mca_btl_udapl_proc_t* mca_btl_udapl_proc_create(ompi_proc_t* ompi_proc)
{
mca_btl_udapl_proc_t* udapl_proc = NULL;
size_t size;
int rc;
/* Check if we have already created a uDAPL proc
* structure for this ompi process */
udapl_proc = mca_btl_udapl_proc_lookup_ompi(ompi_proc);
if(udapl_proc != NULL) {
return udapl_proc;
}
/* create a new udapl proc out of the ompi_proc ... */
udapl_proc = OBJ_NEW(mca_btl_udapl_proc_t);
udapl_proc->proc_endpoint_count = 0;
udapl_proc->proc_ompi = ompi_proc;
/* query for the peer address info */
rc = ompi_modex_recv(
&mca_btl_udapl_component.super.btl_version,
ompi_proc,
(void*)&udapl_proc->proc_addrs,
&size);
if(OMPI_SUCCESS != rc) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("ompi_modex_recv failed for peer %s",
ORTE_NAME_PRINT(&ompi_proc->proc_name)));
OBJ_RELEASE(udapl_proc);
return NULL;
}
if((size % sizeof(mca_btl_udapl_addr_t)) != 0) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("invalid udapl address for peer %s",
ORTE_NAME_PRINT(&ompi_proc->proc_name)));
OBJ_RELEASE(udapl_proc);
return NULL;
}
udapl_proc->proc_addr_count = size/sizeof(mca_btl_udapl_addr_t);
if (0 == udapl_proc->proc_addr_count) {
udapl_proc->proc_endpoints = NULL;
} else {
udapl_proc->proc_endpoints = (mca_btl_base_endpoint_t**)
malloc(udapl_proc->proc_addr_count * sizeof(mca_btl_base_endpoint_t*));
}
if(NULL == udapl_proc->proc_endpoints) {
OBJ_RELEASE(udapl_proc);
return NULL;
}
return udapl_proc;
}
int mca_btl_udapl_get(
mca_btl_base_module_t* btl,
mca_btl_base_endpoint_t* endpoint,
mca_btl_base_descriptor_t* des)
{
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_DEVELOPER, ("udapl_get\n"));
return OMPI_ERR_NOT_IMPLEMENTED;
}
int mca_btl_udapl_finalize(struct mca_btl_base_module_t* base_btl)
{
mca_btl_udapl_module_t* udapl_btl = (mca_btl_udapl_module_t*) base_btl;
int32_t i;
/*
* Cleaning up the endpoints here because mca_btl_udapl_del_procs
* is never called by upper layers.
* Note: this is only looking at those endpoints which are available
* off of the btl module rdma list.
*/
for (i=0; i < udapl_btl->udapl_eager_rdma_endpoint_count; i++) {
mca_btl_udapl_endpoint_t* endpoint =
opal_pointer_array_get_item(udapl_btl->udapl_eager_rdma_endpoints,
i);
OBJ_DESTRUCT(endpoint);
}
/* release uDAPL resources */
dat_evd_free(udapl_btl->udapl_evd_dto);
dat_evd_free(udapl_btl->udapl_evd_conn);
dat_pz_free(udapl_btl->udapl_pz);
dat_ia_close(udapl_btl->udapl_ia, DAT_CLOSE_GRACEFUL_FLAG);
/* destroy objects */
OBJ_DESTRUCT(&udapl_btl->udapl_lock);
OBJ_DESTRUCT(&udapl_btl->udapl_frag_eager);
OBJ_DESTRUCT(&udapl_btl->udapl_frag_eager_recv);
OBJ_DESTRUCT(&udapl_btl->udapl_frag_max);
OBJ_DESTRUCT(&udapl_btl->udapl_frag_max_recv);
OBJ_DESTRUCT(&udapl_btl->udapl_frag_user);
OBJ_DESTRUCT(&udapl_btl->udapl_frag_control);
OBJ_DESTRUCT(&udapl_btl->udapl_eager_rdma_lock);
/* destroy mpool */
if (OMPI_SUCCESS !=
mca_mpool_base_module_destroy(udapl_btl->super.btl_mpool)) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("WARNING: Failed to release mpool"));
return OMPI_ERROR;
}
free(udapl_btl);
return OMPI_SUCCESS;
}
int mca_btl_udapl_free(
struct mca_btl_base_module_t* btl,
mca_btl_base_descriptor_t* des)
{
mca_btl_udapl_frag_t* frag = (mca_btl_udapl_frag_t*)des;
if(0 == frag->size) {
if (NULL != frag->registration) {
btl->btl_mpool->mpool_deregister(btl->btl_mpool,
&(frag->registration->base));
frag->registration = NULL;
}
MCA_BTL_UDAPL_FRAG_RETURN_USER(btl, frag);
} else if(frag->size == mca_btl_udapl_component.udapl_eager_frag_size) {
MCA_BTL_UDAPL_FRAG_RETURN_EAGER(btl, frag);
} else if(frag->size == mca_btl_udapl_component.udapl_max_frag_size) {
MCA_BTL_UDAPL_FRAG_RETURN_MAX(btl, frag);
} else {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_DIAGNOSE,
("mca_btl_udapl_free: invalid descriptor\n"));
return OMPI_ERR_BAD_PARAM;
}
return OMPI_SUCCESS;
}
/*
* Find and assign system netmask for the address of the uDAPL BTL
* module, but only if udapl_if_mask has not been set by the "--mca
* btl_udapl_if_mask" parameter. This routine will either find
* the system netmask or set the value to 0.
