/**
* Send an FIN to the peer. If we fail to send this ack (no more available
* fragments or the send failed) this function automatically add the FIN
* to the list of pending FIN, Which guarantee that the FIN will be sent
* later.
*/
int mca_pml_ob1_send_fin( ompi_proc_t* proc,
mca_bml_base_btl_t* bml_btl,
ompi_ptr_t hdr_des,
uint8_t order,
uint32_t status )
{
mca_btl_base_descriptor_t* fin;
mca_pml_ob1_fin_hdr_t* hdr;
int rc;
mca_bml_base_alloc(bml_btl, &fin, order, sizeof(mca_pml_ob1_fin_hdr_t),
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
if(NULL == fin) {
MCA_PML_OB1_ADD_FIN_TO_PENDING(proc, hdr_des, bml_btl, order, status);
return OMPI_ERR_OUT_OF_RESOURCE;
}
fin->des_cbfunc = mca_pml_ob1_fin_completion;
fin->des_cbdata = NULL;
/* fill in header */
hdr = (mca_pml_ob1_fin_hdr_t*)fin->des_src->seg_addr.pval;
hdr->hdr_common.hdr_flags = 0;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_FIN;
hdr->hdr_des = hdr_des;
hdr->hdr_fail = status;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_FIN, proc);
/* queue request */
rc = mca_bml_base_send( bml_btl,
fin,
MCA_PML_OB1_HDR_TYPE_FIN );
if( OPAL_LIKELY( rc >= 0 ) ) {
if( OPAL_LIKELY( 1 == rc ) ) {
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
return OMPI_SUCCESS;
}
mca_bml_base_free(bml_btl, fin);
MCA_PML_OB1_ADD_FIN_TO_PENDING(proc, hdr_des, bml_btl, order, status);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/**
* Send an FIN to the peer. If we fail to send this ack (no more available
* fragments or the send failed) this function automatically add the FIN
* to the list of pending FIN, Which guarantee that the FIN will be sent
* later.
*/
int mca_pml_ob1_send_fin( ompi_proc_t* proc,
mca_bml_base_btl_t* bml_btl,
opal_ptr_t hdr_frag,
uint64_t rdma_size,
uint8_t order,
int status )
{
mca_btl_base_descriptor_t* fin;
int rc;
mca_bml_base_alloc(bml_btl, &fin, order, sizeof(mca_pml_ob1_fin_hdr_t),
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_FLAGS_SIGNAL);
if(NULL == fin) {
MCA_PML_OB1_ADD_FIN_TO_PENDING(proc, hdr_frag, rdma_size, bml_btl, order, status);
return OMPI_ERR_OUT_OF_RESOURCE;
}
fin->des_cbfunc = mca_pml_ob1_fin_completion;
fin->des_cbdata = NULL;
/* fill in header */
mca_pml_ob1_fin_hdr_prepare ((mca_pml_ob1_fin_hdr_t *) fin->des_segments->seg_addr.pval,
0, hdr_frag.lval, status ? status : (int64_t) rdma_size);
ob1_hdr_hton((mca_pml_ob1_hdr_t *) fin->des_segments->seg_addr.pval, MCA_PML_OB1_HDR_TYPE_FIN, proc);
/* queue request */
rc = mca_bml_base_send( bml_btl, fin, MCA_PML_OB1_HDR_TYPE_FIN );
if( OPAL_LIKELY( rc >= 0 ) ) {
if( OPAL_LIKELY( 1 == rc ) ) {
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
return OMPI_SUCCESS;
}
mca_bml_base_free(bml_btl, fin);
MCA_PML_OB1_ADD_FIN_TO_PENDING(proc, hdr_frag, rdma_size, bml_btl, order, status);
return OMPI_ERR_OUT_OF_RESOURCE;
}
static inline int mca_spml_yoda_put_internal(void *dst_addr,
size_t size,
void *src_addr,
int dst,
int is_nb)
{
int rc = OSHMEM_SUCCESS;
mca_spml_yoda_put_request_t *putreq = NULL;
mca_bml_base_btl_t* bml_btl;
mca_btl_base_descriptor_t* des = NULL;
