NAME

SYNOPSIS

 #include <libnal/nal.h>

 #define NAL_SELECT_FLAG_READ  (unsigned int)0x0001
 #define NAL_SELECT_FLAG_SEND  (unsigned int)0x0002
 #define NAL_SELECT_FLAG_RW    (NAL_SELECT_FLAG_READ ⎪ NAL_SELECT_FLAG_SEND)

 NAL_CONNECTION *NAL_CONNECTION_new(void);
 void NAL_CONNECTION_free(NAL_CONNECTION *conn);
 void NAL_CONNECTION_reset(NAL_CONNECTION *conn);
 int NAL_CONNECTION_create(NAL_CONNECTION *conn, const NAL_ADDRESS *addr);
 int NAL_CONNECTION_accept(NAL_CONNECTION *conn, NAL_LISTENER *list,
                                NAL_SELECTOR *sel);
 int NAL_CONNECTION_create_pair(NAL_CONNECTION *conn1, NAL_CONNECTION *conn2,
                                unsigned int def_buffer_size);
 #if 0
 int NAL_CONNECTION_create_dummy(NAL_CONNECTION *conn,
                                 unsigned int def_buffer_size);
 #endif
 int NAL_CONNECTION_set_size(NAL_CONNECTION *conn, unsigned int size);
 NAL_BUFFER *NAL_CONNECTION_get_read(NAL_CONNECTION *conn);
 NAL_BUFFER *NAL_CONNECTION_get_send(NAL_CONNECTION *conn);
 const NAL_BUFFER *NAL_CONNECTION_get_read_c(const NAL_CONNECTION *conn);
 const NAL_BUFFER *NAL_CONNECTION_get_send_c(const NAL_CONNECTION *conn);
 int NAL_CONNECTION_io(NAL_CONNECTION *conn, NAL_SELECTOR *sel);
 int NAL_CONNECTION_io_cap(NAL_CONNECTION *conn, NAL_SELECTOR *sel,
                           unsigned int max_read, unsigned int max_send);
 int NAL_CONNECTION_is_established(const NAL_CONNECTION *conn);
 void NAL_CONNECTION_add_to_selector(const NAL_CONNECTION *conn,
                                     NAL_SELECTOR *sel);
 void NAL_CONNECTION_add_to_selector_ex(const NAL_CONNECTION *conn,
                                        NAL_SELECTOR *sel,
                                        unsigned int flags);
 void NAL_CONNECTION_del_from_selector(const NAL_CONNECTION *conn,
                                       NAL_SELECTOR *sel);

DESCRIPTION

NAL_CONNECTION_free() destroys a NAL_CONNECTION object.

NAL_CONNECTION_reset() will, if necessary, cleanup any prior state in conn so that it can be reused in NAL_CONNECTION_create(). Internally, there are other optimisations and benefits to using NAL_CONNECTION_reset() instead of NAL_CONNECTION_free() and NAL_CONNECTION_new() - the implementation can try to avoid repeated reallocation and reinitialisation of state, only doing full cleanup and reinitialisation when necessary.

NAL_CONNECTION_create() will attempt to connect to the address represented by addr. If this succeeds, it means either that the underlying connection of conn is established, or that a non-blocking connect was successfully initiated but has not yet completed (it may still be rejected by the peer eventually). Typically, unix domain sockets connect or fail immediately, and usually TCP/IPv4 connect non-blocking, though this may not be true for some interfaces such as `localhost'. NAL_CONNECTION_is_established() can be used to distinguish the difference. The size of the connection's underlying read and send NAL_BUFFERs is initialised to the default that was created in addr. See the "NOTES" section for more discussion of connection semantics.

NAL_CONNECTION_accept() will not block waiting for incoming connection requests on list, but will accept any pending connection request that had already been identified by a previous call to NAL_SELECTOR_select(2) on sel. See "NOTES".

NAL_CONNECTION_create_pair() will initialise conn1 and conn2 to be end-points of a single connection. This is typically implemented using the socketpair(2) function, and is designed to allow for an IPC mechanism that integrates with libnal. def_buffer_size will control the size of the read and send buffers of both connections if the functions succeed. See the EXAMPLES section for some uses of ``pairs''.

