static
mxm_mem_key_t *mca_spml_ikrit_get_mkey_slow(int pe, void *va, int ptl_id, void **rva)
{
sshmem_mkey_t *mkey;
retry:
mkey = mca_memheap_base_get_cached_mkey(pe, va, ptl_id, rva);
if (NULL == mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable", pe, va);
oshmem_shmem_abort(-1);
return NULL;
}
if (MXM_PTL_SHM == ptl_id) {
if (mca_memheap_base_can_local_copy(mkey, va)) {
return NULL;
}
/* if dst addr is on memheap and local copy is not allowed
* disable direct shm transport
*/
if (memheap_is_va_in_segment(va, HEAP_SEG_INDEX)) {
mca_spml_ikrit.mxm_peers[pe].ptl_id = MXM_PTL_RDMA;
}
/* going via mxm must always work */
ptl_id = MXM_PTL_RDMA;
goto retry;
}
return to_mxm_mkey(mkey);
}
static int mca_spml_ikrit_get_helper(mxm_send_req_t *sreq,
void *src_addr,
size_t size,
void *dst_addr,
int src)
{
/* shmem spec states that get() operations are blocking. So it is enough
to have single mxm request. Also we count on mxm doing copy */
void *rva;
sshmem_mkey_t *r_mkey;
int ptl_id;
ptl_id = get_ptl_id(src);
/* already tried to send via shm and failed. go via rdma */
if (ptl_id == MXM_PTL_SHM)
ptl_id = MXM_PTL_RDMA;
/**
* Get the address to the remote rkey.
**/
r_mkey = mca_memheap.memheap_get_cached_mkey(src,
src_addr,
ptl_id,
&rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
src, src_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
SPML_VERBOSE(100,
"get: pe:%d ptl=%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, ptl_id, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
/* mxm does not really cares for get lkey */
sreq->base.mq = mca_spml_ikrit.mxm_mq;
sreq->base.conn = mca_spml_ikrit.mxm_peers[src]->mxm_conn;
sreq->base.data_type = MXM_REQ_DATA_BUFFER;
sreq->base.data.buffer.ptr = dst_addr;
sreq->base.data.buffer.length = size;
#if MXM_API < MXM_VERSION(2,0)
sreq->base.data.buffer.memh = NULL;
sreq->op.mem.remote_memh = NULL;
#else
sreq->op.mem.remote_mkey = to_mxm_mkey(r_mkey);
#endif
sreq->opcode = MXM_REQ_OP_GET;
sreq->op.mem.remote_vaddr = (intptr_t) rva;
sreq->base.state = MXM_REQ_NEW;
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;
}
/**
* TODO: using put request as handle is not good.
*/
static inline int mca_spml_ikrit_put_internal(void* dst_addr,
size_t size,
void* src_addr,
int dst,
void **handle,
int zcopy)
{
void *rva;
mca_spml_ikrit_put_request_t *put_req;
int ptl_id;
sshmem_mkey_t *r_mkey;
static int count;
int need_progress = 0;
if (0 >= size) {
return OSHMEM_SUCCESS;
}
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, 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
put_req = alloc_put_req();
if (NULL == put_req) {
SPML_ERROR("out of put requests - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
if (handle)
*handle = put_req;
/* fill out request */
put_req->mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
/* request immediate responce if we are getting low on send buffers. We only get responce from remote on ack timeout.
* Also request explicit ack once in a while */
#if MXM_API < MXM_VERSION(2,0)
put_req->mxm_req.opcode = MXM_REQ_OP_PUT;
if (mca_spml_ikrit.free_list_max - mca_spml_ikrit.n_active_puts <= SPML_IKRIT_PUT_LOW_WATER ||
(mca_spml_ikrit.mxm_peers[dst]->n_active_puts + 1) % SPML_IKRIT_PACKETS_PER_SYNC == 0) {
put_req->mxm_req.base.flags = MXM_REQ_FLAG_SEND_SYNC;
need_progress = 1;
} else {
put_req->mxm_req.base.flags = MXM_REQ_FLAG_SEND_LAZY|MXM_REQ_FLAG_SEND_SYNC;
}
#else
put_req->mxm_req.flags = 0;
if (mca_spml_ikrit.free_list_max - mca_spml_ikrit.n_active_puts <= SPML_IKRIT_PUT_LOW_WATER ||
(int)opal_list_get_size(&mca_spml_ikrit.active_peers) > mca_spml_ikrit.unsync_conn_max ||
(mca_spml_ikrit.mxm_peers[dst]->n_active_puts + 1) % SPML_IKRIT_PACKETS_PER_SYNC == 0) {
need_progress = 1;
put_req->mxm_req.opcode = MXM_REQ_OP_PUT_SYNC;
} else {
