root/trunk/src/io/bmi/bmi_tcp/bmi-tcp.c @ 7819

/*
 * (C) 2001 Clemson University and The University of Chicago
 *
 * See COPYING in top-level directory.
 */

/* TCP/IP implementation of a BMI method */

#include <errno.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/poll.h>
#include <netinet/tcp.h>
#include <assert.h>
#include <sys/uio.h>
#include <time.h>
#include <sys/time.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include "pint-mem.h"

#include "pvfs2-config.h"
#ifdef HAVE_NETDB_H
#include <netdb.h>
#endif

#include "bmi-method-support.h"
#include "bmi-method-callback.h"
#include "bmi-tcp-addressing.h"
#ifdef __PVFS2_USE_EPOLL__
#include "socket-collection-epoll.h"
#else
#include "socket-collection.h"
#endif
#include "op-list.h"
#include "gossip.h"
#include "sockio.h"
#include "bmi-byteswap.h"
#include "id-generator.h"
#include "pint-event.h"
#include "pvfs2-debug.h"
#ifdef USE_TRUSTED
#include "server-config.h"
#include "bmi-tcp-addressing.h"
#endif
#include "gen-locks.h"
#include "pint-hint.h"
#include "pint-event.h"

static gen_mutex_t interface_mutex = GEN_MUTEX_INITIALIZER;
static gen_cond_t interface_cond = GEN_COND_INITIALIZER;
static int sc_test_busy = 0;

/* function prototypes */
int BMI_tcp_initialize(bmi_method_addr_p listen_addr,
                       int method_id,
                       int init_flags);
int BMI_tcp_finalize(void);
int BMI_tcp_set_info(int option,
                     void *inout_parameter);
int BMI_tcp_get_info(int option,
                     void *inout_parameter);
void *BMI_tcp_memalloc(bmi_size_t size,
                       enum bmi_op_type send_recv);
int BMI_tcp_memfree(void *buffer,
                    bmi_size_t size,
                    enum bmi_op_type send_recv);
int BMI_tcp_unexpected_free(void *buffer);
int BMI_tcp_post_send(bmi_op_id_t * id,
                      bmi_method_addr_p dest,
                      const void *buffer,
                      bmi_size_t size,
                      enum bmi_buffer_type buffer_type,
                      bmi_msg_tag_t tag,
                      void *user_ptr,
                      bmi_context_id context_id,
                      PVFS_hint hints);
int BMI_tcp_post_sendunexpected(bmi_op_id_t * id,
                                bmi_method_addr_p dest,
                                const void *buffer,
                                bmi_size_t size,
                                enum bmi_buffer_type buffer_type,
                                bmi_msg_tag_t tag,
                                void *user_ptr,
                                bmi_context_id context_id,
                                PVFS_hint hints);
int BMI_tcp_post_recv(bmi_op_id_t * id,
                      bmi_method_addr_p src,
                      void *buffer,
                      bmi_size_t expected_size,
                      bmi_size_t * actual_size,
                      enum bmi_buffer_type buffer_type,
                      bmi_msg_tag_t tag,
                      void *user_ptr,
                      bmi_context_id context_id,
                      PVFS_hint hints);
int BMI_tcp_test(bmi_op_id_t id,
                 int *outcount,
                 bmi_error_code_t * error_code,
                 bmi_size_t * actual_size,
                 void **user_ptr,
                 int max_idle_time_ms,
                 bmi_context_id context_id);
int BMI_tcp_testsome(int incount,
                     bmi_op_id_t * id_array,
                     int *outcount,
                     int *index_array,
                     bmi_error_code_t * error_code_array,
                     bmi_size_t * actual_size_array,
                     void **user_ptr_array,
                     int max_idle_time_ms,
                     bmi_context_id context_id);
int BMI_tcp_testunexpected(int incount,
                           int *outcount,
                           struct bmi_method_unexpected_info *info,
                           int max_idle_time_ms);
int BMI_tcp_testcontext(int incount,
                     bmi_op_id_t * out_id_array,
                     int *outcount,
                     bmi_error_code_t * error_code_array,
                     bmi_size_t * actual_size_array,
                     void **user_ptr_array,
                     int max_idle_time_ms,
                     bmi_context_id context_id);
bmi_method_addr_p BMI_tcp_method_addr_lookup(const char *id_string);
const char* BMI_tcp_addr_rev_lookup_unexpected(bmi_method_addr_p map);
int BMI_tcp_query_addr_range(bmi_method_addr_p, const char *, int);
int BMI_tcp_post_send_list(bmi_op_id_t * id,
                           bmi_method_addr_p dest,
                           const void *const *buffer_list,
                           const bmi_size_t *size_list,
                           int list_count,
                           bmi_size_t total_size,
                           enum bmi_buffer_type buffer_type,
                           bmi_msg_tag_t tag,
                           void *user_ptr,
                           bmi_context_id context_id,
                           PVFS_hint hints);
int BMI_tcp_post_recv_list(bmi_op_id_t * id,
                           bmi_method_addr_p src,
                           void *const *buffer_list,
                           const bmi_size_t *size_list,
                           int list_count,
                           bmi_size_t total_expected_size,
                           bmi_size_t * total_actual_size,
                           enum bmi_buffer_type buffer_type,
                           bmi_msg_tag_t tag,
                           void *user_ptr,
                           bmi_context_id context_id,
                           PVFS_hint hints);
int BMI_tcp_post_sendunexpected_list(bmi_op_id_t * id,
                                     bmi_method_addr_p dest,
                                     const void *const *buffer_list,
                                     const bmi_size_t *size_list,
                                     int list_count,
                                     bmi_size_t total_size,
                                     enum bmi_buffer_type buffer_type,
                                     bmi_msg_tag_t tag,
                                     void *user_ptr,
                                     bmi_context_id context_id,
                                     PVFS_hint hints);
int BMI_tcp_open_context(bmi_context_id context_id);
void BMI_tcp_close_context(bmi_context_id context_id);
int BMI_tcp_cancel(bmi_op_id_t id, bmi_context_id context_id);

char BMI_tcp_method_name[] = "bmi_tcp";

/* size of encoded message header */
#define TCP_ENC_HDR_SIZE 24

/* structure internal to tcp for use as a message header */
struct tcp_msg_header
{
    uint32_t magic_nr;          /* magic number */
    uint32_t mode;              /* eager, rendezvous, etc. */
    bmi_msg_tag_t tag;          /* user specified message tag */
    bmi_size_t size;            /* length of trailing message */
    char enc_hdr[TCP_ENC_HDR_SIZE];  /* encoded version of header info */
};

#define BMI_TCP_ENC_HDR(hdr)                                            \
    do {                                                                \
        *((uint32_t*)&((hdr).enc_hdr[0])) = htobmi32((hdr).magic_nr);   \
        *((uint32_t*)&((hdr).enc_hdr[4])) = htobmi32((hdr).mode);       \
        *((uint64_t*)&((hdr).enc_hdr[8])) = htobmi64((hdr).tag);        \
        *((uint64_t*)&((hdr).enc_hdr[16])) = htobmi64((hdr).size);      \
    } while(0)                                              

#define BMI_TCP_DEC_HDR(hdr)                                            \
    do {                                                                \
        (hdr).magic_nr = bmitoh32(*((uint32_t*)&((hdr).enc_hdr[0])));   \
        (hdr).mode = bmitoh32(*((uint32_t*)&((hdr).enc_hdr[4])));       \
        (hdr).tag = bmitoh64(*((uint64_t*)&((hdr).enc_hdr[8])));        \
        (hdr).size = bmitoh64(*((uint64_t*)&((hdr).enc_hdr[16])));      \
    } while(0)                                              

/* enumerate states that we care about */
enum bmi_tcp_state
{
    BMI_TCP_INPROGRESS,
    BMI_TCP_BUFFERING,
    BMI_TCP_COMPLETE
};

/* tcp private portion of operation structure */
struct tcp_op
{
    struct tcp_msg_header env;  /* envelope for this message */
    enum bmi_tcp_state tcp_op_state;
    /* these two fields are used as place holders for the buffer
     * list and size list when we really don't have lists (regular
     * BMI_send or BMI_recv operations); it allows us to use
     * generic code to handle both cases 
     */
    void *buffer_list_stub;
    bmi_size_t size_list_stub;
};

/* static io vector for use with readv and writev; we can only use
 * this because BMI serializes module calls
 */
#define BMI_TCP_IOV_COUNT 10
static struct iovec stat_io_vector[BMI_TCP_IOV_COUNT+1];

/* internal utility functions */
static int tcp_server_init(void);
static void dealloc_tcp_method_addr(bmi_method_addr_p map);
static int tcp_sock_init(bmi_method_addr_p my_method_addr);
static int enqueue_operation(op_list_p target_list,
                             enum bmi_op_type send_recv,
                             bmi_method_addr_p map,
                             void *const *buffer_list,
                             const bmi_size_t *size_list,
                             int list_count,
                             bmi_size_t amt_complete,
                             bmi_size_t env_amt_complete,
                             bmi_op_id_t * id,
                             int tcp_op_state,
                             struct tcp_msg_header header,
                             void *user_ptr,
                             bmi_size_t actual_size,
                             bmi_size_t expected_size,
                             bmi_context_id context_id,
                             int32_t event_id);
static int tcp_cleanse_addr(bmi_method_addr_p map, int error_code);
static int tcp_shutdown_addr(bmi_method_addr_p map);
static int tcp_do_work(int max_idle_time);
static int tcp_do_work_error(bmi_method_addr_p map);
static int tcp_do_work_recv(bmi_method_addr_p map, int* stall_flag);
static int tcp_do_work_send(bmi_method_addr_p map, int* stall_flag);
static int work_on_recv_op(method_op_p my_method_op,
                           int *stall_flag);
static int work_on_send_op(method_op_p my_method_op,
                           int *blocked_flag, int* stall_flag);
static int tcp_accept_init(int *socket, char** peer);
static method_op_p alloc_tcp_method_op(void);
static void dealloc_tcp_method_op(method_op_p old_op);
static int handle_new_connection(bmi_method_addr_p map);
static int tcp_post_send_generic(bmi_op_id_t * id,
                                 bmi_method_addr_p dest,
                                 const void *const *buffer_list,
                                 const bmi_size_t *size_list,
                                 int list_count,
                                 enum bmi_buffer_type buffer_type,
                                 struct tcp_msg_header my_header,
                                 void *user_ptr,
                                 bmi_context_id context_id,
                                 PVFS_hint hints);
static int tcp_post_recv_generic(bmi_op_id_t * id,
                                 bmi_method_addr_p src,
                                 void *const *buffer_list,
                                 const bmi_size_t *size_list,
                                 int list_count,
                                 bmi_size_t expected_size,
                                 bmi_size_t * actual_size,
                                 enum bmi_buffer_type buffer_type,
                                 bmi_msg_tag_t tag,
                                 void *user_ptr,
                                 bmi_context_id context_id,
                                 PVFS_hint hints);
static int payload_progress(int s, void *const *buffer_list, const bmi_size_t* 
    size_list, int list_count, bmi_size_t total_size, int* list_index, 
    bmi_size_t* current_index_complete, enum bmi_op_type send_recv, 
    char* enc_hdr, bmi_size_t* env_amt_complete);

#if defined(USE_TRUSTED) && defined(__PVFS2_CLIENT__)
static int tcp_enable_trusted(struct tcp_addr *tcp_addr_data);
#endif
#if defined(USE_TRUSTED) && defined(__PVFS2_SERVER__)
static int tcp_allow_trusted(struct sockaddr_in *peer_sockaddr);
#endif

static void bmi_set_sock_buffers(int socket);

/* exported method interface */
const struct bmi_method_ops bmi_tcp_ops = {
    .method_name = BMI_tcp_method_name,
    .initialize = BMI_tcp_initialize,
    .finalize = BMI_tcp_finalize,
    .set_info = BMI_tcp_set_info,
    .get_info = BMI_tcp_get_info,
    .memalloc = BMI_tcp_memalloc,
    .memfree  = BMI_tcp_memfree,
    .unexpected_free = BMI_tcp_unexpected_free,
    .post_send = BMI_tcp_post_send,
    .post_sendunexpected = BMI_tcp_post_sendunexpected,
    .post_recv = BMI_tcp_post_recv,
    .test = BMI_tcp_test,
    .testsome = BMI_tcp_testsome,
    .testcontext = BMI_tcp_testcontext,
    .testunexpected = BMI_tcp_testunexpected,
    .method_addr_lookup = BMI_tcp_method_addr_lookup,
    .post_send_list = BMI_tcp_post_send_list,
    .post_recv_list = BMI_tcp_post_recv_list,
    .post_sendunexpected_list = BMI_tcp_post_sendunexpected_list,
    .open_context = BMI_tcp_open_context,
    .close_context = BMI_tcp_close_context,
    .cancel = BMI_tcp_cancel,
    .rev_lookup_unexpected = BMI_tcp_addr_rev_lookup_unexpected,
    .query_addr_range = BMI_tcp_query_addr_range,
};

/* module parameters */
static struct
{
    int method_flags;
    int method_id;
    bmi_method_addr_p listen_addr;
} tcp_method_params;

#if defined(USE_TRUSTED) && defined(__PVFS2_SERVER__)
static struct tcp_allowed_connection_s *gtcp_allowed_connection = NULL;
#endif

static int check_unexpected = 1;

/* op_list_array indices */
enum
{
    NUM_INDICES = 5,
    IND_SEND = 0,
    IND_RECV = 1,
    IND_RECV_INFLIGHT = 2,
    IND_RECV_EAGER_DONE_BUFFERING = 3,
    IND_COMPLETE_RECV_UNEXP = 4,        /* MAKE THIS COMES LAST */
};

/* internal operation lists */
static op_list_p op_list_array[6] = { NULL, NULL, NULL, NULL,
    NULL, NULL
};

/* internal completion queues */
static op_list_p completion_array[BMI_MAX_CONTEXTS] = { NULL };

/* internal socket collection */
static socket_collection_p tcp_socket_collection_p = NULL;

/* tunable parameters */
enum
{
    /* amount of pending connections we'll allow */
    TCP_BACKLOG = 256,
    /* amount of work to be done during a test.  This roughly 
     * translates into the number of sockets that we will perform
     * nonblocking operations on during one function call.
     */
    TCP_WORK_METRIC = 128
};

/* TCP message modes */
enum
{
    TCP_MODE_IMMED = 1,         /* not used for TCP/IP */
    TCP_MODE_UNEXP = 2,
    TCP_MODE_EAGER = 4,
    TCP_MODE_REND = 8
};

/* Allowable sizes for each mode */
enum
{
    TCP_MODE_EAGER_LIMIT = 16384,       /* 16K */
    TCP_MODE_REND_LIMIT = 16777216      /* 16M */
};

/* toggles cancel mode; for bmi_tcp this will result in socket being closed
 * in all cancellation cases
 */
static int forceful_cancel_mode = 0;

/*
  Socket buffer sizes, currently these default values will be used 
  for the clients... (TODO)
 */
static int tcp_buffer_size_receive = 0;
static int tcp_buffer_size_send = 0;

static PINT_event_type bmi_tcp_send_event_id;
static PINT_event_type bmi_tcp_recv_event_id;

static PINT_event_group bmi_tcp_event_group;
static pid_t bmi_tcp_pid;

/*************************************************************************
 * Visible Interface 
 */

/* BMI_tcp_initialize()
 *
 * Initializes the tcp method.  Must be called before any other tcp
 * method functions.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_initialize(bmi_method_addr_p listen_addr,
                       int method_id,
                       int init_flags)
{

    int ret = -1;
    int tmp_errno = bmi_tcp_errno_to_pvfs(-ENOSYS);
    struct tcp_addr *tcp_addr_data = NULL;
    int i = 0;

    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "Initializing TCP/IP module.\n");

