int mca_spml_ikrit_put(shmem_ctx_t ctx, void* dst_addr, size_t size, void* src_addr, int dst)
{
int err;
mca_spml_ikrit_put_request_t *put_req;
mxm_wait_t wait;
put_req = 0;
err = mca_spml_ikrit_put_internal(dst_addr,
size,
src_addr,
dst,
(void **) &put_req,
0);
if (OSHMEM_SUCCESS != err) {
SPML_ERROR("put failed - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
if (!put_req)
return OSHMEM_SUCCESS;
wait.req = &put_req->mxm_req.base;
wait.state = (mxm_req_state_t)(MXM_REQ_SENT | MXM_REQ_COMPLETED);
wait.progress_cb = NULL;
wait.progress_arg = NULL;
mxm_wait(&wait);
return OSHMEM_SUCCESS;
}
int ompi_mtl_mxm_send(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t* comm, int dest, int tag,
struct opal_convertor_t *convertor,
mca_pml_base_send_mode_t mode)
{
mxm_send_req_t mxm_send_req;
mxm_wait_t wait;
mxm_error_t err;
int ret;
/* prepare local send request */
mxm_send_req.base.state = MXM_REQ_NEW;
mxm_send_req.base.mq = ompi_mtl_mxm_mq_lookup(comm);
mxm_send_req.base.conn = ompi_mtl_mxm_conn_lookup(comm, dest);
mxm_send_req.base.flags = MXM_REQ_FLAG_BLOCKING;
mxm_send_req.base.context = NULL;
ret = ompi_mtl_mxm_choose_send_datatype(&mxm_send_req, convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
mxm_send_req.base.data.buffer.mkey = MXM_MKEY_NONE;
mxm_send_req.base.completed_cb = NULL;
if (mode == MCA_PML_BASE_SEND_SYNCHRONOUS) {
mxm_send_req.base.flags |= MXM_REQ_FLAG_SEND_SYNC;
}
mxm_send_req.opcode = MXM_REQ_OP_SEND;
mxm_send_req.op.send.tag = tag;
mxm_send_req.op.send.imm_data = ompi_comm_rank(comm);
/* post-send */
err = mxm_req_send(&mxm_send_req);
if (MXM_OK != err) {
orte_show_help("help-mtl-mxm.txt", "error posting send", true, 0, mxm_error_string(err));
return OMPI_ERROR;
}
/* wait for request completion */
wait.req = &mxm_send_req.base;
wait.state = MXM_REQ_COMPLETED;
wait.progress_cb = ompi_mtl_mxm_send_progress_cb;
wait.progress_arg = NULL;
mxm_wait(&wait);
return OMPI_SUCCESS;
}
static inline void gasnetc_ReqRepWait(mxm_send_req_t * mxm_sreq,
uint8_t is_request, uint8_t msg_num)
{
if (is_request) {
/*
* Ensure progress, but only on requester side (even on async request).
* We don't want to have recursion on responder side - otherwise
* send might be called from context of receive and so on.
*/
#if GASNET_DEBUG_AM
MXM_LOG("[pid %d] polling and testing(%s, msg_num = 0x%02x)...\n",
getpid(), is_request ? "request" : "response", msg_num);
#endif
while (!mxm_req_test(&mxm_sreq->base))
gasnetc_AMPoll();
#if GASNET_DEBUG_AM
MXM_LOG("[pid %d] done polling and testing(%s, msg_num = 0x%02x)...\n",
getpid(), is_request ? "request" : "response", msg_num);
#endif
} else {
mxm_wait_t wait;
wait.progress_cb = NULL;
wait.progress_arg = NULL;
wait.req = &mxm_sreq->base;
wait.state = MXM_REQ_COMPLETED;
/*
* We're on the responder side - just wait for completion.
*/
#if GASNET_DEBUG_AM
MXM_LOG("[pid %d] calling mxm_req_wait(%s, msg_num = 0x%02x)...\n",
getpid(), is_request ? "request" : "response", msg_num);
#endif
mxm_wait(&wait);
#if GASNET_DEBUG_AM
MXM_LOG("[pid %d] mxm_req_wait(%s, msg_num = 0x%02x) returned\n",
getpid(), is_request ? "request" : "response", msg_num);
#endif
}
}
/* simple buffered put implementation. NOT IN USE
* Problems:
* - slighly worse performance than impl based on non buffered put
* - fence complexity is O(n_active_connections) instead of O(n_connections_with_outstanding_puts).
* Later is bounded by the network RTT & mxm ack timer.
