void MPIDI_CH3_Request_destroy(MPID_Request * req)
{
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3_REQUEST_DESTROY);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3_REQUEST_DESTROY);
MPIU_DBG_MSG_P(CH3_CHANNEL,VERBOSE,
"freeing request, handle=0x%08x", req->handle);
#ifdef MPICH_DBG_OUTPUT
/*MPIU_Assert(HANDLE_GET_MPI_KIND(req->handle) == MPID_REQUEST);*/
if (HANDLE_GET_MPI_KIND(req->handle) != MPID_REQUEST)
{
int mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER,
"**invalid_handle", "**invalid_handle %d", req->handle);
MPID_Abort(MPIR_Process.comm_world, mpi_errno, -1, NULL);
}
/* XXX DJG FIXME should we be checking this? */
/*MPIU_Assert(req->ref_count == 0);*/
if (req->ref_count != 0)
{
int mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER,
"**invalid_refcount", "**invalid_refcount %d", req->ref_count);
MPID_Abort(MPIR_Process.comm_world, mpi_errno, -1, NULL);
}
#endif
/* FIXME: We need a better way to handle these so that we
do not always need to initialize these fields and check them
when we destroy a request */
/* FIXME: We need a way to call these routines ONLY when the
related ref count has become zero. */
if (req->comm != NULL) {
MPIR_Comm_release(req->comm, 0);
}
if (req->greq_fns != NULL) {
MPIU_Free(req->greq_fns);
}
if (req->dev.datatype_ptr != NULL) {
MPID_Datatype_release(req->dev.datatype_ptr);
}
if (req->dev.segment_ptr != NULL) {
MPID_Segment_free(req->dev.segment_ptr);
}
if (MPIDI_Request_get_srbuf_flag(req)) {
MPIDI_CH3U_SRBuf_free(req);
}
MPIU_Handle_obj_free(&MPID_Request_mem, req);
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3_REQUEST_DESTROY);
}
int MPIDI_CH3U_Request_load_recv_iov(MPID_Request * const rreq)
{
MPI_Aint last;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_REQUEST_LOAD_RECV_IOV);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_REQUEST_LOAD_RECV_IOV);
if (rreq->dev.segment_first < rreq->dev.segment_size)
{
/* still reading data that needs to go into the user buffer */
if (MPIDI_Request_get_type(rreq) != MPIDI_REQUEST_TYPE_ACCUM_RECV &&
MPIDI_Request_get_type(rreq) != MPIDI_REQUEST_TYPE_GET_ACCUM_RECV &&
MPIDI_Request_get_srbuf_flag(rreq))
{
MPIDI_msg_sz_t data_sz;
MPIDI_msg_sz_t tmpbuf_sz;
/* Once a SRBuf is in use, we continue to use it since a small
amount of data may already be present at the beginning
of the buffer. This data is left over from the previous unpack,
most like a result of alignment issues. NOTE: we
could force the use of the SRBuf only
when (rreq->dev.tmpbuf_off > 0)... */
data_sz = rreq->dev.segment_size - rreq->dev.segment_first -
rreq->dev.tmpbuf_off;
MPIU_Assert(data_sz > 0);
tmpbuf_sz = rreq->dev.tmpbuf_sz - rreq->dev.tmpbuf_off;
if (data_sz > tmpbuf_sz)
{
data_sz = tmpbuf_sz;
}
rreq->dev.iov[0].MPID_IOV_BUF =
(MPID_IOV_BUF_CAST)((char *) rreq->dev.tmpbuf +
rreq->dev.tmpbuf_off);
rreq->dev.iov[0].MPID_IOV_LEN = data_sz;