    /* check args */
    if ((init_flags & BMI_INIT_SERVER) && !listen_addr)
    {
        gossip_lerr("Error: bad parameters given to TCP/IP module.\n");
        return (bmi_tcp_errno_to_pvfs(-EINVAL));
    }

    gen_mutex_lock(&interface_mutex);

    /* zero out our parameter structure and fill it in */
    memset(&tcp_method_params, 0, sizeof(tcp_method_params));
    tcp_method_params.method_id = method_id;
    tcp_method_params.method_flags = init_flags;

    if (init_flags & BMI_INIT_SERVER)
    {
        /* hang on to our local listening address if needed */
        tcp_method_params.listen_addr = listen_addr;
        /* and initialize server functions */
        ret = tcp_server_init();
        if (ret < 0)
        {
            tmp_errno = bmi_tcp_errno_to_pvfs(ret);
            gossip_err("Error: tcp_server_init() failure.\n");
            goto initialize_failure;
        }
    }

    /* set up the operation lists */
    for (i = 0; i < NUM_INDICES; i++)
    {
        op_list_array[i] = op_list_new();
        if (!op_list_array[i])
        {
            tmp_errno = bmi_tcp_errno_to_pvfs(-ENOMEM);
            goto initialize_failure;
        }
    }

    /* set up the socket collection */
    if (tcp_method_params.method_flags & BMI_INIT_SERVER)
    {
        tcp_addr_data = tcp_method_params.listen_addr->method_data;
        tcp_socket_collection_p = BMI_socket_collection_init(tcp_addr_data->socket);
    }
    else
    {
        tcp_socket_collection_p = BMI_socket_collection_init(-1);
    }

    if (!tcp_socket_collection_p)
    {
        tmp_errno = bmi_tcp_errno_to_pvfs(-ENOMEM);
        goto initialize_failure;
    }

    bmi_tcp_pid = getpid();
    PINT_event_define_group("bmi_tcp", &bmi_tcp_event_group);

    /* Define the send event:
     *   START: (client_id, request_id, rank, handle, op_id, send_size)
     *   STOP: (size_sent)
     */
    PINT_event_define_event(
        &bmi_tcp_event_group,
#ifdef __PVFS2_SERVER__
        "bmi_server_send",
#else
        "bmi_client_send",
#endif
        "%d%d%d%llu%d%d",
        "%d", &bmi_tcp_send_event_id);

    /* Define the recv event:
     *   START: (client_id, request_id, rank, handle, op_id, recv_size)
     *   STOP: (size_received)
     */
    PINT_event_define_event(
        &bmi_tcp_event_group,
#ifdef __PVFS2_SERVER__
        "bmi_server_recv",
#else
        "bmi_client_recv",
#endif
        "%d%d%d%llu%d%d",
        "%d", &bmi_tcp_recv_event_id);

    gen_mutex_unlock(&interface_mutex);
    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP,
                  "TCP/IP module successfully initialized.\n");
    return (0);

  initialize_failure:

    /* cleanup data structures and bail out */
    for (i = 0; i < NUM_INDICES; i++)
    {
        if (op_list_array[i])
        {
            op_list_cleanup(op_list_array[i]);
        }
    }
    if (tcp_socket_collection_p)
    {
        BMI_socket_collection_finalize(tcp_socket_collection_p);
    }
    gen_mutex_unlock(&interface_mutex);
    return (tmp_errno);
}


/* BMI_tcp_finalize()
 * 
 * Shuts down the tcp method.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_finalize(void)
{
    int i = 0;

    gen_mutex_lock(&interface_mutex);

    /* shut down our listen addr, if we have one */
    if ((tcp_method_params.method_flags & BMI_INIT_SERVER)
        && tcp_method_params.listen_addr)
    {
        dealloc_tcp_method_addr(tcp_method_params.listen_addr);
    }

    /* note that this forcefully shuts down operations */
    for (i = 0; i < NUM_INDICES; i++)
    {
        if (op_list_array[i])
        {
            op_list_cleanup(op_list_array[i]);
            op_list_array[i] = NULL;
        }
    }

    /* get rid of socket collection */
    if (tcp_socket_collection_p)
    {
        BMI_socket_collection_finalize(tcp_socket_collection_p);
        tcp_socket_collection_p = NULL;
    }

    /* NOTE: we are trusting the calling BMI layer to deallocate 
     * all of the method addresses (this will close any open sockets)
     */
    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "TCP/IP module finalized.\n");
    gen_mutex_unlock(&interface_mutex);
    return (0);
}


/*
 * BMI_tcp_method_addr_lookup()
 *
 * resolves the string representation of an address into a method
 * address structure.  
 *
 * returns a pointer to method_addr on success, NULL on failure
 */
bmi_method_addr_p BMI_tcp_method_addr_lookup(const char *id_string)
{
    char *tcp_string = NULL;
    char *delim = NULL;
    char *hostname = NULL;
    bmi_method_addr_p new_addr = NULL;
    struct tcp_addr *tcp_addr_data = NULL;
    int ret = -1;

    tcp_string = string_key("tcp", id_string);
    if (!tcp_string)
    {
        /* the string doesn't even have our info */
        return (NULL);
    }

    /* start breaking up the method information */
    /* for normal tcp, it is simply hostname:port */
    if ((delim = index(tcp_string, ':')) == NULL)
    {
        gossip_lerr("Error: malformed tcp address.\n");
        free(tcp_string);
        return (NULL);
    }

    /* looks ok, so let's build the method addr structure */
    new_addr = alloc_tcp_method_addr();
    if (!new_addr)
    {
        free(tcp_string);
        return (NULL);
    }
    tcp_addr_data = new_addr->method_data;

    ret = sscanf((delim + 1), "%d", &(tcp_addr_data->port));
    if (ret != 1)
    {
        gossip_lerr("Error: malformed tcp address.\n");
        dealloc_tcp_method_addr(new_addr);
        free(tcp_string);
        return (NULL);
    }

    hostname = (char *) malloc((delim - tcp_string + 1));
    if (!hostname)
    {
        dealloc_tcp_method_addr(new_addr);
        free(tcp_string);
        return (NULL);
    }
    strncpy(hostname, tcp_string, (delim - tcp_string));
    hostname[delim - tcp_string] = '\0';

    tcp_addr_data->hostname = hostname;

    free(tcp_string);
    return (new_addr);
}


/* BMI_tcp_memalloc()
 * 
 * Allocates memory that can be used in native mode by tcp.
 *
 * returns 0 on success, -errno on failure
 */
void *BMI_tcp_memalloc(bmi_size_t size,
                       enum bmi_op_type send_recv)
{
    /* we really don't care what flags the caller uses, TCP/IP has no
     * preferences about how the memory should be configured.
     */

/*    return (calloc(1,(size_t) size)); */
    return PINT_mem_aligned_alloc(size, 4096);
}


/* BMI_tcp_memfree()
 * 
 * Frees memory that was allocated with BMI_tcp_memalloc()
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_memfree(void *buffer,
                    bmi_size_t size,
                    enum bmi_op_type send_recv)
{
    PINT_mem_aligned_free(buffer);
    return (0);
}

/* BMI_tcp_unexpected_free()
 * 
 * Frees memory that was returned from BMI_tcp_test_unexpected()
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_unexpected_free(void *buffer)
{
    if (buffer)
    {
        free(buffer);
    }
    return (0);
}

#ifdef USE_TRUSTED

static struct tcp_allowed_connection_s *
alloc_trusted_connection_info(int network_count)
{
    struct tcp_allowed_connection_s *tcp_allowed_connection_info = NULL;

    tcp_allowed_connection_info = (struct tcp_allowed_connection_s *)
            calloc(1, sizeof(struct tcp_allowed_connection_s));
    if (tcp_allowed_connection_info)
    {
        tcp_allowed_connection_info->network =
            (struct in_addr *) calloc(network_count, sizeof(struct in_addr));
        if (tcp_allowed_connection_info->network == NULL)
        {
            free(tcp_allowed_connection_info);
            tcp_allowed_connection_info = NULL;
        }
        else
        {
            tcp_allowed_connection_info->netmask =
                (struct in_addr *) calloc(network_count, sizeof(struct in_addr));
            if (tcp_allowed_connection_info->netmask == NULL)
            {
                free(tcp_allowed_connection_info->network);
                free(tcp_allowed_connection_info);
                tcp_allowed_connection_info = NULL;
            }
            else {
                tcp_allowed_connection_info->network_count = network_count;
            }
        }
    }
    return tcp_allowed_connection_info;
}

static void 
dealloc_trusted_connection_info(void* ptcp_allowed_connection_info)
{
    struct tcp_allowed_connection_s *tcp_allowed_connection_info =
        (struct tcp_allowed_connection_s *) ptcp_allowed_connection_info;
    if (tcp_allowed_connection_info)
    {
        free(tcp_allowed_connection_info->network);
        tcp_allowed_connection_info->network = NULL;
        free(tcp_allowed_connection_info->netmask);
        tcp_allowed_connection_info->netmask = NULL;
        free(tcp_allowed_connection_info);
    }
    return;
}

#endif

/*
 * This function will convert a mask_bits value to an in_addr
 * representation. i.e for example if
 * mask_bits was 24 then it would be 255.255.255.0
 * if mask_bits was 22 then it would be 255.255.252.0
 * etc
 */
static void convert_mask(int mask_bits, struct in_addr *mask)
{
   uint32_t addr = -1;
   addr = addr & ~~(-1 << (mask_bits ? (32 - mask_bits) : 32));
   mask->s_addr = htonl(addr);
   return;
}

/* BMI_tcp_set_info()
 * 
 * Pass in optional parameters.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_set_info(int option,
                     void *inout_parameter)
{
    int ret = -1;
    bmi_method_addr_p tmp_addr = NULL;

    gen_mutex_lock(&interface_mutex);

    switch (option)
    {
    case BMI_TCP_BUFFER_SEND_SIZE:
       tcp_buffer_size_send = *((int *)inout_parameter);
       ret = 0;
#ifdef __PVFS2_SERVER__
       /* Set the default socket buffer sizes for the server socket */
       bmi_set_sock_buffers(
           ((struct tcp_addr *)
            tcp_method_params.listen_addr->method_data)->socket);
#endif
       break;
    case BMI_TCP_BUFFER_RECEIVE_SIZE:
       tcp_buffer_size_receive = *((int *)inout_parameter);
       ret = 0;
#ifdef __PVFS2_SERVER__
       /* Set the default socket buffer sizes for the server socket */
       bmi_set_sock_buffers(
           ((struct tcp_addr *)
            tcp_method_params.listen_addr->method_data)->socket);
#endif
       break;
    case BMI_TCP_CLOSE_SOCKET: 
        /* this should no longer make it to the bmi_tcp method; see bmi.c */
        ret = 0;
        break;
    case BMI_FORCEFUL_CANCEL_MODE:
        forceful_cancel_mode = 1;
        ret = 0;
        break;
    case BMI_DROP_ADDR:
        if (inout_parameter == NULL)
        {
            ret = bmi_tcp_errno_to_pvfs(-EINVAL);
        }
        else
        {
            tmp_addr = (bmi_method_addr_p) inout_parameter;
            /* take it out of the socket collection */
            tcp_forget_addr(tmp_addr, 1, 0);
            ret = 0;
        }
        break;
#ifdef USE_TRUSTED
    case BMI_TRUSTED_CONNECTION:
    {
        struct tcp_allowed_connection_s *tcp_allowed_connection = NULL;
        if (inout_parameter == NULL)
        {
            ret = bmi_tcp_errno_to_pvfs(-EINVAL);
            break;
        }
        else 
        {
            int    bmi_networks_count = 0;
            char **bmi_networks = NULL;
            int   *bmi_netmasks = NULL;
            struct server_configuration_s *svc_config = NULL;

            svc_config = (struct server_configuration_s *) inout_parameter;
            tcp_allowed_connection = alloc_trusted_connection_info(svc_config->allowed_networks_count);
            if (tcp_allowed_connection == NULL)
            {
                ret = bmi_tcp_errno_to_pvfs(-ENOMEM);
                break;
            }
#ifdef      __PVFS2_SERVER__
            gtcp_allowed_connection = tcp_allowed_connection;
#endif
            /* Stash this in the server_configuration_s structure. freed later on */
            svc_config->security = tcp_allowed_connection;
            svc_config->security_dtor = &dealloc_trusted_connection_info;
            ret = 0;
            /* Fill up the list of allowed ports */
            PINT_config_get_allowed_ports(svc_config, 
                    &tcp_allowed_connection->port_enforce, 
                    tcp_allowed_connection->ports);

            /* if it was enabled, make sure that we know how to deal with it */
            if (tcp_allowed_connection->port_enforce == 1)
            {
                /* illegal ports */
                if (tcp_allowed_connection->ports[0] > 65535 
                        || tcp_allowed_connection->ports[1] > 65535
                        || tcp_allowed_connection->ports[1] < tcp_allowed_connection->ports[0])
                {
                    gossip_lerr("Error: illegal trusted port values\n");
                    ret = bmi_tcp_errno_to_pvfs(-EINVAL);
                    /* don't enforce anything! */
                    tcp_allowed_connection->port_enforce = 0;
                }
            }
            ret = 0;
            /* Retrieve the list of BMI network addresses and masks  */
            PINT_config_get_allowed_networks(svc_config,
                    &tcp_allowed_connection->network_enforce,
                    &bmi_networks_count,
                    &bmi_networks,
                    &bmi_netmasks);

            /* if it was enabled, make sure that we know how to deal with it */
            if (tcp_allowed_connection->network_enforce == 1)
            {
                int i;

                for (i = 0; i < bmi_networks_count; i++)
                {
                    char *tcp_string = NULL;
                    /* Convert the network string into an in_addr_t structure */
                    tcp_string = string_key("tcp", bmi_networks[i]);
                    if (!tcp_string)
                    {
                        /* the string doesn't even have our info */
                        gossip_lerr("Error: malformed tcp network address\n");
                        ret = bmi_tcp_errno_to_pvfs(-EINVAL);
                    }
                    else {
                        /* convert this into an in_addr_t */
                        inet_aton(tcp_string, &tcp_allowed_connection->network[i]);
                        free(tcp_string);
                    }
                    convert_mask(bmi_netmasks[i], &tcp_allowed_connection->netmask[i]);
                }
                /* don't enforce anything if there were any errors */
                if (ret != 0)
                {
                    tcp_allowed_connection->network_enforce = 0;
                }
            }
        }
        break;
    }
#endif
    case BMI_TCP_CHECK_UNEXPECTED:
    {
        check_unexpected = *(int *)inout_parameter;
        ret = 0;
        break;
    }

    default:
        gossip_ldebug(GOSSIP_BMI_DEBUG_TCP,
                      "TCP hint %d not implemented.\n", option);
        ret = 0;
        break;
    }

    gen_mutex_unlock(&interface_mutex);
    return (ret);
}

/* BMI_tcp_get_info()
 * 
 * Query for optional parameters.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_get_info(int option,
                     void *inout_parameter)
{
    struct method_drop_addr_query* query;
    struct tcp_addr* tcp_addr_data;
    int ret = 0;

    gen_mutex_lock(&interface_mutex);

    switch (option)
    {
    case BMI_CHECK_MAXSIZE:
        *((int *) inout_parameter) = TCP_MODE_REND_LIMIT;
        ret = 0;
        break;
    case BMI_DROP_ADDR_QUERY:
        query = (struct method_drop_addr_query*)inout_parameter;
        tcp_addr_data=query->addr->method_data;
        /* only suggest that we discard the address if we have experienced
         * an error and there is no way to reconnect
         */
        if(tcp_addr_data->addr_error != 0 &&
           tcp_addr_data->dont_reconnect == 1)
        {
            query->response = 1;
        }
        else
        {
            query->response = 0;
        }
        ret = 0;
        break;
    case BMI_GET_UNEXP_SIZE:
        *((int *) inout_parameter) = TCP_MODE_EAGER_LIMIT;
        ret = 0;
        break;

    default:
        gossip_ldebug(GOSSIP_BMI_DEBUG_TCP,
                      "TCP hint %d not implemented.\n", option);
        ret = -ENOSYS;
        break;
    }

    gen_mutex_unlock(&interface_mutex);
    return (ret < 0) ? bmi_tcp_errno_to_pvfs(ret) : ret;
}


