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 inline int mca_spml_ikrit_get_shm(void *src_addr,
size_t size,
void *dst_addr,
int src)
{
int ptl_id;
void *rva;
sshmem_mkey_t *r_mkey;
ptl_id = get_ptl_id(src);
/**
* Get the address to the remote rkey.
**/
if (ptl_id != MXM_PTL_SHM)
return OSHMEM_ERROR;
r_mkey = mca_memheap_base_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;
}
if (!mca_memheap_base_can_local_copy(r_mkey, src_addr))
return OSHMEM_ERROR;
SPML_VERBOSE(100,
"shm 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));
memcpy(dst_addr, (void *) (unsigned long) rva, size);
opal_progress();
return OSHMEM_SUCCESS;
}
spml_ucx_mkey_t * mca_spml_ucx_get_mkey_slow(int pe, void *va, void **rva)
{
sshmem_mkey_t *r_mkey;
r_mkey = mca_memheap_base_get_cached_mkey(pe, va, 0, rva);
if (OPAL_UNLIKELY(!r_mkey)) {
SPML_ERROR("pe=%d: %p is not address of symmetric variable",
pe, va);
oshmem_shmem_abort(-1);
return NULL;
}
return (spml_ucx_mkey_t *)(r_mkey->spml_context);
}
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_base_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;
}
int shmem_addr_accessible(const void *addr, int pe)
{
void* rva;
sshmem_mkey_t *mkey;
int i;
RUNTIME_CHECK_INIT();
for (i = 0; i < mca_memheap_base_num_transports(); i++) {
mkey = mca_memheap_base_get_cached_mkey(pe, (void *)addr, i, &rva);
if (mkey) {
return 1;
}
}
return 0;
}
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;
}
/**
* 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;
}
/* 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;
}
void mca_memheap_modex_recv_all(void)
{
int i;
int j;
int nprocs, my_pe;
oshmem_proc_t *proc;
mca_spml_mkey_t *mkey;
void* dummy_rva;
if (!mca_memheap_base_key_exchange)
return;
/* init rkey cache */
nprocs = oshmem_num_procs();
my_pe = oshmem_my_proc_id();
/* Note:
* Doing exchange via rml till we figure out problem with grpcomm.modex and barrier
*/
for (i = 0; i < nprocs; i++) {
if (i == my_pe)
continue;
proc = oshmem_proc_group_find(oshmem_group_all, i);
for (j = 0; j < memheap_map->n_segments; j++) {
mkey =
mca_memheap_base_get_cached_mkey(i,
memheap_map->mem_segs[j].start,
proc->transport_ids[0],
&dummy_rva);
if (!mkey) {
MEMHEAP_ERROR("Failed to receive mkeys");
oshmem_shmem_abort(-1);
}
}
}
/*
* There is an issue with orte_grpcomm.barrier usage as
* ess/pmi directs to use grpcomm/pmi in case slurm srun() call grpcomm/pmi calls PMI_Barrier()
* that is a function of external library.
* There is no opal_progress() in such way. As a result slow PEs send a request (MEMHEAP_RKEY_REQ) to
* fast PEs waiting on barrier and do not get a respond (MEMHEAP_RKEY_RESP).
*
* there are following ways to solve one:
* 1. calculate requests from remote PEs and do ORTE_PROGRESSED_WAIT waiting for expected value;
* 2. use shmem_barrier_all();
* 3. rework pmi/barrier to use opal_progress();
* 4. use orte_grpcomm.barrier carefully;
*
* It seems there is no need to use orte_grpcomm.barrier here
*/
if (memheap_map->mem_segs[HEAP_SEG_INDEX].shmid != MEMHEAP_SHM_INVALID) {
/* unfortunately we must do barrier here to assure that everyone are attached to our segment
* good thing that this code path only invoked on older linuxes (-mca shmalloc_use_hugepages 3|4)
* try to minimize damage here by waiting 5 seconds and doing progress
*/
shmem_barrier_all();
/* keys exchanged, segments attached, now we can safely cleanup */
if (memheap_map->mem_segs[HEAP_SEG_INDEX].type
== MAP_SEGMENT_ALLOC_SHM) {
shmctl(memheap_map->mem_segs[HEAP_SEG_INDEX].shmid,
IPC_RMID,
NULL );
}
}
}
