static int ckpt_cb(void *arg)
{
int rc, ret;
const struct cr_restart_info* ri;
if (MPIDI_Process.my_pg_rank == 0) {
MPIDI_nem_ckpt_start_checkpoint = TRUE;
/* poke the progress engine in case we're waiting in a blocking recv */
MPIDI_CH3_Progress_signal_completion();
}
do {
ret = sem_wait(&ckpt_sem);
} while (ret == -1 && errno == EINTR);
CHECK_ERR(ret, "sem_wait");
if (MPID_nem_netmod_func->ckpt_precheck) {
int mpi_errno;
mpi_errno = MPID_nem_netmod_func->ckpt_precheck();
CHECK_ERR_MPI(mpi_errno, mpi_errno, "ckpt_precheck failed");
}
rc = cr_checkpoint(0);
if (rc < 0) {
ckpt_result = CKPT_ERROR;
} else if (rc) {
ckpt_result = CKPT_RESTART;
ri = cr_get_restart_info();
CHECK_ERR(!ri, "cr_get_restart_info");
ret = restore_env(ri->requester, MPIDI_Process.my_pg_rank);
CHECK_ERR(ret, "restore_env");
ret = restore_stdinouterr(MPIDI_Process.my_pg_rank);
CHECK_ERR(ret, "restore_stdinouterr");
ret = reinit_pmi();
CHECK_ERR(ret, "reinit_pmi");
if (MPID_nem_netmod_func->ckpt_restart) {
int mpi_errno;
mpi_errno = MPID_nem_netmod_func->ckpt_restart();
CHECK_ERR_MPI(mpi_errno, mpi_errno, "ckpt_restart failed");
}
} else {
ckpt_result = CKPT_CONTINUE;
if (MPID_nem_netmod_func->ckpt_continue) {
int mpi_errno;
mpi_errno = MPID_nem_netmod_func->ckpt_continue();
CHECK_ERR_MPI(mpi_errno, mpi_errno, "ckpt_continue failed");
}
}
do {
ret = sem_post(&cont_sem);
} while (ret == -1 && errno == EINTR);
CHECK_ERR(ret, "sem_post");
return 0;
}
static void sigusr1_handler(int sig)
{
++sigusr1_count;
/* poke the progress engine in case we're waiting in a blocking recv */
MPIDI_CH3_Progress_signal_completion();
if (prev_sighandler)
prev_sighandler(sig);
}
static int pkt_ckpt_marker_handler(MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt, MPIDI_msg_sz_t *buflen, MPID_Request **req)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_pkt_ckpt_marker_t * const ckpt_pkt = (MPID_nem_pkt_ckpt_marker_t *)pkt;
MPIDI_STATE_DECL(MPID_STATE_PKT_CKPT_MARKER_HANDLER);
MPIDI_FUNC_ENTER(MPID_STATE_PKT_CKPT_MARKER_HANDLER);
if (!checkpointing) {
mpi_errno = MPIDI_nem_ckpt_start();
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPIU_Assert(current_wave == ckpt_pkt->wave);
--marker_count;
/* We're checkpointing the shared memory region, so we don't need
to flush the channels between local processes, only remote
processes */
if (marker_count == 0) {
MPIDI_nem_ckpt_finish_checkpoint = TRUE;
/* make sure we break out of receive loop into progress loop */
MPIDI_CH3_Progress_signal_completion();
}
*buflen = sizeof(MPIDI_CH3_Pkt_t);
*req = NULL;
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_PKT_CKPT_MARKER_HANDLER);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int MPIDI_CH3I_Progress_handle_sock_event(MPIDU_Sock_event_t * event)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_PROGRESS_HANDLE_SOCK_EVENT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_PROGRESS_HANDLE_SOCK_EVENT);
MPIU_DBG_MSG_D(CH3_OTHER,VERBOSE,"Socket event of type %d", event->op_type );
switch (event->op_type)
{
case MPIDU_SOCK_OP_READ:
{
MPIDI_CH3I_Connection_t * conn =
(MPIDI_CH3I_Connection_t *) event->user_ptr;
MPID_Request * rreq = conn->recv_active;
/* --BEGIN ERROR HANDLING-- */
if (event->error != MPI_SUCCESS)
{