/* BMI_tcp_post_send()
 * 
 * Submits send operations.
 *
 * returns 0 on success that requires later poll, returns 1 on instant
 * completion, -errno on failure
 */
int BMI_tcp_post_send(bmi_op_id_t * id,
                      bmi_method_addr_p dest,
                      const void *buffer,
                      bmi_size_t size,
                      enum bmi_buffer_type buffer_type,
                      bmi_msg_tag_t tag,
                      void *user_ptr,
                      bmi_context_id context_id,
                      PVFS_hint hints)
{
    struct tcp_msg_header my_header;
    int ret = -1;

    /* clear the id field for safety */
    *id = 0;

    /* fill in the TCP-specific message header */
    if (size > TCP_MODE_REND_LIMIT)
    {
        return (bmi_tcp_errno_to_pvfs(-EMSGSIZE));
    }

    if (size <= TCP_MODE_EAGER_LIMIT)
    {
        my_header.mode = TCP_MODE_EAGER;
    }
    else
    {
        my_header.mode = TCP_MODE_REND;
    }
    my_header.tag = tag;
    my_header.size = size;
    my_header.magic_nr = BMI_MAGIC_NR;

    gen_mutex_lock(&interface_mutex);

    ret = tcp_post_send_generic(id, dest, &buffer,
                                &size, 1, buffer_type, my_header,
                                user_ptr, context_id, hints);

    gen_mutex_unlock(&interface_mutex);
    return(ret);
}


/* BMI_tcp_post_sendunexpected()
 * 
 * Submits unexpected send operations.
 *
 * returns 0 on success that requires later poll, returns 1 on instant
 * completion, -errno on failure
 */
int BMI_tcp_post_sendunexpected(bmi_op_id_t * id,
                                bmi_method_addr_p dest,
                                const void *buffer,
                                bmi_size_t size,
                                enum bmi_buffer_type buffer_type,
                                bmi_msg_tag_t tag,
                                void *user_ptr,
                                bmi_context_id context_id,
                                PVFS_hint hints)
{
    struct tcp_msg_header my_header;
    int ret = -1;

    /* clear the id field for safety */
    *id = 0;

    if (size > TCP_MODE_EAGER_LIMIT)
    {
        return (bmi_tcp_errno_to_pvfs(-EMSGSIZE));
    }

    my_header.mode = TCP_MODE_UNEXP;
    my_header.tag = tag;
    my_header.size = size;
    my_header.magic_nr = BMI_MAGIC_NR;

    gen_mutex_lock(&interface_mutex);

    ret = tcp_post_send_generic(id, dest, &buffer,
                                &size, 1, buffer_type, my_header,
                                user_ptr, context_id, hints);
    gen_mutex_unlock(&interface_mutex);
    return(ret);
}



/* BMI_tcp_post_recv()
 * 
 * Submits recv operations.
 *
 * returns 0 on success that requires later poll, returns 1 on instant
 * completion, -errno on failure
 */
int BMI_tcp_post_recv(bmi_op_id_t * id,
                      bmi_method_addr_p src,
                      void *buffer,
                      bmi_size_t expected_size,
                      bmi_size_t * actual_size,
                      enum bmi_buffer_type buffer_type,
                      bmi_msg_tag_t tag,
                      void *user_ptr,
                      bmi_context_id context_id,
                      PVFS_hint hints)
{
    int ret = -1;

    /* A few things could happen here:
     * a) rendez. recv with sender not ready yet
     * b) rendez. recv with sender waiting
     * c) eager recv, data not available yet
     * d) eager recv, some/all data already here
     * e) rendez. recv with sender in eager mode
     *
     * b or d could lead to completion without polling.
     * we don't look for unexpected messages here.
     */

    if (expected_size > TCP_MODE_REND_LIMIT)
    {
        return (bmi_tcp_errno_to_pvfs(-EINVAL));
    }
    gen_mutex_lock(&interface_mutex);

    ret = tcp_post_recv_generic(id, src, &buffer, &expected_size,
                                1, expected_size, actual_size,
                                buffer_type, tag,
                                user_ptr, context_id, hints);

    gen_mutex_unlock(&interface_mutex);
    return (ret);
}


/* BMI_tcp_test()
 * 
 * Checks to see if a particular message has completed.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_test(bmi_op_id_t id,
                 int *outcount,
                 bmi_error_code_t * error_code,
                 bmi_size_t * actual_size,
                 void **user_ptr,
                 int max_idle_time,
                 bmi_context_id context_id)
{
    int ret = -1;
    method_op_p query_op = (method_op_p)id_gen_fast_lookup(id);

    assert(query_op != NULL);

    gen_mutex_lock(&interface_mutex);

    /* do some ``real work'' here */
    ret = tcp_do_work(max_idle_time);
    if (ret < 0)
    {
        gen_mutex_unlock(&interface_mutex);
        return (ret);
    }

    if(((struct tcp_op*)(query_op->method_data))->tcp_op_state ==
        BMI_TCP_COMPLETE)
    {
        assert(query_op->context_id == context_id);
        op_list_remove(query_op);
        if (user_ptr != NULL)
        {
            (*user_ptr) = query_op->user_ptr;
        }
        (*error_code) = query_op->error_code;
        (*actual_size) = query_op->actual_size;
        PINT_EVENT_END(
            (query_op->send_recv == BMI_SEND ?
             bmi_tcp_send_event_id : bmi_tcp_recv_event_id), bmi_tcp_pid, NULL,
             query_op->event_id, id, *actual_size);

        dealloc_tcp_method_op(query_op);
        (*outcount)++;
    }

    gen_mutex_unlock(&interface_mutex);
    return (0);
}

/* BMI_tcp_testsome()
 * 
 * Checks to see if any messages from the specified list have completed.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_testsome(int incount,
                     bmi_op_id_t * id_array,
                     int *outcount,
                     int *index_array,
                     bmi_error_code_t * error_code_array,
                     bmi_size_t * actual_size_array,
                     void **user_ptr_array,
                     int max_idle_time,
                     bmi_context_id context_id)
{
    int ret = -1;
    method_op_p query_op = NULL;
    int i;

    gen_mutex_lock(&interface_mutex);

    /* do some ``real work'' here */
    ret = tcp_do_work(max_idle_time);
    if (ret < 0)
    {
        gen_mutex_unlock(&interface_mutex);
        return (ret);
    }

    for(i=0; i<incount; i++)
    {
        if(id_array[i])
        {
            /* NOTE: this depends on the user passing in valid id's;
             * otherwise we segfault.  
             */
            query_op = (method_op_p)id_gen_fast_lookup(id_array[i]);
            if(((struct tcp_op*)(query_op->method_data))->tcp_op_state ==
               BMI_TCP_COMPLETE)
            {
                assert(query_op->context_id == context_id);
                /* this one's done; pop it out */
                op_list_remove(query_op);
                error_code_array[*outcount] = query_op->error_code;
                actual_size_array[*outcount] = query_op->actual_size;
                index_array[*outcount] = i;
                if (user_ptr_array != NULL)
                {
                    user_ptr_array[*outcount] = query_op->user_ptr;
                }
                PINT_EVENT_END(
                    (query_op->send_recv == BMI_SEND ?
                     bmi_tcp_send_event_id : bmi_tcp_recv_event_id),
                    bmi_tcp_pid, NULL,
                    query_op->event_id, actual_size_array[*outcount]);
                dealloc_tcp_method_op(query_op);
                (*outcount)++;
            }
        }
    }

    gen_mutex_unlock(&interface_mutex);
    return(0);
}


/* BMI_tcp_testunexpected()
 * 
 * Checks to see if any unexpected messages have completed.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_testunexpected(int incount,
                           int *outcount,
                           struct bmi_method_unexpected_info *info,
                           int max_idle_time)
{
    int ret = -1;
    method_op_p query_op = NULL;

    gen_mutex_lock(&interface_mutex);

    if(op_list_empty(op_list_array[IND_COMPLETE_RECV_UNEXP]))
    {
        /* do some ``real work'' here */
        ret = tcp_do_work(max_idle_time);
        if (ret < 0)
        {
            gen_mutex_unlock(&interface_mutex);
            return (ret);
        }
    }

    *outcount = 0;

    /* go through the completed/unexpected list as long as we are finding 
     * stuff and we have room in the info array for it
     */
    while ((*outcount < incount) &&
           (query_op =
            op_list_shownext(op_list_array[IND_COMPLETE_RECV_UNEXP])))
    {
        info[*outcount].error_code = query_op->error_code;
        info[*outcount].addr = query_op->addr;
        info[*outcount].buffer = query_op->buffer;
        info[*outcount].size = query_op->actual_size;
        info[*outcount].tag = query_op->msg_tag;
        op_list_remove(query_op);
        dealloc_tcp_method_op(query_op);
        (*outcount)++;
    }
    gen_mutex_unlock(&interface_mutex);
    return (0);
}


/* BMI_tcp_testcontext()
 * 
 * Checks to see if any messages from the specified context have completed.
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_testcontext(int incount,
                     bmi_op_id_t* out_id_array,
                     int *outcount,
                     bmi_error_code_t * error_code_array,
                     bmi_size_t * actual_size_array,
                     void **user_ptr_array,
                     int max_idle_time,
                     bmi_context_id context_id)
{
    int ret = -1;
    method_op_p query_op = NULL;

    *outcount = 0;

    gen_mutex_lock(&interface_mutex);

    if(op_list_empty(completion_array[context_id]))
    {
        /* if there are unexpected ops ready to go, then short out so
         * that the next testunexpected call can pick it up without
         * delay
         */
        if(check_unexpected &&
           !op_list_empty(op_list_array[IND_COMPLETE_RECV_UNEXP]))
        {
            gen_mutex_unlock(&interface_mutex);
            return(0);
        }

        /* do some ``real work'' here */
        ret = tcp_do_work(max_idle_time);
        if (ret < 0)
        {
            gen_mutex_unlock(&interface_mutex);
            return (ret);
        }
    }

    /* pop as many items off of the completion queue as we can */
    while((*outcount < incount) && 
          (query_op =
           op_list_shownext(completion_array[context_id])))
    {
        assert(query_op);
        assert(query_op->context_id == context_id);

        /* this one's done; pop it out */
        op_list_remove(query_op);
        error_code_array[*outcount] = query_op->error_code;
        actual_size_array[*outcount] = query_op->actual_size;
        out_id_array[*outcount] = query_op->op_id;
        if (user_ptr_array != NULL)
        {
            user_ptr_array[*outcount] = query_op->user_ptr;
        }

        PINT_EVENT_END((query_op->send_recv == BMI_SEND ?
                        bmi_tcp_send_event_id : bmi_tcp_recv_event_id),
                       bmi_tcp_pid, NULL, query_op->event_id,
                       query_op->actual_size);

        dealloc_tcp_method_op(query_op);
        query_op = NULL;
        (*outcount)++;
    }

    gen_mutex_unlock(&interface_mutex);
    return(0);
}



/* BMI_tcp_post_send_list()
 *
 * same as the BMI_tcp_post_send() function, except that it sends
 * from an array of possibly non contiguous buffers
 *
 * returns 0 on success, 1 on immediate successful completion,
 * -errno on failure
 */
int BMI_tcp_post_send_list(bmi_op_id_t * id,
                           bmi_method_addr_p dest,
                           const void *const *buffer_list,
                           const bmi_size_t *size_list,
                           int list_count,
                           bmi_size_t total_size,
                           enum bmi_buffer_type buffer_type,
                           bmi_msg_tag_t tag,
                           void *user_ptr,
                           bmi_context_id context_id,
                           PVFS_hint hints)
{
    struct tcp_msg_header my_header;
    int ret = -1;

    /* clear the id field for safety */
    *id = 0;

    /* fill in the TCP-specific message header */
    if (total_size > TCP_MODE_REND_LIMIT)
    {
        gossip_lerr("Error: BMI message too large!\n");
        return (bmi_tcp_errno_to_pvfs(-EMSGSIZE));
    }

    if (total_size <= TCP_MODE_EAGER_LIMIT)
    {
        my_header.mode = TCP_MODE_EAGER;
    }
    else
    {
        my_header.mode = TCP_MODE_REND;
    }
    my_header.tag = tag;
    my_header.size = total_size;
    my_header.magic_nr = BMI_MAGIC_NR;

    gen_mutex_lock(&interface_mutex);

    ret = tcp_post_send_generic(id, dest, buffer_list,
                                size_list, list_count, buffer_type,
                                my_header, user_ptr, context_id, hints);
    gen_mutex_unlock(&interface_mutex);
    return(ret);
}

/* BMI_tcp_post_recv_list()
 *
 * same as the BMI_tcp_post_recv() function, except that it recvs
 * into an array of possibly non contiguous buffers
 *
 * returns 0 on success, 1 on immediate successful completion,
 * -errno on failure
 */
int BMI_tcp_post_recv_list(bmi_op_id_t * id,
                           bmi_method_addr_p src,
                           void *const *buffer_list,
                           const bmi_size_t *size_list,
                           int list_count,
                           bmi_size_t total_expected_size,
                           bmi_size_t * total_actual_size,
                           enum bmi_buffer_type buffer_type,
                           bmi_msg_tag_t tag,
                           void *user_ptr,
                           bmi_context_id context_id,
                           PVFS_hint hints)
{
    int ret = -1;

    if (total_expected_size > TCP_MODE_REND_LIMIT)
    {
        return (bmi_tcp_errno_to_pvfs(-EINVAL));
    }

    gen_mutex_lock(&interface_mutex);

    ret = tcp_post_recv_generic(id, src, buffer_list, size_list,
                                list_count, total_expected_size,
                                total_actual_size, buffer_type, tag, user_ptr,
                                context_id, hints);

    gen_mutex_unlock(&interface_mutex);
    return (ret);
}


/* BMI_tcp_post_sendunexpected_list()
 *
 * same as the BMI_tcp_post_sendunexpected() function, except that 
 * it sends from an array of possibly non contiguous buffers
 *
 * returns 0 on success, 1 on immediate successful completion,
 * -errno on failure
 */
int BMI_tcp_post_sendunexpected_list(bmi_op_id_t * id,
                                     bmi_method_addr_p dest,
                                     const void *const *buffer_list,
                                     const bmi_size_t *size_list,
                                     int list_count,
                                     bmi_size_t total_size,
                                     enum bmi_buffer_type buffer_type,
                                     bmi_msg_tag_t tag,
                                     void *user_ptr,
                                     bmi_context_id context_id,
                                     PVFS_hint hints)
{
    struct tcp_msg_header my_header;
    int ret = -1;

    /* clear the id field for safety */
    *id = 0;

    if (total_size > TCP_MODE_EAGER_LIMIT)
    {
        return (bmi_tcp_errno_to_pvfs(-EMSGSIZE));
    }

    my_header.mode = TCP_MODE_UNEXP;
    my_header.tag = tag;
    my_header.size = total_size;
    my_header.magic_nr = BMI_MAGIC_NR;

    gen_mutex_lock(&interface_mutex);

    ret = tcp_post_send_generic(id, dest, buffer_list,
                                size_list, list_count, buffer_type,
                                my_header, user_ptr, context_id, hints);

    gen_mutex_unlock(&interface_mutex);
    return(ret);
}


/* BMI_tcp_open_context()
 *
 * opens a new context with the specified context id
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_open_context(bmi_context_id context_id)
{

    gen_mutex_lock(&interface_mutex);