mca_btl_base_segment_t* segment;
mca_spml_yoda_rdma_frag_t* frag;
int nfrags;
int i;
unsigned ncopied = 0;
unsigned int frag_size = 0;
char *p_src, *p_dst;
void* rva;
sshmem_mkey_t *r_mkey;
int btl_id = 0;
struct yoda_btl *ybtl;
int put_via_send;
mca_btl_base_registration_handle_t *local_handle = NULL, *remote_handle = NULL;
/* If nothing to put its OK.*/
if (0 >= size) {
return OSHMEM_SUCCESS;
}
/* Find bml_btl and its global btl_id */
bml_btl = get_next_btl(dst, &btl_id);
if (!bml_btl) {
SPML_ERROR("cannot reach %d pe: no appropriate btl found", oshmem_my_proc_id());
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
/* Check if btl has PUT method. If it doesn't - use SEND*/
put_via_send = !(bml_btl->btl->btl_flags & MCA_BTL_FLAGS_PUT);
/* Get rkey of remote PE (dst proc) which must be on memheap*/
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, btl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
#if SPML_YODA_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
ybtl = &mca_spml_yoda.btl_type_map[btl_id];
if (ybtl->btl->btl_register_mem) {
assert (r_mkey->len == ybtl->btl->btl_registration_handle_size);
remote_handle = (mca_btl_base_registration_handle_t *) r_mkey->u.data;
}
/* check if we doing put into shm attached segment and if so
* just do memcpy
*/
if ((YODA_BTL_SM == ybtl->btl_type || YODA_BTL_VADER == ybtl->btl_type)
&& mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size);
return OSHMEM_SUCCESS;
}
/* We support only blocking PUT now => we always need copy for src buffer*/
calc_nfrags_put (bml_btl, size, &frag_size, &nfrags, put_via_send);
p_src = (char*) src_addr;
p_dst = (char*) (unsigned long) rva;
for (i = 0; i < nfrags; i++) {
/* Allocating send request from free list */
putreq = mca_spml_yoda_putreq_alloc(dst);
frag = &putreq->put_frag;
ncopied = i < nfrags - 1 ? frag_size :(unsigned) ((char *) src_addr + size - p_src);
/* Preparing source buffer */
/* allocate buffer */
mca_spml_yoda_bml_alloc(bml_btl,
&des,
MCA_BTL_NO_ORDER,
ncopied,
MCA_BTL_DES_SEND_ALWAYS_CALLBACK,
put_via_send);
if (OPAL_UNLIKELY(!des || !des->des_segments )) {
SPML_ERROR("src=%p nfrags = %d frag_size=%d",
src_addr, nfrags, frag_size);
SPML_ERROR("shmem OOM error need %d bytes", ncopied);
opal_show_help("help-oshmem-spml-yoda.txt",
"internal_oom_error",
true,
"Put", ncopied, mca_spml_yoda.bml_alloc_threshold);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
/* copy data to allocated buffer*/
segment = des->des_segments;
spml_yoda_prepare_for_put((void*)segment->seg_addr.pval, ncopied,
(void*)p_src, (void*)p_dst, put_via_send);
if (!put_via_send && ybtl->btl->btl_register_mem) {
local_handle = ybtl->btl->btl_register_mem (ybtl->btl, bml_btl->btl_endpoint,
segment->seg_addr.pval, ncopied, 0);
if (NULL == local_handle) {
/* No free resources, Block on completion here */
SPML_ERROR("shmem error: OSHMEM_ERR_OUT_OF_RESOURCE");
oshmem_request_wait_completion(&putreq->req_put.req_base.req_oshmem);
}
}
frag->rdma_segs[0].base_seg.seg_addr.lval = (uintptr_t) p_dst;
frag->rdma_segs[0].base_seg.seg_len = (put_via_send ?
