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

ob1: add support for get fallback on put/send

Просмотр исходного кода 
This commit was SVN r26329.
Этот коммит содержится в:
Nathan Hjelm 2012-04-24 20:18:56 +00:00
родитель 93780c63be
Коммит 9a35f96bda
12 изменённых файлов: 181 добавлений и 148 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

17
ompi/mca/btl/ugni/btl_ugni_get.c
Просмотреть файл

@ -13,19 +13,6 @@
#include "btl_ugni_rdma.h"
#include "btl_ugni_smsg.h"
static int mca_btl_ugni_init_put (struct mca_btl_base_module_t *btl,
mca_btl_ugni_base_frag_t *frag) {
/* off alignment/off size. switch to put */
frag->hdr.rdma.src_seg = frag->base.des_src[0];
frag->hdr.rdma.dst_seg = frag->base.des_dst[0];
frag->hdr.rdma.ctx = (void *) frag;
/* send the fragment header using smsg. ignore local completion */
return ompi_mca_btl_ugni_smsg_send (frag, true, &frag->hdr.rdma,
sizeof (frag->hdr.rdma), NULL, 0,
MCA_BTL_UGNI_TAG_PUT_INIT);
}
/**
* Initiate a get operation.
*
@ -54,7 +41,7 @@ int mca_btl_ugni_get (struct mca_btl_base_module_t *btl,
if (OPAL_UNLIKELY(check || size > mca_btl_ugni_component.ugni_get_limit)) {
/* switch to put */
return mca_btl_ugni_init_put (btl, frag);
return OMPI_ERR_NOT_AVAILABLE;
}
if (NULL != frag->base.des_cbfunc) {
@ -68,7 +55,7 @@ int mca_btl_ugni_get (struct mca_btl_base_module_t *btl,
return mca_btl_ugni_post_bte (frag, GNI_POST_RDMA_GET, des->des_dst, des->des_src);
}
void mca_btl_ugni_callback_rdma_complete (mca_btl_ugni_base_frag_t *frag, int rc)
static void mca_btl_ugni_callback_rdma_complete (mca_btl_ugni_base_frag_t *frag, int rc)
{
BTL_VERBOSE(("rdma operation for rem_ctx %p complete", frag->hdr.rdma.ctx));

48
ompi/mca/btl/ugni/btl_ugni_put.c
Просмотреть файл

@ -46,51 +46,3 @@ int mca_btl_ugni_put (struct mca_btl_base_module_t *btl,
return mca_btl_ugni_post_bte (frag, GNI_POST_RDMA_PUT, des->des_src, des->des_dst);
}
/* reversed get */
static void mca_btl_ugni_callback_put_retry (mca_btl_ugni_base_frag_t *frag, int rc)
{
(void) mca_btl_ugni_start_put(frag->endpoint, frag->hdr.rdma, frag);
}
int mca_btl_ugni_start_put (mca_btl_base_endpoint_t *ep,
mca_btl_ugni_rdma_frag_hdr_t hdr,
mca_btl_ugni_base_frag_t *frag)
{
int rc;
BTL_VERBOSE(("starting reverse get (put) for remote ctx: %p", hdr.ctx));
if (NULL == frag) {
rc = MCA_BTL_UGNI_FRAG_ALLOC_RDMA_INT(ep, frag);
if (OPAL_UNLIKELY(NULL == frag)) {
BTL_ERROR(("error allocating rdma frag for reverse get. rc = %d. fl_num_allocated = %d", rc,
ep->btl->rdma_int_frags.fl_num_allocated));
return rc;
}
}
frag->hdr.rdma = hdr;
frag->base.des_cbfunc = NULL;
frag->base.des_flags = MCA_BTL_DES_FLAGS_BTL_OWNERSHIP;
frag->segments[0] = hdr.src_seg;
frag->base.des_src = frag->segments;
frag->base.des_src_cnt = 1;
frag->segments[1] = hdr.dst_seg;
frag->base.des_dst = frag->segments + 1;
frag->base.des_dst_cnt = 1;
rc = mca_btl_ugni_put (&ep->btl->super, ep, &frag->base);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
frag->cbfunc = mca_btl_ugni_callback_put_retry;
opal_list_append (&ep->btl->failed_frags, (opal_list_item_t *) frag);
return rc;
}
frag->cbfunc = mca_btl_ugni_callback_rdma_complete;
return OMPI_SUCCESS;
}

