int MPIDI_CH3U_Init_sock(int has_parent, MPIDI_PG_t *pg_p, int pg_rank,
char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
int pg_size;
int p;
/* FIXME: Why are these unused? */
MPIU_UNREFERENCED_ARG(has_parent);
MPIU_UNREFERENCED_ARG(pg_rank);
/*
* Initialize the VCs associated with this process group (and thus
* MPI_COMM_WORLD)
*/
/* FIXME: Get the size from the process group */
pg_size = MPIDI_PG_Get_size(pg_p);
/* FIXME: This should probably be the same as MPIDI_VC_InitSock. If
not, why not? */
/* FIXME: Note that MPIDI_CH3_VC_Init sets state, sendq_head and tail.
so this should be MPIDI_CH3_VC_Init( &pg_p->vct[p] );
followed by MPIDI_VC_InitSock( ditto );
In fact, there should be a single VC_Init call here */
/* FIXME: Why isn't this MPIDI_VC_Init( vc, NULL, 0 )? */
for (p = 0; p < pg_size; p++)
{
MPIDI_CH3I_VC *vcch = &pg_p->vct[p].ch;
vcch->sendq_head = NULL;
vcch->sendq_tail = NULL;
vcch->state = MPIDI_CH3I_VC_STATE_UNCONNECTED;
vcch->sock = MPIDU_SOCK_INVALID_SOCK;
vcch->conn = NULL;
}
mpi_errno = MPIDI_CH3U_Get_business_card_sock(pg_rank,
bc_val_p, val_max_sz_p);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER, "**init_buscard");
}
fn_exit:
return mpi_errno;
fn_fail:
/* FIXME: This doesn't belong here, since the pg is not created in
this routine */
/* --BEGIN ERROR HANDLING-- */
if (pg_p != NULL)
{
MPIDI_PG_Destroy(pg_p);
}
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPIDI_PG_Finalize - Finalize the process groups, including freeing all
process group structures
@*/
int MPIDI_PG_Finalize(void)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_PG_t *pg, *pgNext;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_FINALIZE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_FINALIZE);
/* Print the state of the process groups */
if (verbose) {
MPIU_PG_Printall( stdout );
}
/* FIXME - straighten out the use of PMI_Finalize - no use after
PG_Finalize */
if (pg_world->connData) {
#ifdef USE_PMI2_API
mpi_errno = PMI2_Finalize();
if (mpi_errno) MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|pmi_finalize");
#else
int rc;
rc = PMI_Finalize();
if (rc) {
MPIR_ERR_SET1(mpi_errno,MPI_ERR_OTHER,
"**ch3|pmi_finalize",
"**ch3|pmi_finalize %d", rc);
}
#endif
}
/* Free the storage associated with the process groups */
pg = MPIDI_PG_list;
while (pg) {
pgNext = pg->next;
/* In finalize, we free all process group information, even if
the ref count is not zero. This can happen if the user
fails to use MPI_Comm_disconnect on communicators that
were created with the dynamic process routines.*/
/* XXX DJG FIXME-MT should we be checking this? */
if (MPIR_Object_get_ref(pg) == 0 || 1) {
if (pg == MPIDI_Process.my_pg)
MPIDI_Process.my_pg = NULL;
MPIR_Object_set_ref(pg, 0); /* satisfy assertions in PG_Destroy */
MPIDI_PG_Destroy( pg );
}
pg = pgNext;
}
/* If COMM_WORLD is still around (it normally should be),
try to free it here. The reason that we need to free it at this
point is that comm_world (and comm_self) still exist, and
hence the usual process to free the related VC structures will
not be invoked. */
if (MPIDI_Process.my_pg) {
MPIDI_PG_Destroy(MPIDI_Process.my_pg);
}
MPIDI_Process.my_pg = NULL;
/* ifdefing out this check because the list will not be NULL in
Ch3_finalize because
one additional reference is retained in MPIDI_Process.my_pg.
That reference is released
only after ch3_finalize returns. If I release it before ch3_finalize,
the ssm channel crashes. */
#if 0
if (MPIDI_PG_list != NULL)
{
/* --BEGIN ERROR HANDLING-- */
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_INTERN,
"**dev|pg_finalize|list_not_empty", NULL);
/* --END ERROR HANDLING-- */
}
#endif
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_FINALIZE);
return mpi_errno;
}
/**
* Initialize Process Manager Interface and update global_info.
