Explanatory Notes for XTI

Transport Endpoints

To be active, a transport endpoint must have a transport address associated with it by the t_bind() function. A transport connection is characterised by the association of two active endpoints, made by using the functions of establishment of transport connection. The fd is a communication path to a transport provider. There is no direct assignation of the processes to the transport provider, so multiple processes, which obtain the fd by open(), fork() or dup() operations, may access a given communication path. Note that the open() function will work only if the opened character string is a pathname.

Note that in order to guarantee portability, the only operations which the applications may perform on any fd returned by t_open() are those defined by XTI and fcntl(), dup() or dup2(). Other operations are permitted but these will have system-dependent results.

Transport Providers

The transport layer may comprise one or more transport providers at the same time. The identifier parameter of the transport provider passed to the t_open() function determines the required transport provider. To keep the applications portable, the identifier parameter of the transport provider should not be hard-coded into the application source code.

An application which wants to manage multiple transport providers must call t_open() for each provider. For example, a server application which is waiting for incoming connection indications from several transport providers must open a transport endpoint for each provider and listen for connection indications on each of the associated file descriptors.

Association of a UNIX Process to an Endpoint

One process can simultaneously open several fds. However, in synchronous mode, the process must manage the different actions of the associated transport connections sequentially. Conversely, several processes can share the same fd (by fork() or dup() operations) but they have to synchronise themselves so as not to issue a function that is unsuitable to the current state of the transport endpoint.

It is important to remember that the transport provider treats all users of a transport endpoint as a single user. If multiple processes are using the same endpoint, they should coordinate their activities so as not to violate the state of the provider. The t_sync() function returns the current state of the provider to the user, thereby enabling the user to verify the state before taking further action. This coordination is only valid among cooperating processes; it is possible that a process or an incoming event could change the provider's state after a t_sync() is issued.

A process can listen for an incoming connection indication on one fd and accept the connection on a different fd which has been bound with the qlen parameter (see t_bind()) set to zero. This facilitates the writing of a listener application whereby the listener waits for all incoming connection indications on a given Transport Service Access Point (TSAP). The listener will accept the connection on a new fd, and fork() a child process to service the request without blocking other incoming connection indications.

Use of the Same Protocol Address

If several endpoints are bound to the same protocol address, only one at the time may be listening for incoming connections. However, others may be in data transfer state or establishing a transport connection as initiators.

Modes of Service

The connection-mode transport service is circuit-oriented and enables data to be transferred over an established connection in a reliable, sequenced manner. This service enables the negotiation of the parameters and options that govern the transfer of data. It provides an identification mechanism that avoids the overhead of address transmission and resolution during the data transfer phase. It also provides a context in which successive units of data, transferred between peer users, are logically related. This service is attractive to applications that require relatively long-lived, datastream-oriented interactions.

In contrast, the connectionless-mode transport service is message-oriented and supports data transfer in self-contained units with no logical relationship required among multiple units. These units are also known as datagrams. This service requires a pre-existing association between the peer users involved, which determines the characteristics of the data to be transmitted. No dynamic negotiation of parameters and options is supported by this service. All the information required to deliver a unit of data (for example, destination address) is presented to the transport provider, together with the data to be transmitted, in a single service access which need not relate to any other service access. Also, each unit of data transmitted is entirely self-contained, and can be independently routed by the transport provider. This service is attractive to applications that involve short-term request/response interactions, exhibit a high level of redundancy, are dynamically reconfigurable or do not require guaranteed, in-sequence delivery of data.

Error Handling

Two levels of error are defined for the transport interface. The first is the library error level. Each library function has one or more error returns. Failures are indicated by a return value of -1. When header file <xti.h> is included, symbol t_errno is defined as a modifiable lvalue of type int, t_errno,href="#tag_foot_1">¹ and can be used to access the specific error number when such a failure occurs. Applications should not include t_errno in the left operand of assignment statements. This value is set when errors occur but is not cleared on successful library calls, so it should be tested only after an error has been indicated. A diagnostic function, t_error(), prints out information on the current transport error. The state of the transport provider may change if a transport error occurs.

The second level of error is the operating system service routine level. A special library level error number has been defined called [TSYSERR] which is generated by each library function when an operating system service routine fails or some general error occurs. When a function sets t_errno to [TSYSERR], the specific system error may be accessed through the external variable errno.

For example, a system error can be generated by the transport provider when a protocol error has occurred. If the error is severe, it may cause the file descriptor and transport endpoint to be unusable. To continue in this case, all users of the fd must close it. Then the transport endpoint may be re-opened and initialised.

Synchronous and Asynchronous Execution Modes

The transport service interface supports two execution modes for handling asynchronous events: synchronous mode and asynchronous mode. In the synchronous mode of operation, the transport primitives wait for specific events before returning control to the user. While waiting, the user cannot perform other tasks. For example, a function that attempts to receive data in synchronous mode will wait until data arrives before returning control to the user. Synchronous mode is the default mode of execution. It is useful for user processes that want to wait for events to occur, or for user processes that maintain only a single transport connection.

The asynchronous mode of operation, on the other hand, provides a mechanism for notifying a user of some event without forcing the user to wait for the event. The handling of networking events in an asynchronous manner is seen as a desirable capability of the transport interface. This would enable users to perform useful work while expecting a particular event. For example, a function that attempts to receive data in asynchronous mode will return control to the user immediately if no data is available. The user may then periodically poll for incoming data until it arrives. The asynchronous mode is intended for those applications that expect long delays between events and have other tasks that they can perform in the meantime or handle multiple connections concurrently.