ncopied + SPML_YODA_SEND_CONTEXT_SIZE :
ncopied);
frag->rdma_req = putreq;
/* initialize callback data for put*/
des->des_cbdata = frag;
des->des_cbfunc = mca_spml_yoda_put_completion;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_puts, 1);
/* put the data to remote side */
if (!put_via_send) {
rc = mca_bml_base_put (bml_btl, segment->seg_addr.pval, (uint64_t) (intptr_t) p_dst,
local_handle, remote_handle, ncopied, 0, 0, mca_spml_yoda_put_completion_rdma,
des);
} else {
rc = mca_bml_base_send(bml_btl, des, MCA_SPML_YODA_PUT);
if (1 == rc)
rc = OSHMEM_SUCCESS;
}
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != rc)) {
if (OSHMEM_ERR_OUT_OF_RESOURCE == rc) {
/* No free resources, Block on completion here */
SPML_ERROR("shmem error: OSHMEM_ERR_OUT_OF_RESOURCE");
oshmem_request_wait_completion(&putreq->req_put.req_base.req_oshmem);
} else {
SPML_ERROR("shmem error");
}
/* exit with errro */
SPML_ERROR("shmem error: ret = %i, send_pe = %i, dest_pe = %i",
rc, oshmem_my_proc_id(), dst);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
p_src += ncopied;
p_dst += ncopied;
}
return rc;
exit_fatal:
if (OSHMEM_SUCCESS != rc) {
oshmem_shmem_abort(rc);
}
return rc;
}
static void mca_yoda_get_callback(mca_btl_base_module_t* btl,
mca_btl_base_tag_t tag,
mca_btl_base_descriptor_t* des,
void* cbdata )
{
void** p, ** p_src, **p_dst;
size_t* size;
int* dst;
void** p_getreq;
mca_btl_base_descriptor_t* des_loc;
int rc;
mca_bml_base_btl_t* bml_btl;
mca_spml_yoda_rdma_frag_t* frag;
int btl_id;
mca_spml_yoda_put_request_t *putreq;
rc = OSHMEM_SUCCESS;
btl_id = 0;
putreq = NULL;
/* Unpack data */
p = (void **)des->des_segments->seg_addr.pval;
p_src = (void*) p;
size = (size_t*)((char*)p_src + sizeof(*p_src) );
dst = (int*)( (char*)size + sizeof(*size));
p_dst = (void*) ((char*)dst + sizeof(*dst));
p_getreq =(void**) ( (char*)p_dst + sizeof(*p_dst));
/* Prepare put via send*/
bml_btl = get_next_btl(*dst, &btl_id);
putreq = mca_spml_yoda_putreq_alloc(*dst);
frag = &putreq->put_frag;
mca_spml_yoda_bml_alloc(bml_btl,
&des_loc,
MCA_BTL_NO_ORDER,
*size,
MCA_BTL_DES_SEND_ALWAYS_CALLBACK,
1);
if (OPAL_UNLIKELY(!des_loc || !des_loc->des_segments)) {
SPML_ERROR("shmem OOM error need %d bytes", (int)*size);
oshmem_shmem_abort(-1);
}
spml_yoda_prepare_for_get_response((void*)des_loc->des_segments->seg_addr.pval, *size, (void*)*p_src, (void*) *p_dst,(void*)*p_getreq,1);
frag->rdma_req = putreq;
/* Initialize callback data for put*/
des_loc->des_cbdata = frag;
des_loc->des_cbfunc = mca_spml_yoda_put_completion;
des_loc->des_segment_count = 1;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_puts, 1);
/* Put via send*/
rc = mca_bml_base_send(bml_btl, des_loc, MCA_SPML_YODA_GET_RESPONSE);
if (1 == rc) {
rc = OSHMEM_SUCCESS;
}
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != rc)) {
if (OSHMEM_ERR_OUT_OF_RESOURCE == rc) {
/* No free resources, Block on completion here */
SPML_ERROR("shmem error: OSHMEM_ERR_OUT_OF_RESOURCE");
oshmem_request_wait_completion(&putreq->req_put.req_base.req_oshmem);
} else {
SPML_ERROR("shmem error");
}
/* exit with errro */
SPML_ERROR("shmem error: ret = %i, send_pe = %i, dest_pe = %i",
rc, oshmem_my_proc_id(), *dst);
oshmem_shmem_abort(-1);
rc = OSHMEM_ERROR;
}
}
/**
* shmem_get reads data from a remote address
* in the symmetric heap via RDMA READ.