* Called by MPID_nem_ib_init.
*
* -# Initialize the Process Manager Interface;
* -# Set the rank;
* -# Set the progexx group size;
*
* \see MPID_nem_ib_init
*
\todo Need to add more stuff here
Look at InitPG in mvapich2/trunk/src/mpid/ch3/src/mpid_init.c
*/
int MPID_nem_ib_pmi_init()
{
int pmi_errno = 0;
int mpi_errno = MPI_SUCCESS;
int spawned;
/* Process group id size*/
int pg_id_sz, pg_size;
char *pg_id;
MPIDI_PG_t *pg = 0;
assert( global_info!= NULL );
MPIDI_STATE_DECL(MPID_STATE_MPIDI_PMI_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_PMI_INIT);
/* Initialize the Process Manager Interface */
pmi_errno = PMI_Init(&spawned);
if (pmi_errno != PMI_SUCCESS) {
MPIU_ERR_SETANDJUMP1(mpi_errno, MPI_ERR_OTHER, "**pmi_init",
"**pmi_init %d", pmi_errno);
}
/* Set the rank */
pmi_errno = PMI_Get_rank(&global_info->pg_rank);
if (pmi_errno != PMI_SUCCESS) {
MPIU_ERR_SETANDJUMP1(mpi_errno, MPI_ERR_OTHER, "**pmi_get_rank",
"**pmi_get_rank %d", pmi_errno);
}
/* Set the progexx group size */
pmi_errno = PMI_Get_size(&global_info->pg_size);
if (pmi_errno != 0) {
MPIU_ERR_SETANDJUMP1(mpi_errno, MPI_ERR_OTHER, "**pmi_get_size",
"**pmi_get_size %d", pmi_errno);
}
/* -------------------------------------- From InitPG in mvapich2/trunk/src/mpid/ch3/src/mpid_init.c
pmi_errno = PMI_Get_appnum(&appnum);
if (pmi_errno != PMI_SUCCESS) {
MPIU_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_get_appnum",
"**pmi_get_appnum %d", pmi_errno);
}
/ * Note that if pmi is not availble, the value of MPI_APPNUM is
not set * /
if (appnum != -1) {
MPIR_Process.attrs.appnum = appnum;
}
*/
/* Now, initialize the process group information with PMI calls */
/*
* Get the process group id
*/
pmi_errno = PMI_KVS_Get_name_length_max(&pg_id_sz);
if (pmi_errno != PMI_SUCCESS) {
/*
* I don't believe that MPICH2 has updated the error message for this
* yet.
*/
MPIU_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER,
"**pmi_get_id_length_max",
"**pmi_get_id_length_max %d", pmi_errno);
}
/* This memory will be freed by the PG_Destroy if there is an error */
pg_id = MPIU_Malloc(pg_id_sz + 1);
if (pg_id == NULL) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER, "**nomem");
}
/* Note in the singleton init case, the pg_id is a dummy.
We'll want to replace this value if we join a
Process manager */
pmi_errno = PMI_KVS_Get_my_name(pg_id, pg_id_sz);
if (pmi_errno != PMI_SUCCESS) {
/*
* I don't believe the MPICH2 team has updated the error message for
* this change yet.
*/
MPIU_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_get_id",
"**pmi_get_id %d", pmi_errno);
}
/*
* Create a new structure to track the process group for our MPI_COMM_WORLD
*/
mpi_errno = MPIDI_PG_Create(pg_size, pg_id, &pg);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER, "**dev|pg_create");
}
MPIDI_PG_InitConnKVS( pg );
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_PMI_INIT);
return mpi_errno;
fn_fail:
if (pg) {
MPIDI_PG_Destroy( pg );
}
goto fn_exit;
}
int MPIDI_VCRT_Release(struct MPIDI_VCRT *vcrt, int isDisconnect )
{
int in_use;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_VCRT_RELEASE);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_VCRT_RELEASE);
MPIU_Object_release_ref(vcrt, &in_use);
MPIU_DBG_MSG_FMT(REFCOUNT,TYPICAL,(MPIU_DBG_FDEST, "Decr VCRT %p ref count",vcrt));
/* If this VC reference table is no longer in use, we can
decrement the reference count of each of the VCs. If the
count on the VCs goes to zero, then we can decrement the
ref count on the process group and so on.