7
ompi/mca/btl/ugni/btl_ugni_rdma.h
Просмотреть файл

@ -16,17 +16,10 @@
#include "btl_ugni.h"
#include "btl_ugni_frag.h"
/* mca_btl_ugni_start_put: get operation could not be completed. start put instead */
int mca_btl_ugni_start_put (mca_btl_base_endpoint_t *ep,
mca_btl_ugni_rdma_frag_hdr_t hdr,
mca_btl_ugni_base_frag_t *frag);
int mca_btl_ugni_start_eager_get (mca_btl_base_endpoint_t *ep,
mca_btl_ugni_eager_ex_frag_hdr_t hdr,
mca_btl_ugni_base_frag_t *frag);
void mca_btl_ugni_callback_rdma_complete (mca_btl_ugni_base_frag_t *frag, int rc);
static inline int init_gni_post_desc(mca_btl_ugni_base_frag_t *frag,
gni_post_type_t op_type,
uint64_t lcl_addr,

5
ompi/mca/btl/ugni/btl_ugni_smsg.c
Просмотреть файл

@ -77,11 +77,6 @@ int mca_btl_ugni_smsg_process (mca_btl_base_endpoint_t *ep)
reg->cbfunc(&ep->btl->super, tag, &(frag.base), reg->cbdata);
break;
case MCA_BTL_UGNI_TAG_PUT_INIT:
frag.hdr.rdma = ((mca_btl_ugni_rdma_frag_hdr_t *) data_ptr)[0];
mca_btl_ugni_start_put (ep, frag.hdr.rdma, NULL);
break;
case MCA_BTL_UGNI_TAG_GET_INIT:
frag.hdr.eager_ex = ((mca_btl_ugni_eager_ex_frag_hdr_t *) data_ptr)[0];

1
ompi/mca/btl/ugni/btl_ugni_smsg.h
Просмотреть файл

@ -21,7 +21,6 @@
typedef enum {
MCA_BTL_UGNI_TAG_SEND,
MCA_BTL_UGNI_TAG_DISCONNECT,
MCA_BTL_UGNI_TAG_PUT_INIT,
MCA_BTL_UGNI_TAG_GET_INIT,
MCA_BTL_UGNI_TAG_RDMA_COMPLETE
} mca_btl_ugni_smsg_tag_t;

5
ompi/mca/pml/ob1/pml_ob1.c
Просмотреть файл

@ -147,6 +147,7 @@ int mca_pml_ob1_enable(bool enable)
OBJ_CONSTRUCT(&mca_pml_ob1.recv_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.pckt_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.rdma_pending, opal_list_t);
/* missing communicator pending list */
OBJ_CONSTRUCT(&mca_pml_ob1.non_existing_communicator_pending, opal_list_t);
@ -599,8 +600,10 @@ void mca_pml_ob1_process_pending_rdma(void)
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
if(NULL == frag)
break;
frag->retries++;
if(frag->rdma_state == MCA_PML_OB1_RDMA_PUT) {
frag->retries++;
rc = mca_pml_ob1_send_request_put_frag(frag);
} else {
rc = mca_pml_ob1_recv_request_get_frag(frag);

2
ompi/mca/pml/ob1/pml_ob1.h
Просмотреть файл

@ -52,7 +52,7 @@ struct mca_pml_ob1_t {
int free_list_inc; /* number of elements to grow free list */
size_t send_pipeline_depth;
size_t recv_pipeline_depth;
size_t rdma_put_retries_limit;
size_t rdma_retries_limit;
int max_rdma_per_request;
int max_send_per_range;
bool leave_pinned;

4
ompi/mca/pml/ob1/pml_ob1_component.c
Просмотреть файл

@ -112,8 +112,8 @@ static int mca_pml_ob1_component_open(void)
mca_pml_ob1_param_register_int("send_pipeline_depth", 3);
mca_pml_ob1.recv_pipeline_depth =
mca_pml_ob1_param_register_int("recv_pipeline_depth", 4);
mca_pml_ob1.rdma_put_retries_limit =
mca_pml_ob1_param_register_int("rdma_put_retries_limit", 5);
mca_pml_ob1.rdma_retries_limit =
mca_pml_ob1_param_register_int("rdma_retries_limit", 5);
mca_pml_ob1.max_rdma_per_request =
mca_pml_ob1_param_register_int("max_rdma_per_request", 4);
mca_pml_ob1.max_send_per_range =