*
* @param udapl_btl (IN) BTL module
*
* @return OMPI_SUCCESS or OMPI_ERROR
*/
static int mca_btl_udapl_assign_netmask(mca_btl_udapl_module_t* udapl_btl)
{
struct sockaddr *saddr;
struct sockaddr_in *btl_addr;
char btl_addr_string[INET_ADDRSTRLEN];
char btl_ifname[INET_ADDRSTRLEN];
/* Setting if_mask to 0 informs future steps to assume all
* addresses are reachable.
*/
udapl_btl->udapl_if_mask = 0;
if (mca_btl_udapl_component.udapl_compare_subnet) {
/* go get system netmask value */
/* use generic address to find address family */
saddr = (struct sockaddr *)&(udapl_btl->udapl_addr.addr);
if (saddr->sa_family == AF_INET) {
btl_addr = (struct sockaddr_in *)saddr;
/*
* Retrieve the netmask of the udapl btl address. To
* accomplish this requires 4 steps and the use of an opal
* utility. This same utility is used by the tcp oob.
* Steps:
* 1. Get string value of known udapl btl module address.
* 2. Use string value to find the interface name of address.
* 3. Use interface name to find its index.
* 4. From the index get the netmask.
*/
/* retrieve string value of udapl btl address */
inet_ntop(AF_INET, (void *) &btl_addr->sin_addr,
btl_addr_string, INET_ADDRSTRLEN);
/* use address string to retrieve associated interface name */
if (OPAL_SUCCESS !=
opal_ifaddrtoname(btl_addr_string,
btl_ifname, INET_ADDRSTRLEN)) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "interface not found",
true, ompi_process_info.nodename, btl_addr_string));
return OMPI_ERROR;
}
/* use interface name to retrieve index; then
* use index to retrieve udapl btl address netmask
*/
if (OPAL_SUCCESS !=
opal_ifindextomask(opal_ifnametoindex(btl_ifname),
&(udapl_btl->udapl_if_mask), sizeof(udapl_btl->udapl_if_mask))) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "netmask not found",
true, ompi_process_info.nodename, btl_addr_string));
return OMPI_ERROR;
}
/* report if_mask used by address */
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("uDAPL BTL address %s : if_mask = %d",
btl_addr_string, udapl_btl->udapl_if_mask));
} else {
/* current uDAPL BTL does not support IPv6 */
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "IPv4 only",
true, ompi_process_info.nodename));
return OMPI_ERROR;
}
}
return OMPI_SUCCESS;
}
int
mca_btl_udapl_init(DAT_NAME_PTR ia_name, mca_btl_udapl_module_t* btl)
{
mca_mpool_base_resources_t res;
DAT_CONN_QUAL port;
DAT_RETURN rc;
/* open the uDAPL interface */
btl->udapl_evd_async = DAT_HANDLE_NULL;
rc = dat_ia_open(ia_name, btl->udapl_async_evd_qlen,
&btl->udapl_evd_async, &btl->udapl_ia);
if(DAT_SUCCESS != rc) {
char* major;
char* minor;
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
#if defined(__SVR4) && defined(__sun)
if (strcmp(major, "DAT_INVALID_PARAMETER") == 0 &&
strcmp(minor, "DAT_INVALID_RO_COOKIE") == 0) {
/* Some platforms that Solaris runs on implement the PCI
* standard for relaxed ordering(RO). Using RDMA with
* polling on a memory location as the uDAPL (and openib
* by the way) BTL does for short messages with
* relaxed ordering could potentially produce silent data
* corruption. For this reason we need to take extra
* steps and this is accomplished by setting
* "ro_aware_system = 1" and handling as required.
*
* The uDAPL standard does not provide an interface to
* inform users of this scenario so Sun has implemented the
* following: If a platform supports relaxed ordering
* when the interface name is passed into the
* dat_ia_open() call, the call will return
* DAT_INVALID_PARAMETER and DAT_INVALID_RO_COOKIE.
* DAT_INVALID_RO_COOKIE is not part of the uDAPL standard
* at this time. The only way to open this interface is
* to prefix the following cookie "RO_AWARE_" to the ia
* name that was retreived from the dat registry.
*
* Example: ia_name = "ib0", new expected name will be
* "RO_AWARE_ib0".
*
* Here, since our first ia open attempt failed in the
* standard way, add the cookie and try to open again.