* Get operation:
* 1. Get the rkey to the remote address.
* 2. Allocate a get request.
* 3. Allocated a temporary pre-registered buffer
* to copy the data to.
* 4. Init the request descriptor with remote side
* data and local side data.
* 5. Read the remote buffer to a pre-registered
* buffer on the local PE using RDMA READ.
* 6. Copy the received data to dst_addr if an
* intermediate pre-register buffer was used.
* 7. Clear the request and return.
*
* src_addr - address on remote pe.
* size - the amount on bytes to be read.
* dst_addr - address on the local pe.
* src - the pe of remote process.
*/
int mca_spml_yoda_get(void* src_addr, size_t size, void* dst_addr, int src)
{
int rc = OSHMEM_SUCCESS;
sshmem_mkey_t *r_mkey, *l_mkey;
void* rva;
unsigned ncopied = 0;
unsigned int frag_size = 0;
char *p_src, *p_dst;
int i;
int nfrags;
mca_bml_base_btl_t* bml_btl = NULL;
mca_btl_base_segment_t* segment;
mca_btl_base_descriptor_t* des = NULL;
mca_spml_yoda_rdma_frag_t* frag = NULL;
struct mca_spml_yoda_getreq_parent get_holder;
struct yoda_btl *ybtl;
int btl_id = 0;
int get_via_send;
mca_btl_base_registration_handle_t *local_handle, *remote_handle = NULL;
mca_spml_yoda_get_request_t* getreq = NULL;
/*If nothing to get its OK.*/
if (0 >= size) {
return rc;
}
/* Find bml_btl and its global btl_id */
bml_btl = get_next_btl(src, &btl_id);
if (!bml_btl) {
SPML_ERROR("cannot reach %d pe: no appropriate btl found", oshmem_my_proc_id());
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
/* Check if btl has GET method. If it doesn't - use SEND*/
get_via_send = ! ( (bml_btl->btl->btl_flags & (MCA_BTL_FLAGS_GET)) &&
(bml_btl->btl->btl_flags & (MCA_BTL_FLAGS_PUT)) );
/* Get rkey of remote PE (src proc) which must be on memheap*/
r_mkey = mca_memheap_base_get_cached_mkey(src, src_addr, btl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
src, src_addr);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
#if SPML_YODA_DEBUG == 1
SPML_VERBOSE(100, "get: pe:%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
ybtl = &mca_spml_yoda.btl_type_map[btl_id];
if (ybtl->btl->btl_register_mem) {
assert(ybtl->btl->btl_registration_handle_size == r_mkey->len);
remote_handle = (mca_btl_base_registration_handle_t *) r_mkey->u.data;
}
nfrags = 1;
/* check if we doing get into shm attached segment and if so
* just do memcpy
*/
if ((YODA_BTL_SM == ybtl->btl_type || YODA_BTL_VADER == ybtl->btl_type)
&& mca_memheap_base_can_local_copy(r_mkey, src_addr)) {
memcpy(dst_addr, (void *) rva, size);
/* must call progress here to avoid deadlock. Scenarion:
* pe1 pols pe2 via shm get. pe2 tries to get static variable from node one, which goes to sm btl
* In this case pe2 is stuck forever because pe1 never calls opal_progress.
* May be we do not need to call progress on every get() here but rather once in a while.
*/
opal_progress();
return OSHMEM_SUCCESS;
}
l_mkey = mca_memheap.memheap_get_local_mkey(dst_addr,
btl_id);
/*
* Need a copy if local memory has not been registered or
* we make GET via SEND
*/
frag_size = ncopied;
if ((NULL == l_mkey) || get_via_send) {
calc_nfrags_get (bml_btl, size, &frag_size, &nfrags, get_via_send);
}
p_src = (char*) (unsigned long) rva;
p_dst = (char*) dst_addr;
get_holder.active_count = 0;
for (i = 0; i < nfrags; i++) {
/**
* Allocating a get request from a pre-allocated
* and pre-registered free list.
