ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

Just some more refactoring and cleanup of bfo PML.

Просмотр исходного кода 
This commit was SVN r23884.
Этот коммит содержится в:
Rolf vandeVaart 2010-10-12 13:34:35 +00:00
родитель e9a7fea42d
Коммит 44d7006f34
6 изменённых файлов: 208 добавлений и 214 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

3
ompi/mca/pml/bfo/pml_bfo.h
Просмотреть файл

@ -75,9 +75,6 @@ struct mca_pml_bfo_t {
/* List of pending fragments without a matching communicator */
opal_list_t non_existing_communicator_pending;
bool enabled;
/* BFO FAILOVER CODE - begin */
bool fast_failover;
/* BFO FAILOVER CODE - end */
char* allocator_name;
mca_allocator_base_module_t* allocator;
uint32_t unexpected_limit;

4
ompi/mca/pml/bfo/pml_bfo_component.c
Просмотреть файл

@ -121,10 +121,6 @@ static int mca_pml_bfo_component_open(void)
mca_pml_bfo.unexpected_limit =
mca_pml_bfo_param_register_int("unexpected_limit", 128);
/* BFO FAILOVER CODE - begin */
mca_pml_bfo.fast_failover =
mca_pml_bfo_param_register_int("fast_failover", 0);
/* BFO FAILOVER CODE - end */
mca_base_param_reg_string(&mca_pml_bfo_component.pmlm_version,
"allocator",

29
ompi/mca/pml/bfo/pml_bfo_failover.c
Просмотреть файл

@ -1477,11 +1477,10 @@ void mca_pml_bfo_failover_error_handler(struct mca_btl_base_module_t* btl,
btlname = "unknown";
}
/* If the process to map out is not specified or if the fast
* failover flag is specified, then map out the entire BTL.
* Otherwise, only map out the BTL for the specific remote
* process. */
if (NULL == errproc || mca_pml_bfo.fast_failover) {
/* If the process to map out is not specified then map out the
* entire BTL. Otherwise, only map out the BTL for the specific
* remote process. */
if (NULL == errproc) {
for( p = 0; p < num_procs; p++ ) {
mca_pml_bfo_map_out_btl(btl, procs[p], btlname);
}
@ -1936,3 +1935,23 @@ int mca_pml_bfo_register_callbacks(void) {
return rc;
}
/**
* Update a few fields when we are restarting either a RNDV or
* RGET type message.
*/
void mca_pml_bfo_update_rndv_fields(mca_pml_bfo_hdr_t* hdr,
mca_pml_bfo_send_request_t* sendreq, char *type)
{
hdr->hdr_common.hdr_flags |= MCA_PML_BFO_HDR_FLAGS_RESTART;
hdr->hdr_rndv.hdr_dst_req = sendreq->req_recv;
hdr->hdr_rndv.hdr_restartseq = sendreq->req_restartseq;
opal_output_verbose(30, mca_pml_bfo_output,
"%s: restarting: PML=%d, RQS=%d, CTX=%d, SRC=%d, "
"src_req=%p, dst_req=%p, peer=%d",
type, (uint16_t)sendreq->req_send.req_base.req_sequence,
sendreq->req_restartseq,
sendreq->req_send.req_base.req_comm->c_contextid,
sendreq->req_send.req_base.req_comm->c_my_rank, (void *)sendreq,
sendreq->req_recv.pval, sendreq->req_send.req_base.req_peer);
}