*/
DAT_NAME_PTR ro_ia_name;
/* prefix relaxed order cookie to ia_name */
asprintf(&ro_ia_name, "RO_AWARE_%s", ia_name);
if (NULL == ro_ia_name) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* because this is not standard inform user in some way */
BTL_UDAPL_VERBOSE_HELP(VERBOSE_INFORM,
("help-mpi-btl-udapl.txt", "relaxed order support",
true, ia_name, ro_ia_name));
/* try and open again */
btl->udapl_evd_async = DAT_HANDLE_NULL;
rc = dat_ia_open(ro_ia_name, btl->udapl_async_evd_qlen,
&btl->udapl_evd_async, &btl->udapl_ia);
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
if (DAT_SUCCESS == rc) {
mca_btl_udapl_component.ro_aware_system = 1;
free(ro_ia_name);
} else {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt",
"dat_ia_open fail RO", true, ro_ia_name,
major, minor, ia_name));
free(ro_ia_name);
return OMPI_ERROR;
}
} else {
#endif
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP, ("help-mpi-btl-udapl.txt",
"dat_ia_open fail", true, ia_name, major, minor));
return OMPI_ERROR;
#if defined(__SVR4) && defined(__sun)
}
#endif
}
/* create a protection zone */
rc = dat_pz_create(btl->udapl_ia, &btl->udapl_pz);
if(DAT_SUCCESS != rc) {
char* major;
char* minor;
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
BTL_ERROR(("ERROR: %s %s %s\n", "dat_pz_create",
major, minor));
goto failure;
}
/* query to get address information */
rc = dat_ia_query(btl->udapl_ia, &btl->udapl_evd_async,
DAT_IA_ALL, &(btl->udapl_ia_attr), 0, NULL);
if(DAT_SUCCESS != rc) {
char* major;
char* minor;
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
BTL_ERROR(("ERROR: %s %s %s\n", "dat_ia_query",
major, minor));
goto failure;
}
memcpy(&btl->udapl_addr.addr, (btl->udapl_ia_attr).ia_address_ptr,
sizeof(DAT_SOCK_ADDR));
/* determine netmask */
mca_btl_udapl_assign_netmask(btl);
/* check evd qlen against adapter max */
if (btl->udapl_dto_evd_qlen > (btl->udapl_ia_attr).max_evd_qlen) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP, ("help-mpi-btl-udapl.txt",
"evd_qlen adapter max",
true,
"btl_udapl_dto_evd_qlen",
btl->udapl_dto_evd_qlen,
(btl->udapl_ia_attr).max_evd_qlen));
btl->udapl_dto_evd_qlen = btl->udapl_ia_attr.max_evd_qlen;
}
if (btl->udapl_conn_evd_qlen > (btl->udapl_ia_attr).max_evd_qlen) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP, ("help-mpi-btl-udapl.txt",
"evd_qlen adapter max",
true,
"btl_udapl_conn_evd_qlen",
btl->udapl_conn_evd_qlen,
(btl->udapl_ia_attr).max_evd_qlen));
btl->udapl_conn_evd_qlen = btl->udapl_ia_attr.max_evd_qlen;
}
/* set up evd's */
rc = dat_evd_create(btl->udapl_ia,
btl->udapl_dto_evd_qlen, DAT_HANDLE_NULL,
DAT_EVD_DTO_FLAG | DAT_EVD_RMR_BIND_FLAG, &btl->udapl_evd_dto);
if(DAT_SUCCESS != rc) {
char* major;
char* minor;
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
BTL_ERROR(("ERROR: %s %s %s\n", "dat_evd_create (dto)",
major, minor));
goto failure;
}
rc = dat_evd_create(btl->udapl_ia,
btl->udapl_conn_evd_qlen, DAT_HANDLE_NULL,
DAT_EVD_CR_FLAG | DAT_EVD_CONNECTION_FLAG, &btl->udapl_evd_conn);
if(DAT_SUCCESS != rc) {
char* major;
char* minor;
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
BTL_ERROR(("ERROR: %s %s %s\n", "dat_evd_create (conn)",
major, minor));
goto failure;
}
/* create our public service point */
rc = dat_psp_create_any(btl->udapl_ia, &port, btl->udapl_evd_conn,
DAT_PSP_CONSUMER_FLAG, &btl->udapl_psp);
if(DAT_SUCCESS != rc) {
char* major;
char* minor;
dat_strerror(rc, (const char**)&major,
(const char**)&minor);
BTL_ERROR(("ERROR: %s %s %s\n", "dat_psp_create_any",
major, minor));
goto failure;
}
/* establish endpoint parameters */
rc = mca_btl_udapl_endpoint_get_params(btl, &(btl->udapl_ep_param));
if(OMPI_SUCCESS != rc) {
/* by not erroring out here we can try to continue with
* the default endpoint parameter values
*/
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP, ("help-mpi-btl-udapl.txt",
"use default endpoint params",
true));
}
/* Save the port with the address information */
/* TODO - since we're doing the hack below, do we need our own port? */
btl->udapl_addr.port = port;
/* Using dat_ep_query to obtain the remote port would be ideal but
* since the current udapl implementations don't seem to support
* this we store the port in udapl_addr and explictly exchange the
* information later.