15
ompi/mca/pml/ob1/pml_ob1_recvfrag.c
Просмотреть файл

@ -294,15 +294,22 @@ void mca_pml_ob1_recv_frag_callback_ack(mca_btl_base_module_t* btl,
if( OPAL_UNLIKELY(segments->seg_len < sizeof(mca_pml_ob1_common_hdr_t)) ) {
return;
}
ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_ACK);
sendreq = (mca_pml_ob1_send_request_t*)hdr->hdr_ack.hdr_src_req.pval;
sendreq->req_recv = hdr->hdr_ack.hdr_dst_req;
/* if the request should be delivered entirely by copy in/out
* then throttle sends */
if(hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NORDMA)
if(hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NORDMA) {
if (NULL != sendreq->src_des) {
/* release registered memory */
mca_bml_base_free (sendreq->req_rdma[0].bml_btl, sendreq->src_des);
sendreq->src_des = NULL;
}
sendreq->req_throttle_sends = true;
}
mca_pml_ob1_send_request_copy_in_out(sendreq,
hdr->hdr_ack.hdr_send_offset,
@ -324,7 +331,7 @@ void mca_pml_ob1_recv_frag_callback_ack(mca_btl_base_module_t* btl,
if(send_request_pml_complete_check(sendreq) == false)
mca_pml_ob1_send_request_schedule(sendreq);
return;
}