167
ompi/mca/pml/bfo/pml_bfo_failover.h
Просмотреть файл

@ -76,6 +76,10 @@ extern void mca_pml_bfo_map_out( mca_btl_base_module_t *btl,
int mca_pml_bfo_register_callbacks(void);
void mca_pml_bfo_update_rndv_fields(mca_pml_bfo_hdr_t* hdr,
mca_pml_bfo_send_request_t*, char *type);
/**
* Four new callbacks for the four new message types.
@ -191,6 +195,169 @@ extern void mca_pml_bfo_recv_frag_callback_recverrnotify( mca_btl_base_module_t
mca_pml_bfo_recv_request_recverrnotify(recvreq, MCA_PML_BFO_HDR_TYPE_RGET, status); \
}
#define MCA_PML_BFO_SECOND_ERROR_CHECK_ON_RDMA_READ_COMPLETION(recvreq, status, btl) \
/* See if the request has received a RNDVRESTARTNOTIFY */ \
if( OPAL_UNLIKELY(recvreq->req_errstate)) { \
if (recvreq->req_errstate & RECVREQ_RNDVRESTART_RECVED) { \
opal_output_verbose(30, mca_pml_bfo_output, \
"RDMA read: completion: recvreq has error, outstanding events=%d " \
"PML=%d, RQS=%d, src_req=%lx, dst_req=%lx, status=%d, peer=%d", \
recvreq->req_events, recvreq->req_msgseq, recvreq->req_restartseq, \
(unsigned long)recvreq->remote_req_send.pval, \
(unsigned long)recvreq, status, \
recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE); \
if (0 == recvreq->req_events) { \
mca_pml_bfo_recv_request_rndvrestartack(recvreq, MCA_PML_BFO_HDR_TYPE_RGET, \
status, btl); \
} \
} \
MCA_PML_BFO_RDMA_FRAG_RETURN(frag); \
return; \
}
/**
* Macros for pml_bfo_sendreq.c file.
*/
/* The completion event for the RNDV message has returned with an
* error. We know that the send request we are looking at is valid
* because it cannot be completed until the sendreq->req_state value
* reaches 0. And for the sendreq->req_state to reach 0, the
* completion event on the RNDV message must occur. So, we do not
* bother checking whether the send request is valid, because we know
* it is, but we put a few asserts in for good measure. We then check
* a few fields in the request to decide what to do. If the
* sendreq->req_error is set, that means that something has happend
* already to the request and we do not want to restart it.
* Presumably, we may have received a RECVERRNOTIFY message from the
* receiver. We also check the sendreq->req_acked field to see if it
* has been acked. If it has, then again we do not restart everything
* because obviously the RNDV message has made it to the other side. */
#define MCA_PML_BFO_ERROR_ON_RNDV_COMPLETION(sendreq, des) \
do { \
assert(((mca_pml_bfo_hdr_t*)((des)->des_src->seg_addr.pval))->hdr_match.hdr_ctx == \
(sendreq)->req_send.req_base.req_comm->c_contextid); \
assert(((mca_pml_bfo_hdr_t*)((des)->des_src->seg_addr.pval))->hdr_match.hdr_src == \
(sendreq)->req_send.req_base.req_comm->c_my_rank); \
assert(((mca_pml_bfo_hdr_t*)((des)->des_src->seg_addr.pval))->hdr_match.hdr_seq == \
(uint16_t)(sendreq)->req_send.req_base.req_sequence); \
if ((!(sendreq)->req_error) && (!(sendreq)->req_acked)) { \
(sendreq)->req_events--; \
/* Assume RNDV did not make it, so restart from the beginning. */ \
mca_pml_bfo_send_request_restart(sendreq, true, MCA_PML_BFO_HDR_TYPE_RNDV); \
return; \
} \
} while (0)
/* Now check the error state. This request can be in error if the
* RNDV message made it over, but the receiver got an error trying to
* send the ACK back and therefore sent a RECVERRNOTIFY message. In
* that case, we want to start the restart dance as the receiver has
* matched this message already. Only restart if there are no
* outstanding events on send request. */
#define MCA_PML_BFO_CHECK_SENDREQ_ERROR_ON_RNDV_COMPLETION(sendreq, status, btl) \
if ((sendreq)->req_error) { \
opal_output_verbose(30, mca_pml_bfo_output, \
"RNDV: completion: sendreq has error, outstanding events=%d, " \
"PML=%d, RQS=%d, src_req=%lx, dst_req=%lx, status=%d, peer=%d", \
(sendreq)->req_events, \
(uint16_t)(sendreq)->req_send.req_base.req_sequence, \
(sendreq)->req_restartseq, (unsigned long)(sendreq), \