NAL_CONNECTION_create_dummy() will implement a virtual FIFO that has no underlying network resource associated with it. Writing data to the connection amounts to pushing data onto the front of the FIFO, and reading data from the connection amounts to popping data off the end of the FIFO. The size of the FIFO is specified by def_buffer_size. See the "BUGS" section for a note on using these connection types with NAL_SELECTOR.

NAL_CONNECTION_set_size() will resize the read and send buffers of conn to size. The default size of those buffers is inherited from the setting created in the NAL_ADDRESS that initialised conn, or if conn was accepted from a NAL_LISTENER object, then from the address that created the listener. The individual buffers can be resized independantly by using the following two functions to obtain the buffesr and using NAL_BUFFER functions directly.

NAL_CONNECTION_get_read() and NAL_CONNECTION_get_send() return the read and send buffers of conn. This is how reading and writing is performed on conn, as NAL_BUFFER functions may be used on these buffers directly. NAL_CONNECTION_get_read_c() and NAL_CONNECTION_get_send_c() perform the same function but on a constant conn parameter and returning constant pointers to the corresponding buffers.

NAL_CONNECTION_io() will perform any network input/output that is possible given the state in sel. Unless conn had been added to sel via NAL_SELECTOR_add_conn() (or its `_ex' variant) and a resulting call to NAL_SELECTOR_select() had revealed readability and/or writability on conn, this function will silently succeed. Otherwise it will attempt to perform whatever reading or writing was required. If this function fails, that indicates that the connection is no longer valid - this represents a disconnection by the peer, the result of a non-blocking connect that had been initiated but was unable to connect, or some network error that makes conn unusable. See the "NOTES" section.

NAL_CONNECTION_io_cap() is a version of NAL_CONNECTION_io() that allows the caller to specify a limit on the maximum amount conn should read from, or send to, the network. Whether this amount is read or sent (or even whether reading or sending takes place at all) depends on; the data (and space) available is in the connection's buffers, what the results of the last select on sel were, and how much data the host system's networking support will accept or provide to conn.

NAL_CONNECTION_is_established() is useful for determining when a non-blocking connect has completed. See the "NOTES" section.

NAL_CONNECTION_add_to_selector() registers conn with the selector sel for any events relevant to it. NAL_CONNECTION_del_from_selector() can be used to reverse this if called before any subsequent call to NAL_SELECTOR_select(). NAL_CONNECTION_add_to_selector_ex() extends NAL_CONNECTION_add_to_selector() by allowing a bit-mask to be supplied to control what events the connection can be selected on, these flags are indicated above prefixed with NAL_SELECT_FLAG_.

RETURN VALUES

NAL_CONNECTION_free(), NAL_CONNECTION_reset(), NAL_CONNECTION_add_to_selector(), NAL_CONNECTION_add_to_selector_ex(), and NAL_CONNECTION_del_from_selector() have no return value.

NAL_CONNECTION_get_read(), NAL_CONNECTION_get_send(), NAL_CONNECTION_get_read_c(), and NAL_CONNECTION_get_send_c() return pointers to the connection's buffer objects or NULL for failure.

NAL_CONNECTION_accept() returns non-zero if a connection was accepted and is represented by the provided NAL_CONNECTION object, or zero if no connection attempt was pending (or if there was but an error prevented the accept operation).

All other NAL_CONNECTION functions return zero for failure or false, and non-zero for success or true.

NOTES

NAL_CONNECTION_add_to_selector() uses the following logic; the connection is always selected for exception events, and will be selected for readability if its read buffer is not full and writability if its send buffer is not empty.

NAL_CONNECTION_io() is used after calling NAL_CONNECTION_add_to_selector() and a subsequent call to NAL_SELECTOR_select(). It observes the following logic; if an exception event has occured it returns failure, if readability is indicated it will read incoming data up to the limit of the available space in the read buffer, and if writability is indicated it will send as much of the send buffer's data as possible. If NAL_CONNECTION_io() returns failure, the connection is considered broken for some reason and no further I/O operations should be attempted (the behaviour is undefined). NB: The connection object is not automatically cleaned up so as to allow the caller to continue reading any data in the read buffer and/or examine any unsent data in the send buffer.