rreq->dev.iov_offset = 0;
rreq->dev.iov_count = 1;
MPIU_Assert(rreq->dev.segment_first + data_sz +
rreq->dev.tmpbuf_off <= rreq->dev.recv_data_sz);
if (rreq->dev.segment_first + data_sz + rreq->dev.tmpbuf_off ==
rreq->dev.recv_data_sz)
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"updating rreq to read the remaining data into the SRBuf");
rreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_UnpackSRBufComplete;
}
else
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"updating rreq to read more data into the SRBuf");
rreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_UnpackSRBufReloadIOV;
}
goto fn_exit;
}
last = rreq->dev.segment_size;
rreq->dev.iov_count = MPID_IOV_LIMIT;
rreq->dev.iov_offset = 0;
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"pre-upv: first=" MPIDI_MSG_SZ_FMT ", last=" MPIDI_MSG_SZ_FMT ", iov_n=%d",
rreq->dev.segment_first, last, rreq->dev.iov_count));
MPIU_Assert(rreq->dev.segment_first < last);
MPIU_Assert(last > 0);
MPID_Segment_unpack_vector(rreq->dev.segment_ptr,
rreq->dev.segment_first,
&last, &rreq->dev.iov[0], &rreq->dev.iov_count);
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"post-upv: first=" MPIDI_MSG_SZ_FMT ", last=" MPIDI_MSG_SZ_FMT ", iov_n=%d, iov_offset=%lld",
rreq->dev.segment_first, last, rreq->dev.iov_count, (long long)rreq->dev.iov_offset));
MPIU_Assert(rreq->dev.iov_count >= 0 && rreq->dev.iov_count <=
MPID_IOV_LIMIT);
/* --BEGIN ERROR HANDLING-- */
if (rreq->dev.iov_count == 0)
{
/* If the data can't be unpacked, the we have a mis-match between
the datatype and the amount of data received. Adjust
the segment info so that the remaining data is received and
thrown away. */
rreq->status.MPI_ERROR = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_TYPE,
"**dtypemismatch", 0);
MPIR_STATUS_SET_COUNT(rreq->status, rreq->dev.segment_first);
rreq->dev.segment_size = rreq->dev.segment_first;
mpi_errno = MPIDI_CH3U_Request_load_recv_iov(rreq);
goto fn_exit;
}
else
{
MPIU_Assert(rreq->dev.iov_offset < rreq->dev.iov_count);
}
/* --END ERROR HANDLING-- */
if (last == rreq->dev.recv_data_sz)
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"updating rreq to read the remaining data directly into the user buffer");
/* Eventually, use OnFinal for this instead */
rreq->dev.OnDataAvail = rreq->dev.OnFinal;
}
else if (MPIDI_Request_get_type(rreq) == MPIDI_REQUEST_TYPE_ACCUM_RECV ||
MPIDI_Request_get_type(rreq) == MPIDI_REQUEST_TYPE_GET_ACCUM_RECV ||
(last == rreq->dev.segment_size ||
(last - rreq->dev.segment_first) / rreq->dev.iov_count >= MPIDI_IOV_DENSITY_MIN))
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"updating rreq to read more data directly into the user buffer");
rreq->dev.segment_first = last;
rreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_ReloadIOV;
}
else
{
/* Too little data would have been received using an IOV.