(unsigned long)(sendreq)->req_recv.pval, \
status, (sendreq)->req_send.req_base.req_peer); \
if (0 == (sendreq)->req_events) { \
mca_pml_bfo_send_request_rndvrestartnotify(sendreq, false, \
MCA_PML_BFO_HDR_TYPE_RNDV, \
status, btl); \
} \
return; \
}
/* This can happen if a FIN message arrives after the request was
* marked in error. So, just drop the message. Note that the status
* field is not checked here. That is because that is the value
* returned in the FIN hdr.hdr_fail field and may be used for other
* things. */
#define MCA_PML_BFO_CHECK_SENDREQ_ERROR_ON_RGET_COMPLETION(sendreq, btl, des) \
if( OPAL_UNLIKELY(sendreq->req_error)) { \
opal_output_verbose(30, mca_pml_bfo_output, \
"FIN: received on broken request, skipping, " \
"PML=%d, src_req=%lx, dst_req=%lx, peer=%d", \
(uint16_t)sendreq->req_send.req_base.req_sequence, \
(unsigned long)sendreq, (unsigned long)sendreq->req_recv.pval, \
sendreq->req_send.req_base.req_peer); \
btl->btl_free(btl, des); \
return; \
}
/* If we get an error on the RGET message, then first make sure that
* header matches the send request that we are pointing to. This is
* necessary, because even though the sending side got an error, the
* RGET may have made it to the receiving side and the message transfer
* may have completed. This would then mean the send request has been
* completed and perhaps in use by another communication. So there is
* no need to restart this request. Therefore, ensure that we are
* looking at the same request that the header thinks we are looking
* at. If not, then there is nothing else to be done. */
#define MCA_PML_BFO_ERROR_ON_SEND_CTL_COMPLETION(sendreq, des) \
do { \
mca_pml_bfo_hdr_t* hdr = des->des_src->seg_addr.pval; \
mca_pml_bfo_send_request_t* sendreq = (mca_pml_bfo_send_request_t*)des->des_cbdata; \
switch (hdr->hdr_common.hdr_type) { \
case MCA_PML_BFO_HDR_TYPE_RGET: \
if ((hdr->hdr_match.hdr_ctx != sendreq->req_send.req_base.req_comm->c_contextid) || \
(hdr->hdr_match.hdr_src != sendreq->req_send.req_base.req_comm->c_my_rank) || \
(hdr->hdr_match.hdr_seq != (uint16_t)sendreq->req_send.req_base.req_sequence)) { \
opal_output_verbose(30, mca_pml_bfo_output, \
"RGET: completion event: dropping because no valid request " \
"PML:exp=%d,act=%d CTX:exp=%d,act=%d SRC:exp=%d,act=%d " \
"RQS:exp=%d,act=%d, dst_req=%p", \
(uint16_t)sendreq->req_send.req_base.req_sequence, \
hdr->hdr_match.hdr_seq, \
sendreq->req_send.req_base.req_comm->c_contextid, \
hdr->hdr_match.hdr_ctx, \
sendreq->req_send.req_base.req_comm->c_my_rank, \
hdr->hdr_match.hdr_src, \
sendreq->req_restartseq, hdr->hdr_rndv.hdr_restartseq, \
(void *)sendreq); \
return; \
} \
mca_pml_bfo_send_request_restart(sendreq, true, MCA_PML_BFO_HDR_TYPE_RGET); \
return; \
default: \
opal_output(0, "%s:%d FATAL ERROR, unknown header (hdr=%d)", \
__FILE__, __LINE__, hdr->hdr_common.hdr_type); \
orte_errmgr.abort(-1, NULL); \
} \
} while (0)
/* Check if there has been an error on the send request when we get
* a completion event on the RDMA write. */
#define MCA_PML_BFO_CHECK_SENDREQ_ERROR_ON_PUT_COMPLETION(sendreq, status, btl) \
if ( OPAL_UNLIKELY(sendreq->req_error)) { \
opal_output_verbose(30, mca_pml_bfo_output, \
"RDMA write: completion: sendreq has error, outstanding events=%d," \
" PML=%d, RQS=%d, src_req=%p, dst_req=%p, status=%d, peer=%d", \
sendreq->req_events, \
(uint16_t)sendreq->req_send.req_base.req_sequence, \
sendreq->req_restartseq, (void *)sendreq, \
sendreq->req_recv.pval, \
status, sendreq->req_send.req_base.req_peer); \
if (0 == sendreq->req_events) { \
mca_pml_bfo_send_request_rndvrestartnotify(sendreq, false, \
MCA_PML_BFO_HDR_TYPE_PUT, \
status, btl); \
} \
MCA_PML_BFO_RDMA_FRAG_RETURN(frag); \
return; \
}
#define MCA_PML_BFO_CHECK_FOR_RNDV_RESTART(hdr, sendreq, type) \
if (0 < sendreq->req_restartseq) { \
mca_pml_bfo_update_rndv_fields(hdr, sendreq, type); \
}
END_C_DECLS