*/
if (!in_use) {
int i, inuse;
for (i = 0; i < vcrt->size; i++)
{
MPIDI_VC_t * const vc = vcrt->vcr_table[i];
MPIDI_VC_release_ref(vc, &in_use);
/* Dynamic connections start with a refcount of 2 instead of 1.
* That way we can distinguish between an MPI_Free and an
* MPI_Comm_disconnect. */
/* XXX DJG FIXME-MT should we be checking this? */
/* probably not, need to do something like the following instead: */
#if 0
if (isDisconnect) {
MPIU_Assert(in_use);
/* FIXME this is still bogus, the VCRT may contain a mix of
* dynamic and non-dynamic VCs, so the ref_count isn't
* guaranteed to have started at 2. The best thing to do might
* be to avoid overloading the reference counting this way and
* use a separate check for dynamic VCs (another flag? compare
* PGs?) */
MPIU_Object_release_ref(vc, &in_use);
}
#endif
if (isDisconnect && MPIU_Object_get_ref(vc) == 1) {
MPIDI_VC_release_ref(vc, &in_use);
}
if (!in_use)
{
/* If the VC is myself then skip the close message */
if (vc->pg == MPIDI_Process.my_pg &&
vc->pg_rank == MPIDI_Process.my_pg_rank)
{
MPIDI_PG_release_ref(vc->pg, &inuse);
if (inuse == 0)
{
MPIDI_PG_Destroy(vc->pg);
}
continue;
}
/* FIXME: the correct test is ACTIVE or REMOTE_CLOSE */
/*if (vc->state != MPIDI_VC_STATE_INACTIVE) { */
if (vc->state == MPIDI_VC_STATE_ACTIVE ||
vc->state == MPIDI_VC_STATE_REMOTE_CLOSE)
{
MPIDI_CH3U_VC_SendClose( vc, i );
}
else
{
MPIDI_PG_release_ref(vc->pg, &inuse);
if (inuse == 0)
{
MPIDI_PG_Destroy(vc->pg);
}
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"vc=%p: not sending a close to %d, vc in state %s",
vc, i, MPIDI_VC_GetStateString(vc->state)));
}
/* NOTE: we used to * MPIDI_CH3_VC_Destroy(&(pg->vct[i])))
here but that is incorrect. According to the standard, it's
entirely possible (likely even) that this VC might still be
connected. VCs are now destroyed when the PG that "owns"
them is destroyed (see MPIDI_PG_Destroy). [goodell@ 2008-06-13] */
}
}
MPIU_Free(vcrt);
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_VCRT_RELEASE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPIDI_CH3U_Handle_connection - handle connection event
Input Parameters:
+ vc - virtual connection
. event - connection event
NOTE:
This routine is used to transition the VC state.
The only events currently handled are TERMINATED events. This
routine should be called (with TERMINATED) whenever a connection is
terminated whether normally (in MPIDI_CH3_Connection_terminate() ),
or abnormally.
FIXME: Currently state transitions resulting from receiving CLOSE
packets are performed in MPIDI_CH3_PktHandler_Close(). Perhaps that
should move here.
@*/
int MPIDI_CH3U_Handle_connection(MPIDI_VC_t * vc, MPIDI_VC_Event_t event)
{
int inuse;
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3U_HANDLE_CONNECTION);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3U_HANDLE_CONNECTION);
switch (event)
{
case MPIDI_VC_EVENT_TERMINATED:
{
switch (vc->state)
{
case MPIDI_VC_STATE_CLOSED:
/* Normal termination. */
MPIDI_CHANGE_VC_STATE(vc, INACTIVE);
/* MT: this is not thread safe */
MPIDI_Outstanding_close_ops -= 1;
MPL_DBG_MSG_D(MPIDI_CH3_DBG_DISCONNECT,TYPICAL,
"outstanding close operations = %d", MPIDI_Outstanding_close_ops);
if (MPIDI_Outstanding_close_ops == 0)
{
MPIDI_CH3_Progress_signal_completion();
mpi_errno = MPIDI_CH3_Channel_close();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
break;
case MPIDI_VC_STATE_INACTIVE:
/* VC was terminated before it was activated.