*/
((struct sockaddr_in*)&btl->udapl_addr.addr)->sin_port = htons(port);
/* initialize the memory pool */
res.pool_name = "udapl";
res.reg_data = btl;
res.sizeof_reg = sizeof(mca_btl_udapl_reg_t);
res.register_mem = udapl_reg_mr;
res.deregister_mem = udapl_dereg_mr;
btl->super.btl_mpool = mca_mpool_base_module_create(
mca_btl_udapl_component.udapl_mpool_name, &btl->super, &res);
if (NULL == btl->super.btl_mpool) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("WARNING: Failed to create mpool."));
goto failure;
}
/* initialize objects */
OBJ_CONSTRUCT(&btl->udapl_frag_eager, ompi_free_list_t);
OBJ_CONSTRUCT(&btl->udapl_frag_eager_recv, ompi_free_list_t);
OBJ_CONSTRUCT(&btl->udapl_frag_max, ompi_free_list_t);
OBJ_CONSTRUCT(&btl->udapl_frag_max_recv, ompi_free_list_t);
OBJ_CONSTRUCT(&btl->udapl_frag_user, ompi_free_list_t);
OBJ_CONSTRUCT(&btl->udapl_frag_control, ompi_free_list_t);
OBJ_CONSTRUCT(&btl->udapl_lock, opal_mutex_t);
/* check buffer alignment against dat library */
if (mca_btl_udapl_component.udapl_buffer_alignment !=
DAT_OPTIMAL_ALIGNMENT) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP, ("help-mpi-btl-udapl.txt",
"optimal buffer alignment mismatch",
true,
DAT_OPTIMAL_ALIGNMENT,
mca_btl_udapl_component.udapl_buffer_alignment,
DAT_OPTIMAL_ALIGNMENT));
}
/* initialize free lists */
ompi_free_list_init_ex_new(&btl->udapl_frag_eager,
sizeof(mca_btl_udapl_frag_eager_t) +
mca_btl_udapl_component.udapl_eager_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
OBJ_CLASS(mca_btl_udapl_frag_eager_t),
mca_btl_udapl_component.udapl_eager_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
mca_btl_udapl_component.udapl_free_list_num,
mca_btl_udapl_component.udapl_free_list_max,
mca_btl_udapl_component.udapl_free_list_inc,
btl->super.btl_mpool,
NULL,
NULL);
ompi_free_list_init_ex_new(&btl->udapl_frag_eager_recv,
sizeof(mca_btl_udapl_frag_eager_t) +
mca_btl_udapl_component.udapl_eager_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
OBJ_CLASS(mca_btl_udapl_frag_eager_t),
mca_btl_udapl_component.udapl_eager_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
mca_btl_udapl_component.udapl_free_list_num,
mca_btl_udapl_component.udapl_free_list_max,
mca_btl_udapl_component.udapl_free_list_inc,
btl->super.btl_mpool,
NULL,
NULL);
ompi_free_list_init_ex_new(&btl->udapl_frag_max,
sizeof(mca_btl_udapl_frag_max_t) +
mca_btl_udapl_component.udapl_max_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
OBJ_CLASS(mca_btl_udapl_frag_max_t),
mca_btl_udapl_component.udapl_max_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
mca_btl_udapl_component.udapl_free_list_num,
mca_btl_udapl_component.udapl_free_list_max,
mca_btl_udapl_component.udapl_free_list_inc,
btl->super.btl_mpool,
NULL,
NULL);
ompi_free_list_init_ex_new(&btl->udapl_frag_max_recv,
sizeof(mca_btl_udapl_frag_max_t) +
mca_btl_udapl_component.udapl_max_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
OBJ_CLASS(mca_btl_udapl_frag_max_t),
mca_btl_udapl_component.udapl_max_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
mca_btl_udapl_component.udapl_free_list_num,
mca_btl_udapl_component.udapl_free_list_max,
mca_btl_udapl_component.udapl_free_list_inc,
btl->super.btl_mpool,
NULL,
NULL);
ompi_free_list_init_ex_new(&btl->udapl_frag_user,
sizeof(mca_btl_udapl_frag_user_t),
mca_btl_udapl_component.udapl_buffer_alignment,
OBJ_CLASS(mca_btl_udapl_frag_user_t),
0,0,
mca_btl_udapl_component.udapl_free_list_num,
mca_btl_udapl_component.udapl_free_list_max,
mca_btl_udapl_component.udapl_free_list_inc,
NULL,
NULL,
NULL);
ompi_free_list_init_ex_new(&btl->udapl_frag_control,
sizeof(mca_btl_udapl_frag_eager_t) +
mca_btl_udapl_component.udapl_eager_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
OBJ_CLASS(mca_btl_udapl_frag_eager_t),
