:port_control routine).
- received - How many bytes that are read (received) from the socket, used in similar ways as sent.
- partner - A pointer to another port structure, which is either the listen port from which this port is accepting a connection or the other way around. The "partner relation" is always bidirectional.
- next - Pointer to next structure in a linked list of all port structures. This list is used when accepting connections and when the driver is unloaded.
- buffer_size, buffer_pos, header_pos, buffer - data for input buffering. Refer to the source code (in the kernel/examples directory) for details about the input buffering. That certainly goes beyond the scope of this document.
Selected parts of the distribution driver implementation
The distribution drivers implementation is not completely covered in this text, details about buffering and other things unrelated to driver writing are not explained. Likewise are some peculiarities of the UDS protocol not explained in detail. The chosen protocol is not important.
Prototypes for the driver call-back routines can be found in the erl_driver.h header file.
The driver initialization routine is (usually) declared with a macro to make the driver easier to port between different operating systems (and flavours of systems). This is the only routine that has to have a well defined name. All other call-backs are reached through the driver structure. The macro to use is named DRIVER_INIT and takes the driver name as parameter.
(1) /* Beginning of linked list of ports */
(2) static UdsData *first_data;
(3) DRIVER_INIT(uds_drv)
(4) {
(5) first_data = NULL;
(6) return &uds_driver_entry;
(7) } The routine initializes the single global data structure and returns a pointer to the driver entry. The routine will be called when erl_ddll:load_driver is called from Erlang.
The uds_start routine is called when a port is opened from Erlang. In our case, we only allocate a structure and initialize it. Creating the actual socket is left to the uds_command routine.
( 1) static ErlDrvData uds_start(ErlDrvPort port, char *buff)
( 2) {
( 3) UdsData *ud;
( 4)
( 5) ud = ALLOC(sizeof(UdsData));
( 6) ud->fd = -1;
( 7) ud->lockfd = -1;
( 8) ud->creation = 0;
( 9) ud->port = port;
(10) ud->type = portTypeUnknown;
(11) ud->name = NULL;
(12) ud->buffer_size = 0;
(13) ud->buffer_pos = 0;
(14) ud->header_pos = 0;
(15) ud->buffer = NULL;
(16) ud->sent = 0;
(17) ud->received = 0;
(18) ud->partner = NULL;
(19) ud->next = first_data;
(20) first_data = ud;
(21)
(22) return((ErlDrvData) ud);
(23) } Every data item is initialized, so that no problems will arise when a newly created port is closed (without there being any corresponding socket). This routine is called when open_port({spawn, "uds_drv"},[]) is called from Erlang.
The uds_command routine is the routine called when an Erlang process sends data to the port. All asynchronous commands when the port is in command mode as well as the sending of all data when the port is in data mode is handled in this9s routine. Let's have a look at it:
( 1) static void uds_command(ErlDrvData handle, char *buff, int bufflen)
( 2) {
( 3) UdsData *ud = (UdsData *) handle;
( 4) if (ud->type == portTypeData || ud->type == portTypeIntermediate) {
( 5) DEBUGF(("Passive do_send %d",bufflen));
( 6) do_send(ud, buff + 1, bufflen - 1); /* XXX */
( 7) return;
( 8) }
( 9) if (bufflen == 0) {
(10) return;
(11) }
(12) switch (*buff) {
(13) case 'L':
(14) if (ud->type != portTypeUnknown) {
(15) driver_failure_posix(ud->port, ENOTSUP);
(16) return;
(17) }
(18) uds_command_listen(ud,buff,bufflen);
(19) return;
(20) case 'A':
(21) if (ud->type != portTypeUnknown) {
(22) driver_failure_posix(ud->port, ENOTSUP);
(23) return;
(24) }
(25) uds_command_accept(ud,buff,bufflen);
(26) return;
(27) case 'C':
(28) if (ud->type != portTypeUnknown) {
(29) driver_failure_posix(ud->port, ENOTSUP);
(30) return;
(31) }
(32) uds_command_connect(ud,buff,bufflen);
(33) return;
(34) case 'S':
(35) if (ud->type != portTypeCommand) {
(36) driver_failure_posix(ud->port, ENOTSUP);
(37) return;
(38) }
(39) do_send(ud, buff + 1, bufflen - 1);
(40) return;
(41) case 'R':
(42) if (ud->type != portTypeCommand) {
(43) driver_failure_posix(ud->port, ENOTSUP);
(44) return;
(45) }
(46) do_recv(ud);
(47) return;
(48) default:
(49) return;
(50) }
(51) } The command routine takes three parameters; the handle returned for the port by uds_start, which is a pointer to the internal port structure, the data buffer and the length of the data buffer. The buffer is the data sent from Erlang (a list of bytes) converted to an C array (of bytes).
