int mca_btl_self_rdma( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_btl_base_descriptor_t* des )
{
mca_btl_base_segment_t* src = des->des_src;
mca_btl_base_segment_t* dst = des->des_dst;
size_t src_cnt = des->des_src_cnt;
size_t dst_cnt = des->des_dst_cnt;
unsigned char* src_addr = (unsigned char*)src->seg_addr.pval;
size_t src_len = src->seg_len;
unsigned char* dst_addr = (unsigned char*)ompi_ptr_ltop(dst->seg_addr.lval);
size_t dst_len = dst->seg_len;
while(src_len && dst_len) {
if(src_len == dst_len) {
memcpy(dst_addr, src_addr, src_len);
/* advance src */
if(--src_cnt != 0) {
src++;
src_addr = (unsigned char*)src->seg_addr.pval;
src_len = src->seg_len;
} else {
src_len = 0;
}
/* advance dst */
if(--dst_cnt != 0) {
dst++;
dst_addr = (unsigned char*)dst->seg_addr.pval;
dst_len = dst->seg_len;
} else {
dst_len = 0;
}
} else {
size_t bytes = src_len < dst_len ? src_len : dst_len;
memcpy(dst_addr, src_addr, bytes);
/* advance src */
src_len -= bytes;
if(src_len == 0) {
if(--src_cnt != 0) {
src++;
src_addr = (unsigned char*)src->seg_addr.pval;
src_len = src->seg_len;
}
} else {
src_addr += bytes;
}
/* advance dst */
dst_len -= bytes;
if(dst_len == 0) {
if(--dst_cnt != 0) {
dst++;
dst_addr = (unsigned char*)src->seg_addr.pval;
dst_len = src->seg_len;
}
} else {
dst_addr += bytes;
}
}
}
/* rdma completion */
des->des_cbfunc(btl,endpoint,des,OMPI_SUCCESS);
return OMPI_SUCCESS;
}
bool mca_btl_tcp_frag_recv(mca_btl_tcp_frag_t* frag, int sd)
{
int cnt, dont_copy_data = 0;
size_t i, num_vecs;
mca_btl_base_endpoint_t* btl_endpoint = frag->endpoint;
repeat:
num_vecs = frag->iov_cnt;
#if MCA_BTL_TCP_ENDPOINT_CACHE
if( 0 != btl_endpoint->endpoint_cache_length ) {
size_t length;
/* It's strange at the first look but cnt have to be set to the full amount of data
* available. After going to advance_iov_position we will use cnt to detect if there
* is still some data pending.
*/
cnt = length = btl_endpoint->endpoint_cache_length;
for( i = 0; i < frag->iov_cnt; i++ ) {
if( length > frag->iov_ptr[i].iov_len )
length = frag->iov_ptr[i].iov_len;
if( (0 == dont_copy_data) || (length < frag->iov_ptr[i].iov_len) ) {
memcpy( frag->iov_ptr[i].iov_base, btl_endpoint->endpoint_cache_pos, length );
} else {
frag->segments[0].seg_addr.pval = btl_endpoint->endpoint_cache_pos;
frag->iov_ptr[i].iov_base = btl_endpoint->endpoint_cache_pos;
}
btl_endpoint->endpoint_cache_pos += length;
btl_endpoint->endpoint_cache_length -= length;
length = btl_endpoint->endpoint_cache_length;
if( 0 == length ) {
btl_endpoint->endpoint_cache_pos = btl_endpoint->endpoint_cache;
break;
}
}
goto advance_iov_position;
}
/* What's happens if all iovecs are used by the fragment ? It still work, as we reserve one
* iovec for the caching in the fragment structure (the +1).
*/
frag->iov_ptr[num_vecs].iov_base = btl_endpoint->endpoint_cache_pos;
frag->iov_ptr[num_vecs].iov_len =
mca_btl_tcp_component.tcp_endpoint_cache - btl_endpoint->endpoint_cache_length;
num_vecs++;
#endif /* MCA_BTL_TCP_ENDPOINT_CACHE */
/* non-blocking read, but continue if interrupted */
cnt = -1;
while( cnt < 0 ) {
cnt = readv(sd, frag->iov_ptr, num_vecs);
if( 0 < cnt ) goto advance_iov_position;
if( cnt == 0 ) {
mca_btl_tcp_endpoint_close(btl_endpoint);
return false;
}
switch(opal_socket_errno) {
case EINTR:
continue;
case EWOULDBLOCK:
return false;
case EFAULT:
BTL_ERROR(("mca_btl_tcp_frag_recv: readv error (%p, %d)\n\t%s(%d)\n",
frag->iov_ptr[0].iov_base, frag->iov_ptr[0].iov_len,
strerror(opal_socket_errno), frag->iov_cnt));
mca_btl_tcp_endpoint_close(btl_endpoint);
return false;
default:
BTL_ERROR(("mca_btl_tcp_frag_recv: readv failed: %s (%d)",
strerror(opal_socket_errno),
opal_socket_errno));
mca_btl_tcp_endpoint_close(btl_endpoint);
return false;
}
};
advance_iov_position:
/* if the read didn't complete - update the iovec state */
num_vecs = frag->iov_cnt;
for( i = 0; i < num_vecs; i++ ) {
if( cnt < (int)frag->iov_ptr->iov_len ) {
frag->iov_ptr->iov_base = (ompi_iov_base_ptr_t)
(((unsigned char*)frag->iov_ptr->iov_base) + cnt);
frag->iov_ptr->iov_len -= cnt;
cnt = 0;
break;
}
cnt -= frag->iov_ptr->iov_len;
frag->iov_idx++;
frag->iov_ptr++;
frag->iov_cnt--;
}
#if MCA_BTL_TCP_ENDPOINT_CACHE
btl_endpoint->endpoint_cache_length = cnt;
#endif /* MCA_BTL_TCP_ENDPOINT_CACHE */
/* read header */
if(frag->iov_cnt == 0) {
if (btl_endpoint->endpoint_nbo && frag->iov_idx == 1) MCA_BTL_TCP_HDR_NTOH(frag->hdr);
switch(frag->hdr.type) {
case MCA_BTL_TCP_HDR_TYPE_SEND:
if(frag->iov_idx == 1 && frag->hdr.size) {
frag->segments[0].seg_addr.pval = frag+1;
frag->segments[0].seg_len = frag->hdr.size;
frag->iov[1].iov_base = (IOVBASE_TYPE*)(frag->segments[0].seg_addr.pval);
frag->iov[1].iov_len = frag->hdr.size;
frag->iov_cnt++;
#ifndef __sparc
/* The following cannot be done for sparc code
* because it causes alignment errors when accessing
* structures later on in the btl and pml code.
*/
dont_copy_data = 1;
#endif
goto repeat;
}
break;
case MCA_BTL_TCP_HDR_TYPE_PUT:
if(frag->iov_idx == 1) {
frag->iov[1].iov_base = (IOVBASE_TYPE*)frag->segments;
frag->iov[1].iov_len = frag->hdr.count * sizeof(mca_btl_base_segment_t);
frag->iov_cnt++;
goto repeat;
} else if (frag->iov_idx == 2) {
for( i = 0; i < frag->hdr.count; i++ ) {
frag->iov[i+2].iov_base = (IOVBASE_TYPE*)ompi_ptr_ltop(frag->segments[i].seg_addr.lval);
frag->iov[i+2].iov_len = frag->segments[i].seg_len;
}
frag->iov_cnt += frag->hdr.count;
goto repeat;
}
break;
case MCA_BTL_TCP_HDR_TYPE_GET:
default:
break;
}
return true;
}
return false;
}