The two execution modes are not provided through separate interfaces or different functions. Instead, functions that process incoming events have two modes of operation: synchronous and asynchronous. The desired mode is specified through the O_NONBLOCK flag, which may be set when the transport provider is initially opened, or before any specific function or group of functions is executed using the fcntl() operating system service routine. The effect of this flag is local to this process and is completely specified in the description of each function.

Nine (only eight if the orderly release is not supported) asynchronous events are defined in the transport service interface to cover both connection-mode and connectionless-mode service. They are represented as separate bits in a bit-mask using the following defined symbolic names:

• T_LISTEN

• T_CONNECT

• T_DATA

• T_EXDATA

• T_DISCONNECT

• T_ORDREL

• T_UDERR

• T_GODATA

• T_GOEXDATA.

These are described in Event Management .

A process that issues functions in synchronous mode must still be able to recognise certain asynchronous events and act on them if necessary. This is handled through a special transport error [TLOOK] which is returned by a function when an asynchronous event occurs. The t_look() function is then invoked to identify the specific event that has occurred when this error is returned.

Another means to notify a process that an asynchronous event has occurred is polling. The polling capability enables processes to do useful work and periodically poll for one of the above asynchronous events. This facility is provided by setting O_NONBLOCK for the appropriate primitive(s).

Events and t_look()

All events that occur at a transport endpoint are stored by XTI. These events are retrievable one at a time via the t_look() function. If multiple events occur, it is implementation-dependent in what order t_look() will return the events. An event is outstanding on a transport endpoint until it is consumed. Every event has a corresponding consuming function which handles the event and consumes it. In addition, the abortive T_DISCONNECT consumes other pending events. Both T_DATA and T_EXDATA events are consumed when the corresponding consuming function has read all the corresponding data associated with that event. The intention of this is that T_DATA should always indicate that there is data to receive. Two events, T_GODATA and T_GOEXDATA, are also cleared as they are returned by t_look(). summarises this.

Effect of Signals

If a synchronous XTI call is blocking under circumstances where an asynchronous call would have returned because no event was available, then the call returns -1 with t_errno set to [TSYSERR] and errno set to [EINTR]. The state of the endpoint is unchanged.

In addition an [EINTR] error may be returned by all XTI calls (except t_error() and t_strerror()) under implementation defined conditions. In these cases the state of the endpoint will not have been changed, and no data will have been sent or received. Any buffers provided by the user for return values may have been overwritten.

A "well written" application will itself mask out signals except during specific code sequences (typically only its idle point) to avoid having to handle an [EINTR] return from all system calls.

Application writers should be aware that XTI calls may be implemented in a library as multiple system calls. In order to maintain the endpoint and associated library data areas in a consistent state, some of these system calls may be repeated when interrupted by a signal.

Applications should not call XTI functions from within a signal handler or using the longjmp() or siglongjmp() interfaces (see reference XSH) to exit a signal handler, as either may leave XTI data areas in an inconsistent state.

Applications may be able to cause the XTI library itself to generate signals that interrupt its internal actions (for example, by issuing ioctl( fd, I_SETSIG, S_INPUT ) on a UNIX system); this may cause the user's signal handler to be scheduled, but will not stop the XTI call from completing.

Event Management

Each XTI call deals with one transport endpoint at a time. It is not possible to wait for several events from different sources, particularly from several transport connections at a time. We recognise the need for this functionality which may be available today in a system-dependent fashion.

Throughout the document we refer to an event management service called Event Management (EM) which provides those functions useful to XTI. This Event Management will allow a process to be notified of the following events:

T_LISTEN

A connection request from a remote user was received by a transport provider (connection-mode service only); this event may occur under the following conditions:

1. The file descriptor is bound to a valid address.

2. No transport connection is established at this time.

T_CONNECT

In connection mode only; a connection response was received by the transport provider; occurs after a t_connect() has been issued.

T_DATA

Normal data (whole or part of Transport Service Data Unit (TSDU)) was received by the transport provider.

T_EXDATA

Expedited data was received by the transport provider.

T_DISCONNECT

In connection mode only; a disconnection request was received by the transport provider. It may be reported on both data transfer functions and connection establishment functions and on the t_snddis() function.

T_ORDREL

An orderly release request was received by a transport provider (connection mode with orderly release only).

T_UDERR

In connectionless-mode only; an error was found in a previously sent datagram. It may be notified on the t_rcvudata() or t_unbind() function calls.

T_GODATA

Flow control restrictions on normal data flow that led to a [TFLOW] error have been lifted. Normal data may be sent again.

T_GOEXDATA

Flow control restrictions on expedited data flow that led to a [TFLOW] error have been lifted. Expedited data may be sent again.

Footnotes

1.

This may be implemented as a macro. In addition the name _t_errno is an XTI library-reserved-name for use within such a macro. A typical definition of t_errno for a multithreaded implementation is:

extern int *_t_errno(void);
#define t_errno (*(_t_errno()))

1.

In the case of the t_connect() function the T_CONNECT event is both generated and consumed by the execution of the function and is therefore not visible to the application.