/* FIXME: the following should be handled by the close
protocol */
if (MPIR_ERR_GET_CLASS(event->error) != MPIDU_SOCK_ERR_CONN_CLOSED) {
mpi_errno = event->error;
MPIU_ERR_POP(mpi_errno);
}
break;
}
/* --END ERROR HANDLING-- */
if (conn->state == CONN_STATE_CONNECTED)
{
if (conn->recv_active == NULL)
{
MPIDI_msg_sz_t buflen = sizeof (MPIDI_CH3_Pkt_t);
MPIU_Assert(conn->pkt.type < MPIDI_CH3_PKT_END_CH3);
mpi_errno = pktArray[conn->pkt.type]( conn->vc, &conn->pkt,
&buflen, &rreq );
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
MPIU_Assert(buflen == sizeof (MPIDI_CH3_Pkt_t));
if (rreq == NULL)
{
if (conn->state != CONN_STATE_CLOSING)
{
/* conn->recv_active = NULL; --
already set to NULL */
mpi_errno = connection_post_recv_pkt(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
}
}
else
{
mpi_errno = ReadMoreData( conn, rreq );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
}
else /* incoming data */
{
int (*reqFn)(MPIDI_VC_t *, MPID_Request *, int *);
int complete;
reqFn = rreq->dev.OnDataAvail;
if (!reqFn) {
MPIU_Assert(MPIDI_Request_get_type(rreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
MPIDI_CH3U_Request_complete(rreq);
complete = TRUE;
}
else {
mpi_errno = reqFn( conn->vc, rreq, &complete );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (complete)
{
conn->recv_active = NULL;
mpi_errno = connection_post_recv_pkt(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
}
else /* more data to be read */
{
mpi_errno = ReadMoreData( conn, rreq );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
}
}
else if (conn->state == CONN_STATE_OPEN_LRECV_DATA)
{
mpi_errno = MPIDI_CH3_Sockconn_handle_connopen_event( conn );
if (mpi_errno) { MPIU_ERR_POP( mpi_errno ); }
}
else /* Handling some internal connection establishment or
tear down packet */
{
mpi_errno = MPIDI_CH3_Sockconn_handle_conn_event( conn );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
break;
}
/* END OF SOCK_OP_READ */
case MPIDU_SOCK_OP_WRITE:
{
MPIDI_CH3I_Connection_t * conn =
(MPIDI_CH3I_Connection_t *) event->user_ptr;
/* --BEGIN ERROR HANDLING-- */
if (event->error != MPI_SUCCESS) {
mpi_errno = event->error;
MPIU_ERR_POP(mpi_errno);
}
/* --END ERROR HANDLING-- */
if (conn->send_active)
{
MPID_Request * sreq = conn->send_active;
int (*reqFn)(MPIDI_VC_t *, MPID_Request *, int *);
int complete;
reqFn = sreq->dev.OnDataAvail;
if (!reqFn) {
MPIU_Assert(MPIDI_Request_get_type(sreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
MPIDI_CH3U_Request_complete(sreq);
complete = TRUE;
}
else {
mpi_errno = reqFn( conn->vc, sreq, &complete );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (complete)
{
mpi_errno = connection_pop_sendq_req(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
}
else /* more data to send */
{
for(;;)
{
MPID_IOV * iovp;
MPIU_Size_t nb;
iovp = sreq->dev.iov;
mpi_errno = MPIDU_Sock_writev(conn->sock, iovp, sreq->dev.iov_count, &nb);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER,
"**ch3|sock|immedwrite", "ch3|sock|immedwrite %p %p %p",
sreq, conn, conn->vc);
goto fn_fail;
}
/* --END ERROR HANDLING-- */
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,
(MPIU_DBG_FDEST,"immediate writev, vc=%p, sreq=0x%08x, nb=" MPIDI_MSG_SZ_FMT,
conn->vc, sreq->handle, nb));