    /* start a new queue for tracking completions in this context */
    completion_array[context_id] = op_list_new();
    if (!completion_array[context_id])
    {
        gen_mutex_unlock(&interface_mutex);
        return(bmi_tcp_errno_to_pvfs(-ENOMEM));
    }

    gen_mutex_unlock(&interface_mutex);
    return(0);
}


/* BMI_tcp_close_context()
 *
 * shuts down a context, previously opened with BMI_tcp_open_context()
 *
 * no return value
 */
void BMI_tcp_close_context(bmi_context_id context_id)
{
    
    gen_mutex_lock(&interface_mutex);

    /* tear down completion queue for this context */
    op_list_cleanup(completion_array[context_id]);

    gen_mutex_unlock(&interface_mutex);
    return;
}


/* BMI_tcp_cancel()
 *
 * attempt to cancel a pending bmi tcp operation
 *
 * returns 0 on success, -errno on failure
 */
int BMI_tcp_cancel(bmi_op_id_t id, bmi_context_id context_id)
{
    method_op_p query_op = NULL;
    
    gen_mutex_lock(&interface_mutex);

    query_op = (method_op_p)id_gen_fast_lookup(id);
    if(!query_op)
    {
        /* if we can't find the operattion, then assume that it has already
         * completed naturally
         */
        gen_mutex_unlock(&interface_mutex);
        return(0);
    }

    /* easy case: is the operation already completed? */
    if(((struct tcp_op*)(query_op->method_data))->tcp_op_state ==
        BMI_TCP_COMPLETE)
    {
        /* only close socket in forceful cancel mode */
        if(forceful_cancel_mode)
            tcp_forget_addr(query_op->addr, 0, -BMI_ECANCEL);
        /* we are done! status will be collected during test */
        gen_mutex_unlock(&interface_mutex);
        return(0);
    }

    /* has the operation started moving data yet? */
    if(query_op->env_amt_complete)
    {
        /* be pessimistic and kill the socket, even if not in forceful
         * cancel mode */
        /* NOTE: this may place other operations beside this one into
         * EINTR error state 
         */
        tcp_forget_addr(query_op->addr, 0, -BMI_ECANCEL);
        gen_mutex_unlock(&interface_mutex);
        return(0);
    }

    /* if we fall to this point, op has been posted, but no data has moved
     * for it yet as far as we know
     */

    /* mark op as canceled, move to completion queue */
    query_op->error_code = -BMI_ECANCEL;
    if(query_op->send_recv == BMI_SEND)
    {
        BMI_socket_collection_remove_write_bit(tcp_socket_collection_p,
                                           query_op->addr);
    }
    op_list_remove(query_op);
    ((struct tcp_op*)(query_op->method_data))->tcp_op_state = 
        BMI_TCP_COMPLETE;
    /* only close socket in forceful cancel mode */
    if(forceful_cancel_mode)
        tcp_forget_addr(query_op->addr, 0, -BMI_ECANCEL);
    op_list_add(completion_array[query_op->context_id], query_op);
    gen_mutex_unlock(&interface_mutex);
    return(0);
}

/*
 * For now, we only support wildcard strings that are IP addresses
 * and not *hostnames*!
 */
static int check_valid_wildcard(const char *wildcard_string, unsigned long *octets)
{
    int i, len = strlen(wildcard_string), last_dot = -1, octet_count = 0;
    char str[16];
    for (i = 0; i < len; i++)
    {
        char c = wildcard_string[i];
        memset(str, 0, 16);
        if ((c < '0' || c > '9') && c != '*' && c != '.')
            return -EINVAL;
        if (c == '*') {
            if (octet_count >= 4)
                return -EINVAL;
            octets[octet_count++] = 256;
        }
        else if (c == '.')
        {
            char *endptr = NULL;
            if (octet_count >= 4)
                return -EINVAL;
            strncpy(str, &wildcard_string[last_dot + 1], (i - last_dot - 1));
            octets[octet_count++] = strtol(str, &endptr, 10);
            if (*endptr != '\0' || octets[octet_count-1] >= 256)
                return -EINVAL;
            last_dot = i;
        }
    }
    for (i = octet_count; i < 4; i++)
    {
         octets[i] = 256;
    }
    return 0;
}

/*
 * return 1 if the addr specified is part of the wildcard specification of octet
 * return 0 otherwise.
 */
static int check_octets(struct in_addr addr, unsigned long *octets)
{
#define B1_MASK  0xff000000
#define B1_SHIFT 24
#define B2_MASK  0x00ff0000
#define B2_SHIFT 16
#define B3_MASK  0x0000ff00
#define B3_SHIFT 8
#define B4_MASK  0x000000ff
    uint32_t host_addr = ntohl(addr.s_addr);
    /* * stands for all clients */
    if (octets[0] == 256)
    {
        return 1;
    }
    if (((host_addr & B1_MASK) >> B1_SHIFT) != octets[0])
    {
        return 0;
    }
    if (octets[1] == 256)
    {
        return 1;
    }
    if (((host_addr & B2_MASK) >> B2_SHIFT) != octets[1])
    {
        return 0;
    }
    if (octets[2] == 256)
    {
        return 1;
    }
    if (((host_addr & B3_MASK) >> B3_SHIFT) != octets[2])
    {
        return 0;
    }
    if (octets[3] == 256)
    {
        return 1;
    }
    if ((host_addr & B4_MASK) != octets[3])
    {
        return 0;
    }
    return 1;
#undef B1_MASK
#undef B1_SHIFT 
#undef B2_MASK 
#undef B2_SHIFT
#undef B3_MASK
#undef B3_SHIFT
#undef B4_MASK
}
/* BMI_tcp_query_addr_range()
 * Check if a given address is within the network specified by the wildcard string!
 * or if it is part of the subnet mask specified
 */
int BMI_tcp_query_addr_range(bmi_method_addr_p map, const char *wildcard_string, int netmask)
{
    struct tcp_addr *tcp_addr_data = map->method_data;
    struct sockaddr_in map_addr;
    socklen_t map_addr_len = sizeof(map_addr);
    const char *tcp_wildcard = wildcard_string + 6 /* strlen("tcp://") */;
    int ret = -1;

    memset(&map_addr, 0, sizeof(map_addr));
    if(getpeername(tcp_addr_data->socket, (struct sockaddr *) &map_addr, &map_addr_len) < 0)
    {
        ret =  bmi_tcp_errno_to_pvfs(-EINVAL);
        gossip_err("Error: failed to retrieve peer name for client.\n");
        return(ret);
    }
    /* Wildcard specification */
    if (netmask == -1)
    {
        unsigned long octets[4];
        if (check_valid_wildcard(tcp_wildcard, octets) < 0)
        {
            gossip_lerr("Invalid wildcard specification: %s\n", tcp_wildcard);
            return -EINVAL;
        }
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, "Map Address is : %s, Wildcard Octets: %lu.%lu.%lu.%lu\n", inet_ntoa(map_addr.sin_addr),
                octets[0], octets[1], octets[2], octets[3]);
        if (check_octets(map_addr.sin_addr, octets) == 1)
        {
            return 1;
        }
    }
    /* Netmask specification */
    else {
        struct sockaddr_in mask_addr, network_addr;
        memset(&mask_addr, 0, sizeof(mask_addr));
        memset(&network_addr, 0, sizeof(network_addr));
        /* Convert the netmask address */
        convert_mask(netmask, &mask_addr.sin_addr);
        /* Invalid network address */
        if (inet_aton(tcp_wildcard, &network_addr.sin_addr) == 0)
        {
            gossip_err("Invalid network specification: %s\n", tcp_wildcard);
            return -EINVAL;
        }
        /* Matches the subnet mask! */
        if ((map_addr.sin_addr.s_addr & mask_addr.sin_addr.s_addr)
                == (network_addr.sin_addr.s_addr & mask_addr.sin_addr.s_addr))
        {
            return 1;
        }
    }
    return 0;
}

/* BMI_tcp_addr_rev_lookup_unexpected()
 *
 * looks up an address that was initialized unexpectedly and returns a string
 * hostname
 *
 * returns string on success, "UNKNOWN" on failure
 */
const char* BMI_tcp_addr_rev_lookup_unexpected(bmi_method_addr_p map)
{
    struct tcp_addr *tcp_addr_data = map->method_data;
    int debug_on;
    uint64_t mask;
    socklen_t peerlen;
    struct sockaddr_in peer;
    int ret;
    struct hostent *peerent;
    char* tmp_peer;

    /* return default response if we don't have support for the right socket
     * calls 
     */
#if !defined(HAVE_GETHOSTBYADDR)
    return(tcp_addr_data->peer);
#else 

    /* Only resolve hostnames if a gossip mask is set to request it.
     * Otherwise we leave it at ip address 
     */
    gossip_get_debug_mask(&debug_on, &mask);

    if(!debug_on || (!(mask & GOSSIP_ACCESS_HOSTNAMES)))
    {
        return(tcp_addr_data->peer);
    }

    peerlen = sizeof(struct sockaddr_in);

    if(tcp_addr_data->peer_type == BMI_TCP_PEER_HOSTNAME)
    {
        /* full hostname already cached; return now */
        return(tcp_addr_data->peer);
    }

    /* if we hit this point, we need to resolve hostname */
    ret = getpeername(tcp_addr_data->socket, (struct sockaddr*)&(peer), &peerlen);
    if(ret < 0)
    {
        /* default to use IP address */
        return(tcp_addr_data->peer);
    }

    peerent = gethostbyaddr((void*)&peer.sin_addr.s_addr, 
        sizeof(struct in_addr), AF_INET);
    if(peerent == NULL)
    {
        /* default to use IP address */
        return(tcp_addr_data->peer);
    }
 
    tmp_peer = (char*)malloc(strlen(peerent->h_name) + 1);
    if(!tmp_peer)
    {
        /* default to use IP address */
        return(tcp_addr_data->peer);
    }
    strcpy(tmp_peer, peerent->h_name);
    if(tcp_addr_data->peer)
    {
        free(tcp_addr_data->peer);
    }
    tcp_addr_data->peer = tmp_peer;
    tcp_addr_data->peer_type = BMI_TCP_PEER_HOSTNAME;
    return(tcp_addr_data->peer);

#endif

}

/* tcp_forget_addr()
 *
 * completely removes a tcp method address from use, and aborts any
 * operations that use the address.  If the
 * dealloc_flag is set, the memory used by the address will be
 * deallocated as well.
 *
 * no return value
 */
void tcp_forget_addr(bmi_method_addr_p map,
                     int dealloc_flag,
                     int error_code)
{
    struct tcp_addr* tcp_addr_data = map->method_data;
    BMI_addr_t bmi_addr = tcp_addr_data->bmi_addr;
    int tmp_outcount;
    bmi_method_addr_p tmp_addr;
    int tmp_status;

    if (tcp_socket_collection_p)
    {
        BMI_socket_collection_remove(tcp_socket_collection_p, map);
        /* perform a test to force the socket collection to act on the remove
         * request before continuing
         */
        if(!sc_test_busy)
        {
            BMI_socket_collection_testglobal(tcp_socket_collection_p,
                0, &tmp_outcount, &tmp_addr, &tmp_status, 0);
        }
    }

    tcp_shutdown_addr(map);
    tcp_cleanse_addr(map, error_code);
    tcp_addr_data->addr_error = error_code;
    if (dealloc_flag)
    {
        dealloc_tcp_method_addr(map);
    }
    else
    {
        /* this will cause the bmi control layer to check to see if 
         * this address can be completely forgotten
         */
        bmi_method_addr_forget_callback(bmi_addr);
    }
    return;
};

/******************************************************************
 * Internal support functions
 */


/*
 * dealloc_tcp_method_addr()
 *
 * destroys method address structures generated by the TCP/IP module.
 *
 * no return value
 */
static void dealloc_tcp_method_addr(bmi_method_addr_p map)
{

    struct tcp_addr *tcp_addr_data = NULL;

    tcp_addr_data = map->method_data;
    /* close the socket, as long as it is not the one we are listening on
     * as a server.
     */
    if (!tcp_addr_data->server_port)
    {
        if (tcp_addr_data->socket > -1)
        {
            close(tcp_addr_data->socket);
        }
    }

    if (tcp_addr_data->hostname)
        free(tcp_addr_data->hostname);
    if (tcp_addr_data->peer)
        free(tcp_addr_data->peer);

    bmi_dealloc_method_addr(map);

    return;
}


/*
 * alloc_tcp_method_addr()
 *
 * creates a new method address with defaults filled in for TCP/IP.
 *
 * returns pointer to struct on success, NULL on failure
 */
bmi_method_addr_p alloc_tcp_method_addr(void)
{

    struct bmi_method_addr *my_method_addr = NULL;
    struct tcp_addr *tcp_addr_data = NULL;

    my_method_addr =
        bmi_alloc_method_addr(tcp_method_params.method_id, sizeof(struct tcp_addr));
    if (!my_method_addr)
    {
        return (NULL);
    }

    /* note that we trust the alloc_method_addr() function to have zeroed
     * out the structures for us already 
     */

    tcp_addr_data = my_method_addr->method_data;
    tcp_addr_data->socket = -1;
    tcp_addr_data->port = -1;
    tcp_addr_data->map = my_method_addr;
    tcp_addr_data->sc_index = -1;

    return (my_method_addr);
}


/*
 * tcp_server_init()
 *
 * this function is used to prepare a node to recieve incoming
 * connections if it is initialized in a server configuration.   
 *
 * returns 0 on succes, -errno on failure
 */
static int tcp_server_init(void)
{

    int oldfl = 0;              /* old socket flags */
    struct tcp_addr *tcp_addr_data = NULL;
    int tmp_errno = bmi_tcp_errno_to_pvfs(-EINVAL);
    int ret = 0;

    /* create a socket */
    tcp_addr_data = tcp_method_params.listen_addr->method_data;
    if ((tcp_addr_data->socket = BMI_sockio_new_sock()) < 0)
    {
        tmp_errno = errno;
        gossip_err("Error: BMI_sockio_new_sock: %s\n", strerror(tmp_errno));
        return (bmi_tcp_errno_to_pvfs(-tmp_errno));
    }

    /* set it to non-blocking operation */
    oldfl = fcntl(tcp_addr_data->socket, F_GETFL, 0);
    if (!(oldfl & O_NONBLOCK))
    {
        fcntl(tcp_addr_data->socket, F_SETFL, oldfl | O_NONBLOCK);
    }

    /* setup for a fast restart to avoid bind addr in use errors */
    BMI_sockio_set_sockopt(tcp_addr_data->socket, SO_REUSEADDR, 1);

    /* bind it to the appropriate port */
    if(tcp_method_params.method_flags & BMI_TCP_BIND_SPECIFIC)
    {
        ret = BMI_sockio_bind_sock_specific(tcp_addr_data->socket,
            tcp_addr_data->hostname,
            tcp_addr_data->port);
        /* NOTE: this particular function converts errno in advance */
        if(ret < 0)
        {
            PVFS_perror_gossip("BMI_sockio_bind_sock_specific", ret);
            return(ret);
        }
    }
    else
    {
        ret = BMI_sockio_bind_sock(tcp_addr_data->socket,
            tcp_addr_data->port);
    }
    
    if (ret < 0)
    {
        tmp_errno = errno;
        gossip_err("Error: BMI_sockio_bind_sock: %s\n", strerror(tmp_errno));
        return (bmi_tcp_errno_to_pvfs(-tmp_errno));
    }