If Erlang sends i.e. the list [$a,$b,$c] to the port, the bufflen variable will be 3 ant the buff variable will contain {'a','b','c'} (no null termination). Usually the first byte is used as an opcode, which is the case in our driver to (at least when the port is in command mode). The opcodes are defined as:
- 'L'<socketname>: Create and listen on socket with the given name.
- 'A'<listennumber as 32 bit bigendian>: Accept from the listen socket identified by the given identification number. The identification number is retrieved with the uds_control routine.
- 'C'<socketname>: Connect to the socket named <socketname>.
- 'S'<data>: Send the data <data> on the connected/accepted socket (in command mode). The sending is acked when the data has left this process.
- 'R': Receive one packet of data.
One may wonder what is meant by "one packet of data" in the 'R' command. This driver always sends data packeted with a 4 byte header containing a big endian 32 bit integer that represents the length of the data in the packet. There is no need for different packet sizes or some kind of streamed mode, as this driver is for the distribution only. One may wonder why the header word is coded explicitly in big endian when an UDS socket is local to the host. The answer simply is that I see it as a good practice when writing a distribution driver, as distribution in practice usually cross the host boundaries.
On line 4-8 we handle the case where the port is in data or intermediate mode, the rest of the routine handles the different commands. We see (first on line 15) that the routine uses the driver_failure_posix() routine to report errors. One important thing to remember is that the failure routines make a call to our uds_stop routine, which will remove the internal port data. The handle (and the casted handle ud) is therefore invalid pointers after a driver_failure call and we should immediately return. The runtime system will send exit signals to all linked processes.
The uds_input routine gets called when data is available on a file descriptor previously passed to the driver_select routine. Typically this happens when a read command is issued and no data is available. Lets look at the do_recv routine:
( 1) static void do_recv(UdsData *ud)
( 2) {
( 3) int res;
( 4) char *ibuf;
( 5) for(;;) {
( 6) if ((res = buffered_read_package(ud,&ibuf)) < 0) {
( 7) if (res == NORMAL_READ_FAILURE) {
( 8) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_READ, 1);
( 9) } else {
(10) driver_failure_eof(ud->port);
(11) }
(12) return;
(13) }
(14) /* Got a package */
(15) if (ud->type == portTypeCommand) {
(16) ibuf[-1] = 'R'; /* There is always room for a single byte
(17) opcode before the actual buffer
(18) (where the packet header was) */
(19) driver_output(ud->port,ibuf - 1, res + 1);
(20) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_READ,0);
(21) return;
(22) } else {
(23) ibuf[-1] = DIST_MAGIC_RECV_TAG; /* XXX */
(24) driver_output(ud->port,ibuf - 1, res + 1);
(25) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_READ,1);
(26) }
(27) }
(28) } The routine tries to read data until a packet is read or the buffered_read_package routine returns a NORMAL_READ_FAILURE (an internally defined constant for the module that means that the read operation resulted in an EWOULDBLOCK). If the port is in command mode, the reading stops when one package is read, but if it is in data mode, the reading continues until the socket buffer is empty (read failure). If no more data can be read and more is wanted (always the case when socket is in data mode) driver_select is called to make the uds_input call-back be called when more data is available for reading.
When the port is in data mode, all data is sent to Erlang in a format that suits the distribution, in fact the raw data will never reach any Erlang process, but will be translated/interpreted by the emulator itself and then delivered in the correct format to the correct processes. In the current emulator version, received data should be tagged with a single byte of 100. Thats what the macro DIST_MAGIC_RECV_TAG is defined to. The tagging of data in the distribution will possibly change in the future.