This can happen if a failed process was
detected before the process used the VC. */
MPL_DBG_MSG(MPIDI_CH3_DBG_DISCONNECT,TYPICAL, "VC terminated before it was activated. We probably got a failed"
" process notification.");
MPIDI_CH3U_Complete_posted_with_error(vc);
++MPIDI_Failed_vc_count;
MPIDI_CHANGE_VC_STATE(vc, MORIBUND);
break;
case MPIDI_VC_STATE_ACTIVE:
case MPIDI_VC_STATE_REMOTE_CLOSE:
/* This is a premature termination. This process
has not started the close protocol. There may
be outstanding sends or receives on the local
side, remote side or both. */
MPL_DBG_MSG(MPIDI_CH3_DBG_DISCONNECT,TYPICAL, "Connection closed prematurely.");
MPIDI_CH3U_Complete_posted_with_error(vc);
++MPIDI_Failed_vc_count;
MPIDU_Ftb_publish_vc(MPIDU_FTB_EV_UNREACHABLE, vc);
MPIDI_CHANGE_VC_STATE(vc, MORIBUND);
break;
case MPIDI_VC_STATE_LOCAL_CLOSE:
/* This is a premature termination. This process
has started the close protocol, but hasn't
received a CLOSE packet from the remote side.
This process may not have been able to send the
CLOSE ack=F packet to the remote side. There
may be outstanding sends or receives on the
local or remote sides. */
case MPIDI_VC_STATE_CLOSE_ACKED:
/* This is a premature termination. Both sides
have started the close protocol. This process
has received CLOSE ack=F, but not CLOSE ack=t.
This process may not have been able to send
CLOSE ack=T. There should not be any
outstanding sends or receives on either
side. */
MPL_DBG_MSG_D(MPIDI_CH3_DBG_DISCONNECT,TYPICAL, "Connection closed prematurely during close protocol. "
"Outstanding close operations = %d", MPIDI_Outstanding_close_ops);
MPIDI_CH3U_Complete_posted_with_error(vc);
++MPIDI_Failed_vc_count;
MPIDU_Ftb_publish_vc(MPIDU_FTB_EV_UNREACHABLE, vc);
MPIDI_CHANGE_VC_STATE(vc, MORIBUND);
/* MT: this is not thread safe */
MPIDI_Outstanding_close_ops -= 1;
if (MPIDI_Outstanding_close_ops == 0) {
MPIDI_CH3_Progress_signal_completion();
mpi_errno = MPIDI_CH3_Channel_close();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
break;
default:
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_DISCONNECT,TYPICAL, "Unhandled connection state %d when closing connection",vc->state);
mpi_errno = MPIR_Err_create_code(
MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__,
MPI_ERR_INTERN, "**ch3|unhandled_connection_state",
"**ch3|unhandled_connection_state %p %d", vc, vc->state);
goto fn_fail;
break;
}
}
/* FIXME: Decrement the reference count? Who increments? */
/* FIXME: The reference count is often already 0. But
not always */
/* MPIR_Object_set_ref(vc, 0); ??? */
/*
* FIXME: The VC used in connect accept has a NULL
* process group
*/
/* XXX DJG FIXME-MT should we be checking this ref_count? */
if (vc->pg != NULL && (MPIR_Object_get_ref(vc) == 0))
{
/* FIXME: Who increments the reference count that
this is decrementing? */
/* When the reference count for a vc becomes zero,
decrement the reference count
of the associated process group. */
/* FIXME: This should be done when the reference
count of the vc is first decremented */
MPIDI_PG_release_ref(vc->pg, &inuse);
if (inuse == 0) {
MPIDI_PG_Destroy(vc->pg);
}
}
break;
}
default:
{
break;
}
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3U_HANDLE_CONNECTION);
return mpi_errno;
fn_fail:
goto fn_exit;
}