mca_btl_udapl_component.udapl_eager_frag_size,
mca_btl_udapl_component.udapl_buffer_alignment,
mca_btl_udapl_component.udapl_free_list_num,
-1,
mca_btl_udapl_component.udapl_free_list_inc,
btl->super.btl_mpool,
NULL,
NULL);
/* initialize eager rdma buffer info */
btl->udapl_eager_rdma_endpoints = OBJ_NEW(opal_pointer_array_t);
opal_pointer_array_init(btl->udapl_eager_rdma_endpoints,
mca_btl_udapl_component.udapl_max_eager_rdma_peers,
mca_btl_udapl_component.udapl_max_eager_rdma_peers,
0);
btl->udapl_eager_rdma_endpoint_count = 0;
OBJ_CONSTRUCT(&btl->udapl_eager_rdma_lock, opal_mutex_t);
/* initialize miscellaneous variables */
btl->udapl_async_events = 0;
btl->udapl_connect_inprogress = 0;
btl->udapl_num_peers = 0;
/* TODO - Set up SRQ when it is supported */
return OMPI_SUCCESS;
failure:
dat_ia_close(btl->udapl_ia, DAT_CLOSE_ABRUPT_FLAG);
return OMPI_ERROR;
}
/*
* Find an address on the peer_process which matches stated criteria
* to the udapl btl module address information. Return in peer_addr_idx
* the index to the peer_process address that matches the btl module
* address. Where match criteria is:
* - the address in not already in use
* - compare addresses using netmask, the netmask value can be modified with
* "--mca btl_udapl_if_mask"
*
* Note: since this is called from mca_btl_udapl_proc_insert() it
* is assumed that the process lock is locked when entered.
*
* @param udapl_btl (IN) BTL module
* @param peer_process (IN) BTL peer process
* @param peer_addr_idx(IN/OUT) Index of address on peer_process
* which matches the udapl_btl address data.
* On success should be >= 0.
* @return OMPI_SUCCESS or error status on failure
*/
static int mca_btl_udapl_proc_address_match(
mca_btl_udapl_module_t* udapl_btl,
mca_btl_udapl_proc_t* peer_proc,
int* peer_addr_idx)
{
int i;
struct sockaddr *saddr;
struct sockaddr_in *btl_addr;
struct sockaddr_in *peer_addr;
char btl_addr_string[INET_ADDRSTRLEN];
char peer_addr_string[INET_ADDRSTRLEN];
*peer_addr_idx = MCA_BTL_UDAPL_INVALID_PEER_ADDR_IDX;
/* use generic address to find address family */
saddr = (struct sockaddr *)&(udapl_btl->udapl_addr.addr);
if (saddr->sa_family == AF_INET) {
btl_addr = (struct sockaddr_in *)saddr;
/* Loop thru peer process addresses looking for match.
* Match criteria:
* - address should not be "inuse"
* - both udapl btl module and peer address should be on
* the same subnet (compare with if_mask value)
*/
for(i = 0; i < (int) peer_proc->proc_addr_count; i++) {
peer_addr =
(struct sockaddr_in *)&(peer_proc->proc_addrs[i].addr);
if (VERBOSE_INFORM <=
mca_btl_udapl_component.udapl_verbosity) {
/* retrieve udapl btl and peer address string for reporting */
inet_ntop(AF_INET, (void *) &btl_addr->sin_addr,
btl_addr_string, INET_ADDRSTRLEN);
inet_ntop(AF_INET, (void *) &peer_addr->sin_addr,
peer_addr_string, INET_ADDRSTRLEN);
}
if ((false == peer_proc->proc_addrs[i].inuse) &&
(opal_net_samenetwork((struct sockaddr *)btl_addr,
(struct sockaddr *)peer_addr, udapl_btl->udapl_if_mask))) {
/* capture index of remote address where match found */
*peer_addr_idx = i;
/* mark this address as now being used */
peer_proc->proc_addrs[i].inuse = true;
/* report what address was found to match */
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("uDAPL BTL module(%s) matched %s",
btl_addr_string, peer_addr_string));
break;
} else {
/* peer address already used by another udapl btl
* module or netmask check not successful so skip
*/
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("uDAPL BTL module(%s) either skipped because it "
"is already in use or match criteria not successful "
"for peer address %s",
btl_addr_string, peer_addr_string));
}
}
} else {
/* current uDAPL BTL only supports IPv4 */
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "IPv4 only",