put_req->mxm_req.opcode = MXM_REQ_OP_PUT;
}
if (!zcopy) {
if (size < mca_spml_ikrit.put_zcopy_threshold) {
put_req->mxm_req.flags |= MXM_REQ_SEND_FLAG_BLOCKING;
} else {
put_req->mxm_req.opcode = MXM_REQ_OP_PUT_SYNC;
}
}
#endif
put_req->mxm_req.base.conn = mca_spml_ikrit.mxm_peers[dst]->mxm_conn;
put_req->mxm_req.base.data_type = MXM_REQ_DATA_BUFFER;
put_req->mxm_req.base.data.buffer.ptr = src_addr;
put_req->mxm_req.base.data.buffer.length = size;
put_req->mxm_req.base.completed_cb = put_completion_cb;
put_req->mxm_req.base.context = put_req;
put_req->mxm_req.op.mem.remote_vaddr = (intptr_t) rva;
put_req->mxm_req.base.state = MXM_REQ_NEW;
put_req->pe = dst;
#if MXM_API < MXM_VERSION(2,0)
put_req->mxm_req.base.data.buffer.memh = NULL;
put_req->mxm_req.op.mem.remote_memh = NULL;
#else
put_req->mxm_req.op.mem.remote_mkey = to_mxm_mkey(r_mkey);
#endif
OPAL_THREAD_ADD32(&mca_spml_ikrit.n_active_puts, 1);
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;
}
mca_spml_ikrit.mxm_peers[dst]->n_active_puts++;
SPML_IKRIT_MXM_POST_SEND(put_req->mxm_req);
if (need_progress)
mxm_progress(mca_spml_ikrit.mxm_context);
return OSHMEM_SUCCESS;
}
int mca_atomic_mxm_fadd(void *target,
void *prev,
const void *value,
size_t nlong,
int pe,
struct oshmem_op_t *op)
{
unsigned my_pe;
uint8_t nlong_order;
void *remote_addr;
int ptl_id;
mxm_send_req_t sreq;
mxm_error_t mxm_err;
sshmem_mkey_t *r_mkey;
static char dummy_buf[8];
my_pe = oshmem_my_proc_id();
ptl_id = -1;
mxm_err = MXM_OK;
if (!target || !value) {
ATOMIC_ERROR("[#%d] target or value are not defined", my_pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
if ((pe < 0) || (pe >= oshmem_num_procs())) {
ATOMIC_ERROR("[#%d] PE=%d not valid", my_pe, pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
switch (nlong) {
case 1:
nlong_order = 0;
break;
case 2:
nlong_order = 1;
break;
case 4:
nlong_order = 2;
break;
case 8:
nlong_order = 3;
break;
default:
ATOMIC_ERROR("[#%d] Type size must be 1/2/4 or 8 bytes.", my_pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
ptl_id = oshmem_proc_group_all(pe)->transport_ids[0];
if (MXM_PTL_SHM == ptl_id) {
ptl_id = MXM_PTL_RDMA;
}
r_mkey = mca_memheap.memheap_get_cached_mkey(pe,
target,
ptl_id,
&remote_addr);
if (!r_mkey) {
ATOMIC_ERROR("[#%d] %p is not address of symmetric variable",
my_pe, target);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
/* mxm request init */
sreq.base.state = MXM_REQ_NEW;
sreq.base.mq = mca_spml_self->mxm_mq;
sreq.base.conn = mca_spml_self->mxm_peers[pe]->mxm_hw_rdma_conn;
sreq.base.completed_cb = NULL;
sreq.base.data_type = MXM_REQ_DATA_BUFFER;
sreq.op.atomic.remote_vaddr = (uintptr_t) remote_addr;
#if MXM_API < MXM_VERSION(2,0)
sreq.op.atomic.remote_memh = MXM_INVALID_MEM_HANDLE;
memcpy(&sreq.op.atomic.value8, value, nlong);
#else
sreq.op.atomic.remote_mkey = to_mxm_mkey(r_mkey);
memcpy(&sreq.op.atomic.value, value, nlong);
#endif
sreq.op.atomic.order = nlong_order;
/* Do we need atomic 'add' or atomic 'fetch and add'? */
if (NULL == prev) {
sreq.base.data.buffer.ptr = dummy_buf;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = MXM_REQ_FLAG_SEND_SYNC;
sreq.opcode = MXM_REQ_OP_ATOMIC_ADD;
#else
sreq.flags = 0;
sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
#endif
} else {
sreq.base.data.buffer.ptr = prev;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = 0;
#else
sreq.flags = 0;
#endif
sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
}
if (MXM_OK != (mxm_err = mxm_req_send(&sreq))) {
ATOMIC_ERROR("[#%d] mxm_req_send failed, mxm_error = %d",
my_pe, mxm_err);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
mxm_req_wait(&sreq.base);
if (MXM_OK != sreq.base.error) {
ATOMIC_ERROR("[#%d] mxm_req_wait got non MXM_OK error: %d",
my_pe, sreq.base.error);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