The above is almost true, BTW :-) The special case is that of non-blocking connects. If NAL_CONNECTION_create() cannot immediately connect without blocking, it will return success but subsequent calls to NAL_CONNECTION_is_established() will reveal that the connection is not yet complete. Any connection that is not complete will request selection for sendability inside NAL_CONNECTION_add_to_selector(), whether the application has provided data to send or not. The completion (or failure) of the non-blocking connect will thus cause any subsequent NAL_SELECTOR_select() operation to break. As with all other semantics, it is the follow up call to NAL_CONNECTION_io() that changes the state of the connection object - if it returns failure, the non-blocking connect failed. If it returns success, you should still call NAL_CONNECTION_is_established() to determine if the connection is complete, as the selector could have broken because of signals or network events on other objects.

NAL_CONNECTION_accept() will return immediately, and will only succeed if the NAL_LISTENER object had already been added to the selector using NAL_LISTENER_add_to_select(), the selector had been subsequently selected using NAL_SELECTOR_select(2), and this indicated an incoming connection request waiting on the listener.

It should be noted that the actual transport in use is virtualised to allow for multiple transports and, because of this, multiple semantics for how the network functionality behaves. TCP/IPv4 and unix domain socket based connections, as well as connection pairs from NAL_CONNECTION_create_pair(), operate very much as described here. The FIFO connection type, created by NAL_CONNECTION_create_dummy() is not yet consistent with this and is described in the "BUGS" section.

BUGS

EXAMPLES

    NAL_BUFFER *c_read, *c_send;
    NAL_SELECTOR *sel = NAL_SELECTOR_new();
    NAL_CONNECTION *conn = NAL_CONNECTION_new();
    NAL_ADDRESS *addr = retrieve_the_desired_address();

    /* Setup */
    NAL_CONNECTION_create(conn, addr);
    c_read = NAL_CONNECTION_get_read(conn);
    c_send = NAL_CONNECTION_get_send(conn);

    /* Loop */
    do {
        /* This is where the state-machine code should process as much data as
         * possible from 'c_read' and/or produce as much output to 'c_send' as
         * it can. */
        ...
        ... user code
        ...
        /* block on (relevant) network events for 'conn' */
        NAL_CONNECTION_add_to_selector(conn, sel);
        NAL_SELECTOR_select(sel, 0, 0);
        /* Do network I/O after the above blocking select and continue looping
         * only if the connection is still alive. */
    } while(NAL_CONNECTION_io(conn, sel));

An example of using a connection pair (with 2 Kb read and send buffers for each connection) to create IPC between a parent process and its child (again, no error checking);

    NAL_CONNECTION *ipc_to_parent = NAL_CONNECTION_new();
    NAL_CONNECTION *ipc_to_child = NAL_CONNECTION_new();

    /* Setup */
    NAL_CONNECTION_create_pair(ipc_to_parent, ipc_to_child, 2048);

    /* Create child process */
    switch(fork()) {
        case 0:
            /* Inside the child process, close our copy of the parent's side */
            NAL_CONNECTION_free(ipc_to_child);
            /* Do child process things, and use 'ipc_to_parent' to communicate
             * with the parent. */
            do_child_logic(ipc_to_parent);
            exit(0);
        default:
            /* Inside the parent process, close our copy of the child's side */
            NAL_CONNECTION_free(ipc_to_parent);
            break;
    }
    /* Continue in the parent process, and use 'ipc_to_child' to communicate
     * with the child. */
    do_parent_logic(ipc_to_child);

Note that these connection pairs can also be a useful way of handling process termination that allow you to bypass signal handling altogether. If a child process terminates, the connection between the pair will be broken and so this will be noticed in the parent process by any selector selecting on the ipc_to_child connection - the subsequent NAL_CONNECTION_io() operation will fail indicating that the child process is dead (or in the process of dying) and so the parent could immediately call wait(2) or waitpid(2). Whether the SIGCHLD signal arrives before the NAL_CONNECTION_io() call or not is not too important, at worst it might prematurely interrupt NAL_SELECTOR_select() (causing it to return zero) so that a redundant loop of the state-machine runs before the next select operation will notice the disconnection. If you already need IPC between the parent and child for exchange of data anyway, this mechanism could be useful in avoiding global variables, signal handlers, and the associated difficulties.

SEE ALSO

NAL_LISTENER_new(2) - Functions for the NAL_LISTENER type.

NAL_SELECTOR_new(2) - Functions for the NAL_SELECTOR type.

NAL_BUFFER_new(2) - Functions for the NAL_BUFFER type.

distcache(8) - Overview of the distcache architecture.

http://www.distcache.org/ - Distcache home page.

AUTHOR

Home Page: http://www.distcache.org