true, orte_process_info.nodename));
return OMPI_ERROR;
}
if (MCA_BTL_UDAPL_INVALID_PEER_ADDR_IDX == *peer_addr_idx) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "no network match",
true, btl_addr_string, orte_process_info.nodename,
peer_proc->proc_ompi->proc_hostname));
return OMPI_ERR_OUT_OF_RESOURCE;
}
return OMPI_SUCCESS;
}
int mca_btl_udapl_component_progress()
{
mca_btl_udapl_module_t* btl;
static int32_t inprogress = 0;
DAT_EVENT event;
size_t i;
int32_t j, rdma_ep_count;
int count = 0, btl_ownership;
mca_btl_udapl_frag_t* frag;
mca_btl_base_endpoint_t* endpoint;
/* prevent deadlock - only one thread should be 'progressing' at a time */
if(OPAL_THREAD_ADD32(&inprogress, 1) > 1) {
OPAL_THREAD_ADD32(&inprogress, -1);
return OMPI_SUCCESS;
}
/* check for work to do on each uDAPL btl */
OPAL_THREAD_LOCK(&mca_btl_udapl_component.udapl_lock);
for(i = 0; i < mca_btl_udapl_component.udapl_num_btls; i++) {
btl = mca_btl_udapl_component.udapl_btls[i];
/* Check DTO EVD */
while(DAT_SUCCESS ==
dat_evd_dequeue(btl->udapl_evd_dto, &event)) {
DAT_DTO_COMPLETION_EVENT_DATA* dto;
switch(event.event_number) {
case DAT_DTO_COMPLETION_EVENT:
dto = &event.event_data.dto_completion_event_data;
frag = dto->user_cookie.as_ptr;
/* Was the DTO successful? */
if(DAT_DTO_SUCCESS != dto->status) {
if (DAT_DTO_ERR_FLUSHED == dto->status) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("DAT_DTO_ERR_FLUSHED: probably OK if occurs during MPI_Finalize().\n"));
} else {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("ERROR: DAT_DTO_COMPLETION_EVENT: %d %d %lu %p.\n",
dto->status, frag->type,
(unsigned long)frag->size, dto->ep_handle));
}
return OMPI_ERROR;
}
endpoint = frag->endpoint;
btl_ownership = (frag->base.des_flags & MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
switch(frag->type) {
case MCA_BTL_UDAPL_RDMA_WRITE:
{
assert(frag->base.des_src == &frag->segment);
assert(frag->base.des_src_cnt == 1);
assert(frag->base.des_dst == NULL);
assert(frag->base.des_dst_cnt == 0);
assert(frag->type == MCA_BTL_UDAPL_RDMA_WRITE);
frag->base.des_cbfunc(&btl->super, endpoint,
&frag->base, OMPI_SUCCESS);
if( btl_ownership ) {
mca_btl_udapl_free(&btl->super,
&frag->base);
}
OPAL_THREAD_ADD32(&(endpoint->endpoint_lwqe_tokens[BTL_UDAPL_EAGER_CONNECTION]), 1);
mca_btl_udapl_frag_progress_pending(btl,
endpoint, BTL_UDAPL_EAGER_CONNECTION);
break;
}
case MCA_BTL_UDAPL_SEND:
{
int connection = BTL_UDAPL_EAGER_CONNECTION;
assert(frag->base.des_src == &frag->segment);
assert(frag->base.des_src_cnt == 1);
assert(frag->base.des_dst == NULL);
assert(frag->base.des_dst_cnt == 0);
assert(frag->type == MCA_BTL_UDAPL_SEND);
if(frag->size !=
mca_btl_udapl_component.udapl_eager_frag_size) {
assert(frag->size ==
mca_btl_udapl_component.udapl_max_frag_size);
connection = BTL_UDAPL_MAX_CONNECTION;
}
frag->base.des_cbfunc(&btl->super, endpoint,
&frag->base, OMPI_SUCCESS);
if( btl_ownership ) {
mca_btl_udapl_free(&btl->super,
&frag->base);
}
OPAL_THREAD_ADD32(&(endpoint->endpoint_lwqe_tokens[connection]), 1);
mca_btl_udapl_frag_progress_pending(btl,
endpoint, connection);
break;
}
case MCA_BTL_UDAPL_RECV:
{
mca_btl_active_message_callback_t* reg;
int cntrl_msg = -1;
assert(frag->base.des_dst == &frag->segment);
assert(frag->base.des_dst_cnt == 1);
assert(frag->base.des_src == NULL);
assert(frag->base.des_src_cnt == 0);
assert(frag->type == MCA_BTL_UDAPL_RECV);
assert(frag->triplet.virtual_address ==
(DAT_VADDR)(uintptr_t)frag->segment.seg_addr.pval);
assert(frag->triplet.segment_length == frag->size);
assert(frag->btl == btl);
/* setup frag ftr location and do callback */
frag->segment.seg_len = dto->transfered_length -
sizeof(mca_btl_udapl_footer_t);
frag->ftr = (mca_btl_udapl_footer_t *)
((char *)frag->segment.seg_addr.pval +
frag->segment.seg_len);