94
ompi/mca/pml/ob1/pml_ob1_recvreq.c
Просмотреть файл

@ -352,6 +352,66 @@ static void mca_pml_ob1_rget_completion( mca_btl_base_module_t* btl,
}
static int mca_pml_ob1_init_get_fallback (mca_pml_ob1_rdma_frag_t *frag,
mca_btl_base_descriptor_t *dst) {
mca_pml_ob1_recv_request_t *recvreq = (mca_pml_ob1_recv_request_t *) frag->rdma_req;
mca_bml_base_btl_t *bml_btl = frag->rdma_bml;
mca_btl_base_descriptor_t *ctl;
mca_pml_ob1_rdma_hdr_t *hdr;
size_t hdr_size;
unsigned int i;
int rc;
/* prepare a descriptor for rdma control message */
hdr_size = sizeof (mca_pml_ob1_rdma_hdr_t);
if (dst->des_dst_cnt > 1) {
hdr_size += (sizeof (mca_btl_base_segment_t) *
(dst->des_dst_cnt-1));
}
mca_bml_base_alloc (bml_btl, &ctl, MCA_BTL_NO_ORDER, hdr_size,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP |
MCA_BTL_DES_SEND_ALWAYS_CALLBACK);
if (OPAL_UNLIKELY(NULL == ctl)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
ctl->des_cbfunc = mca_pml_ob1_recv_ctl_completion;
/* fill in rdma header */
hdr = (mca_pml_ob1_rdma_hdr_t *) ctl->des_src->seg_addr.pval;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_PUT;
hdr->hdr_common.hdr_flags =
(!recvreq->req_ack_sent) ? MCA_PML_OB1_HDR_TYPE_ACK : 0;
hdr->hdr_req = frag->rdma_hdr.hdr_rget.hdr_rndv.hdr_src_req;
hdr->hdr_rdma_offset = recvreq->req_rdma_offset;
hdr->hdr_des.pval = dst;
hdr->hdr_seg_cnt = dst->des_dst_cnt;
for (i = 0 ; i < dst->des_dst_cnt ; ++i) {
hdr->hdr_segs[i].seg_addr.lval = ompi_ptr_ptol(dst->des_dst[i].seg_addr.pval);
hdr->hdr_segs[i].seg_len = dst->des_dst[i].seg_len;
hdr->hdr_segs[i].seg_key.key64[0] = dst->des_dst[i].seg_key.key64[0];
hdr->hdr_segs[i].seg_key.key64[1] = dst->des_dst[i].seg_key.key64[1];
}
dst->des_cbfunc = mca_pml_ob1_put_completion;
dst->des_cbdata = recvreq;
if (!recvreq->req_ack_sent)
recvreq->req_ack_sent = true;
/* send rdma request to peer */
rc = mca_bml_base_send (bml_btl, ctl, MCA_PML_OB1_HDR_TYPE_PUT);
if (OPAL_UNLIKELY(rc < 0)) {
mca_bml_base_free (bml_btl, ctl);
return rc;
}
return OMPI_SUCCESS;
}
/*
*
*/
@ -371,14 +431,25 @@ int mca_pml_ob1_recv_request_get_frag( mca_pml_ob1_rdma_frag_t* frag )
0,
&frag->rdma_length,
MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_SEND_ALWAYS_CALLBACK |
MCA_BTL_DES_FLAGS_GET,
MCA_BTL_DES_FLAGS_GET,
&descriptor );
if( OPAL_UNLIKELY(NULL == descriptor) ) {
frag->rdma_length = save_size;
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
return OMPI_ERR_OUT_OF_RESOURCE;
if (frag->retries < mca_pml_ob1.rdma_retries_limit) {
frag->rdma_length = save_size;
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
return OMPI_ERR_OUT_OF_RESOURCE;
} else {
ompi_proc_t *proc = (ompi_proc_t *) recvreq->req_recv.req_base.req_proc;
/* tell peer to fall back on send */
recvreq->req_send_offset = 0;
rc = mca_pml_ob1_recv_request_ack_send(proc, frag->rdma_hdr.hdr_rget.hdr_rndv.hdr_src_req.lval,
recvreq, recvreq->req_send_offset, true);
MCA_PML_OB1_RDMA_FRAG_RETURN(frag);
return rc;
}
}
descriptor->des_src = frag->rdma_segs;
@ -393,6 +464,11 @@ int mca_pml_ob1_recv_request_get_frag( mca_pml_ob1_rdma_frag_t* frag )
/* queue up get request */
rc = mca_bml_base_get(bml_btl,descriptor);
if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
if (OPAL_UNLIKELY(OMPI_ERR_NOT_AVAILABLE == rc)) {
/* get isn't supported for this transfer. tell peer to fallback on put */
rc = mca_pml_ob1_init_get_fallback (frag, descriptor);
}
if(OMPI_ERR_OUT_OF_RESOURCE == rc) {
mca_bml_base_free(bml_btl, descriptor);
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
@ -400,7 +476,7 @@ int mca_pml_ob1_recv_request_get_frag( mca_pml_ob1_rdma_frag_t* frag )
(opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
return OMPI_ERR_OUT_OF_RESOURCE;
} else {
} else if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
ORTE_ERROR_LOG(rc);
orte_errmgr.abort(-1, NULL);
}
@ -551,7 +627,9 @@ void mca_pml_ob1_recv_request_progress_rget( mca_pml_ob1_recv_request_t* recvreq
orte_errmgr.abort(-1, NULL);
}
#endif /* OMPI_CUDA_SUPPORT */
frag->rdma_hdr.hdr_rget = *hdr;
frag->retries = 0;
frag->rdma_req = recvreq;
frag->rdma_ep = bml_endpoint;
frag->rdma_length = size;
@ -792,7 +870,7 @@ int mca_pml_ob1_recv_request_schedule_once( mca_pml_ob1_recv_request_t* recvreq,
mca_bml_base_prepare_dst(bml_btl, reg,
&recvreq->req_recv.req_base.req_convertor,
MCA_BTL_NO_ORDER, 0, &size, MCA_BTL_DES_FLAGS_BTL_OWNERSHIP |
MCA_BTL_DES_FLAGS_PUT, &dst);
MCA_BTL_DES_FLAGS_PUT, &dst);
OPAL_THREAD_UNLOCK(&recvreq->lock);
if(OPAL_UNLIKELY(dst == NULL)) {