if (nb > 0 && adjust_iov(&iovp, &sreq->dev.iov_count, nb))
{
reqFn = sreq->dev.OnDataAvail;
if (!reqFn) {
MPIU_Assert(MPIDI_Request_get_type(sreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
MPIDI_CH3U_Request_complete(sreq);
complete = TRUE;
}
else {
mpi_errno = reqFn( conn->vc, sreq, &complete );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (complete)
{
mpi_errno = connection_pop_sendq_req(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
break;
}
}
else
{
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,
(MPIU_DBG_FDEST,"posting writev, vc=%p, conn=%p, sreq=0x%08x",
conn->vc, conn, sreq->handle));
mpi_errno = MPIDU_Sock_post_writev(conn->sock, iovp, sreq->dev.iov_count, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**ch3|sock|postwrite",
"ch3|sock|postwrite %p %p %p", sreq, conn, conn->vc);
goto fn_fail;
}
/* --END ERROR HANDLING-- */
break;
}
}
}
}
else /* finished writing internal packet header */
{
/* the connection is not active yet */
mpi_errno = MPIDI_CH3_Sockconn_handle_connwrite( conn );
if (mpi_errno) { MPIU_ERR_POP( mpi_errno ); }
}
break;
}
/* END OF SOCK_OP_WRITE */
case MPIDU_SOCK_OP_ACCEPT:
{
mpi_errno = MPIDI_CH3_Sockconn_handle_accept_event();
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
break;
}
case MPIDU_SOCK_OP_CONNECT:
{
mpi_errno = MPIDI_CH3_Sockconn_handle_connect_event(
(MPIDI_CH3I_Connection_t *) event->user_ptr,
event->error );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
break;
}
case MPIDU_SOCK_OP_CLOSE:
{
mpi_errno = MPIDI_CH3_Sockconn_handle_close_event(
(MPIDI_CH3I_Connection_t *) event->user_ptr );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
break;
}
case MPIDU_SOCK_OP_WAKEUP:
{
MPIDI_CH3_Progress_signal_completion();
/* MPIDI_CH3I_progress_completion_count++; */
break;
}
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_PROGRESS_HANDLE_SOCK_EVENT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int MPIDI_CH3i_Progress_wait(MPID_Progress_state * progress_state)
{
MPIDU_Sock_event_t event;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_PROGRESS_WAIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_PROGRESS_WAIT);
/*
* MT: the following code will be needed if progress can occur between
* MPIDI_CH3_Progress_start() and
* MPIDI_CH3_Progress_wait(), or iterations of MPIDI_CH3_Progress_wait().
*
* This is presently not possible, and thus the code is commented out.
*/
# if 0
/* FIXME: Was (USE_THREAD_IMPL == MPICH_THREAD_IMPL_NOT_IMPLEMENTED),
which really meant not-using-global-mutex-thread model . This
was true for the single threaded case, but was probably not intended
for that case*/
{
if (progress_state->ch.completion_count != MPIDI_CH3I_progress_completion_count)
{
goto fn_exit;
}
}
# endif
# ifdef MPICH_IS_THREADED
MPIU_THREAD_CHECK_BEGIN
{
if (MPIDI_CH3I_progress_blocked == TRUE)
{
/*
* Another thread is already blocking in the progress engine.
*
* MT: Another thread is already blocking in poll. Right now,
* calls to MPIDI_CH3_Progress_wait() are effectively
* serialized by the device. The only way another thread may
* enter this function is if MPIDU_Sock_wait() blocks. If
* this changes, a flag other than MPIDI_CH3I_Progress_blocked
* may be required to determine if another thread is in
* the progress engine.
