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; }
static int pack_local_mkeys(opal_buffer_t *msg, int pe, int seg) { int i, n; sshmem_mkey_t *mkey; /* go over all transports and pack mkeys */ n = memheap_map->num_transports; opal_dss.pack(msg, &n, 1, OPAL_UINT32); MEMHEAP_VERBOSE(5, "found %d transports to %d", n, pe); for (i = 0; i < n; i++) { mkey = mca_memheap_base_get_mkey(mca_memheap_seg2base_va(seg), i); if (!mkey) { MEMHEAP_ERROR("seg#%d tr_id: %d failed to find local mkey", seg, i); return OSHMEM_ERROR; } opal_dss.pack(msg, &i, 1, OPAL_UINT32); opal_dss.pack(msg, &mkey->va_base, 1, OPAL_UINT64); if (0 == mkey->va_base) { opal_dss.pack(msg, &mkey->u.key, 1, OPAL_UINT64); } else { opal_dss.pack(msg, &mkey->len, 1, OPAL_UINT16); if (0 < mkey->len) { opal_dss.pack(msg, mkey->u.data, mkey->len, OPAL_BYTE); } } MEMHEAP_VERBOSE(5, "seg#%d tr_id: %d %s", seg, i, mca_spml_base_mkey2str(mkey)); } return OSHMEM_SUCCESS; }
static int memheap_oob_get_mkeys(int pe, uint32_t seg, sshmem_mkey_t *mkeys) { opal_buffer_t *msg; uint8_t cmd; int i; int rc; if (OSHMEM_SUCCESS == MCA_SPML_CALL(oob_get_mkeys(pe, seg, mkeys))) { for (i = 0; i < memheap_map->num_transports; i++) { mkeys[i].va_base = mca_memheap_seg2base_va(seg); MEMHEAP_VERBOSE(5, "MKEY CALCULATED BY LOCAL SPML: pe: %d tr_id: %d %s", pe, i, mca_spml_base_mkey2str(&mkeys[i])); } return OSHMEM_SUCCESS; } OPAL_THREAD_LOCK(&memheap_oob.lck); memheap_oob.mkeys = mkeys; memheap_oob.mkeys_rcvd = 0; msg = OBJ_NEW(opal_buffer_t); if (!msg) { OPAL_THREAD_UNLOCK(&memheap_oob.lck); MEMHEAP_ERROR("failed to get msg buffer"); return OSHMEM_ERROR; } OPAL_THREAD_LOCK(&memheap_oob.lck); cmd = MEMHEAP_RKEY_REQ; opal_dss.pack(msg, &cmd, 1, OPAL_UINT8); opal_dss.pack(msg, &seg, 1, OPAL_UINT32); rc = send_buffer(pe, msg); if (MPI_SUCCESS != rc) { OPAL_THREAD_UNLOCK(&memheap_oob.lck); MEMHEAP_ERROR("FAILED to send rml message %d", rc); return OSHMEM_ERROR; } while (!memheap_oob.mkeys_rcvd) { opal_condition_wait(&memheap_oob.cond, &memheap_oob.lck); } if (MEMHEAP_RKEY_RESP == memheap_oob.mkeys_rcvd) { rc = OSHMEM_SUCCESS; } else { MEMHEAP_ERROR("failed to get rkey seg#%d pe=%d", seg, pe); rc = OSHMEM_ERROR; } OPAL_THREAD_UNLOCK(&memheap_oob.lck); return rc; }
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; }
static void unpack_remote_mkeys(opal_buffer_t *msg, int remote_pe) { int32_t cnt; int32_t n; int32_t tr_id; int i; oshmem_proc_t *proc; proc = oshmem_proc_group_find(oshmem_group_all, remote_pe); cnt = 1; opal_dss.unpack(msg, &n, &cnt, OPAL_UINT32); for (i = 0; i < n; i++) { cnt = 1; opal_dss.unpack(msg, &tr_id, &cnt, OPAL_UINT32); cnt = 1; opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].va_base, &cnt, OPAL_UINT64); if (0 == memheap_oob.mkeys[tr_id].va_base) { cnt = 1; opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].u.key, &cnt, OPAL_UINT64); if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) { memheap_attach_segment(&memheap_oob.mkeys[tr_id], tr_id); } } else { cnt = 1; opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].len, &cnt, OPAL_UINT16); if (0 < memheap_oob.mkeys[tr_id].len) { memheap_oob.mkeys[tr_id].u.data = malloc(memheap_oob.mkeys[tr_id].len); if (NULL == memheap_oob.mkeys[tr_id].u.data) { MEMHEAP_ERROR("Failed allocate %d bytes", memheap_oob.mkeys[tr_id].len); oshmem_shmem_abort(-1); } cnt = memheap_oob.mkeys[tr_id].len; opal_dss.unpack(msg, memheap_oob.mkeys[tr_id].u.data, &cnt, OPAL_BYTE); MCA_SPML_CALL(rmkey_unpack(&memheap_oob.mkeys[tr_id], remote_pe)); } else { memheap_oob.mkeys[tr_id].u.key = MAP_SEGMENT_SHM_INVALID; } } MEMHEAP_VERBOSE(5, "tr_id: %d %s", tr_id, mca_spml_base_mkey2str(&memheap_oob.mkeys[tr_id])); } }
/** * @param all_trs * 0 - pack mkeys for transports to given pe * 1 - pack mkeys for ALL possible transports. value of pe is ignored */ static int pack_local_mkeys(opal_buffer_t *msg, int pe, int seg, int all_trs) { oshmem_proc_t *proc; int i, n, tr_id; sshmem_mkey_t *mkey; /* go over all transports to remote pe and pack mkeys */ if (!all_trs) { n = oshmem_get_transport_count(pe); proc = oshmem_proc_group_find(oshmem_group_all, pe); } else { proc = NULL; n = memheap_map->num_transports; } opal_dss.pack(msg, &n, 1, OPAL_UINT32); MEMHEAP_VERBOSE(5, "found %d transports to %d", n, pe); for (i = 0; i < n; i++) { if (!all_trs) { tr_id = proc->transport_ids[i]; } else { tr_id = i; } mkey = mca_memheap_base_get_mkey(mca_memheap_seg2base_va(seg), tr_id); if (!mkey) { MEMHEAP_ERROR("seg#%d tr_id: %d failed to find local mkey", seg, tr_id); return OSHMEM_ERROR; } opal_dss.pack(msg, &tr_id, 1, OPAL_UINT32); opal_dss.pack(msg, &mkey->va_base, 1, OPAL_UINT64); if (0 == mkey->va_base) { opal_dss.pack(msg, &mkey->u.key, 1, OPAL_UINT64); } else { opal_dss.pack(msg, &mkey->len, 1, OPAL_UINT16); if (0 < mkey->len) { opal_dss.pack(msg, mkey->u.data, mkey->len, OPAL_BYTE); } } MEMHEAP_VERBOSE(5, "seg#%d tr_id: %d %s", seg, tr_id, mca_spml_base_mkey2str(mkey)); } return OSHMEM_SUCCESS; }
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; }
sshmem_mkey_t *mca_spml_ikrit_register(void* addr, size_t size, uint64_t shmid, int *count) { int i; sshmem_mkey_t *mkeys; mxm_error_t err; mxm_mem_key_t *m_key; int my_rank = oshmem_my_proc_id(); *count = 0; mkeys = (sshmem_mkey_t *) calloc(1, MXM_PTL_LAST * sizeof(*mkeys)); if (!mkeys) { return NULL ; } for (i = 0; i < MXM_PTL_LAST; i++) { mkeys[i].u.key = MAP_SEGMENT_SHM_INVALID; switch (i) { case MXM_PTL_SHM: if ((int)shmid != MAP_SEGMENT_SHM_INVALID) { mkeys[i].u.key = shmid; mkeys[i].va_base = 0; } else { mkeys[i].len = 0; mkeys[i].va_base = addr; } mkeys[i].spml_context = 0; break; case MXM_PTL_RDMA: mkeys[i].va_base = addr; mkeys[i].spml_context = 0; if (mca_spml_ikrit.ud_only) { mkeys[i].len = 0; break; } err = mxm_mem_map(mca_spml_ikrit.mxm_context, &addr, &size, 0, 0, 0); if (MXM_OK != err) { SPML_ERROR("Failed to register memory: %s", mxm_error_string(err)); goto error_out; } mkeys[i].spml_context = (void *)(unsigned long)size; m_key = malloc(sizeof(*m_key)); if (NULL == m_key) { SPML_ERROR("Failed to allocate m_key memory"); goto error_out; } mkeys[i].len = sizeof(*m_key); mkeys[i].u.data = m_key; err = mxm_mem_get_key(mca_spml_ikrit.mxm_context, addr, m_key); if (MXM_OK != err) { SPML_ERROR("Failed to get memory key: %s", mxm_error_string(err)); goto error_out; } break; default: SPML_ERROR("unsupported PTL: %d", i); goto error_out; } SPML_VERBOSE(5, "rank %d ptl %d addr %p size %llu %s", my_rank, i, addr, (unsigned long long)size, mca_spml_base_mkey2str(&mkeys[i])); mca_spml_ikrit_cache_mkeys(&mkeys[i], memheap_find_segnum(addr), my_rank, i); } *count = MXM_PTL_LAST; return mkeys; error_out: mca_spml_ikrit_deregister(mkeys); return NULL; }
/* 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; }
sshmem_mkey_t *mca_spml_ikrit_register(void* addr, size_t size, uint64_t shmid, int *count) { int i; sshmem_mkey_t *mkeys; #if MXM_API >= MXM_VERSION(2,0) mxm_error_t err; mxm_mem_key_t *m_key; #endif *count = 0; mkeys = (sshmem_mkey_t *) calloc(1, MXM_PTL_LAST * sizeof(*mkeys)); if (!mkeys) { return NULL ; } for (i = 0; i < MXM_PTL_LAST; i++) { mkeys[i].u.key = MAP_SEGMENT_SHM_INVALID; switch (i) { case MXM_PTL_SHM: if ((int)shmid != MAP_SEGMENT_SHM_INVALID) { mkeys[i].u.key = shmid; mkeys[i].va_base = 0; } else { mkeys[i].len = 0; mkeys[i].va_base = addr; } mkeys[i].spml_context = 0; break; #if MXM_API < MXM_VERSION(2,0) case MXM_PTL_SELF: mkeys[i].len = 0; mkeys[i].spml_context = 0; mkeys[i].va_base = addr; break; #endif case MXM_PTL_RDMA: mkeys[i].va_base = addr; mkeys[i].spml_context = 0; #if MXM_API < MXM_VERSION(2,0) mkeys[i].len = 0; #else if (mca_spml_ikrit.ud_only && !mca_spml_ikrit.hw_rdma_channel) { mkeys[i].len = 0; break; } err = mxm_mem_map(mca_spml_ikrit.mxm_context, &addr, &size, 0, 0, 0); if (MXM_OK != err) { SPML_ERROR("Failed to register memory: %s", mxm_error_string(err)); goto error_out; } mkeys[i].spml_context = (void *)(unsigned long)size; m_key = malloc(sizeof(*m_key)); if (NULL == m_key) { SPML_ERROR("Failed to allocate m_key memory"); goto error_out; } mkeys[i].len = sizeof(*m_key); mkeys[i].u.data = m_key; err = mxm_mem_get_key(mca_spml_ikrit.mxm_context, addr, m_key); if (MXM_OK != err) { SPML_ERROR("Failed to get memory key: %s", mxm_error_string(err)); goto error_out; } #endif break; default: SPML_ERROR("unsupported PTL: %d", i); goto error_out; } SPML_VERBOSE(5, "rank %d ptl %d addr %p size %llu %s", oshmem_proc_local_proc->super.proc_name.vpid, i, addr, (unsigned long long)size, mca_spml_base_mkey2str(&mkeys[i])); } *count = MXM_PTL_LAST; return mkeys; error_out: mca_spml_ikrit_deregister(mkeys); return NULL; }