cntrl_msg = frag->ftr->tag;
reg = mca_btl_base_active_message_trigger + frag->ftr->tag;
OPAL_THREAD_UNLOCK(&mca_btl_udapl_component.udapl_lock);
reg->cbfunc(&btl->super,
frag->ftr->tag, &frag->base, reg->cbdata);
OPAL_THREAD_LOCK(&mca_btl_udapl_component.udapl_lock);
/* Repost the frag */
frag->ftr = frag->segment.seg_addr.pval;
frag->segment.seg_len =
(frag->size - sizeof(mca_btl_udapl_footer_t) -
sizeof(mca_btl_udapl_rdma_footer_t));
frag->base.des_flags = 0;
if(frag->size ==
mca_btl_udapl_component.udapl_eager_frag_size) {
OPAL_THREAD_ADD32(&(frag->endpoint->endpoint_sr_credits[BTL_UDAPL_EAGER_CONNECTION]), 1);
dat_ep_post_recv(frag->endpoint->endpoint_eager,
1, &frag->triplet, dto->user_cookie,
DAT_COMPLETION_DEFAULT_FLAG);
if (frag->endpoint->endpoint_sr_credits[BTL_UDAPL_EAGER_CONNECTION] >=
mca_btl_udapl_component.udapl_sr_win) {
mca_btl_udapl_endpoint_send_sr_credits(frag->endpoint,
BTL_UDAPL_EAGER_CONNECTION);
}
if (MCA_BTL_TAG_UDAPL == cntrl_msg) {
mca_btl_udapl_frag_progress_pending(btl,
frag->endpoint,
BTL_UDAPL_EAGER_CONNECTION);
}
} else {
assert(frag->size ==
mca_btl_udapl_component.udapl_max_frag_size);
OPAL_THREAD_ADD32(&(frag->endpoint->endpoint_sr_credits[BTL_UDAPL_MAX_CONNECTION]), 1);
dat_ep_post_recv(frag->endpoint->endpoint_max,
1, &frag->triplet, dto->user_cookie,
DAT_COMPLETION_DEFAULT_FLAG);
if (frag->endpoint->endpoint_sr_credits[BTL_UDAPL_MAX_CONNECTION] >=
mca_btl_udapl_component.udapl_sr_win) {
mca_btl_udapl_endpoint_send_sr_credits(frag->endpoint,
BTL_UDAPL_MAX_CONNECTION);
}
if (MCA_BTL_TAG_UDAPL == cntrl_msg) {
mca_btl_udapl_frag_progress_pending(btl,
frag->endpoint,
BTL_UDAPL_MAX_CONNECTION);
}
}
break;
}
case MCA_BTL_UDAPL_PUT:
{
assert(frag->base.des_src == &frag->segment);
assert(frag->base.des_src_cnt == 1);
assert(frag->base.des_dst_cnt == 1);
assert(frag->type == MCA_BTL_UDAPL_PUT);
frag->base.des_cbfunc(&btl->super, endpoint,
&frag->base, OMPI_SUCCESS);
if( btl_ownership ) {
mca_btl_udapl_free(&btl->super,
&frag->base);
}
OPAL_THREAD_ADD32(&(endpoint->endpoint_lwqe_tokens[BTL_UDAPL_MAX_CONNECTION]), 1);
OPAL_THREAD_ADD32(&(endpoint->endpoint_sr_tokens[BTL_UDAPL_MAX_CONNECTION]), 1);
mca_btl_udapl_frag_progress_pending(btl,
endpoint, BTL_UDAPL_MAX_CONNECTION);
break;
}
case MCA_BTL_UDAPL_CONN_RECV:
mca_btl_udapl_endpoint_finish_connect(btl,
frag->segment.seg_addr.pval,
(int32_t *)((char *)frag->segment.seg_addr.pval +
sizeof(mca_btl_udapl_addr_t)),
event.event_data.connect_event_data.ep_handle);
/* No break - fall through to free */
case MCA_BTL_UDAPL_CONN_SEND:
frag->segment.seg_len =
mca_btl_udapl_module.super.btl_eager_limit;
mca_btl_udapl_free(&btl->super, &frag->base);
break;
default:
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_DIAGNOSE,
("WARNING: unknown frag type: %d\n",
frag->type));
}
count++;
break;
default:
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_DIAGNOSE,
("WARNING: DTO event: %s (%d)\n",
mca_btl_udapl_dat_event_to_string(event.event_number),
event.event_number));
}
}
/* Check connection EVD */
while((btl->udapl_connect_inprogress > 0) && (DAT_SUCCESS ==
dat_evd_dequeue(btl->udapl_evd_conn, &event))) {
switch(event.event_number) {
case DAT_CONNECTION_REQUEST_EVENT:
/* Accept a new connection */
mca_btl_udapl_accept_connect(btl,
event.event_data.cr_arrival_event_data.cr_handle);
count++;
break;
case DAT_CONNECTION_EVENT_ESTABLISHED:
/* Both the client and server side of a connection generate
this event */
if (mca_btl_udapl_component.udapl_conn_priv_data) {
/* private data is only valid at this point if this
* event is from a dat_ep_connect call, not an accept
*/
mca_btl_udapl_endpoint_pd_established_conn(btl,
event.event_data.connect_event_data.ep_handle);
} else {
/* explicitly exchange process data */
mca_btl_udapl_sendrecv(btl,