*/
MPIDI_CH3I_Progress_delay(MPIDI_CH3I_progress_completion_count);
goto fn_exit;
}
}
MPIU_THREAD_CHECK_END
# endif
do
{
int made_progress = FALSE;
/* make progress on NBC schedules, must come before we block on sock_wait */
mpi_errno = MPIDU_Sched_progress(&made_progress);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (made_progress) {
MPIDI_CH3_Progress_signal_completion();
break;
}
# ifdef MPICH_IS_THREADED
/* The logic for this case is just complicated enough that
we write separate code for each possibility */
# ifdef HAVE_RUNTIME_THREADCHECK
if (MPIR_ThreadInfo.isThreaded) {
MPIDI_CH3I_progress_blocked = TRUE;
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
MPIDI_CH3I_progress_blocked = FALSE;
MPIDI_CH3I_progress_wakeup_signalled = FALSE;
}
else {
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
}
# else
MPIDI_CH3I_progress_blocked = TRUE;
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
MPIDI_CH3I_progress_blocked = FALSE;
MPIDI_CH3I_progress_wakeup_signalled = FALSE;
# endif /* HAVE_RUNTIME_THREADCHECK */
# else
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
# endif
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPIU_Assert(MPIR_ERR_GET_CLASS(mpi_errno) != MPIDU_SOCK_ERR_TIMEOUT);
MPIU_ERR_SET(mpi_errno,MPI_ERR_OTHER,"**progress_sock_wait");
goto fn_fail;
}
/* --END ERROR HANDLING-- */
mpi_errno = MPIDI_CH3I_Progress_handle_sock_event(&event);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|sock|handle_sock_event");
}
}
while (progress_state->ch.completion_count == MPIDI_CH3I_progress_completion_count);
/*
* We could continue to call MPIU_Sock_wait in a non-blocking fashion
* and process any other events; however, this would not
* give the application a chance to post new receives, and thus could
* result in an increased number of unexpected messages
* that would need to be buffered.
*/
# if MPICH_IS_THREADED
{
/*
* Awaken any threads which are waiting for the progress that just
* occurred
*/
MPIDI_CH3I_Progress_continue(MPIDI_CH3I_progress_completion_count);
}
# endif
fn_exit:
/*
* Reset the progress state so it is fresh for the next iteration
*/
progress_state->ch.completion_count = MPIDI_CH3I_progress_completion_count;
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_PROGRESS_WAIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_CH3_PktHandler_RndvReqToSend( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt,
MPIDI_msg_sz_t *buflen, MPID_Request **rreqp )
{
MPID_Request * rreq;
int found;
MPIDI_CH3_Pkt_rndv_req_to_send_t * rts_pkt = &pkt->rndv_req_to_send;
int mpi_errno = MPI_SUCCESS;
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"received rndv RTS pkt, sreq=0x%08x, rank=%d, tag=%d, context=%d, data_sz=" MPIDI_MSG_SZ_FMT,
rts_pkt->sender_req_id, rts_pkt->match.parts.rank,
rts_pkt->match.parts.tag,
rts_pkt->match.parts.context_id, rts_pkt->data_sz));
MPIU_DBG_MSGPKT(vc,rts_pkt->match.parts.tag,rts_pkt->match.parts.context_id,
rts_pkt->match.parts.rank,rts_pkt->data_sz,
"ReceivedRndv");
MPIU_THREAD_CS_ENTER(MSGQUEUE,);
rreq = MPIDI_CH3U_Recvq_FDP_or_AEU(&rts_pkt->match, &found);
MPIU_ERR_CHKANDJUMP1(!rreq, mpi_errno,MPI_ERR_OTHER, "**nomemreq", "**nomemuereq %d", MPIDI_CH3U_Recvq_count_unexp());
set_request_info(rreq, rts_pkt, MPIDI_REQUEST_RNDV_MSG);
MPIU_THREAD_CS_EXIT(MSGQUEUE,);
*buflen = sizeof(MPIDI_CH3_Pkt_t);
if (found)
{
MPID_Request * cts_req;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_rndv_clr_to_send_t * cts_pkt = &upkt.rndv_clr_to_send;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"posted request found");
/* FIXME: What if the receive user buffer is not big enough to
hold the data about to be cleared for sending? */
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending rndv CTS packet");
MPIDI_Pkt_init(cts_pkt, MPIDI_CH3_PKT_RNDV_CLR_TO_SEND);
cts_pkt->sender_req_id = rts_pkt->sender_req_id;
cts_pkt->receiver_req_id = rreq->handle;
MPIU_THREAD_CS_ENTER(CH3COMM,vc);
mpi_errno = MPIU_CALL(MPIDI_CH3,iStartMsg(vc, cts_pkt,
sizeof(*cts_pkt), &cts_req));
MPIU_THREAD_CS_EXIT(CH3COMM,vc);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|ctspkt");
}
if (cts_req != NULL) {
MPID_Request_release(cts_req);
}
}
else
{
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"unexpected request allocated");
/*
* A MPID_Probe() may be waiting for the request we just
* inserted, so we need to tell the progress engine to exit.