The uds_input routine will handle other input events (like nonblocking accept), but most importantly handle data arriving at the socket by calling do_recv:
( 1) static void uds_input(ErlDrvData handle, ErlDrvEvent event)
( 2) {
( 3) UdsData *ud = (UdsData *) handle;
( 4) if (ud->type == portTypeListener) {
( 5) UdsData *ad = ud->partner;
( 6) struct sockaddr_un peer;
( 7) int pl = sizeof(struct sockaddr_un);
( 8) int fd;
( 9) if ((fd = accept(ud->fd, (struct sockaddr *) &peer, &pl)) < 0) {
(10) if (errno != EWOULDBLOCK) {
(11) driver_failure_posix(ud->port, errno);
(12) return;
(13) }
(14) return;
(15) }
(16) SET_NONBLOCKING(fd);
(17) ad->fd = fd;
(18) ad->partner = NULL;
(19) ad->type = portTypeCommand;
(20) ud->partner = NULL;
(21) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_READ, 0);
(22) driver_output(ad->port, "Aok",3);
(23) return;
(24) }
(25) do_recv(ud);
(26) } The important line here is the last line in the function, the do_read routine is called to handle new input. The rest of the function handles input on a listen socket, which means that there should be possible to do an accept on the socket, which is also recognized as a read event.
The output mechanisms are similar to the input. Lets first look at the do_send routine:
( 1) static void do_send(UdsData *ud, char *buff, int bufflen)
( 2) {
( 3) char header[4];
( 4) int written;
( 5) SysIOVec iov[2];
( 6) ErlIOVec eio;
( 7) ErlDrvBinary *binv[] = {NULL,NULL};
( 8) put_packet_length(header, bufflen);
( 9) iov[0].iov_base = (char *) header;
(10) iov[0].iov_len = 4;
(11) iov[1].iov_base = buff;
(12) iov[1].iov_len = bufflen;
(13) eio.iov = iov;
(14) eio.binv = binv;
(15) eio.vsize = 2;
(16) eio.size = bufflen + 4;
(17) written = 0;
(18) if (driver_sizeq(ud->port) == 0) {
(19) if ((written = writev(ud->fd, iov, 2)) == eio.size) {
(20) ud->sent += written;
(21) if (ud->type == portTypeCommand) {
(22) driver_output(ud->port, "Sok", 3);
(23) }
(24) return;
(25) } else if (written < 0) {
(26) if (errno != EWOULDBLOCK) {
(27) driver_failure_eof(ud->port);
(28) return;
(29) } else {
(30) written = 0;
(31) }
(32) } else {
(33) ud->sent += written;
(34) }
(35) /* Enqueue remaining */
(36) }
(37) driver_enqv(ud->port, &eio, written);
(38) send_out_queue(ud);
(39) } This driver uses the writev system call to send data onto the socket. A combination of writev and the driver output queues is very convenient. An ErlIOVec structure contains a SysIOVec (which is equivalent to the struct iovec structure defined in uio.h. The ErlIOVec also contains an array of ErlDrvBinary pointers, of the same length as the number of buffers in the I/O vector itself. One can use this to allocate the binaries for the queue "manually" in the driver, but we'll just fill the binary array with NULL values (line 7) , which will make the runtime system allocate its own buffers when we call driver_enqv (line 37).
The routine builds an I/O vector containing the header bytes and the buffer (the opcode has been removed and the buffer length decreased by the output routine). If the queue is empty, we'll write the data directly to the socket (or at least try to). If any data is left, it is stored in the queue and then we try to send the queue (line 38). An ack is sent when the message is delivered completely (line 22). The send_out_queue will send acks if the sending is completed there. If the port is in command mode, the Erlang code serializes the send operations so that only one packet can be waiting for delivery at a time. Therefore the ack can be sent simply whenever the queue is empty.
