ssize_t psmx2_write_generic(struct fid_ep *ep, const void *buf, size_t len, void *desc, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags, uint64_t data) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int nargs; int am_flags = PSM2_AM_FLAG_ASYNC; int chunk_size; psm2_epaddr_t psm2_epaddr; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag; size_t idx; void *psm2_context; int no_event; int err; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) return psmx2_trigger_queue_write(ep, buf, len, desc, dest_addr, addr, key, context, flags, data); if (!buf) return -FI_EINVAL; av = ep_priv->av; if (av && PSMX2_SEP_ADDR_TEST(dest_addr)) { psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, dest_addr); } else if (av && av->type == FI_AV_TABLE) { idx = dest_addr; if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx))) return err; psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx]; } else { if (!dest_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->tx->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_WRITE, ep_priv, (void *)buf, len, desc, addr, key, context, flags, data); no_event = (flags & PSMX2_NO_COMPLETION) || (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)); req = psmx2_am_request_alloc(ep_priv->tx); if (!req) return -FI_ENOMEM; if (flags & FI_INJECT) { if (len > psmx2_env.inject_size) { psmx2_am_request_free(ep_priv->tx, req); return -FI_EMSGSIZE; } req->tmpbuf = malloc(len); if (!req->tmpbuf) { psmx2_am_request_free(ep_priv->tx, req); return -FI_ENOMEM; } memcpy(req->tmpbuf, (void *)buf, len); buf = req->tmpbuf; } else { PSMX2_CTXT_TYPE(&req->fi_context) = no_event ? PSMX2_NOCOMP_WRITE_CONTEXT : PSMX2_WRITE_CONTEXT; } req->no_event = no_event; req->op = PSMX2_AM_REQ_WRITE; req->write.buf = (void *)buf; req->write.len = len; req->write.addr = addr; /* needed? */ req->write.key = key; /* needed? */ req->write.context = context; req->ep = ep_priv; req->cq_flags = FI_WRITE | FI_RMA; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; chunk_size = ep_priv->tx->psm2_am_param.max_request_short; args[0].u32w0 = 0; if (psmx2_env.tagged_rma && len > chunk_size) { PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_WRITE); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG); args[0].u32w1 = len; args[1].u64 = (uint64_t)req; args[2].u64 = addr; args[3].u64 = key; nargs = 4; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA); args[4].u64 = data; nargs++; } if (flags & FI_DELIVERY_COMPLETE) { args[0].u32w0 |= PSMX2_AM_FORCE_ACK; psm2_context = NULL; } else { psm2_context = (void *)&req->fi_context; } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, NULL, 0, am_flags, NULL, NULL); psm2_mq_isend2(ep_priv->tx->psm2_mq, psm2_epaddr, 0, &psm2_tag, buf, len, psm2_context, &psm2_req); return 0; } PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE); nargs = 4; while (len > chunk_size) { args[0].u32w1 = chunk_size; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, chunk_size, am_flags, NULL, NULL); buf = (const uint8_t *)buf + chunk_size; addr += chunk_size; len -= chunk_size; } args[0].u32w1 = len; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM); args[4].u64 = data; nargs++; } else { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, len, am_flags, NULL, NULL); return 0; }
ssize_t psmx2_read_generic(struct fid_ep *ep, void *buf, size_t len, void *desc, fi_addr_t src_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int chunk_size; size_t offset = 0; psm2_epaddr_t psm2_epaddr; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag, psm2_tagsel; size_t idx; int err; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) return psmx2_trigger_queue_read(ep, buf, len, desc, src_addr, addr, key, context, flags); if (!buf) return -FI_EINVAL; av = ep_priv->av; if (av && PSMX2_SEP_ADDR_TEST(src_addr)) { psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, src_addr); } else if (av && av->type == FI_AV_TABLE) { idx = src_addr; if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx))) return err; psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx]; } else { if (!src_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->tx->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_READ, ep_priv, buf, len, desc, addr, key, context, flags, 0); req = psmx2_am_request_alloc(ep_priv->tx); if (!req) return -FI_ENOMEM; req->op = PSMX2_AM_REQ_READ; req->read.buf = buf; req->read.len = len; req->read.addr = addr; /* needed? */ req->read.key = key; /* needed? */ req->read.context = context; req->ep = ep_priv; req->cq_flags = FI_READ | FI_RMA; PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) { PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT; req->no_event = 1; } chunk_size = ep_priv->tx->psm2_am_param.max_reply_short; args[0].u32w0 = 0; if (psmx2_env.tagged_rma && len > chunk_size) { PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_READ); PSMX2_SET_MASK(psm2_tagsel, PSMX2_MATCH_ALL, PSMX2_RMA_TYPE_MASK); psm2_mq_irecv2(ep_priv->tx->psm2_mq, psm2_epaddr, &psm2_tag, &psm2_tagsel, 0, buf, len, (void *)&req->fi_context, &psm2_req); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG); args[0].u32w1 = len; args[1].u64 = (uint64_t)req; args[2].u64 = addr; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 4, NULL, 0, 0, NULL, NULL); return 0; } PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ); args[1].u64 = (uint64_t)(uintptr_t)req; args[3].u64 = key; while (len > chunk_size) { args[0].u32w1 = chunk_size; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); addr += chunk_size; len -= chunk_size; offset += chunk_size; } PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); args[0].u32w1 = len; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); return 0; }
ssize_t psmx2_readv_generic(struct fid_ep *ep, const struct iovec *iov, void *desc, size_t count, fi_addr_t src_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int chunk_size; size_t offset = 0; psm2_epaddr_t psm2_epaddr; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag, psm2_tagsel; size_t idx; size_t total_len, long_len = 0, short_len; void *long_buf = NULL; int i; int err; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) return psmx2_trigger_queue_readv(ep, iov, desc, count, src_addr, addr, key, context, flags); av = ep_priv->av; if (av && PSMX2_SEP_ADDR_TEST(src_addr)) { psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, src_addr); } else if (av && av->type == FI_AV_TABLE) { idx = src_addr; if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx))) return err; psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx]; } else { if (!src_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->tx->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_READV, ep_priv, (void *)iov, count, desc, addr, key, context, flags, 0); total_len = 0; for (i=0; i<count; i++) total_len += iov[i].iov_len; req = psmx2_am_request_alloc(ep_priv->tx); if (!req) return -FI_ENOMEM; req->tmpbuf = malloc(count * sizeof(struct iovec)); if (!req->tmpbuf) { psmx2_am_request_free(ep_priv->tx, req); return -FI_ENOMEM; } req->iov = req->tmpbuf; memcpy(req->iov, iov, count * sizeof(struct iovec)); req->op = PSMX2_AM_REQ_READV; req->read.iov_count = count; req->read.len = total_len; req->read.addr = addr; /* needed? */ req->read.key = key; /* needed? */ req->read.context = context; req->ep = ep_priv; req->cq_flags = FI_READ | FI_RMA; PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) { PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT; req->no_event = 1; } chunk_size = ep_priv->tx->psm2_am_param.max_reply_short; if (psmx2_env.tagged_rma) { for (i=count-1; i>=0; i--) { if (iov[i].iov_len > chunk_size) { long_buf = iov[i].iov_base; long_len = iov[i].iov_len; break; } else if (iov[i].iov_len) { break; } } } short_len = total_len - long_len; /* Use short protocol for all but the last segment (long_len) */ args[0].u32w0 = 0; PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ); args[1].u64 = (uint64_t)(uintptr_t)req; args[3].u64 = key; while (short_len > chunk_size) { args[0].u32w1 = chunk_size; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); addr += chunk_size; short_len -= chunk_size; offset += chunk_size; } if (!long_len) PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); args[0].u32w1 = short_len; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); /* Use the long protocol for the last segment */ if (long_len) { PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_READ); PSMX2_SET_MASK(psm2_tagsel, PSMX2_MATCH_ALL, PSMX2_RMA_TYPE_MASK); psm2_mq_irecv2(ep_priv->tx->psm2_mq, psm2_epaddr, &psm2_tag, &psm2_tagsel, 0, long_buf, long_len, (void *)&req->fi_context, &psm2_req); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG); args[0].u32w1 = long_len; args[1].u64 = (uint64_t)req; args[2].u64 = addr + short_len; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 4, NULL, 0, 0, NULL, NULL); } return 0; }
ssize_t psmx2_writev_generic(struct fid_ep *ep, const struct iovec *iov, void **desc, size_t count, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags, uint64_t data) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int nargs; int am_flags = PSM2_AM_FLAG_ASYNC; int chunk_size; psm2_epaddr_t psm2_epaddr; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag; size_t idx; void *psm2_context; int no_event; size_t total_len, len, len_sent; uint8_t *buf, *p; int i; int err; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) return psmx2_trigger_queue_writev(ep, iov, desc, count, dest_addr, addr, key, context, flags, data); av = ep_priv->av; if (av && PSMX2_SEP_ADDR_TEST(dest_addr)) { psm2_epaddr = psmx2_av_translate_sep(av, ep_priv->tx, dest_addr); } else if (av && av->type == FI_AV_TABLE) { idx = dest_addr; if ((err = psmx2_av_check_table_idx(av, ep_priv->tx, idx))) return err; psm2_epaddr = av->tables[ep_priv->tx->id].epaddrs[idx]; } else { if (!dest_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->tx->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_WRITEV, ep_priv, (void *)iov, count, desc, addr, key, context, flags, data); no_event = (flags & PSMX2_NO_COMPLETION) || (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)); total_len = 0; for (i=0; i<count; i++) total_len += iov[i].iov_len; chunk_size = ep_priv->tx->psm2_am_param.max_request_short; req = psmx2_am_request_alloc(ep_priv->tx); if (!req) return -FI_ENOMEM; /* Case 1: fit into a AM message, then pack and send */ if (total_len <= chunk_size) { req->tmpbuf = malloc(total_len); if (!req->tmpbuf) { psmx2_am_request_free(ep_priv->tx, req); return -FI_ENOMEM; } p = req->tmpbuf; for (i=0; i<count; i++) { if (iov[i].iov_len) { memcpy(p, iov[i].iov_base, iov[i].iov_len); p += iov[i].iov_len; } } buf = req->tmpbuf; len = total_len; req->no_event = no_event; req->op = PSMX2_AM_REQ_WRITE; req->write.buf = (void *)buf; req->write.len = len; req->write.addr = addr; /* needed? */ req->write.key = key; /* needed? */ req->write.context = context; req->ep = ep_priv; req->cq_flags = FI_WRITE | FI_RMA; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; args[0].u32w0 = 0; PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE); args[0].u32w1 = len; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; nargs = 4; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM); args[4].u64 = data; nargs++; } else { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, len, am_flags, NULL, NULL); return 0; } if (flags & FI_INJECT) { psmx2_am_request_free(ep_priv->tx, req); return -FI_EMSGSIZE; } PSMX2_CTXT_TYPE(&req->fi_context) = no_event ? PSMX2_NOCOMP_WRITE_CONTEXT : PSMX2_WRITE_CONTEXT; req->no_event = no_event; req->op = PSMX2_AM_REQ_WRITE; req->write.buf = (void *)iov[0].iov_base; req->write.len = total_len; req->write.addr = addr; /* needed? */ req->write.key = key; /* needed? */ req->write.context = context; req->ep = ep_priv; req->cq_flags = FI_WRITE | FI_RMA; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; /* Case 2: send iov in sequence */ args[0].u32w0 = 0; len_sent = 0; for (i=0; i<count; i++) { if (!iov[i].iov_len) continue; /* Case 2.1: use long protocol for the last segment if it is large */ if (psmx2_env.tagged_rma && iov[i].iov_len > chunk_size && len_sent + iov[i].iov_len == total_len) { PSMX2_SET_TAG(psm2_tag, (uint64_t)req, 0, PSMX2_RMA_TYPE_WRITE); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG); args[0].u32w1 = iov[i].iov_len; args[1].u64 = (uint64_t)req; args[2].u64 = addr; args[3].u64 = key; nargs = 4; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA); args[4].u64 = data; nargs++; } if (flags & FI_DELIVERY_COMPLETE) { args[0].u32w0 |= PSMX2_AM_FORCE_ACK; psm2_context = NULL; } else { psm2_context = (void *)&req->fi_context; } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, NULL, 0, am_flags, NULL, NULL); psm2_mq_isend2(ep_priv->tx->psm2_mq, psm2_epaddr, 0, &psm2_tag, iov[i].iov_base, iov[i].iov_len, psm2_context, &psm2_req); return 0; } /* Case 2.2: use short protocol all other segments */ PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE); nargs = 4; buf = iov[i].iov_base; len = iov[i].iov_len; while (len > chunk_size) { args[0].u32w1 = chunk_size; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, chunk_size, am_flags, NULL, NULL); buf += chunk_size; addr += chunk_size; len -= chunk_size; len_sent += chunk_size; } args[0].u32w1 = len; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; if (len_sent + len == total_len) { if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM); args[4].u64 = data; nargs++; } else { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); } } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, len, am_flags, NULL, NULL); addr += len; len_sent += len; } return 0; }
ssize_t psmx2_write_generic(struct fid_ep *ep, const void *buf, size_t len, void *desc, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags, uint64_t data) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int nargs; int am_flags = PSM2_AM_FLAG_ASYNC; int chunk_size; psm2_epaddr_t psm2_epaddr; uint8_t vlane; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag; uint32_t tag32; size_t idx; void *psm2_context; int no_event; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) { struct psmx2_trigger *trigger; struct fi_triggered_context *ctxt = context; trigger = calloc(1, sizeof(*trigger)); if (!trigger) return -FI_ENOMEM; trigger->op = PSMX2_TRIGGERED_WRITE; trigger->cntr = container_of(ctxt->trigger.threshold.cntr, struct psmx2_fid_cntr, cntr); trigger->threshold = ctxt->trigger.threshold.threshold; trigger->write.ep = ep; trigger->write.buf = buf; trigger->write.len = len; trigger->write.desc = desc; trigger->write.dest_addr = dest_addr; trigger->write.addr = addr; trigger->write.key = key; trigger->write.context = context; trigger->write.flags = flags & ~FI_TRIGGER; trigger->write.data = data; psmx2_cntr_add_trigger(trigger->cntr, trigger); return 0; } if (!buf) return -FI_EINVAL; av = ep_priv->av; if (av && av->type == FI_AV_TABLE) { idx = dest_addr; if (idx >= av->last) return -FI_EINVAL; psm2_epaddr = av->epaddrs[idx]; vlane = av->vlanes[idx]; } else { if (!dest_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr); vlane = PSMX2_ADDR_TO_VL(dest_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->domain->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_WRITE, ep_priv, ep_priv->domain->eps[vlane], (void *)buf, len, desc, addr, key, context, flags, data); no_event = (flags & PSMX2_NO_COMPLETION) || (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)); if (flags & FI_INJECT) { if (len > PSMX2_INJECT_SIZE) return -FI_EMSGSIZE; req = malloc(sizeof(*req) + len); if (!req) return -FI_ENOMEM; memset(req, 0, sizeof(*req)); memcpy((uint8_t *)req + sizeof(*req), (void *)buf, len); buf = (uint8_t *)req + sizeof(*req); } else { req = calloc(1, sizeof(*req)); if (!req) return -FI_ENOMEM; PSMX2_CTXT_TYPE(&req->fi_context) = no_event ? PSMX2_NOCOMP_WRITE_CONTEXT : PSMX2_WRITE_CONTEXT; } req->no_event = no_event; req->op = PSMX2_AM_REQ_WRITE; req->write.buf = (void *)buf; req->write.len = len; req->write.addr = addr; /* needed? */ req->write.key = key; /* needed? */ req->write.context = context; req->ep = ep_priv; req->cq_flags = FI_WRITE | FI_RMA; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; chunk_size = psmx2_am_param.max_request_short; args[0].u32w0 = 0; PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane); PSMX2_AM_SET_DST(args[0].u32w0, vlane); if (psmx2_env.tagged_rma && len > chunk_size) { tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, ep_priv->vlane, vlane); PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG); args[0].u32w1 = len; args[1].u64 = (uint64_t)req; args[2].u64 = addr; args[3].u64 = key; nargs = 4; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA); args[4].u64 = data; nargs++; } if (flags & FI_DELIVERY_COMPLETE) { args[0].u32w0 |= PSMX2_AM_FORCE_ACK; psm2_context = NULL; } else { psm2_context = (void *)&req->fi_context; } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, NULL, 0, am_flags, NULL, NULL); psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr, 0, &psm2_tag, buf, len, psm2_context, &psm2_req); return 0; } PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE); nargs = 4; while (len > chunk_size) { args[0].u32w1 = chunk_size; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, chunk_size, am_flags, NULL, NULL); buf = (const uint8_t *)buf + chunk_size; addr += chunk_size; len -= chunk_size; } args[0].u32w1 = len; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM); args[4].u64 = data; nargs++; } else { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, len, am_flags, NULL, NULL); return 0; }
ssize_t psmx2_readv_generic(struct fid_ep *ep, const struct iovec *iov, void *desc, size_t count, fi_addr_t src_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int chunk_size; size_t offset = 0; psm2_epaddr_t psm2_epaddr; uint8_t vlane; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag, psm2_tagsel; uint32_t tag32; size_t idx; size_t total_len, long_len, short_len; void *long_buf; int i; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) { struct psmx2_trigger *trigger; struct fi_triggered_context *ctxt = context; trigger = calloc(1, sizeof(*trigger)); if (!trigger) return -FI_ENOMEM; trigger->op = PSMX2_TRIGGERED_READV; trigger->cntr = container_of(ctxt->trigger.threshold.cntr, struct psmx2_fid_cntr, cntr); trigger->threshold = ctxt->trigger.threshold.threshold; trigger->readv.ep = ep; trigger->readv.iov = iov; trigger->readv.count = count; trigger->readv.desc = desc; trigger->readv.src_addr = src_addr; trigger->readv.addr = addr; trigger->readv.key = key; trigger->readv.context = context; trigger->readv.flags = flags & ~FI_TRIGGER; psmx2_cntr_add_trigger(trigger->cntr, trigger); return 0; } av = ep_priv->av; if (av && av->type == FI_AV_TABLE) { idx = src_addr; if (idx >= av->last) return -FI_EINVAL; psm2_epaddr = av->epaddrs[idx]; vlane = av->vlanes[idx]; } else { if (!src_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr); vlane = PSMX2_ADDR_TO_VL(src_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->domain->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_READV, ep_priv, ep_priv->domain->eps[vlane], (void *)iov, count, desc, addr, key, context, flags, 0); total_len = 0; for (i=0; i<count; i++) total_len += iov[i].iov_len; req = calloc(1, sizeof(*req) + count * sizeof(struct iovec)); if (!req) return -FI_ENOMEM; memcpy(req->iov, iov, count * sizeof(struct iovec)); req->op = PSMX2_AM_REQ_READV; req->read.iov_count = count; req->read.len = total_len; req->read.addr = addr; /* needed? */ req->read.key = key; /* needed? */ req->read.context = context; req->ep = ep_priv; req->cq_flags = FI_READ | FI_RMA; PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) { PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT; req->no_event = 1; } chunk_size = psmx2_am_param.max_reply_short; long_len = 0; if (psmx2_env.tagged_rma) { for (i=count-1; i>=0; i--) { if (iov[i].iov_len > chunk_size) { long_buf = iov[i].iov_base; long_len = iov[i].iov_len; break; } else if (iov[i].iov_len) { break; } } } short_len = total_len - long_len; /* Use short protocol for all but the last segment (long_len) */ args[0].u32w0 = 0; PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane); PSMX2_AM_SET_DST(args[0].u32w0, vlane); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ); args[1].u64 = (uint64_t)(uintptr_t)req; args[3].u64 = key; while (short_len > chunk_size) { args[0].u32w1 = chunk_size; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); addr += chunk_size; short_len -= chunk_size; offset += chunk_size; } if (!long_len) PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); args[0].u32w1 = short_len; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); /* Use the long protocol for the last segment */ if (long_len) { tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, vlane, ep_priv->vlane); PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32); PSMX2_SET_TAG(psm2_tagsel, -1ULL, -1); psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr, &psm2_tag, &psm2_tagsel, 0, long_buf, long_len, (void *)&req->fi_context, &psm2_req); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG); args[0].u32w1 = long_len; args[1].u64 = (uint64_t)req; args[2].u64 = addr + short_len; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 4, NULL, 0, 0, NULL, NULL); } return 0; }
ssize_t psmx2_read_generic(struct fid_ep *ep, void *buf, size_t len, void *desc, fi_addr_t src_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int chunk_size; size_t offset = 0; psm2_epaddr_t psm2_epaddr; uint8_t vlane; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag, psm2_tagsel; uint32_t tag32; size_t idx; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) { struct psmx2_trigger *trigger; struct fi_triggered_context *ctxt = context; trigger = calloc(1, sizeof(*trigger)); if (!trigger) return -FI_ENOMEM; trigger->op = PSMX2_TRIGGERED_READ; trigger->cntr = container_of(ctxt->trigger.threshold.cntr, struct psmx2_fid_cntr, cntr); trigger->threshold = ctxt->trigger.threshold.threshold; trigger->read.ep = ep; trigger->read.buf = buf; trigger->read.len = len; trigger->read.desc = desc; trigger->read.src_addr = src_addr; trigger->read.addr = addr; trigger->read.key = key; trigger->read.context = context; trigger->read.flags = flags & ~FI_TRIGGER; psmx2_cntr_add_trigger(trigger->cntr, trigger); return 0; } if (!buf) return -FI_EINVAL; av = ep_priv->av; if (av && av->type == FI_AV_TABLE) { idx = src_addr; if (idx >= av->last) return -FI_EINVAL; psm2_epaddr = av->epaddrs[idx]; vlane = av->vlanes[idx]; } else { if (!src_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(src_addr); vlane = PSMX2_ADDR_TO_VL(src_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->domain->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_READ, ep_priv, ep_priv->domain->eps[vlane], buf, len, desc, addr, key, context, flags, 0); req = calloc(1, sizeof(*req)); if (!req) return -FI_ENOMEM; req->op = PSMX2_AM_REQ_READ; req->read.buf = buf; req->read.len = len; req->read.addr = addr; /* needed? */ req->read.key = key; /* needed? */ req->read.context = context; req->ep = ep_priv; req->cq_flags = FI_READ | FI_RMA; PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_READ_CONTEXT; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; if (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)) { PSMX2_CTXT_TYPE(&req->fi_context) = PSMX2_NOCOMP_READ_CONTEXT; req->no_event = 1; } chunk_size = psmx2_am_param.max_reply_short; args[0].u32w0 = 0; PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane); PSMX2_AM_SET_DST(args[0].u32w0, vlane); if (psmx2_env.tagged_rma && len > chunk_size) { tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, vlane, ep_priv->vlane); PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32); PSMX2_SET_TAG(psm2_tagsel, -1ULL, -1); psm2_mq_irecv2(ep_priv->domain->psm2_mq, psm2_epaddr, &psm2_tag, &psm2_tagsel, 0, buf, len, (void *)&req->fi_context, &psm2_req); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ_LONG); args[0].u32w1 = len; args[1].u64 = (uint64_t)req; args[2].u64 = addr; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 4, NULL, 0, 0, NULL, NULL); return 0; } PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_READ); args[1].u64 = (uint64_t)(uintptr_t)req; args[3].u64 = key; while (len > chunk_size) { args[0].u32w1 = chunk_size; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); addr += chunk_size; len -= chunk_size; offset += chunk_size; } PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); args[0].u32w1 = len; args[2].u64 = addr; args[4].u64 = offset; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, 5, NULL, 0, 0, NULL, NULL); return 0; }