    /* go ahead and listen to the socket */
    if (listen(tcp_addr_data->socket, TCP_BACKLOG) != 0)
    {
        tmp_errno = errno;
        gossip_err("Error: listen: %s\n", strerror(tmp_errno));
        return (bmi_tcp_errno_to_pvfs(-tmp_errno));
    }

    return (0);
}


/* find_recv_inflight()
 *
 * checks to see if there is a recv operation in flight (when in flight
 * means that some of the data or envelope has been read) for a 
 * particular address. 
 *
 * returns pointer to operation on success, NULL if nothing found.
 */
static method_op_p find_recv_inflight(bmi_method_addr_p map)
{
    struct op_list_search_key key;
    method_op_p query_op = NULL;

    memset(&key, 0, sizeof(struct op_list_search_key));
    key.method_addr = map;
    key.method_addr_yes = 1;

    query_op = op_list_search(op_list_array[IND_RECV_INFLIGHT], &key);

    return (query_op);
}


/* tcp_sock_init()
 *
 * this is an internal function which is used to build up a TCP/IP
 * connection in the situation of a client side operation.
 * addressing information to determine which fields need to be set.
 * If the connection is already established then it does no work.
 *
 * NOTE: this is safe to call repeatedly.  However, always check the
 * value of the not_connected field in the tcp address before using the
 * address.
 *
 * returns 0 on success, -errno on failure
 */
static int tcp_sock_init(bmi_method_addr_p my_method_addr)
{

    int oldfl = 0;              /* socket flags */
    int ret = -1;
    struct pollfd poll_conn;
    struct tcp_addr *tcp_addr_data = my_method_addr->method_data;
    int tmp_errno = 0;

    /* check for obvious problems */
    assert(my_method_addr);
    assert(my_method_addr->method_type == tcp_method_params.method_id);
    assert(tcp_addr_data->server_port == 0);

    /* fail immediately if the address is in failure mode and we have no way
     * to reconnect
     */
    if(tcp_addr_data->addr_error && tcp_addr_data->dont_reconnect)
    {
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, 
        "Warning: BMI communication attempted on an address in failure mode.\n");
        return(tcp_addr_data->addr_error);
    }

    if(tcp_addr_data->addr_error)
    {
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, "%s: attempting reconnect.\n",
          __func__);
        tcp_addr_data->addr_error = 0;
        assert(tcp_addr_data->socket < 0);
        tcp_addr_data->not_connected = 1;
    }

    /* is there already a socket? */
    if (tcp_addr_data->socket > -1)
    {
        /* check to see if we still need to work on the connect.. */
        if (tcp_addr_data->not_connected)
        {
            /* this is a little weird, but we complete the nonblocking
             * connection by polling */
            poll_conn.fd = tcp_addr_data->socket;
            poll_conn.events = POLLOUT;
            ret = poll(&poll_conn, 1, 2);
            if ((ret < 0) || (poll_conn.revents & POLLERR))
            {
                tmp_errno = errno;
                gossip_lerr("Error: poll: %s\n", strerror(tmp_errno));
                return (bmi_tcp_errno_to_pvfs(-tmp_errno));
            }
            if (poll_conn.revents & POLLOUT)
            {
                tcp_addr_data->not_connected = 0;
            }
        }
        /* return.  the caller should check the "not_connected" flag to
         * see if the socket is usable yet. */
        return (0);
    }
    
    bmi_set_sock_buffers(tcp_addr_data->socket);

    /* at this point there is no socket.  try to build it */
    if (tcp_addr_data->port < 1)
    {
        return (bmi_tcp_errno_to_pvfs(-EINVAL));
    }

    /* make a socket */
    if ((tcp_addr_data->socket = BMI_sockio_new_sock()) < 0)
    {
        tmp_errno = errno;
        return (bmi_tcp_errno_to_pvfs(-tmp_errno));
    }

    /* set it to non-blocking operation */
    oldfl = fcntl(tcp_addr_data->socket, F_GETFL, 0);
    if (!(oldfl & O_NONBLOCK))
    {
        fcntl(tcp_addr_data->socket, F_SETFL, oldfl | O_NONBLOCK);
    }

#if defined(USE_TRUSTED) && defined(__PVFS2_CLIENT__)
    /* make sure if we need to bind or not to some local port ranges */
    tcp_enable_trusted(tcp_addr_data);
#endif

    /* turn off Nagle's algorithm */
    if (BMI_sockio_set_tcpopt(tcp_addr_data->socket, TCP_NODELAY, 1) < 0)
    {
        tmp_errno = errno;
        gossip_lerr("Error: failed to set TCP_NODELAY option.\n");
        close(tcp_addr_data->socket);
        return (bmi_tcp_errno_to_pvfs(-tmp_errno));
    }

       bmi_set_sock_buffers(tcp_addr_data->socket);

    if (tcp_addr_data->hostname)
    {
        gossip_ldebug(GOSSIP_BMI_DEBUG_TCP,
                      "Connect: socket=%d, hostname=%s, port=%d\n",
                      tcp_addr_data->socket, tcp_addr_data->hostname,
                      tcp_addr_data->port);
        ret = BMI_sockio_connect_sock(tcp_addr_data->socket,
                      tcp_addr_data->hostname,
                      tcp_addr_data->port);
    }
    else
    {
        return (bmi_tcp_errno_to_pvfs(-EINVAL));
    }

    if (ret < 0)
    {
        if (ret == -EINPROGRESS)
        {
            tcp_addr_data->not_connected = 1;
            /* this will have to be connected later with a poll */
        }
        else
        {
            /* NOTE: BMI_sockio_connect_sock returns a PVFS error */
            char buff[300];

            snprintf(buff, 300, "Error: BMI_sockio_connect_sock: (%s):", 
                     tcp_addr_data->hostname);

            PVFS_perror_gossip(buff, ret);
            return (ret);
        }
    }

    return (0);
}


/* enqueue_operation()
 *
 * creates a new operation based on the arguments to the function.  It
 * then makes sure that the address is added to the socket collection,
 * and the operation is added to the appropriate operation queue.
 *
 * Damn, what a big prototype!
 *
 * returns 0 on success, -errno on failure
 */
static int enqueue_operation(op_list_p target_list,
                             enum bmi_op_type send_recv,
                             bmi_method_addr_p map,
                             void *const *buffer_list,
                             const bmi_size_t *size_list,
                             int list_count,
                             bmi_size_t amt_complete,
                             bmi_size_t env_amt_complete,
                             bmi_op_id_t * id,
                             int tcp_op_state,
                             struct tcp_msg_header header,
                             void *user_ptr,
                             bmi_size_t actual_size,
                             bmi_size_t expected_size,
                             bmi_context_id context_id,
                             int32_t eid)
{
    method_op_p new_method_op = NULL;
    struct tcp_op *tcp_op_data = NULL;
    struct tcp_addr* tcp_addr_data = NULL;
    int i;

    /* allocate the operation structure */
    new_method_op = alloc_tcp_method_op();
    if (!new_method_op)
    {
        return (bmi_tcp_errno_to_pvfs(-ENOMEM));
    }

    *id = new_method_op->op_id;
    new_method_op->event_id = eid;

    /* set the fields */
    new_method_op->send_recv = send_recv;
    new_method_op->addr = map;
    new_method_op->user_ptr = user_ptr;
    /* this is on purpose; we want to use the buffer_list all of
     * the time, no special case for one contig buffer
     */
    new_method_op->buffer = NULL;
    new_method_op->actual_size = actual_size;
    new_method_op->expected_size = expected_size;
    new_method_op->send_recv = send_recv;
    new_method_op->amt_complete = amt_complete;
    new_method_op->env_amt_complete = env_amt_complete;
    new_method_op->msg_tag = header.tag;
    new_method_op->mode = header.mode;
    new_method_op->list_count = list_count;
    new_method_op->context_id = context_id;

    /* set our current position in list processing */
    i=0;
    new_method_op->list_index = 0;
    new_method_op->cur_index_complete = 0;
    while(amt_complete > 0)
    {
        if(amt_complete >= size_list[i])
        {
            amt_complete -= size_list[i];
            new_method_op->list_index++;
            i++;
        }
        else
        {
            new_method_op->cur_index_complete = amt_complete;
            amt_complete = 0;
        }
    }

    tcp_op_data = new_method_op->method_data;
    tcp_op_data->tcp_op_state = tcp_op_state;
    tcp_op_data->env = header;

    /* if there is only one item in the list, then keep the list stored
     * in the op structure.  This allows us to use the same code for send
     * and recv as we use for send_list and recv_list, without having to 
     * malloc lists for those special cases
     */
    if (list_count == 1)
    {
        new_method_op->buffer_list = &tcp_op_data->buffer_list_stub;
        new_method_op->size_list = &tcp_op_data->size_list_stub;
        ((void**)new_method_op->buffer_list)[0] = buffer_list[0];
        ((bmi_size_t*)new_method_op->size_list)[0] = size_list[0];
    }
    else
    {
        new_method_op->size_list = size_list;
        new_method_op->buffer_list = buffer_list;
    }

    tcp_addr_data = map->method_data;

    if(tcp_addr_data->addr_error)
    {
        /* server should always fail here, client should let receives queue
         * as if nothing were wrong
         */
        if(tcp_addr_data->dont_reconnect || send_recv == BMI_SEND)
        {
            gossip_debug(GOSSIP_BMI_DEBUG_TCP, 
                       "Warning: BMI communication attempted on an "
                       "address in failure mode.\n");
            new_method_op->error_code = tcp_addr_data->addr_error;
            op_list_add(op_list_array[new_method_op->context_id],
                        new_method_op);
            return(tcp_addr_data->addr_error);
        }
    }

#if 0
    if(tcp_addr_data->addr_error)
    {
        /* this address is bad, don't try to do anything with it */
        gossip_err("Warning: BMI communication attempted on an "
                   "address in failure mode.\n");

        new_method_op->error_code = tcp_addr_data->addr_error;
        op_list_add(op_list_array[new_method_op->context_id],
                    new_method_op);
        return(tcp_addr_data->addr_error);
    }
#endif

    /* add the socket to poll on */
    BMI_socket_collection_add(tcp_socket_collection_p, map);
    if(send_recv == BMI_SEND)
    {
        BMI_socket_collection_add_write_bit(tcp_socket_collection_p, map);
    }

    /* keep up with the operation */
    op_list_add(target_list, new_method_op);

    return (0);
}


/* tcp_post_recv_generic()
 *
 * does the real work of posting an operation - works for both
 * eager and rendezvous messages
 *
 * returns 0 on success that requires later poll, returns 1 on instant
 * completion, -errno on failure
 */
static int tcp_post_recv_generic(bmi_op_id_t * id,
                                 bmi_method_addr_p src,
                                 void *const *buffer_list,
                                 const bmi_size_t *size_list,
                                 int list_count,
                                 bmi_size_t expected_size,
                                 bmi_size_t * actual_size,
                                 enum bmi_buffer_type buffer_type,
                                 bmi_msg_tag_t tag,
                                 void *user_ptr,
                                 bmi_context_id context_id,
                                 PVFS_hint hints)
{
    method_op_p query_op = NULL;
    int ret = -1;
    struct tcp_addr *tcp_addr_data = NULL;
    struct tcp_op *tcp_op_data = NULL;
    struct tcp_msg_header bogus_header;
    struct op_list_search_key key;
    bmi_size_t copy_size = 0;
    bmi_size_t total_copied = 0;
    int i;
    PINT_event_id eid = 0;

    PINT_EVENT_START(
        bmi_tcp_recv_event_id, bmi_tcp_pid, NULL, &eid,
        PINT_HINT_GET_CLIENT_ID(hints),
        PINT_HINT_GET_REQUEST_ID(hints),
        PINT_HINT_GET_RANK(hints),
        PINT_HINT_GET_HANDLE(hints),
        PINT_HINT_GET_OP_ID(hints),
        expected_size);

    tcp_addr_data = src->method_data;

    /* short out immediately if the address is bad and we have no way to
     * reconnect
     */
    if(tcp_addr_data->addr_error && tcp_addr_data->dont_reconnect)
    {
        gossip_debug(
            GOSSIP_BMI_DEBUG_TCP,
            "Warning: BMI communication attempted "
            "on an address in failure mode.\n");
        return(tcp_addr_data->addr_error);
    }

    /* lets make sure that the message hasn't already been fully
     * buffered in eager mode before doing anything else
     */
    memset(&key, 0, sizeof(struct op_list_search_key));
    key.method_addr = src;
    key.method_addr_yes = 1;
    key.msg_tag = tag;
    key.msg_tag_yes = 1;

    query_op =
        op_list_search(op_list_array[IND_RECV_EAGER_DONE_BUFFERING], &key);
    if (query_op)
    {
        /* make sure it isn't too big */
        if (query_op->actual_size > expected_size)
        {
            gossip_err("Error: message ordering violation;\n");
            gossip_err("Error: message too large for next buffer.\n");
            return (bmi_tcp_errno_to_pvfs(-EPROTO));
        }

        /* whoohoo- it is already done! */
        /* copy buffer out to list segments; handle short case */
        for (i = 0; i < list_count; i++)
        {
            copy_size = size_list[i];
            if (copy_size + total_copied > query_op->actual_size)
            {
                copy_size = query_op->actual_size - total_copied;
            }
            memcpy(buffer_list[i], (void *) ((char *) query_op->buffer +
                                             total_copied), copy_size);
            total_copied += copy_size;
            if (total_copied == query_op->actual_size)
            {
                break;
            }
        }
        /* copy out to correct memory regions */
        (*actual_size) = query_op->actual_size;
        free(query_op->buffer);
        *id = 0;
        op_list_remove(query_op);
        dealloc_tcp_method_op(query_op);
        PINT_EVENT_END(bmi_tcp_recv_event_id, bmi_tcp_pid, NULL, eid, 0,
                       *actual_size);

        return (1);
    }

    /* look for a message that is already being received */
    query_op = op_list_search(op_list_array[IND_RECV_INFLIGHT], &key);
    if (query_op)
    {
        tcp_op_data = query_op->method_data;
    }

    /* see if it is being buffered into a temporary memory region */
    if (query_op && tcp_op_data->tcp_op_state == BMI_TCP_BUFFERING)
    {
        /* make sure it isn't too big */
        if (query_op->actual_size > expected_size)
        {
            gossip_err("Error: message ordering violation;\n");
            gossip_err("Error: message too large for next buffer.\n");
            return (bmi_tcp_errno_to_pvfs(-EPROTO));
        }

        /* copy what we have so far into the correct buffers */
        total_copied = 0;
        for (i = 0; i < list_count; i++)
        {
            copy_size = size_list[i];
            if (copy_size + total_copied > query_op->amt_complete)
            {
                copy_size = query_op->amt_complete - total_copied;
            }
            if (copy_size > 0)
            {
                memcpy(buffer_list[i], (void *) ((char *) query_op->buffer +
                                                 total_copied), copy_size);
            }
            total_copied += copy_size;
            if (total_copied == query_op->amt_complete)
            {
                query_op->list_index = i;
                query_op->cur_index_complete = copy_size;
                break;
            }
        }

        /* see if we ended on a buffer boundary */
        if (query_op->cur_index_complete ==
            query_op->size_list[query_op->list_index])
        {
            query_op->list_index++;
            query_op->cur_index_complete = 0;
        }

        /* release the old buffer */
        if (query_op->buffer)
        {
            free(query_op->buffer);
        }