We will start receiving data into a SRBuf and unpacking it
later. */
MPIU_Assert(MPIDI_Request_get_srbuf_flag(rreq) == FALSE);
MPIDI_CH3U_SRBuf_alloc(rreq,
rreq->dev.segment_size - rreq->dev.segment_first);
rreq->dev.tmpbuf_off = 0;
/* --BEGIN ERROR HANDLING-- */
if (rreq->dev.tmpbuf_sz == 0)
{
/* FIXME - we should drain the data off the pipe here, but we
don't have a buffer to drain it into. should this be
a fatal error? */
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,"SRBuf allocation failure");
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem",
"**nomem %d",
rreq->dev.segment_size - rreq->dev.segment_first);
rreq->status.MPI_ERROR = mpi_errno;
goto fn_exit;
}
/* --END ERROR HANDLING-- */
/* fill in the IOV using a recursive call */
mpi_errno = MPIDI_CH3U_Request_load_recv_iov(rreq);
}
}
else
{
/* receive and toss any extra data that does not fit in the user's
buffer */
MPIDI_msg_sz_t data_sz;
data_sz = rreq->dev.recv_data_sz - rreq->dev.segment_first;
if (!MPIDI_Request_get_srbuf_flag(rreq))
{
MPIDI_CH3U_SRBuf_alloc(rreq, data_sz);
/* --BEGIN ERROR HANDLING-- */
if (rreq->dev.tmpbuf_sz == 0)
{
MPIU_DBG_MSG(CH3_CHANNEL,TYPICAL,"SRBuf allocation failure");
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0);
rreq->status.MPI_ERROR = mpi_errno;
goto fn_exit;
}
/* --END ERROR HANDLING-- */
}
if (data_sz <= rreq->dev.tmpbuf_sz)
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"updating rreq to read overflow data into the SRBuf and complete");
rreq->dev.iov[0].MPID_IOV_LEN = data_sz;
MPIU_Assert(MPIDI_Request_get_type(rreq) == MPIDI_REQUEST_TYPE_RECV);
/* Eventually, use OnFinal for this instead */
rreq->dev.OnDataAvail = rreq->dev.OnFinal;
}
else
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"updating rreq to read overflow data into the SRBuf and reload IOV");
rreq->dev.iov[0].MPID_IOV_LEN = rreq->dev.tmpbuf_sz;
rreq->dev.segment_first += rreq->dev.tmpbuf_sz;
rreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_ReloadIOV;
}
rreq->dev.iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)rreq->dev.tmpbuf;
rreq->dev.iov_count = 1;
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_REQUEST_LOAD_RECV_IOV);
return mpi_errno;
}
int MPIDI_CH3U_Request_load_send_iov(MPID_Request * const sreq,
MPID_IOV * const iov, int * const iov_n)
{
MPI_Aint last;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_REQUEST_LOAD_SEND_IOV);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_REQUEST_LOAD_SEND_IOV);
MPIU_Assert(sreq->dev.segment_ptr != NULL);
last = sreq->dev.segment_size;
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"pre-pv: first=" MPIDI_MSG_SZ_FMT ", last=" MPIDI_MSG_SZ_FMT ", iov_n=%d",
sreq->dev.segment_first, last, *iov_n));
MPIU_Assert(sreq->dev.segment_first < last);
MPIU_Assert(last > 0);
MPIU_Assert(*iov_n > 0 && *iov_n <= MPID_IOV_LIMIT);
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first,
&last, iov, iov_n);
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"post-pv: first=" MPIDI_MSG_SZ_FMT ", last=" MPIDI_MSG_SZ_FMT ", iov_n=%d",
sreq->dev.segment_first, last, *iov_n));
MPIU_Assert(*iov_n > 0 && *iov_n <= MPID_IOV_LIMIT);
if (last == sreq->dev.segment_size)
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,"remaining data loaded into IOV");
sreq->dev.OnDataAvail = sreq->dev.OnFinal;
}
else if ((last - sreq->dev.segment_first) / *iov_n >= MPIDI_IOV_DENSITY_MIN)
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,"more data loaded into IOV");
sreq->dev.segment_first = last;
sreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_SendReloadIOV;
}
else
{
MPIDI_msg_sz_t data_sz;
int i, iov_data_copied;
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,"low density. using SRBuf.");
data_sz = sreq->dev.segment_size - sreq->dev.segment_first;
if (!MPIDI_Request_get_srbuf_flag(sreq))
{
MPIDI_CH3U_SRBuf_alloc(sreq, data_sz);
/* --BEGIN ERROR HANDLING-- */
if (sreq->dev.tmpbuf_sz == 0)
{
MPIU_DBG_MSG(CH3_CHANNEL,TYPICAL,"SRBuf allocation failure");
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem",
"**nomem %d", data_sz);
sreq->status.MPI_ERROR = mpi_errno;
goto fn_exit;
}
/* --END ERROR HANDLING-- */
}
iov_data_copied = 0;
for (i = 0; i < *iov_n; i++) {
MPIU_Memcpy((char*) sreq->dev.tmpbuf + iov_data_copied,
iov[i].MPID_IOV_BUF, iov[i].MPID_IOV_LEN);
iov_data_copied += iov[i].MPID_IOV_LEN;
}
sreq->dev.segment_first = last;
last = (data_sz <= sreq->dev.tmpbuf_sz - iov_data_copied) ?