*/
int mca_spml_ikrit_put_simple(void* dst_addr,
size_t size,
void* src_addr,
int dst)
{
void *rva;
mxm_send_req_t mxm_req;
mxm_wait_t wait;
int ptl_id;
mxm_mem_key_t *mkey;
static int count;
ptl_id = get_ptl_id(dst);
mkey = mca_spml_ikrit_get_mkey(dst, dst_addr, ptl_id, &rva, &mca_spml_ikrit);
SPML_VERBOSE_FASTPATH(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva);
if (NULL == mkey) {
memcpy((void *) (unsigned long) rva, src_addr, size);
/* call progress as often as we would have with regular put */
if (++count % SPML_IKRIT_PACKETS_PER_SYNC == 0)
mxm_progress(mca_spml_ikrit.mxm_context);
return OSHMEM_SUCCESS;
}
SPML_VERBOSE_FASTPATH(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, MXM_PTL_RDMA, dst_addr, src_addr, (int)size, (void *)rva);
/* fill out request */
mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
mxm_req.flags = MXM_REQ_SEND_FLAG_BLOCKING;
mxm_req.base.conn = mca_spml_ikrit.mxm_peers[dst].mxm_conn;
mxm_req.base.data_type = MXM_REQ_DATA_BUFFER;
mxm_req.base.data.buffer.ptr = src_addr;
mxm_req.base.data.buffer.length = size;
mxm_req.base.completed_cb = 0;
mxm_req.base.context = 0;
mxm_req.opcode = MXM_REQ_OP_PUT;
mxm_req.op.mem.remote_vaddr = (intptr_t) rva;
mxm_req.base.state = MXM_REQ_NEW;
mxm_req.base.error = MXM_OK;
mxm_req.op.mem.remote_mkey = mkey;
if (mca_spml_ikrit.mxm_peers[dst].need_fence == 0) {
opal_list_append(&mca_spml_ikrit.active_peers,
&mca_spml_ikrit.mxm_peers[dst].link);
mca_spml_ikrit.mxm_peers[dst].need_fence = 1;
}
SPML_IKRIT_MXM_POST_SEND(mxm_req);
wait.req = &mxm_req.base;
wait.state = (mxm_req_state_t)(MXM_REQ_SENT | MXM_REQ_COMPLETED);
wait.progress_cb = NULL;
wait.progress_arg = NULL;
mxm_wait(&wait);
return OSHMEM_SUCCESS;
}
/* simple buffered put implementation. NOT IN USE
* Problems:
* - slighly worse performance than impl based on non buffered put
* - fence complexity is O(n_active_connections) instead of O(n_connections_with_outstanding_puts).
* Later is bounded by the network RTT & mxm ack timer.
*/
int mca_spml_ikrit_put_simple(void* dst_addr,
size_t size,
void* src_addr,
int dst)
{
void *rva;
mxm_send_req_t mxm_req;
mxm_wait_t wait;
int ptl_id;
sshmem_mkey_t *r_mkey;
static int count;
ptl_id = get_ptl_id(dst);
/* Get rkey of remote PE (dst proc) which must be on memheap */
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, ptl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
#if SPML_IKRIT_PUT_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
if (ptl_id == MXM_PTL_SHM) {
if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size);
/* call progress as often as we would have with regular put */
if (++count % SPML_IKRIT_PACKETS_PER_SYNC == 0)
mxm_progress(mca_spml_ikrit.mxm_context);
return OSHMEM_SUCCESS;
}
/* segment not mapped - fallback to rmda */
ptl_id = MXM_PTL_RDMA;
r_mkey = mca_memheap_base_get_cached_mkey(dst,
//(unsigned long) dst_addr,
dst_addr,
ptl_id,
&rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
}
#if SPML_IKRIT_PUT_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
/* fill out request */
mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
#if MXM_API < MXM_VERSION(2,0)
mxm_req.base.flags = MXM_REQ_FLAG_BLOCKING;
#else
mxm_req.flags = MXM_REQ_SEND_FLAG_BLOCKING;
#endif
mxm_req.base.conn = mca_spml_ikrit.mxm_peers[dst]->mxm_conn;
mxm_req.base.data_type = MXM_REQ_DATA_BUFFER;
mxm_req.base.data.buffer.ptr = src_addr;
mxm_req.base.data.buffer.length = size;
mxm_req.base.completed_cb = 0;
mxm_req.base.context = 0;
mxm_req.opcode = MXM_REQ_OP_PUT;
mxm_req.op.mem.remote_vaddr = (intptr_t) rva;
mxm_req.base.state = MXM_REQ_NEW;
mxm_req.base.error = MXM_OK;
#if MXM_API < MXM_VERSION(2, 0)
mxm_req.base.data.buffer.memh = NULL;
mxm_req.op.mem.remote_memh = NULL;
#else
mxm_req.op.mem.remote_mkey = to_mxm_mkey(r_mkey);
#endif
if (mca_spml_ikrit.mxm_peers[dst]->need_fence == 0) {
opal_list_append(&mca_spml_ikrit.active_peers,
&mca_spml_ikrit.mxm_peers[dst]->super);
mca_spml_ikrit.mxm_peers[dst]->need_fence = 1;
}
SPML_IKRIT_MXM_POST_SEND(mxm_req);
wait.req = &mxm_req.base;
wait.state = (mxm_req_state_t)(MXM_REQ_SENT | MXM_REQ_COMPLETED);
wait.progress_cb = NULL;
wait.progress_arg = NULL;
mxm_wait(&wait);
return OSHMEM_SUCCESS;
}