26
ompi/mca/pml/bfo/pml_bfo_recvreq.c
Просмотреть файл

@ -198,7 +198,7 @@ static void mca_pml_bfo_put_completion( mca_btl_base_module_t* btl,
btl->btl_free(btl, des);
/* BFO FAILOVER CODE - begin */
MCA_PML_BFO_ERROR_CHECK_ON_FIN_FOR_PUT(recvreq)
MCA_PML_BFO_ERROR_CHECK_ON_FIN_FOR_PUT(recvreq);
/* BFO FAILOVER CODE - end */
/* check completion status */
@ -351,29 +351,13 @@ static void mca_pml_bfo_rget_completion( mca_btl_base_module_t* btl,
/* check completion status */
if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
MCA_PML_BFO_ERROR_CHECK_ON_RDMA_READ_COMPLETION(recvreq)
MCA_PML_BFO_ERROR_CHECK_ON_RDMA_READ_COMPLETION(recvreq);
}
/* BFO FAILOVER CODE - end */
/* BFO FAILOVER CODE - begin */
/* See if the request has received a RNDVRESTARTNOTIFY */
if( OPAL_UNLIKELY(recvreq->req_errstate)) {
if (recvreq->req_errstate & RECVREQ_RNDVRESTART_RECVED) {
opal_output_verbose(30, mca_pml_bfo_output,
"RDMA read: completion: recvreq has error, outstanding events=%d "
"PML=%d, RQS=%d, src_req=%lx, dst_req=%lx, status=%d, peer=%d",
recvreq->req_events, recvreq->req_msgseq, recvreq->req_restartseq,
(unsigned long)recvreq->remote_req_send.pval,
(unsigned long)recvreq, status,
recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE);
if (0 == recvreq->req_events) {
mca_pml_bfo_recv_request_rndvrestartack(recvreq, MCA_PML_BFO_HDR_TYPE_RGET,
status, btl);
}
}
MCA_PML_BFO_RDMA_FRAG_RETURN(frag);
return;
}
MCA_PML_BFO_SECOND_ERROR_CHECK_ON_RDMA_READ_COMPLETION(recvreq, status, btl);
/* BFO FAILOVER CODE - end */
/* BFO FAILOVER CODE - begin */
/* Find back the bml_btl that this btl belongs to. If we cannot
* find it, then it may have been removed from underneath us, so