        *id = query_op->op_id;
        tcp_op_data = query_op->method_data;
        tcp_op_data->tcp_op_state = BMI_TCP_INPROGRESS;

        query_op->list_count = list_count;
        query_op->user_ptr = user_ptr;
        query_op->context_id = context_id;
        /* if there is only one item in the list, then keep the list stored
         * in the op structure.  This allows us to use the same code for send
         * and recv as we use for send_list and recv_list, without having to 
         * malloc lists for those special cases
         */
        if (list_count == 1)
        {
            query_op->buffer_list = &tcp_op_data->buffer_list_stub;
            query_op->size_list = &tcp_op_data->size_list_stub;
            ((void **)query_op->buffer_list)[0] = buffer_list[0];
            ((bmi_size_t *)query_op->size_list)[0] = size_list[0];
        }
        else
        {
            query_op->buffer_list = buffer_list;
            query_op->size_list = size_list;
        }

        if (query_op->amt_complete < query_op->actual_size)
        {
            /* try to recv some more data */
            tcp_addr_data = query_op->addr->method_data;
            ret = payload_progress(tcp_addr_data->socket,
                                   query_op->buffer_list,
                                   query_op->size_list,
                                   query_op->list_count,
                                   query_op->actual_size,
                                   &(query_op->list_index),
                                   &(query_op->cur_index_complete),
                                   BMI_RECV,
                                   NULL,
                                   0);
            if (ret < 0)
            {
                PVFS_perror_gossip("Error: payload_progress", ret);
                /* payload_progress() returns BMI error codes */
                tcp_forget_addr(query_op->addr, 0, ret);
                return (ret);
            }

            query_op->amt_complete += ret;
        }
        assert(query_op->amt_complete <= query_op->actual_size);
        if (query_op->amt_complete == query_op->actual_size)
        {
            /* we are done */
            op_list_remove(query_op);
            *id = 0;
            (*actual_size) = query_op->actual_size;
            dealloc_tcp_method_op(query_op);
            PINT_EVENT_END(
                bmi_tcp_recv_event_id, bmi_tcp_pid, NULL, eid,
                0, *actual_size);

            return (1);
        }
        else
        {
            /* there is still more work to do */
            tcp_op_data->tcp_op_state = BMI_TCP_INPROGRESS;
            return (0);
        }
    }

    /* NOTE: if the message was in flight, but not buffering, then
     * that means that it has already matched an earlier receive
     * post or else is an unexpected message that doesn't require a
     * matching receive post - at any rate it shouldn't be handled
     * here
     */

    /* if we hit this point we must enqueue */
    if (expected_size <= TCP_MODE_EAGER_LIMIT)
    {
        bogus_header.mode = TCP_MODE_EAGER;
    }
    else
    {
        bogus_header.mode = TCP_MODE_REND;
    }
    bogus_header.tag = tag;
    ret = enqueue_operation(op_list_array[IND_RECV],
                            BMI_RECV, src, buffer_list, size_list,
                            list_count, 0, 0, id, BMI_TCP_INPROGRESS,
                            bogus_header, user_ptr, 0,
                            expected_size, context_id, eid);
    /* just for safety; this field isn't valid to the caller anymore */
    (*actual_size) = 0;
    /* TODO: figure out why this causes deadlocks; observable in 2
     * scenarios:
     * - pvfs2-client-core with threaded library and nptl
     * - pvfs2-server threaded with nptl sending messages to itself
     */
#if 0
    if (ret >= 0)
    {
        /* go ahead and try to do some work while we are in this
         * function since we appear to be backlogged.  Make sure that
         * we do not wait in the poll, however.
         */
        ret = tcp_do_work(0);
    }
#endif
    return (ret);
}


/* tcp_cleanse_addr()
 *
 * finds all active operations matching the given address, places them
 * in an error state, and moves them to the completed queue.
 *
 * NOTE: this function does not shut down the address.  That should be
 * handled separately
 *
 * returns 0 on success, -errno on failure
 */
static int tcp_cleanse_addr(bmi_method_addr_p map, int error_code)
{
    int i = 0;
    struct op_list_search_key key;
    method_op_p query_op = NULL;

    memset(&key, 0, sizeof(struct op_list_search_key));
    key.method_addr = map;
    key.method_addr_yes = 1;

    /* NOTE: we know the unexpected completed queue is the last index! */
    for (i = 0; i < (NUM_INDICES - 1); i++)
    {
        if (op_list_array[i])
        {
            while ((query_op = op_list_search(op_list_array[i], &key)))
            {
                op_list_remove(query_op);
                query_op->error_code = error_code;
                if (query_op->mode == TCP_MODE_UNEXP && query_op->send_recv
                    == BMI_RECV)
                {
                    op_list_add(op_list_array[IND_COMPLETE_RECV_UNEXP],
                                query_op);
                }
                else
                {
                    ((struct tcp_op*)(query_op->method_data))->tcp_op_state = 
                        BMI_TCP_COMPLETE;
                    op_list_add(completion_array[query_op->context_id], query_op);
                }
            }
        }
    }

    return (0);
}


/* tcp_shutdown_addr()
 *
 * closes connections associated with a tcp method address
 *
 * returns 0 on success, -errno on failure
 */
static int tcp_shutdown_addr(bmi_method_addr_p map)
{

    struct tcp_addr *tcp_addr_data = map->method_data;
    if (tcp_addr_data->socket > -1)
    {
        close(tcp_addr_data->socket);
    }
    tcp_addr_data->socket = -1;
    tcp_addr_data->not_connected = 1;

    return (0);
}


/* tcp_do_work()
 *
 * this is the function that actually does communication work during
 * BMI_tcp_testXXX and BMI_tcp_waitXXX functions.  The amount of work 
 * that it does is tunable.
 *
 * returns 0 on success, -errno on failure.
 */
static int tcp_do_work(int max_idle_time)
{
    int ret = -1;
    bmi_method_addr_p addr_array[TCP_WORK_METRIC];
    int status_array[TCP_WORK_METRIC];
    int socket_count = 0;
    int i = 0;
    int stall_flag = 0;
    int busy_flag = 1;
    struct timespec req;
    struct tcp_addr* tcp_addr_data = NULL;
    struct timespec wait_time;
    struct timeval start;

    if(sc_test_busy)
    {
        /* another thread is already polling or working on sockets */
        if(max_idle_time == 0)
        {
            /* we don't want to spend time waiting on it; return
             * immediately.
             */
            return(0);
        }

        /* Sleep until working thread thread signals that it has finished
         * its work and then return.  No need for this thread to poll;
         * the other thread may have already finished what we wanted.
         * This condition wait is used strictly as a best effort to
         * prevent busy spin.  We'll sort out the results later.
         */
        gettimeofday(&start, NULL);
        wait_time.tv_sec = start.tv_sec + max_idle_time / 1000;
        wait_time.tv_nsec = (start.tv_usec + ((max_idle_time % 1000)*1000))*1000;
        if (wait_time.tv_nsec > 1000000000)
        {
            wait_time.tv_nsec = wait_time.tv_nsec - 1000000000;
            wait_time.tv_sec++;
        }
        gen_cond_timedwait(&interface_cond, &interface_mutex, &wait_time);
        return(0);
    }

    /* this thread has gained control of the polling.  */
    sc_test_busy = 1;
    gen_mutex_unlock(&interface_mutex);

    /* our turn to look at the socket collection */
    ret = BMI_socket_collection_testglobal(tcp_socket_collection_p,
                                       TCP_WORK_METRIC, &socket_count,
                                       addr_array, status_array,
                                       max_idle_time);

    gen_mutex_lock(&interface_mutex);
    sc_test_busy = 0;

    if (ret < 0)
    {
        /* wake up anyone else who might have been waiting */
        gen_cond_broadcast(&interface_cond);
        PVFS_perror_gossip("Error: socket collection:", ret);
        /* BMI_socket_collection_testglobal() returns BMI error code */
        return (ret);
    }

    if(socket_count == 0)
        busy_flag = 0;

    /* do different kinds of work depending on results */
    for (i = 0; i < socket_count; i++)
    {
        tcp_addr_data = addr_array[i]->method_data;
        /* skip working on addresses in failure mode */
        if(tcp_addr_data->addr_error)
        {
            /* addr_error field is in BMI error code format */
            tcp_forget_addr(addr_array[i], 0, tcp_addr_data->addr_error);
            continue;
        }

        if (status_array[i] & SC_ERROR_BIT)
        {
            ret = tcp_do_work_error(addr_array[i]);
            if (ret < 0)
            {
                PVFS_perror_gossip("Warning: BMI error handling failure, continuing", ret);
            }
        }
        else
        {
            if (status_array[i] & SC_WRITE_BIT)
            {
                ret = tcp_do_work_send(addr_array[i], &stall_flag);
                if (ret < 0)
                {
                    PVFS_perror_gossip("Warning: BMI send error, continuing", ret);
                }
                if(!stall_flag)
                    busy_flag = 0;
            }
            if (status_array[i] & SC_READ_BIT)
            {
                ret = tcp_do_work_recv(addr_array[i], &stall_flag);
                if (ret < 0)
                {
                    PVFS_perror_gossip("Warning: BMI recv error, continuing", ret);
                }
                if(!stall_flag)
                    busy_flag = 0;
            }
        }
    }

    /* IMPORTANT NOTE: if we have set the following flag, then it indicates that
     * poll() is finding data on our sockets, yet we are not able to move
     * any of it right now.  This means that the sockets are backlogged, and
     * BMI is in danger of busy spinning during test functions.  Let's sleep
     * for a millisecond here in hopes of letting the rest of the system
     * catch up somehow (either by clearing a backlog in another I/O
     * component, or by posting more matching BMI recieve operations)
     */
    if(busy_flag)
    {
        req.tv_sec = 0;
        req.tv_nsec = 1000;
        gen_mutex_unlock(&interface_mutex);
        nanosleep(&req, NULL);
        gen_mutex_lock(&interface_mutex);
    }

    /* wake up anyone else who might have been waiting */
    gen_cond_broadcast(&interface_cond);
    return (0);
}


/* tcp_do_work_send()
 *
 * does work on a TCP address that is ready to send data.
 *
 * returns 0 on success, -errno on failure
 */
static int tcp_do_work_send(bmi_method_addr_p map, int* stall_flag)
{
    method_op_p active_method_op = NULL;
    struct op_list_search_key key;
    int blocked_flag = 0;
    int ret = 0;
    int tmp_stall_flag;

    *stall_flag = 1;

    while (blocked_flag == 0 && ret == 0)
    {
        /* what we want to do here is find the first operation in the send
         * queue for this address.
         */
        memset(&key, 0, sizeof(struct op_list_search_key));
        key.method_addr = map;
        key.method_addr_yes = 1;
        active_method_op = op_list_search(op_list_array[IND_SEND], &key);
        if (!active_method_op)
        {
            /* ran out of queued sends to work on */
            return (0);
        }

        ret = work_on_send_op(active_method_op, &blocked_flag, &tmp_stall_flag);
        if(!tmp_stall_flag)
            *stall_flag = 0;
    }

    return (ret);
}


/* handle_new_connection()
 *
 * this function should be called only on special tcp method addresses
 * that represent local server ports.  It will attempt to accept a new
 * connection and create a new method address for the remote host.
 *
 * side effect: destroys the temporary method_address that is passed in
 * to it.
 *
 * returns 0 on success, -errno on failure
 */
static int handle_new_connection(bmi_method_addr_p map)
{
    struct tcp_addr *tcp_addr_data = NULL;
    int accepted_socket = -1;
    bmi_method_addr_p new_addr = NULL;
    int ret = -1;
    char* tmp_peer = NULL;

    ret = tcp_accept_init(&accepted_socket, &tmp_peer);
    if (ret < 0)
    {
        return (ret);
    }
    if (accepted_socket < 0)
    {
        /* guess it wasn't ready after all */
        return (0);
    }

    /* ok, we have a new socket.  what now?  Probably simplest
     * thing to do is to create a new method_addr, add it to the
     * socket collection, and return.  It will get caught the next
     * time around */
    new_addr = alloc_tcp_method_addr();
    if (!new_addr)
    {
        return (bmi_tcp_errno_to_pvfs(-ENOMEM));
    }
    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP,
                  "Assigning socket %d to new method addr.\n",
                  accepted_socket);
    tcp_addr_data = new_addr->method_data;
    tcp_addr_data->socket = accepted_socket;
    tcp_addr_data->peer = tmp_peer;
    tcp_addr_data->peer_type = BMI_TCP_PEER_IP;

    /* set a flag to make sure that we never try to reconnect this address
     * in the future
     */
    tcp_addr_data->dont_reconnect = 1;
    /* register this address with the method control layer */
    tcp_addr_data->bmi_addr = bmi_method_addr_reg_callback(new_addr);
    if (ret < 0)
    {
        tcp_shutdown_addr(new_addr);
        dealloc_tcp_method_addr(new_addr);
        dealloc_tcp_method_addr(map);
        return (ret);
    }
    BMI_socket_collection_add(tcp_socket_collection_p, new_addr);

    dealloc_tcp_method_addr(map);
    return (0);

}


/* tcp_do_work_recv()
 * 
 * does work on a TCP address that is ready to recv data.
 *
 * returns 0 on success, -errno on failure
 */
static int tcp_do_work_recv(bmi_method_addr_p map, int* stall_flag)
{

    method_op_p active_method_op = NULL;
    int ret = -1;
    void *new_buffer = NULL;
    struct op_list_search_key key;
    struct tcp_msg_header new_header;
    struct tcp_addr *tcp_addr_data = map->method_data;
    struct tcp_op *tcp_op_data = NULL;
    int tmp_errno;
    int tmp;
    bmi_size_t old_amt_complete = 0;
    time_t current_time;

    *stall_flag = 1;

    /* figure out if this is a new connection */
    if (tcp_addr_data->server_port)
    {
        /* just try to accept connection- no work yet */
        *stall_flag = 0;
        return (handle_new_connection(map));
    }

    /* look for a recv for this address that is already in flight */
    active_method_op = find_recv_inflight(map);
    /* see if we found one in progress... */
    if (active_method_op)
    {
        tcp_op_data = active_method_op->method_data;
        if (active_method_op->mode == TCP_MODE_REND &&
            tcp_op_data->tcp_op_state == BMI_TCP_BUFFERING)
        {
            /* we must wait for recv post */
            return (0);
        }
        else
        {
            old_amt_complete = active_method_op->amt_complete;
            ret = work_on_recv_op(active_method_op, stall_flag);
            gossip_debug(GOSSIP_BMI_DEBUG_TCP, "actual_size=%d, "
                         "amt_complete=%d, old_amt_complete=%d\n",
                         (int)active_method_op->actual_size,
                         (int)active_method_op->amt_complete,
                         (int)old_amt_complete);

            if ((ret == 0) &&
                (old_amt_complete == active_method_op->amt_complete) &&
                active_method_op->actual_size &&
                (active_method_op->amt_complete <
                 active_method_op->actual_size))
            {
                gossip_debug(
                    GOSSIP_BMI_DEBUG_TCP, "Warning: bmi_tcp unable "
                    "to recv any data reported by poll(). [1]\n");

                if (tcp_addr_data->zero_read_limit++ ==
                    BMI_TCP_ZERO_READ_LIMIT)
                {
                    gossip_debug(GOSSIP_BMI_DEBUG_TCP,
                                 "...dropping connection.\n");
                    tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-EPIPE));
                }
            }
            else
            {
                tcp_addr_data->zero_read_limit = 0;
            }
            return(ret);
        }
    }

    /* let's see if a the entire header is ready to be received.  If so
     * we will go ahead and pull it.  Otherwise, we will try again later.
     * It isn't worth the complication of reading only a partial message
     * header - we really want it atomically
     */
    ret = BMI_sockio_nbpeek(tcp_addr_data->socket,
                            new_header.enc_hdr, TCP_ENC_HDR_SIZE);
    if (ret < 0)
    {
        tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-errno));
        return (0);
    }

    if (ret == 0)
    {
        gossip_debug(
            GOSSIP_BMI_DEBUG_TCP, "Warning: bmi_tcp unable "
            "to recv any data reported by poll(). [2]\n");

        if (tcp_addr_data->zero_read_limit++ ==
            BMI_TCP_ZERO_READ_LIMIT)
        {
            gossip_debug(GOSSIP_BMI_DEBUG_TCP,
                         "...dropping connection.\n");
            tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-EPIPE));
        }
        return(0);
    }
    else
    {
        tcp_addr_data->zero_read_limit = 0;
    }

    if (ret < TCP_ENC_HDR_SIZE)
    {
        current_time = time(NULL);
        if(!tcp_addr_data->short_header_timer)
        {
            tcp_addr_data->short_header_timer = current_time;
        }
        else if((current_time - tcp_addr_data->short_header_timer) > 
            BMI_TCP_HEADER_WAIT_SECONDS)
        {
            gossip_err("Error: incomplete BMI TCP header after %d seconds, closing connection.\n",
                BMI_TCP_HEADER_WAIT_SECONDS);
            tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-EPIPE));
            return (0);
        }