A short look at the send_out_queue routine:
( 1) static int send_out_queue(UdsData *ud)
( 2) {
( 3) for(;;) {
( 4) int vlen;
( 5) SysIOVec *tmp = driver_peekq(ud->port, &vlen);
( 6) int wrote;
( 7) if (tmp == NULL) {
( 8) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_WRITE, 0);
( 9) if (ud->type == portTypeCommand) {
(10) driver_output(ud->port, "Sok", 3);
(11) }
(12) return 0;
(13) }
(14) if (vlen > IO_VECTOR_MAX) {
(15) vlen = IO_VECTOR_MAX;
(16) }
(17) if ((wrote = writev(ud->fd, tmp, vlen)) < 0) {
(18) if (errno == EWOULDBLOCK) {
(19) driver_select(ud->port, (ErlDrvEvent) ud->fd,
(20) DO_WRITE, 1);
(21) return 0;
(22) } else {
(23) driver_failure_eof(ud->port);
(24) return -1;
(25) }
(26) }
(27) driver_deq(ud->port, wrote);
(28) ud->sent += wrote;
(29) }
(30) } What we do is simply to pick out an I/O vector from the queue (which is the whole queue as an SysIOVec). If the I/O vector is to long (IO_VECTOR_MAX is defined to 16), the vector length is decreased (line 15), otherwise the writev (line 17) call will fail. Writing is tried and anything written is dequeued (line 27). If the write fails with EWOULDBLOCK (note that all sockets are in nonblocking mode), driver_select is called to make the uds_output routine be called when there is space to write again.
We will continue trying to write until the queue is empty or the writing would block.
The routine above are called from the uds_output routine, which looks like this:
( 1) static void uds_output(ErlDrvData handle, ErlDrvEvent event)
( 2) {
( 3) UdsData *ud = (UdsData *) handle;
( 4) if (ud->type == portTypeConnector) {
( 5) ud->type = portTypeCommand;
( 6) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_WRITE, 0);
( 7) driver_output(ud->port, "Cok",3);
( 8) return;
( 9) }
(10) send_out_queue(ud);
(11) } The routine is simple, it first handles the fact that the output select will concern a socket in the business of connecting (and the connecting blocked). If the socket is in a connected state it simply sends the output queue, this routine is called when there is possible to write to a socket where we have an output queue, so there is no question what to do.
The driver implements a control interface, which is a synchronous interface called when Erlang calls erlang:port_control/3. This is the only interface that can control the driver when it is in data mode and it may be called with the following opcodes:
- 'C': Set port in command mode.
- 'I': Set port in intermediate mode.
- 'D': Set port in data mode.
- 'N': Get identification number for listen port, this identification number is used in an accept command to the driver, it is returned as a big endian 32 bit integer, which happens to be the file identifier for the listen socket.
- 'S': Get statistics, which is the number of bytes received, the number of bytes sent and the number of bytes pending in the output queue. This data is used when the distribution checks that a connection is alive (ticking). The statistics is returned as 3 32 bit big endian integers.
- 'T': Send a tick message, which is a packet of length 0. Ticking is done when the port is in data mode, so the command for sending data cannot be used (besides it ignores zero length packages in command mode). This is used by the ticker to send dummy data when no other traffic is present. Note that it is important that the interface for sending ticks is not blocking. This implementation uses erlang:port_control/3 which does not block the caller. If erlang:port_command is used, use erlang:port_command/3 and pass [force] as option list; otherwise, the caller can be blocked indefinitely on a busy port and prevent the system from taking down a connection that is not functioning.
- 'R': Get creation number of listen socket, which is used to dig out the number stored in the lock file to differentiate between invocations of Erlang nodes with the same name.