191
ompi/mca/pml/bfo/pml_bfo_sendreq.c
Просмотреть файл

@ -224,61 +224,12 @@ mca_pml_bfo_rndv_completion( mca_btl_base_module_t* btl,
/* check completion status */
if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* BFO FAILOVER CODE - begin */
/* The completion event for the RNDV message has returned with
* an error. We know that the send request we are looking at is
* valid because it cannot be completed until the sendreq->req_state
* value reaches 0. And for the sendreq->req_state to reach 0,
* the completion event on the RNDV message must occur. So, we
* do not bother checking whether the send request is valid,
* because we know it is, but we put a few asserts in for good
* measure. We then check a few fields in the request to decide what
* to do. If the sendreq->req_error is set, that means that something
* has happend already to the request and we do not want to restart
* it. Presumably, we may have received a RECVERRNOTIFY
* message from the receiver. We also check the sendreq->req_acked
* field to see if it has been acked. If it has, then again we
* do not restart everything because obviously the RNDV message
* has made it to the other side. */
assert(((mca_pml_bfo_hdr_t*)(des->des_src->seg_addr.pval))->hdr_match.hdr_ctx ==
sendreq->req_send.req_base.req_comm->c_contextid);
assert(((mca_pml_bfo_hdr_t*)(des->des_src->seg_addr.pval))->hdr_match.hdr_src ==
sendreq->req_send.req_base.req_comm->c_my_rank);
assert(((mca_pml_bfo_hdr_t*)(des->des_src->seg_addr.pval))->hdr_match.hdr_seq ==
(uint16_t)sendreq->req_send.req_base.req_sequence);
if ((!sendreq->req_error) && (!sendreq->req_acked)) {
sendreq->req_events--;
/* Assume RNDV did not make it, so restart from the beginning. */
mca_pml_bfo_send_request_restart(sendreq, true, MCA_PML_BFO_HDR_TYPE_RNDV);
return;
}
/* BFO FAILOVER CODE - end */
MCA_PML_BFO_ERROR_ON_RNDV_COMPLETION(sendreq, des);
}
/* BFO FAILOVER CODE - begin */
sendreq->req_events--;
/* Now check the error state. This request can be in error if the
* RNDV message made it over, but the receiver got an error trying
* to send the ACK back and therefore sent a RECVERRNOTIFY message.
* In that case, we want to start the restart dance as the receiver
* has matched this message already. Only restart if there are no
* outstanding events on send request. */
if (sendreq->req_error) {
opal_output_verbose(30, mca_pml_bfo_output,
"RNDV: completion: sendreq has error, outstanding events=%d, "
"PML=%d, RQS=%d, src_req=%lx, dst_req=%lx, status=%d, peer=%d",
sendreq->req_events, (uint16_t)sendreq->req_send.req_base.req_sequence,
sendreq->req_restartseq, (unsigned long)sendreq,
(unsigned long)sendreq->req_recv.pval,
status, sendreq->req_send.req_base.req_peer);
if (0 == sendreq->req_events) {
mca_pml_bfo_send_request_rndvrestartnotify(sendreq, false,
MCA_PML_BFO_HDR_TYPE_RNDV,
status, btl);
}
return;
}
MCA_PML_BFO_CHECK_SENDREQ_ERROR_ON_RNDV_COMPLETION(sendreq, status, btl);
/* BFO FAILOVER CODE - end */
/* count bytes of user data actually delivered. As the rndv completion only
@ -306,23 +257,8 @@ mca_pml_bfo_rget_completion( mca_btl_base_module_t* btl,
{
mca_pml_bfo_send_request_t* sendreq = (mca_pml_bfo_send_request_t*)des->des_cbdata;
size_t req_bytes_delivered = 0;
/* BFO FAILOVER CODE - begin */
/* This can happen if a FIN message arrives after the request was
* marked in error. So, just drop the message. Note that the
* status field is not checked here. That is because that is the
* value returned in the FIN hdr.hdr_fail field and may be used for
* other things. */
if( OPAL_UNLIKELY(sendreq->req_error)) {