        /* header not ready yet, but we will keep hoping */
        return (0);
    }

    tcp_addr_data->short_header_timer = 0;
    *stall_flag = 0;
    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "Reading header for new op.\n");
    ret = BMI_sockio_nbrecv(tcp_addr_data->socket,
                           new_header.enc_hdr, TCP_ENC_HDR_SIZE);
    if (ret < TCP_ENC_HDR_SIZE)
    {
        tmp_errno = errno;
        gossip_err("Error: BMI_sockio_nbrecv: %s\n", strerror(tmp_errno));
        tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-tmp_errno));
        return (0);
    }

    /* decode the header */
    BMI_TCP_DEC_HDR(new_header);

    /* so we have the header. now what?  These are the possible
     * scenarios:
     * a) unexpected message
     * b) eager message for which a recv has been posted
     * c) eager message for which a recv has not been posted
     * d) rendezvous messsage for which a recv has been posted
     * e) rendezvous messsage for which a recv has not been posted
     * f) eager message for which a rend. recv has been posted
     */

    /* check magic number of message */
    if(new_header.magic_nr != BMI_MAGIC_NR)
    {
        gossip_err("Error: bad magic in BMI TCP message.\n");
        tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-EBADMSG));
        return(0);
    }

    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "Received new message; mode: %d.\n",
                  (int) new_header.mode);
    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "tag: %d\n", (int) new_header.tag);

    if (new_header.mode == TCP_MODE_UNEXP)
    {
        /* allocate the operation structure */
        active_method_op = alloc_tcp_method_op();
        if (!active_method_op)
        {
            tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-ENOMEM));
            return (bmi_tcp_errno_to_pvfs(-ENOMEM));
        }
        /* create data buffer */
        new_buffer = malloc(new_header.size);
        if (!new_buffer)
        {
            dealloc_tcp_method_op(active_method_op);
            tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-ENOMEM));
            return (bmi_tcp_errno_to_pvfs(-ENOMEM));
        }

        /* set the fields */
        active_method_op->send_recv = BMI_RECV;
        active_method_op->addr = map;
        active_method_op->actual_size = new_header.size;
        active_method_op->expected_size = 0;
        active_method_op->amt_complete = 0;
        active_method_op->env_amt_complete = TCP_ENC_HDR_SIZE;
        active_method_op->msg_tag = new_header.tag;
        active_method_op->buffer = new_buffer;
        active_method_op->mode = TCP_MODE_UNEXP;
        active_method_op->buffer_list = &(active_method_op->buffer);
        active_method_op->size_list = &(active_method_op->actual_size);
        active_method_op->list_count = 1;
        tcp_op_data = active_method_op->method_data;
        tcp_op_data->tcp_op_state = BMI_TCP_INPROGRESS;
        tcp_op_data->env = new_header;

        op_list_add(op_list_array[IND_RECV_INFLIGHT], active_method_op);
        /* grab some data if we can */
        return (work_on_recv_op(active_method_op, &tmp));
    }

    memset(&key, 0, sizeof(struct op_list_search_key));
    key.method_addr = map;
    key.method_addr_yes = 1;
    key.msg_tag = new_header.tag;
    key.msg_tag_yes = 1;

    /* look for a match within the posted operations */
    active_method_op = op_list_search(op_list_array[IND_RECV], &key);

    if (active_method_op)
    {
        /* make sure it isn't too big */
        if (new_header.size > active_method_op->expected_size)
        {
            gossip_err("Error: message ordering violation;\n");
            gossip_err("Error: message too large for next buffer.\n");
            gossip_err("Error: incoming size: %ld, expected size: %ld\n",
                        (long) new_header.size,
                        (long) active_method_op->expected_size);
            /* TODO: return error here or do something else? */
            return (bmi_tcp_errno_to_pvfs(-EPROTO));
        }

        /* we found a match.  go work on it and return */
        op_list_remove(active_method_op);
        active_method_op->env_amt_complete = TCP_ENC_HDR_SIZE;
        active_method_op->actual_size = new_header.size;
        op_list_add(op_list_array[IND_RECV_INFLIGHT], active_method_op);
        return (work_on_recv_op(active_method_op, &tmp));
    }

    /* no match anywhere.  Start a new operation */
    /* allocate the operation structure */
    active_method_op = alloc_tcp_method_op();
    if (!active_method_op)
    {
        tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-ENOMEM));
        return (bmi_tcp_errno_to_pvfs(-ENOMEM));
    }

    if (new_header.mode == TCP_MODE_EAGER)
    {
        /* create data buffer for eager messages */
        new_buffer = malloc(new_header.size);
        if (!new_buffer)
        {
            dealloc_tcp_method_op(active_method_op);
            tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-ENOMEM));
            return (bmi_tcp_errno_to_pvfs(-ENOMEM));
        }
    }
    else
    {
        new_buffer = NULL;
    }

    /* set the fields */
    active_method_op->send_recv = BMI_RECV;
    active_method_op->addr = map;
    active_method_op->actual_size = new_header.size;
    active_method_op->expected_size = 0;
    active_method_op->amt_complete = 0;
    active_method_op->env_amt_complete = TCP_ENC_HDR_SIZE;
    active_method_op->msg_tag = new_header.tag;
    active_method_op->buffer = new_buffer;
    active_method_op->mode = new_header.mode;
    active_method_op->buffer_list = &(active_method_op->buffer);
    active_method_op->size_list = &(active_method_op->actual_size);
    active_method_op->list_count = 1;
    tcp_op_data = active_method_op->method_data;
    tcp_op_data->tcp_op_state = BMI_TCP_BUFFERING;
    tcp_op_data->env = new_header;

    op_list_add(op_list_array[IND_RECV_INFLIGHT], active_method_op);

    /* grab some data if we can */
    if (new_header.mode == TCP_MODE_EAGER)
    {
        return (work_on_recv_op(active_method_op, &tmp));
    }

    return (0);
}


/*
 * work_on_send_op()
 *
 * used to perform work on a send operation.  this is called by the poll
 * function.
 * 
 * sets blocked_flag if no more work can be done on socket without
 * blocking
 * returns 0 on success, -errno on failure.
 */
static int work_on_send_op(method_op_p my_method_op,
                           int *blocked_flag, int* stall_flag)
{
    int ret = -1;
    struct tcp_addr *tcp_addr_data = my_method_op->addr->method_data;
    struct tcp_op *tcp_op_data = my_method_op->method_data;

    *blocked_flag = 1;
    *stall_flag = 0;

    /* make sure that the connection is done before we continue */
    if (tcp_addr_data->not_connected)
    {
        ret = tcp_sock_init(my_method_op->addr);
        if (ret < 0)
        {
            PVFS_perror_gossip("Error: socket failed to init", ret);
            /* tcp_sock_init() returns BMI error code */
            tcp_forget_addr(my_method_op->addr, 0, ret);
            return (0);
        }
        if (tcp_addr_data->not_connected)
        {
            /* try again later- still could not connect */
            tcp_op_data->tcp_op_state = BMI_TCP_INPROGRESS;
            return (0);
        }
    }

    ret = payload_progress(tcp_addr_data->socket,
        my_method_op->buffer_list,
        my_method_op->size_list,
        my_method_op->list_count,
        my_method_op->actual_size,
        &(my_method_op->list_index),
        &(my_method_op->cur_index_complete),
        BMI_SEND,
        tcp_op_data->env.enc_hdr,
        &my_method_op->env_amt_complete);
    if (ret < 0)
    {
        PVFS_perror_gossip("Error: payload_progress", ret);
        /* payload_progress() returns BMI error codes */
        tcp_forget_addr(my_method_op->addr, 0, ret);
        return (0);
    }

    if(ret == 0)
        *stall_flag = 1;

    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "Sent: %d bytes of data.\n", ret);
    my_method_op->amt_complete += ret;
    assert(my_method_op->amt_complete <= my_method_op->actual_size);

    if (my_method_op->amt_complete == my_method_op->actual_size && my_method_op->env_amt_complete == TCP_ENC_HDR_SIZE)
    {
        /* we are done */
        my_method_op->error_code = 0;
        BMI_socket_collection_remove_write_bit(tcp_socket_collection_p,
                                           my_method_op->addr);
        op_list_remove(my_method_op);
        ((struct tcp_op*)(my_method_op->method_data))->tcp_op_state = 
            BMI_TCP_COMPLETE;
        op_list_add(completion_array[my_method_op->context_id], my_method_op);
        *blocked_flag = 0;
    }
    else
    {
        /* there is still more work to do */
        tcp_op_data->tcp_op_state = BMI_TCP_INPROGRESS;
    }

    return (0);
}


/*
 * work_on_recv_op()
 *
 * used to perform work on a recv operation.  this is called by the poll
 * function.
 * NOTE: this function assumes the method header has already been read.
 *
 * returns 0 on success, -errno on failure.
 */
static int work_on_recv_op(method_op_p my_method_op, int* stall_flag)
{

    int ret = -1;
    struct tcp_addr *tcp_addr_data = my_method_op->addr->method_data;
    struct tcp_op *tcp_op_data = my_method_op->method_data;

    *stall_flag = 1;

    if (my_method_op->actual_size != 0)
    {
        /* now let's try to recv some actual data */
        ret = payload_progress(tcp_addr_data->socket,
            my_method_op->buffer_list,
            my_method_op->size_list,
            my_method_op->list_count,
            my_method_op->actual_size,
            &(my_method_op->list_index),
            &(my_method_op->cur_index_complete),
            BMI_RECV,
            NULL,
            0);
        if (ret < 0)
        {
            PVFS_perror_gossip("Error: payload_progress", ret);
            /* payload_progress() returns BMI error codes */
            tcp_forget_addr(my_method_op->addr, 0, ret);
            return (0);
        }
    }
    else
    {
        ret = 0;
    }

    if(ret > 0)
        *stall_flag = 0;

    my_method_op->amt_complete += ret;
    assert(my_method_op->amt_complete <= my_method_op->actual_size);

    if (my_method_op->amt_complete == my_method_op->actual_size)
    {
        /* we are done */
        op_list_remove(my_method_op);
        if (tcp_op_data->tcp_op_state == BMI_TCP_BUFFERING)
        {
            /* queue up to wait on matching post recv */
            op_list_add(op_list_array[IND_RECV_EAGER_DONE_BUFFERING],
                        my_method_op);
        }
        else
        {
            my_method_op->error_code = 0;
            if (my_method_op->mode == TCP_MODE_UNEXP)
            {
                op_list_add(op_list_array[IND_COMPLETE_RECV_UNEXP],
                            my_method_op);
            }
            else
            {
                ((struct tcp_op*)(my_method_op->method_data))->tcp_op_state = 
                    BMI_TCP_COMPLETE;
                op_list_add(completion_array[my_method_op->context_id], my_method_op);
            }
        }
    }

    return (0);
}


/* tcp_do_work_error()
 * 
 * handles a tcp address that has indicated an error during polling.
 *
 * returns 0 on success, -errno on failure
 */
static int tcp_do_work_error(bmi_method_addr_p map)
{
    struct tcp_addr *tcp_addr_data = NULL;
    int buf;
    int ret;
    int tmp_errno;

    tcp_addr_data = map->method_data;

    /* perform a read on the socket so that we can get a real errno */
    ret = read(tcp_addr_data->socket, &buf, sizeof(int));
    if (ret == 0)
        tmp_errno = EPIPE;  /* report other side closed socket with this */
    else
        tmp_errno = errno;

    gossip_debug(GOSSIP_BMI_DEBUG_TCP, "Error: bmi_tcp: %s\n",
      strerror(tmp_errno));

    if (tcp_addr_data->server_port)
    {
        /* Ignore this and hope it goes away... we don't want to lose
         * our local socket */
        dealloc_tcp_method_addr(map);
        gossip_lerr("Warning: error polling on server socket, continuing.\n");
        return (0);
    }

    if(tmp_errno == 0)
        tmp_errno = EPROTO;

    tcp_forget_addr(map, 0, bmi_tcp_errno_to_pvfs(-tmp_errno));

    return (0);
}

#if defined(USE_TRUSTED) && defined(__PVFS2_CLIENT__)
/*
 * tcp_enable_trusted()
 * Ideally, this function should look up the security configuration of
 * the server and determines
 * if it needs to bind to any specific port locally or not..
 * For now look at the FIXME below.
 */
static int tcp_enable_trusted(struct tcp_addr *tcp_addr_data)
{
    /*
     * FIXME:
     * For now, there is no way for us to check if a given
     * server is actually using port protection or not.
     * For now we unconditionally use a trusted port range
     * as long as USE_TRUSTED is #defined.
     *
     * Although most of the time we expect users
     * to be using a range of 0-1024, it is hard to keep probing
     * until one gets a port in the range specified.
     * Hence this is a temporary fix. we will see if this
     * requirement even needs to be met at all.
     */
    static unsigned short my_requested_port = 1023;
    unsigned short my_local_port = 0;
    struct sockaddr_in my_local_sockaddr;
    socklen_t len = sizeof(struct sockaddr_in);
    memset(&my_local_sockaddr, 0, sizeof(struct sockaddr_in));

    /* setup for a fast restart to avoid bind addr in use errors */
    if (BMI_sockio_set_sockopt(tcp_addr_data->socket, SO_REUSEADDR, 1) < 0)
    {
        gossip_lerr("Could not set SO_REUSEADDR on local socket (port %hd)\n", my_local_port);
    }
    if (BMI_sockio_bind_sock(tcp_addr_data->socket, my_requested_port) < 0)
    {
        gossip_lerr("Could not bind to local port %hd: %s\n", 
                my_requested_port, strerror(errno));
    }
    else {
        my_requested_port--;
    }
    my_local_sockaddr.sin_family = AF_INET;
    if (getsockname(tcp_addr_data->socket, 
                (struct sockaddr *)&my_local_sockaddr, &len) == 0)
    {
        my_local_port = ntohs(my_local_sockaddr.sin_port);
    }
    gossip_debug(GOSSIP_BMI_DEBUG_TCP, "Bound locally to port: %hd\n", my_local_port);
    return 0;
}

#endif

#if defined(USE_TRUSTED) && defined(__PVFS2_SERVER__)

static char *bad_errors[] = {
    "invalid network address",
    "invalid port",
    "invalid network address and port"
};

/*
 * tcp_allow_trusted()
 * if trusted ports was enabled make sure
 * that we can accept a particular connection from a given
 * client
 */
static int tcp_allow_trusted(struct sockaddr_in *peer_sockaddr)
{
    char *peer_hostname = inet_ntoa(peer_sockaddr->sin_addr);
    unsigned short peer_port = ntohs(peer_sockaddr->sin_port);
    int   i, what_failed   = -1;