*/
getreq = mca_spml_yoda_getreq_alloc(src);
assert(getreq);
getreq->p_dst = NULL;
frag = &getreq->get_frag;
getreq->parent = &get_holder;
ncopied = i < nfrags - 1 ? frag_size :(unsigned) ((char *) dst_addr + size - p_dst);
frag->allocated = 0;
/* Prepare destination descriptor*/
memcpy(&frag->rdma_segs[0].base_seg,
r_mkey->u.data,
r_mkey->len);
frag->rdma_segs[0].base_seg.seg_len = (get_via_send ? ncopied + SPML_YODA_SEND_CONTEXT_SIZE : ncopied);
if (get_via_send) {
frag->use_send = 1;
frag->allocated = 1;
/**
* Allocate a temporary buffer on the local PE.
* The local buffer will store the data read
* from the remote address.
*/
mca_spml_yoda_bml_alloc(bml_btl,
&des,
MCA_BTL_NO_ORDER,
(int)frag_size,
MCA_BTL_DES_SEND_ALWAYS_CALLBACK,
get_via_send);
if (OPAL_UNLIKELY(!des || !des->des_segments)) {
SPML_ERROR("shmem OOM error need %d bytes", ncopied);
SPML_ERROR("src=%p nfrags = %d frag_size=%d",
src_addr, nfrags, frag_size);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
segment = des->des_segments;
spml_yoda_prepare_for_get((void*)segment->seg_addr.pval, ncopied, (void*)p_src, oshmem_my_proc_id(), (void*)p_dst, (void*) getreq);
des->des_cbfunc = mca_spml_yoda_get_response_completion;
des->des_cbdata = frag;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_gets, 1);
}
else {
/*
* Register src memory if do GET via GET
*/
if (NULL == l_mkey && ybtl->btl->btl_register_mem) {
local_handle = ybtl->btl->btl_register_mem (ybtl->btl, bml_btl->btl_endpoint, p_dst, ncopied,
MCA_BTL_REG_FLAG_LOCAL_WRITE);
if (NULL == local_handle) {
SPML_ERROR("%s: failed to register destination memory %p.",
btl_type2str(ybtl->btl_type), p_dst);
}
frag->local_handle = local_handle;
} else {
local_handle = ((mca_spml_yoda_context_t*)l_mkey->spml_context)->registration;
frag->local_handle = NULL;
}
frag->rdma_segs[0].base_seg.seg_addr.lval = (uintptr_t) p_src;
getreq->p_dst = (uint64_t*) p_dst;
frag->size = ncopied;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_gets, 1);
}
/**
* Initialize the remote data fragment
* with remote address data required for
* executing RDMA READ from a remote buffer.
*/
frag->rdma_req = getreq;
/**
* Do GET operation
*/
if (get_via_send) {
rc = mca_bml_base_send(bml_btl, des, MCA_SPML_YODA_GET);
if (1 == rc)
rc = OSHMEM_SUCCESS;
} else {
rc = mca_bml_base_get(bml_btl, p_dst, (uint64_t) (intptr_t) p_src, local_handle,
remote_handle, ncopied, 0, 0, mca_spml_yoda_get_completion, frag);
}
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != rc)) {
if (OSHMEM_ERR_OUT_OF_RESOURCE == rc) {
/* No free resources, Block on completion here */
oshmem_request_wait_completion(&getreq->req_get.req_base.req_oshmem);
return OSHMEM_SUCCESS;
} else {
SPML_ERROR("oshmem_get: error %d", rc);
goto exit_fatal;
}
}
p_dst += ncopied;
p_src += ncopied;
OPAL_THREAD_ADD32(&get_holder.active_count, 1);
}
/* revisit if we really need this for self and sm */
/* if (YODA_BTL_SELF == ybtl->btl_type) */
opal_progress();
/* Wait for completion on request */
while (get_holder.active_count > 0)
oshmem_request_wait_completion(&getreq->req_get.req_base.req_oshmem);
return rc;
exit_fatal:
if (OSHMEM_SUCCESS != rc) {
oshmem_shmem_abort(rc);
}
return rc;
}