*
* FIXME: This will cause MPID_Progress_wait() to return to the
* MPI layer each time an unexpected RTS packet is
* received. MPID_Probe() should atomically increment a
* counter and MPIDI_CH3_Progress_signal_completion()
* should only be called if that counter is greater than zero.
*/
MPIDI_CH3_Progress_signal_completion();
}
*rreqp = NULL;
fn_fail:
return mpi_errno;
}
/*@
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;
}
int MPIDI_CH3_PktHandler_RndvReqToSend( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt,
intptr_t *buflen, MPIR_Request **rreqp )
{
MPIR_Request * rreq;
int found;
MPIDI_CH3_Pkt_rndv_req_to_send_t * rts_pkt = &pkt->rndv_req_to_send;
int mpi_errno = MPI_SUCCESS;
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_OTHER,VERBOSE,(MPL_DBG_FDEST,
"received rndv RTS pkt, sreq=0x%08x, rank=%d, tag=%d, context=%d, data_sz=%" PRIdPTR,
rts_pkt->sender_req_id, rts_pkt->match.parts.rank,
rts_pkt->match.parts.tag,
rts_pkt->match.parts.context_id, rts_pkt->data_sz));
MPL_DBG_MSGPKT(vc,rts_pkt->match.parts.tag,rts_pkt->match.parts.context_id,
rts_pkt->match.parts.rank,rts_pkt->data_sz,
"ReceivedRndv");
MPID_THREAD_CS_ENTER(POBJ, MPIR_THREAD_POBJ_MSGQ_MUTEX);
rreq = MPIDI_CH3U_Recvq_FDP_or_AEU(&rts_pkt->match, &found);
MPIR_ERR_CHKANDJUMP1(!rreq, mpi_errno,MPI_ERR_OTHER, "**nomemreq", "**nomemuereq %d", MPIDI_CH3U_Recvq_count_unexp());
/* If the completion counter is 0, that means that the communicator to
* which this message is being sent has been revoked and we shouldn't
* bother finishing this. */
if (!found && MPIR_cc_get(rreq->cc) == 0) {
*rreqp = NULL;
goto fn_fail;
}
set_request_info(rreq, rts_pkt, MPIDI_REQUEST_RNDV_MSG);
MPID_THREAD_CS_EXIT(POBJ, MPIR_THREAD_POBJ_MSGQ_MUTEX);
*buflen = sizeof(MPIDI_CH3_Pkt_t);
if (found)
{
MPIR_Request * cts_req;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_rndv_clr_to_send_t * cts_pkt = &upkt.rndv_clr_to_send;
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"posted request found");
/* FIXME: What if the receive user buffer is not big enough to
hold the data about to be cleared for sending? */
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"sending rndv CTS packet");
MPIDI_Pkt_init(cts_pkt, MPIDI_CH3_PKT_RNDV_CLR_TO_SEND);
cts_pkt->sender_req_id = rts_pkt->sender_req_id;
cts_pkt->receiver_req_id = rreq->handle;
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = MPIDI_CH3_iStartMsg(vc, cts_pkt, sizeof(*cts_pkt), &cts_req);
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|ctspkt");
}
if (cts_req != NULL) {
MPIR_Request_free(cts_req);
}
}
else
{
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"unexpected request allocated");
/*
* A MPID_Probe() may be waiting for the request we just
* inserted, so we need to tell the progress engine to exit.
*
* FIXME: This will cause MPID_Progress_wait() to return to the
* MPI layer each time an unexpected RTS packet is
* received. MPID_Probe() should atomically increment a
* counter and MPIDI_CH3_Progress_signal_completion()
* should only be called if that counter is greater than zero.
*/
MPIDI_CH3_Progress_signal_completion();
}
*rreqp = NULL;
fn_fail:
return mpi_errno;
}