sreq->dev.segment_size :
sreq->dev.segment_first + sreq->dev.tmpbuf_sz - iov_data_copied;
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"pre-pack: first=" MPIDI_MSG_SZ_FMT ", last=" MPIDI_MSG_SZ_FMT,
sreq->dev.segment_first, last));
MPID_Segment_pack(sreq->dev.segment_ptr, sreq->dev.segment_first,
&last, (char*) sreq->dev.tmpbuf + iov_data_copied);
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"post-pack: first=" MPIDI_MSG_SZ_FMT ", last=" MPIDI_MSG_SZ_FMT,
sreq->dev.segment_first, last));
iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)sreq->dev.tmpbuf;
iov[0].MPID_IOV_LEN = last - sreq->dev.segment_first + iov_data_copied;
*iov_n = 1;
if (last == sreq->dev.segment_size)
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,"remaining data packed into SRBuf");
sreq->dev.OnDataAvail = sreq->dev.OnFinal;
}
else
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,"more data packed into SRBuf");
sreq->dev.segment_first = last;
sreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_SendReloadIOV;
}
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_REQUEST_LOAD_SEND_IOV);
return mpi_errno;
}
int MPID_nem_send_iov(MPIDI_VC_t *vc, MPIR_Request **sreq_ptr, MPL_IOV *iov, int n_iov)
{
int mpi_errno = MPI_SUCCESS;
intptr_t data_sz;
int i;
int iov_data_copied;
MPIR_Request *sreq = *sreq_ptr;
MPL_IOV *data_iov = &iov[1]; /* iov of just the data, not the header */
int data_n_iov = n_iov - 1;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_SEND_IOV);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_SEND_IOV);
if (*sreq_ptr == NULL)
{
/* create a request */
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__UNDEFINED);
MPIR_Assert(sreq != NULL);
MPIR_Object_set_ref(sreq, 2);
sreq->kind = MPIR_REQUEST_KIND__SEND;
sreq->dev.OnDataAvail = 0;
}
data_sz = 0;
for (i = 0; i < data_n_iov; ++i)
data_sz += data_iov[i].MPL_IOV_LEN;
if (!MPIDI_Request_get_srbuf_flag(sreq))
{
MPIDI_CH3U_SRBuf_alloc(sreq, data_sz);
/* --BEGIN ERROR HANDLING-- */
if (sreq->dev.tmpbuf_sz == 0)
{
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,TYPICAL,"SRBuf allocation failure");
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0);
sreq->status.MPI_ERROR = mpi_errno;
goto fn_exit;
}
/* --END ERROR HANDLING-- */
}
MPIR_Assert(sreq->dev.tmpbuf_sz >= data_sz);
iov_data_copied = 0;
for (i = 0; i < data_n_iov; ++i) {
MPIR_Memcpy((char*) sreq->dev.tmpbuf + iov_data_copied, data_iov[i].MPL_IOV_BUF, data_iov[i].MPL_IOV_LEN);
iov_data_copied += data_iov[i].MPL_IOV_LEN;
}
mpi_errno = vc->ch.iSendContig(vc, sreq, iov[0].MPL_IOV_BUF, iov[0].MPL_IOV_LEN, sreq->dev.tmpbuf, data_sz);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
*sreq_ptr = sreq;
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_SEND_IOV);
return mpi_errno;
fn_fail:
goto fn_exit;
}