opal_output_verbose(30, mca_pml_bfo_output,
"FIN: received on broken request, skipping, "
"PML=%d, src_req=%lx, dst_req=%lx, peer=%d",
(uint16_t)sendreq->req_send.req_base.req_sequence,
(unsigned long)sendreq, (unsigned long)sendreq->req_recv.pval,
sendreq->req_send.req_base.req_peer);
btl->btl_free(btl, des);
return;
}
/* BFO FAILOVER CODE - end */
MCA_PML_BFO_CHECK_SENDREQ_ERROR_ON_RGET_COMPLETION(sendreq, btl, des);
/* count bytes of user data actually delivered and check for request completion */
MCA_PML_BFO_COMPUTE_SEGMENT_LENGTH( des->des_src, des->des_src_cnt,
@ -351,47 +287,8 @@ mca_pml_bfo_send_ctl_completion( mca_btl_base_module_t* btl,
/* check for pending requests */
MCA_PML_BFO_PROGRESS_PENDING(btl);
} else {
mca_pml_bfo_hdr_t* hdr = des->des_src->seg_addr.pval;
/* If we get an error on the RGET message, then first make
* sure that header matches the send request that we are
* pointing to. This is necessary, because even though the
* sending side got an error, the RGET may have made it to the
* receiving side and the message transfer may have completed.
* This would then mean the send request has been completed and
* perhaps in use by another communication. So there is no need
* to restart this request. Therefore, ensure that we are
* looking at the same request that the header thinks we are
* looking at. If not, then there is nothing else to be done. */
mca_pml_bfo_send_request_t* sendreq = (mca_pml_bfo_send_request_t*)des->des_cbdata;
switch (hdr->hdr_common.hdr_type) {
case MCA_PML_BFO_HDR_TYPE_RGET:
if ((hdr->hdr_match.hdr_ctx != sendreq->req_send.req_base.req_comm->c_contextid) ||
(hdr->hdr_match.hdr_src != sendreq->req_send.req_base.req_comm->c_my_rank) ||
(hdr->hdr_match.hdr_seq != (uint16_t)sendreq->req_send.req_base.req_sequence)) {
opal_output_verbose(20, mca_pml_bfo_output,
"RGET: completion event: dropping because no valid request "
"PML:exp=%d,act=%d CTX:exp=%d,act=%d SRC:exp=%d,act=%d "
"RQS:exp=%d,act=%d, dst_req=%p",
(uint16_t)sendreq->req_send.req_base.req_sequence,
hdr->hdr_match.hdr_seq,
sendreq->req_send.req_base.req_comm->c_contextid,
hdr->hdr_match.hdr_ctx,
sendreq->req_send.req_base.req_comm->c_my_rank,
hdr->hdr_match.hdr_src,
sendreq->req_restartseq, hdr->hdr_rndv.hdr_restartseq,
(void *)sendreq);
return;
}
mca_pml_bfo_send_request_restart(sendreq, true, MCA_PML_BFO_HDR_TYPE_RGET);
return;
default:
opal_output(0, "%s:%d FATAL ERROR, unknown header (hdr=%d)",
__FILE__, __LINE__, hdr->hdr_common.hdr_type);
orte_errmgr.abort(-1, NULL);
}
MCA_PML_BFO_ERROR_ON_SEND_CTL_COMPLETION(sendreq, des);
}
/* BFO FAILOVER CODE - end */
}
/**
@ -527,20 +424,7 @@ int mca_pml_bfo_send_request_start_buffered(
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
/* BFO FAILOVER CODE - begin */
if (0 < sendreq->req_restartseq) {
hdr->hdr_common.hdr_flags |= MCA_PML_BFO_HDR_FLAGS_RESTART;
hdr->hdr_rndv.hdr_dst_req = sendreq->req_recv;
hdr->hdr_rndv.hdr_restartseq = sendreq->req_restartseq;
opal_output_verbose(30, mca_pml_bfo_output,
"RNDV(buffered): restarting: PML=%d, RQS=%d, CTX=%d, SRC=%d, "
"src_req=%p, dst_req=%p, peer=%d",
(uint16_t)sendreq->req_send.req_base.req_sequence, sendreq->req_restartseq,
sendreq->req_send.req_base.req_comm->c_contextid,
sendreq->req_send.req_base.req_comm->c_my_rank, (void *)sendreq,
sendreq->req_recv.pval, sendreq->req_send.req_base.req_peer);
}
/* BFO FAILOVER CODE - end */
MCA_PML_BFO_CHECK_FOR_RNDV_RESTART(hdr, sendreq, "RNDV(buffered)");
bfo_hdr_hton(hdr, MCA_PML_BFO_HDR_TYPE_RNDV,
sendreq->req_send.req_base.req_proc);