The control interface gets a buffer to return its value in, but is free to allocate its own buffer is the provided one is to small. Here is the code for uds_control:
( 1) static int uds_control(ErlDrvData handle, unsigned int command,
( 2) char* buf, int count, char** res, int res_size)
( 3) {
( 4) /* Local macro to ensure large enough buffer. */
( 5) #define ENSURE(N) \
( 6) do { \
( 7) if (res_size < N) { \
( 8) *res = ALLOC(N); \
( 9) } \
(10) } while(0)
(11) UdsData *ud = (UdsData *) handle;
(12) switch (command) {
(13) case 'S':
(14) {
(15) ENSURE(13);
(16) **res = 0;
(17) put_packet_length((*res) + 1, ud->received);
(18) put_packet_length((*res) + 5, ud->sent);
(19) put_packet_length((*res) + 9, driver_sizeq(ud->port));
(20) return 13;
(21) }
(22) case 'C':
(23) if (ud->type < portTypeCommand) {
(24) return report_control_error(res, res_size, "einval");
(25) }
(26) ud->type = portTypeCommand;
(27) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_READ, 0);
(28) ENSURE(1);
(29) **res = 0;
(30) return 1;
(31) case 'I':
(32) if (ud->type < portTypeCommand) {
(33) return report_control_error(res, res_size, "einval");
(34) }
(35) ud->type = portTypeIntermediate;
(36) driver_select(ud->port, (ErlDrvEvent) ud->fd, DO_READ, 0);
(37) ENSURE(1);
(38) **res = 0;
(39) return 1;
(40) case 'D':
(41) if (ud->type < portTypeCommand) {
(42) return report_control_error(res, res_size, "einval");
(43) }
(44) ud->type = portTypeData;
(45) do_recv(ud);
(46) ENSURE(1);
(47) **res = 0;
(48) return 1;
(49) case 'N':
(50) if (ud->type != portTypeListener) {
(51) return report_control_error(res, res_size, "einval");
(52) }
(53) ENSURE(5);
(54) (*res)[0] = 0;
(55) put_packet_length((*res) + 1, ud->fd);
(56) return 5;
(57) case 'T': /* tick */
(58) if (ud->type != portTypeData) {
(59) return report_control_error(res, res_size, "einval");
(60) }
(61) do_send(ud,"",0);
(62) ENSURE(1);
(63) **res = 0;
(64) return 1;
(65) case 'R':
(66) if (ud->type != portTypeListener) {
(67) return report_control_error(res, res_size, "einval");
(68) }
(69) ENSURE(2);
(70) (*res)[0] = 0;
(71) (*res)[1] = ud->creation;
(72) return 2;
(73) default:
(74) return report_control_error(res, res_size, "einval");
(75) }
(76) #undef ENSURE
(77) } The macro ENSURE (line 5 to 10) is used to ensure that the buffer is large enough for our answer. We switch on the command and take actions, there is not much to say about this routine. Worth noting is that we always has read select active on a port in data mode (achieved by calling do_recv on line 45), but turn off read selection in intermediate and command modes (line 27 and 36).
The rest of the driver is more or less UDS specific and not of general interest.
3.3 Putting it all together
To test the distribution, one can use the net_kernel:start/1 function, which is useful as it starts the distribution on a running system, where tracing/debugging can be performed. The net_kernel:start/1 routine takes a list as its single argument. The lists first element should be the node name (without the "@hostname") as an atom, and the second (and last) element should be one of the atoms shortnames or longnames. In the example case shortnames is preferred.
For net kernel to find out which distribution module to use, the command line argument -proto_dist is used. The argument is followed by one or more distribution module names, with the "_dist" suffix removed, i.e. uds_dist as a distribution module is specified as -proto_dist uds.
If no epmd (TCP port mapper daemon) is used, one should also specify the command line option -no_epmd, which will make Erlang skip the epmd startup, both as a OS process and as an Erlang ditto.
The path to the directory where the distribution modules reside must be known at boot, which can either be achieved by specifying -pa <path> on the command line or by building a boot script containing the applications used for your distribution protocol (in the uds_dist protocol, it's only the uds_dist application that needs to be added to the script).
The distribution will be started at boot if all the above is specified and an -sname <name> flag is present at the command line, here follows two examples:
$ erl -pa $ERL_TOP/lib/kernel/examples/uds_dist/ebin -proto_dist uds -no_epmd Erlang (BEAM) emulator version 5.0 Eshell V5.0 (abort with ^G) 1> net_kernel:start([bing,shortnames]). {ok,<0.30.0>} (bing@hador)2>
...
$ erl -pa $ERL_TOP/lib/kernel/examples/uds_dist/ebin -proto_dist uds \ -no_epmd -sname bong Erlang (BEAM) emulator version 5.0 Eshell V5.0 (abort with ^G) (bong@hador)1>
One can utilize the ERL_FLAGS environment variable to store the complicated parameters in:
$ ERL_FLAGS=-pa $ERL_TOP/lib/kernel/examples/uds_dist/ebin \ -proto_dist uds -no_epmd $ export ERL_FLAGS $ erl -sname bang Erlang (BEAM) emulator version 5.0 Eshell V5.0 (abort with ^G) (bang@hador)1>
The ERL_FLAGS should preferably not include the name of the node.