event.event_data.connect_event_data.ep_handle);
}
count++;
break;
case DAT_CONNECTION_EVENT_PEER_REJECTED:
case DAT_CONNECTION_EVENT_NON_PEER_REJECTED:
case DAT_CONNECTION_EVENT_ACCEPT_COMPLETION_ERROR:
case DAT_CONNECTION_EVENT_DISCONNECTED:
case DAT_CONNECTION_EVENT_BROKEN:
case DAT_CONNECTION_EVENT_TIMED_OUT:
/* handle this case specially? if we have finite timeout,
we might want to try connecting again here. */
case DAT_CONNECTION_EVENT_UNREACHABLE:
/* Need to set the BTL endpoint to MCA_BTL_UDAPL_FAILED
See dat_ep_connect documentation pdf pg 198 */
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("WARNING: connection event not handled : %s (%d)\n",
mca_btl_udapl_dat_event_to_string(event.event_number),
event.event_number));
break;
default:
BTL_ERROR(("ERROR: connection event : %s (%d)",
mca_btl_udapl_dat_event_to_string(event.event_number),
event.event_number));
}
}
/* Check async EVD */
if (btl->udapl_async_events == mca_btl_udapl_component.udapl_async_events) {
btl->udapl_async_events = 0;
while(DAT_SUCCESS ==
dat_evd_dequeue(btl->udapl_evd_async, &event)) {
switch(event.event_number) {
case DAT_ASYNC_ERROR_EVD_OVERFLOW:
case DAT_ASYNC_ERROR_IA_CATASTROPHIC:
case DAT_ASYNC_ERROR_EP_BROKEN:
case DAT_ASYNC_ERROR_TIMED_OUT:
case DAT_ASYNC_ERROR_PROVIDER_INTERNAL_ERROR:
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("WARNING: async event ignored : %s (%d)",
mca_btl_udapl_dat_event_to_string(event.event_number),
event.event_number));
break;
default:
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("WARNING: %s (%d)\n",
mca_btl_udapl_dat_event_to_string(event.event_number),
event.event_number));
}
}
} else {
btl->udapl_async_events++;
}
/*
* Check eager rdma segments
*/
/* find the number of endpoints with rdma buffers */
rdma_ep_count = btl->udapl_eager_rdma_endpoint_count;
for (j = 0; j < rdma_ep_count; j++) {
mca_btl_udapl_endpoint_t* endpoint;
mca_btl_udapl_frag_t *local_rdma_frag;
endpoint =
opal_pointer_array_get_item(btl->udapl_eager_rdma_endpoints, j);
OPAL_THREAD_LOCK(&endpoint->endpoint_eager_rdma_local.lock);
local_rdma_frag =
MCA_BTL_UDAPL_GET_LOCAL_RDMA_FRAG(endpoint,
endpoint->endpoint_eager_rdma_local.head);
if (local_rdma_frag->rdma_ftr->active == 1) {
int pad = 0;
mca_btl_active_message_callback_t* reg;
MCA_BTL_UDAPL_RDMA_NEXT_INDEX(endpoint->endpoint_eager_rdma_local.head);
OPAL_THREAD_UNLOCK(&endpoint->endpoint_eager_rdma_local.lock);
/* compute pad as needed */
MCA_BTL_UDAPL_FRAG_CALC_ALIGNMENT_PAD(pad,
(local_rdma_frag->rdma_ftr->size +
sizeof(mca_btl_udapl_footer_t)));
/* set fragment information */
local_rdma_frag->ftr = (mca_btl_udapl_footer_t *)
((char *)local_rdma_frag->rdma_ftr -
pad -
sizeof(mca_btl_udapl_footer_t));
local_rdma_frag->segment.seg_len =
local_rdma_frag->rdma_ftr->size;
local_rdma_frag->segment.seg_addr.pval = (unsigned char *)
((char *)local_rdma_frag->ftr -
local_rdma_frag->segment.seg_len);
/* trigger callback */
reg = mca_btl_base_active_message_trigger + local_rdma_frag->ftr->tag;
reg->cbfunc(&btl->super,
local_rdma_frag->ftr->tag, &local_rdma_frag->base, reg->cbdata);
/* repost */
local_rdma_frag->rdma_ftr->active = 0;
local_rdma_frag->segment.seg_len =
mca_btl_udapl_module.super.btl_eager_limit;
local_rdma_frag->base.des_flags = 0;
/* increment local rdma credits */
OPAL_THREAD_ADD32(&(endpoint->endpoint_eager_rdma_local.credits),
1);
if (endpoint->endpoint_eager_rdma_local.credits >=
mca_btl_udapl_component.udapl_eager_rdma_win) {
mca_btl_udapl_endpoint_send_eager_rdma_credits(endpoint);
}
count++;
} else {
OPAL_THREAD_UNLOCK(&endpoint->endpoint_eager_rdma_local.lock);
}
} /* end of rdma_count loop */
}
/* unlock and return */
OPAL_THREAD_UNLOCK(&mca_btl_udapl_component.udapl_lock);
OPAL_THREAD_ADD32(&inprogress, -1);
return count;
}