@ -877,20 +761,7 @@ int mca_pml_bfo_send_request_start_rdma( mca_pml_bfo_send_request_t* sendreq,
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
/* BFO FAILOVER CODE - begin */
if (0 < sendreq->req_restartseq) {
hdr->hdr_common.hdr_flags |= MCA_PML_BFO_HDR_FLAGS_RESTART;
hdr->hdr_rndv.hdr_dst_req = sendreq->req_recv;
hdr->hdr_rndv.hdr_restartseq = sendreq->req_restartseq;
opal_output_verbose(30, mca_pml_bfo_output,
"RGET: restarting: PML=%d, RQS=%d, CTX=%d, SRC=%d, "
"src_req=%p, dst_req=%p, peer=%d",
(uint16_t)sendreq->req_send.req_base.req_sequence,
sendreq->req_restartseq,
sendreq->req_send.req_base.req_comm->c_contextid,
sendreq->req_send.req_base.req_comm->c_my_rank,
(void *)sendreq, sendreq->req_recv.pval,
sendreq->req_send.req_base.req_peer);
}
MCA_PML_BFO_CHECK_FOR_RNDV_RESTART(hdr, sendreq, "RGET");
/* BFO FAILOVER CODE - end */
hdr->hdr_rget.hdr_des.pval = src;
hdr->hdr_rget.hdr_seg_cnt = src->des_src_cnt;
@ -942,20 +813,7 @@ int mca_pml_bfo_send_request_start_rdma( mca_pml_bfo_send_request_t* sendreq,
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
/* BFO FAILOVER CODE - begin */
if (0 < sendreq->req_restartseq) {
hdr->hdr_common.hdr_flags |= MCA_PML_BFO_HDR_FLAGS_RESTART;
hdr->hdr_rndv.hdr_dst_req = sendreq->req_recv;
hdr->hdr_rndv.hdr_restartseq = sendreq->req_restartseq;
opal_output_verbose(30, mca_pml_bfo_output,
"RNDV: restarting: PML=%d, RQS=%d, CTX=%d, SRC=%d, "
"src_req=%p, dst_req=%p, peer=%d",
(uint16_t)sendreq->req_send.req_base.req_sequence,
sendreq->req_restartseq,
sendreq->req_send.req_base.req_comm->c_contextid,
sendreq->req_send.req_base.req_comm->c_my_rank,
(void *)sendreq, sendreq->req_recv.pval,
sendreq->req_send.req_base.req_peer);
}
MCA_PML_BFO_CHECK_FOR_RNDV_RESTART(hdr, sendreq, "RNDV");
/* BFO FAILOVER CODE - end */
bfo_hdr_hton(hdr, MCA_PML_BFO_HDR_TYPE_RNDV,
@ -1054,19 +912,7 @@ int mca_pml_bfo_send_request_start_rndv( mca_pml_bfo_send_request_t* sendreq,
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
/* BFO FAILOVER CODE - begin */
if (0 < sendreq->req_restartseq) {
hdr->hdr_common.hdr_flags |= MCA_PML_BFO_HDR_FLAGS_RESTART;
hdr->hdr_rndv.hdr_dst_req = sendreq->req_recv;
hdr->hdr_rndv.hdr_restartseq = sendreq->req_restartseq;
opal_output_verbose(30, mca_pml_bfo_output,
"RNDV: restarting: PML=%d, RQS=%d, CTX=%d, SRC=%d, "
"src_req=%p, dst_req=%p, peer=%d",
(uint16_t)sendreq->req_send.req_base.req_sequence, sendreq->req_restartseq,
sendreq->req_send.req_base.req_comm->c_contextid,
sendreq->req_send.req_base.req_comm->c_my_rank,
(void *)sendreq, sendreq->req_recv.pval,
sendreq->req_send.req_base.req_peer);
}
MCA_PML_BFO_CHECK_FOR_RNDV_RESTART(hdr, sendreq, "RNDV");
/* BFO FAILOVER CODE - end */
bfo_hdr_hton(hdr, MCA_PML_BFO_HDR_TYPE_RNDV,
@ -1361,22 +1207,7 @@ static void mca_pml_bfo_put_completion( mca_btl_base_module_t* btl,
}
/* BFO FAILOVER CODE - begin */
if ( OPAL_UNLIKELY(sendreq->req_error)) {
opal_output_verbose(30, mca_pml_bfo_output,
"RDMA write: completion: sendreq has error, outstanding events=%d, "
"PML=%d, RQS=%d, src_req=%p, dst_req=%p, status=%d, peer=%d",
sendreq->req_events, (uint16_t)sendreq->req_send.req_base.req_sequence,
sendreq->req_restartseq, (void *)sendreq,
sendreq->req_recv.pval,
status, sendreq->req_send.req_base.req_peer);
if (0 == sendreq->req_events) {
mca_pml_bfo_send_request_rndvrestartnotify(sendreq, false,
MCA_PML_BFO_HDR_TYPE_PUT,
status, btl);
}
MCA_PML_BFO_RDMA_FRAG_RETURN(frag);
return;
}
MCA_PML_BFO_CHECK_SENDREQ_ERROR_ON_PUT_COMPLETION(sendreq, status, btl);
/* BFO FAILOVER CODE - end */
/* BFO FAILOVER CODE - begin */