129
ompi/mca/pml/ob1/pml_ob1_sendreq.c
Просмотреть файл

@ -264,6 +264,7 @@ mca_pml_ob1_rget_completion( mca_btl_base_module_t* btl,
MCA_PML_OB1_COMPUTE_SEGMENT_LENGTH( des->des_src, des->des_src_cnt,
0, req_bytes_delivered );
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, req_bytes_delivered);
sendreq->src_des = NULL;
send_request_pml_complete_check(sendreq);
/* free the descriptor */
@ -639,6 +640,8 @@ int mca_pml_ob1_send_request_start_rdma( mca_pml_ob1_send_request_t* sendreq,
bool need_local_cb = false;
int rc;
sendreq->src_des = NULL;
bml_btl = sendreq->req_rdma[0].bml_btl;
if((sendreq->req_rdma_cnt == 1) && (bml_btl->btl_flags & (MCA_BTL_FLAGS_GET | MCA_BTL_FLAGS_CUDA_GET))) {
mca_mpool_base_registration_t* reg = sendreq->req_rdma[0].btl_reg;
@ -657,10 +660,8 @@ int mca_pml_ob1_send_request_start_rdma( mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_prepare_src( bml_btl,
reg,
&sendreq->req_send.req_base.req_convertor,
MCA_BTL_NO_ORDER,
0,
&size,
MCA_BTL_DES_FLAGS_GET,
MCA_BTL_NO_ORDER, 0, &size,
MCA_BTL_DES_FLAGS_GET | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP,
&src );
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_noaccess,
@ -676,6 +677,8 @@ int mca_pml_ob1_send_request_start_rdma( mca_pml_ob1_send_request_t* sendreq,
src->des_cbfunc = mca_pml_ob1_rget_completion;
src->des_cbdata = sendreq;
sendreq->src_des = src;
/* allocate space for get hdr + segment list */
mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER,
sizeof(mca_pml_ob1_rget_hdr_t) +
@ -782,8 +785,9 @@ int mca_pml_ob1_send_request_start_rdma( mca_pml_ob1_send_request_t* sendreq,
return OMPI_SUCCESS;
}
mca_bml_base_free(bml_btl, des);
if (NULL != src) {
mca_bml_base_free (bml_btl, src);
if (sendreq->src_des) {
mca_bml_base_free (bml_btl, sendreq->src_des);
sendreq->src_des = NULL;
}
return rc;
@ -1133,63 +1137,71 @@ static void mca_pml_ob1_put_completion( mca_btl_base_module_t* btl,
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t* frag )
{
mca_mpool_base_registration_t* reg = NULL;
mca_bml_base_btl_t* bml_btl = frag->rdma_bml;
mca_btl_base_descriptor_t* des;
int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t *frag )
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)frag->rdma_req;
mca_mpool_base_registration_t *reg = NULL;
mca_bml_base_btl_t *bml_btl = frag->rdma_bml;
mca_btl_base_descriptor_t *des;
size_t save_size = frag->rdma_length;
int rc;
/* setup descriptor */
mca_bml_base_prepare_src( bml_btl,
reg,
&frag->convertor,
MCA_BTL_NO_ORDER,
0,
&frag->rdma_length,
MCA_BTL_DES_FLAGS_BTL_OWNERSHIP |
MCA_BTL_DES_FLAGS_PUT,
&des );
if (OPAL_LIKELY(NULL == sendreq->src_des)) {
/* setup descriptor */
mca_bml_base_prepare_src( bml_btl,
reg,
&frag->convertor,
MCA_BTL_NO_ORDER,
0,
&frag->rdma_length,
MCA_BTL_DES_FLAGS_BTL_OWNERSHIP |
MCA_BTL_DES_FLAGS_PUT,
&des );
if( OPAL_UNLIKELY(NULL == des) ) {
if(frag->retries < mca_pml_ob1.rdma_put_retries_limit) {
size_t offset = (size_t)frag->rdma_hdr.hdr_rdma.hdr_rdma_offset;
frag->rdma_length = save_size;
opal_convertor_set_position(&frag->convertor, &offset);
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
} else {
mca_pml_ob1_send_request_t *sendreq =
(mca_pml_ob1_send_request_t*)frag->rdma_req;
if( OPAL_UNLIKELY(NULL == des) ) {
if(frag->retries < mca_pml_ob1.rdma_retries_limit) {
size_t offset = (size_t)frag->rdma_hdr.hdr_rdma.hdr_rdma_offset;
frag->rdma_length = save_size;
opal_convertor_set_position(&frag->convertor, &offset);
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