    /* Don't refuse connects if there were any
     * parse errors or if it is not enabled in the config file
     */
    if (gtcp_allowed_connection->port_enforce == 0
            && gtcp_allowed_connection->network_enforce == 0)
    {
        return 0;
    }
    /* make sure that the client is within the allowed network */
    if (gtcp_allowed_connection->network_enforce == 1)
    {
        /* Always allow localhost to connect */
        if (ntohl(peer_sockaddr->sin_addr.s_addr) == INADDR_LOOPBACK)
        {
            goto port_check;
        }
        for (i = 0; i < gtcp_allowed_connection->network_count; i++)
        {
            /* check with all the masks */
            if ((peer_sockaddr->sin_addr.s_addr & gtcp_allowed_connection->netmask[i].s_addr) 
                    != (gtcp_allowed_connection->network[i].s_addr & gtcp_allowed_connection->netmask[i].s_addr ))
            {
                continue;
            }
            else {
                goto port_check;
            }
        }
        /* not from a trusted network */
        what_failed = 0;
    }
port_check:
    /* make sure that the client port numbers are within specified limits */
    if (gtcp_allowed_connection->port_enforce == 1)
    {
        if (peer_port < gtcp_allowed_connection->ports[0]
                || peer_port > gtcp_allowed_connection->ports[1])
        {
            what_failed = (what_failed < 0) ? 1 : 2;
        }
    }
    /* okay, we are good to go */
    if (what_failed < 0)
    {
        return 0;
    }
    /* no good */
    gossip_err("Rejecting client %s on port %d: %s\n",
           peer_hostname, peer_port, bad_errors[what_failed]);
    return -1;
}

#endif

/* 
 * tcp_accept_init()
 * 
 * used to establish a connection from the server side.  Attempts an
 * accept call and provides the socket if it succeeds.
 *
 * returns 0 on success, -errno on failure.
 */
static int tcp_accept_init(int *socket, char** peer)
{

    int ret = -1;
    int tmp_errno = 0;
    struct tcp_addr *tcp_addr_data = tcp_method_params.listen_addr->method_data;
    int oldfl = 0;
    struct sockaddr_in peer_sockaddr;
    int peer_sockaddr_size = sizeof(struct sockaddr_in);
    char* tmp_peer;

    /* do we have a socket on this end yet? */
    if (tcp_addr_data->socket < 0)
    {
        ret = tcp_server_init();
        if (ret < 0)
        {
            return (ret);
        }
    }

    *socket = accept(tcp_addr_data->socket, (struct sockaddr*)&peer_sockaddr,
              (socklen_t *)&peer_sockaddr_size);

    if (*socket < 0)
    {
        if ((errno == EAGAIN) ||
            (errno == EWOULDBLOCK) ||
            (errno == ENETDOWN) ||
            (errno == EPROTO) ||
            (errno == ENOPROTOOPT) ||
            (errno == EHOSTDOWN) ||
            (errno == ENONET) ||
            (errno == EHOSTUNREACH) ||
            (errno == EOPNOTSUPP) ||
            (errno == ENETUNREACH) ||
            (errno == ENFILE) ||
            (errno == EMFILE))
        {
            /* try again later */
            if ((errno == ENFILE) || (errno == EMFILE))
            {
                gossip_err("Error: accept: %s (continuing)\n",strerror(errno));
                bmi_method_addr_drop_callback(BMI_tcp_method_name);
            }
            return (0);
        }
        else
        {
            gossip_err("Error: accept: %s\n", strerror(errno));
            return (bmi_tcp_errno_to_pvfs(-errno));
        }
    }

#if defined(USE_TRUSTED) && defined(__PVFS2_SERVER__)

    /* make sure that we are allowed to accept this connection */
    if (tcp_allow_trusted(&peer_sockaddr) < 0)
    {
        /* Force closure of the connection */
        close(*socket);
        return (bmi_tcp_errno_to_pvfs(-EACCES));
    }

#endif

    /* we accepted a new connection.  turn off Nagle's algorithm. */
    if (BMI_sockio_set_tcpopt(*socket, TCP_NODELAY, 1) < 0)
    {
        tmp_errno = errno;
        gossip_lerr("Error: failed to set TCP_NODELAY option.\n");
        close(*socket);
        return (bmi_tcp_errno_to_pvfs(-tmp_errno));
    }

    /* set it to non-blocking operation */
    oldfl = fcntl(*socket, F_GETFL, 0);
    if (!(oldfl & O_NONBLOCK))
    {
        fcntl(*socket, F_SETFL, oldfl | O_NONBLOCK);
    }

    /* allocate ip address string */
    tmp_peer = inet_ntoa(peer_sockaddr.sin_addr);
    *peer = (char*)malloc(strlen(tmp_peer)+1);
    if(!(*peer))
    {
        close(*socket);
        return(bmi_tcp_errno_to_pvfs(-BMI_ENOMEM));
    }
    strcpy(*peer, tmp_peer);

    return (0);
}


/* alloc_tcp_method_op()
 *
 * creates a new method op with defaults filled in for tcp.
 *
 * returns pointer to structure on success, NULL on failure
 */
static method_op_p alloc_tcp_method_op(void)
{
    method_op_p my_method_op = NULL;

    my_method_op = bmi_alloc_method_op(sizeof(struct tcp_op));

    /* we trust alloc_method_op to zero it out */

    return (my_method_op);
}


/* dealloc_tcp_method_op()
 *
 * destroys an existing tcp method op, freeing segment lists if
 * needed
 *
 * no return value
 */
static void dealloc_tcp_method_op(method_op_p old_op)
{
    bmi_dealloc_method_op(old_op);
    return;
}

/* tcp_post_send_generic()
 * 
 * Submits send operations (low level).
 *
 * returns 0 on success that requires later poll, returns 1 on instant
 * completion, -errno on failure
 */
static int tcp_post_send_generic(bmi_op_id_t * id,
                                 bmi_method_addr_p dest,
                                 const void *const *buffer_list,
                                 const bmi_size_t *size_list,
                                 int list_count,
                                 enum bmi_buffer_type buffer_type,
                                 struct tcp_msg_header my_header,
                                 void *user_ptr,
                                 bmi_context_id context_id,
                                 PVFS_hint hints)
{
    struct tcp_addr *tcp_addr_data = dest->method_data;
    method_op_p query_op = NULL;
    int ret = -1;
    bmi_size_t total_size = 0;
    bmi_size_t amt_complete = 0;
    bmi_size_t env_amt_complete = 0;
    struct op_list_search_key key;
    int list_index = 0;
    bmi_size_t cur_index_complete = 0;
    PINT_event_id eid = 0;

    if(PINT_EVENT_ENABLED)
    {
        int i = 0;
        for(; i < list_count; ++i)
        {
            total_size += size_list[i];
        }
    }

    PINT_EVENT_START(
        bmi_tcp_send_event_id, bmi_tcp_pid, NULL, &eid,
        PINT_HINT_GET_CLIENT_ID(hints),
        PINT_HINT_GET_REQUEST_ID(hints),
        PINT_HINT_GET_RANK(hints),
        PINT_HINT_GET_HANDLE(hints),
        PINT_HINT_GET_OP_ID(hints),
        total_size);

    /* Three things can happen here:
     * a) another op is already in queue for the address, so we just
     * queue up
     * b) we can send the whole message and return
     * c) we send part of the message and queue the rest
     */

    /* NOTE: on the post_send side of an operation, it doesn't really
     * matter whether the op is going to be eager or rendezvous.  It is
     * handled the same way (except for how the header is filled in).
     * The difference is in the recv processing for TCP.
     */

    /* NOTE: we also don't care what the buffer_type says, TCP could care
     * less what buffers it is using.
     */

    /* encode the message header */
    BMI_TCP_ENC_HDR(my_header);

    /* the first thing we must do is find out if another send is queued
     * up for this address so that we don't mess up our ordering.    */
    memset(&key, 0, sizeof(struct op_list_search_key));
    key.method_addr = dest;
    key.method_addr_yes = 1;
    query_op = op_list_search(op_list_array[IND_SEND], &key);
    if (query_op)
    {
        /* queue up operation */
        ret = enqueue_operation(op_list_array[IND_SEND], BMI_SEND,
                                dest, (void **) buffer_list,
                                size_list, list_count, 0, 0,
                                id, BMI_TCP_INPROGRESS, my_header, user_ptr,
                                my_header.size, 0,
                                context_id,
                                eid);

        /* TODO: is this causing deadlocks?  See similar call in recv
         * path for another example.  This particular one seems to be an
         * issue under a heavy bonnie++ load that Neill has been
         * debugging.  Comment out for now to see if the problem goes
         * away.
         */
#if 0
        if (ret >= 0)
        {
            /* go ahead and try to do some work while we are in this
             * function since we appear to be backlogged.  Make sure that
             * we do not wait in the poll, however.
             */
            ret = tcp_do_work(0);
        }
#endif
        if (ret < 0)
        {
            gossip_err("Error: enqueue_operation() or tcp_do_work() returned: %d\n", ret);
        }
        return (ret);
    }

    /* make sure the connection is established */
    ret = tcp_sock_init(dest);
    if (ret < 0)
    {
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, "tcp_sock_init() failure.\n");
        /* tcp_sock_init() returns BMI error code */
        tcp_forget_addr(dest, 0, ret);
        PINT_EVENT_END(bmi_tcp_send_event_id, bmi_tcp_pid, NULL, 0, ret);
        return (ret);
    }

    tcp_addr_data = dest->method_data;

#if 0
    /* TODO: this is a hack for testing! */
    /* disables immediate send completion... */
    ret = enqueue_operation(op_list_array[IND_SEND], BMI_SEND,
                            dest, buffer_list, size_list, list_count, 0, 0,
                            id, BMI_TCP_INPROGRESS, my_header, user_ptr,
                            my_header.size, 0,
                            context_id);
    return(ret);
#endif

    if (tcp_addr_data->not_connected)
    {
        /* if the connection is not completed, queue up for later work */
        ret = enqueue_operation(op_list_array[IND_SEND], BMI_SEND,
                                dest, (void **) buffer_list, size_list,
                                list_count, 0, 0,
                                id, BMI_TCP_INPROGRESS, my_header, user_ptr,
                                my_header.size, 0,
                                context_id,
                                eid);
        if(ret < 0)
        {
            gossip_err("Error: enqueue_operation() returned: %d\n", ret);
        }
        return (ret);
    }

    /* try to send some data */
    env_amt_complete = 0;
    ret = payload_progress(tcp_addr_data->socket,
        (void **) buffer_list,
        size_list, list_count, my_header.size, &list_index,
        &cur_index_complete, BMI_SEND, my_header.enc_hdr, &env_amt_complete);
    if (ret < 0)
    {
        PVFS_perror_gossip("Error: payload_progress", ret);
        /* payload_progress() returns BMI error codes */
        tcp_forget_addr(dest, 0, ret);
        PINT_EVENT_END(bmi_tcp_send_event_id, bmi_tcp_pid, NULL, eid, 0, ret);
        return (ret);
    }

    gossip_ldebug(GOSSIP_BMI_DEBUG_TCP, "Sent: %d bytes of data.\n", ret);
    amt_complete = ret;
    assert(amt_complete <= my_header.size);
    if (amt_complete == my_header.size && env_amt_complete == TCP_ENC_HDR_SIZE)
    {
        /* we are already done */
        PINT_EVENT_END(bmi_tcp_send_event_id, bmi_tcp_pid,
                       NULL, eid, 0, amt_complete);
        return (1);
    }

    /* queue up the remainder */
    ret = enqueue_operation(op_list_array[IND_SEND], BMI_SEND,
                            dest, (void **) buffer_list,
                            size_list, list_count,
                            amt_complete, env_amt_complete, id,
                            BMI_TCP_INPROGRESS, my_header, user_ptr,
                            my_header.size, 0, context_id, eid);

    if(ret < 0)
    {
        gossip_err("Error: enqueue_operation() returned: %d\n", ret);
    }
    return (ret);
}


/* payload_progress()
 *
 * makes progress on sending/recving data payload portion of a message
 *
 * returns amount completed on success, -errno on failure
 */
static int payload_progress(int s, void *const *buffer_list, const bmi_size_t* 
    size_list, int list_count, bmi_size_t total_size, int* list_index, 
    bmi_size_t* current_index_complete, enum bmi_op_type send_recv, 
    char* enc_hdr, bmi_size_t* env_amt_complete)
{
    int i;
    int count = 0;
    int ret;
    int completed;
    /* used for finding the stopping point on short receives */
    int final_index = list_count-1;
    bmi_size_t final_size = size_list[list_count-1];
    bmi_size_t sum = 0;
    int vector_index = 0;
    int header_flag = 0;
    int tmp_env_done = 0;

    if(send_recv == BMI_RECV)
    {
        /* find out if we should stop short in list processing */
        for(i=0; i<list_count; i++)
        {
            sum += size_list[i];
            if(sum >= total_size)
            {
                final_index = i;
                final_size = size_list[i] - (sum-total_size);
                break;
            }
        }
    }

    assert(list_count > *list_index);

    /* make sure we don't overrun our preallocated iovec array */
    if((list_count - (*list_index)) > BMI_TCP_IOV_COUNT)
    {
        list_count = (*list_index) + BMI_TCP_IOV_COUNT;
    }

    /* do we need to send any of the header? */
    if(send_recv == BMI_SEND && *env_amt_complete < TCP_ENC_HDR_SIZE)
    {
        stat_io_vector[vector_index].iov_base = &enc_hdr[*env_amt_complete];
        stat_io_vector[vector_index].iov_len = TCP_ENC_HDR_SIZE - *env_amt_complete;
        count++;
        vector_index++;
        header_flag = 1;
    }

    /* setup vector */
    stat_io_vector[vector_index].iov_base = 
        (char*)buffer_list[*list_index] + *current_index_complete;
    count++;
    if(final_index == 0)
    {
        stat_io_vector[vector_index].iov_len = final_size - *current_index_complete;
    }
    else
    {
        stat_io_vector[vector_index].iov_len = 
            size_list[*list_index] - *current_index_complete;
        for(i = (*list_index + 1); i < list_count; i++)
        {
            vector_index++;
            count++;
            stat_io_vector[vector_index].iov_base = buffer_list[i];
            if(i == final_index)
            {
                stat_io_vector[vector_index].iov_len = final_size;
                break;
            }
            else
            {
                stat_io_vector[vector_index].iov_len = size_list[i];
            }
        }
    }

    assert(count > 0);

    if(send_recv == BMI_RECV)
    {
        ret = BMI_sockio_nbvector(s, stat_io_vector, count, 1);
    }
    else
    {
        ret = BMI_sockio_nbvector(s, stat_io_vector, count, 0);
    }

    /* if error or nothing done, return now */
    if(ret == 0)
        return(0);
    if(ret <= 0)
        return(bmi_tcp_errno_to_pvfs(-errno));

    completed = ret;
    if(header_flag && (completed >= 0))
    {
        /* take care of completed header status */
        tmp_env_done = TCP_ENC_HDR_SIZE - *env_amt_complete;
        if(tmp_env_done > completed)
            tmp_env_done = completed;
        completed -= tmp_env_done;
        ret -= tmp_env_done;
        (*env_amt_complete) += tmp_env_done;
    }

    i=header_flag;
    while(completed > 0)
    {
        /* take care of completed data payload */
        if(completed >= stat_io_vector[i].iov_len)
        {
            completed -= stat_io_vector[i].iov_len;
            *current_index_complete = 0;
            (*list_index)++;
            i++;
        }
        else
        {
            *current_index_complete += completed;
            completed = 0;
        }
    }

    return(ret);
}

static void bmi_set_sock_buffers(int socket){
        //Set socket buffer sizes:
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, "Default socket buffers send:%d receive:%d\n",
                GET_SENDBUFSIZE(socket), GET_RECVBUFSIZE(socket));
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, "Setting socket buffer size for send:%d receive:%d \n",
                tcp_buffer_size_send, tcp_buffer_size_receive);
    if( tcp_buffer_size_receive != 0)
         SET_RECVBUFSIZE(socket,tcp_buffer_size_receive);
    if( tcp_buffer_size_send != 0)
         SET_SENDBUFSIZE(socket,tcp_buffer_size_send);
        gossip_debug(GOSSIP_BMI_DEBUG_TCP, "Reread socket buffers send:%d receive:%d\n",
                GET_SENDBUFSIZE(socket), GET_RECVBUFSIZE(socket));
}

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ts=8 sts=4 sw=4 expandtab
 */