} else {
mca_pml_ob1_send_request_t *sendreq =
(mca_pml_ob1_send_request_t*)frag->rdma_req;
/* tell receiver to unregister memory */
mca_pml_ob1_send_fin(sendreq->req_send.req_base.req_proc,
bml_btl, frag->rdma_hdr.hdr_rdma.hdr_des,
MCA_BTL_NO_ORDER, 1);
/* tell receiver to unregister memory */
mca_pml_ob1_send_fin(sendreq->req_send.req_base.req_proc,
bml_btl, frag->rdma_hdr.hdr_rdma.hdr_des,
MCA_BTL_NO_ORDER, 1);
/* send fragment by copy in/out */
mca_pml_ob1_send_request_copy_in_out(sendreq,
frag->rdma_hdr.hdr_rdma.hdr_rdma_offset, frag->rdma_length);
/* if a pointer to a receive request is not set it means that
* ACK was not yet received. Don't schedule sends before ACK */
if(NULL != sendreq->req_recv.pval)
mca_pml_ob1_send_request_schedule(sendreq);
/* send fragment by copy in/out */
mca_pml_ob1_send_request_copy_in_out(sendreq,
frag->rdma_hdr.hdr_rdma.hdr_rdma_offset, frag->rdma_length);
/* if a pointer to a receive request is not set it means that
* ACK was not yet received. Don't schedule sends before ACK */
if(NULL != sendreq->req_recv.pval)
mca_pml_ob1_send_request_schedule(sendreq);
}
return OMPI_ERR_OUT_OF_RESOURCE;
}
return OMPI_ERR_OUT_OF_RESOURCE;
} else {
/* already have a source descriptor */
des = sendreq->src_des;
sendreq->src_des = NULL;
}
des->des_dst = frag->rdma_segs;
des->des_dst = frag->rdma_segs;
des->des_dst_cnt = frag->rdma_hdr.hdr_rdma.hdr_seg_cnt;
des->des_cbfunc = mca_pml_ob1_put_completion;
des->des_cbdata = frag;
des->des_cbfunc = mca_pml_ob1_put_completion;
des->des_cbdata = frag;
PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE,
&(((mca_pml_ob1_send_request_t*)frag->rdma_req)->req_send.req_base), save_size, PERUSE_SEND );
rc = mca_bml_base_put(bml_btl, des);
if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
mca_bml_base_free(bml_btl, des);
frag->rdma_length = save_size;
if(OMPI_ERR_OUT_OF_RESOURCE == rc) {
@ -1203,6 +1215,7 @@ int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t* frag )
orte_errmgr.abort(-1, NULL);
}
}
return OMPI_SUCCESS;
}
@ -1261,21 +1274,25 @@ void mca_pml_ob1_send_request_put( mca_pml_ob1_send_request_t* sendreq,
frag->reg = NULL;
frag->retries = 0;
if (OPAL_UNLIKELY(NULL != sendreq->src_des)) {
/* get fallback path */
sendreq->req_state = 0;
}
/* lookup the corresponding registration */
for(i=0; i<sendreq->req_rdma_cnt; i++) {
if(sendreq->req_rdma[i].bml_btl == frag->rdma_bml) {
frag->reg = sendreq->req_rdma[i].btl_reg;
break;
}
}
if(sendreq->req_rdma[i].bml_btl == frag->rdma_bml) {
frag->reg = sendreq->req_rdma[i].btl_reg;
break;
}
}
/* RDMA writes may proceed in parallel to send and to each other, so
* create clone of the convertor for each RDMA fragment
*/
size = hdr->hdr_rdma_offset;
opal_convertor_clone_with_position(&sendreq->req_send.req_base.req_convertor,
&frag->convertor, 0, &size);
&frag->convertor, 0, &size);
mca_pml_ob1_send_request_put_frag(frag);
}

2
ompi/mca/pml/ob1/pml_ob1_sendreq.h
Просмотреть файл

@ -54,6 +54,7 @@ struct mca_pml_ob1_send_request_t {
mca_pml_ob1_send_pending_t req_pending;
opal_mutex_t req_send_range_lock;
opal_list_t req_send_ranges;
mca_btl_base_descriptor_t *src_des;
mca_pml_ob1_com_btl_t req_rdma[1];
};
typedef struct mca_pml_ob1_send_request_t mca_pml_ob1_send_request_t;
@ -129,6 +130,7 @@ get_request_from_send_pending(mca_pml_ob1_send_pending_t *type)
OMPI_FREE_LIST_WAIT(&mca_pml_base_send_requests, item, rc); \
sendreq = (mca_pml_ob1_send_request_t*)item; \
sendreq->req_send.req_base.req_proc = proc; \
sendreq->src_des = NULL; \
} \
}