ssize_t psmx2_writev_generic(struct fid_ep *ep, const struct iovec *iov, void **desc, size_t count, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context, uint64_t flags, uint64_t data) { struct psmx2_fid_ep *ep_priv; struct psmx2_fid_av *av; struct psmx2_epaddr_context *epaddr_context; struct psmx2_am_request *req; psm2_amarg_t args[8]; int nargs; int am_flags = PSM2_AM_FLAG_ASYNC; int chunk_size; psm2_epaddr_t psm2_epaddr; uint8_t vlane; psm2_mq_req_t psm2_req; psm2_mq_tag_t psm2_tag; uint32_t tag32; size_t idx; void *psm2_context; int no_event; size_t total_len, len, len_sent; uint8_t *buf, *p; int i; ep_priv = container_of(ep, struct psmx2_fid_ep, ep); if (flags & FI_TRIGGER) { struct psmx2_trigger *trigger; struct fi_triggered_context *ctxt = context; trigger = calloc(1, sizeof(*trigger)); if (!trigger) return -FI_ENOMEM; trigger->op = PSMX2_TRIGGERED_WRITEV; trigger->cntr = container_of(ctxt->trigger.threshold.cntr, struct psmx2_fid_cntr, cntr); trigger->threshold = ctxt->trigger.threshold.threshold; trigger->writev.ep = ep; trigger->writev.iov = iov; trigger->writev.count = count; trigger->writev.desc = desc; trigger->writev.dest_addr = dest_addr; trigger->writev.addr = addr; trigger->writev.key = key; trigger->writev.context = context; trigger->writev.flags = flags & ~FI_TRIGGER; trigger->writev.data = data; psmx2_cntr_add_trigger(trigger->cntr, trigger); return 0; } av = ep_priv->av; if (av && av->type == FI_AV_TABLE) { idx = dest_addr; if (idx >= av->last) return -FI_EINVAL; psm2_epaddr = av->epaddrs[idx]; vlane = av->vlanes[idx]; } else { if (!dest_addr) return -FI_EINVAL; psm2_epaddr = PSMX2_ADDR_TO_EP(dest_addr); vlane = PSMX2_ADDR_TO_VL(dest_addr); } epaddr_context = psm2_epaddr_getctxt((void *)psm2_epaddr); if (epaddr_context->epid == ep_priv->domain->psm2_epid) return psmx2_rma_self(PSMX2_AM_REQ_WRITEV, ep_priv, ep_priv->domain->eps[vlane], (void *)iov, count, desc, addr, key, context, flags, data); no_event = (flags & PSMX2_NO_COMPLETION) || (ep_priv->send_selective_completion && !(flags & FI_COMPLETION)); total_len = 0; for (i=0; i<count; i++) total_len += iov[i].iov_len; chunk_size = psmx2_am_param.max_request_short; /* Case 1: fit into a AM message, then pack and send */ if (total_len <= chunk_size) { req = malloc(sizeof(*req) + total_len); if (!req) return -FI_ENOMEM; memset(req, 0, sizeof(*req)); p = (uint8_t *)req + sizeof(*req); for (i=0; i<count; i++) { if (iov[i].iov_len) { memcpy(p, iov[i].iov_base, iov[i].iov_len); p += iov[i].iov_len; } } buf = (uint8_t *)req + sizeof(*req); len = total_len; req->no_event = no_event; req->op = PSMX2_AM_REQ_WRITE; req->write.buf = (void *)buf; req->write.len = len; req->write.addr = addr; /* needed? */ req->write.key = key; /* needed? */ req->write.context = context; req->ep = ep_priv; req->cq_flags = FI_WRITE | FI_RMA; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; args[0].u32w0 = 0; PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane); PSMX2_AM_SET_DST(args[0].u32w0, vlane); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE); args[0].u32w1 = len; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; nargs = 4; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM); args[4].u64 = data; nargs++; } else { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, len, am_flags, NULL, NULL); return 0; } if (flags & FI_INJECT) return -FI_EMSGSIZE; req = calloc(1, sizeof(*req)); if (!req) return -FI_ENOMEM; PSMX2_CTXT_TYPE(&req->fi_context) = no_event ? PSMX2_NOCOMP_WRITE_CONTEXT : PSMX2_WRITE_CONTEXT; req->no_event = no_event; req->op = PSMX2_AM_REQ_WRITE; req->write.buf = (void *)iov[0].iov_base; req->write.len = total_len; req->write.addr = addr; /* needed? */ req->write.key = key; /* needed? */ req->write.context = context; req->ep = ep_priv; req->cq_flags = FI_WRITE | FI_RMA; PSMX2_CTXT_USER(&req->fi_context) = context; PSMX2_CTXT_EP(&req->fi_context) = ep_priv; /* Case 2: send iov in sequence */ args[0].u32w0 = 0; PSMX2_AM_SET_SRC(args[0].u32w0, ep_priv->vlane); PSMX2_AM_SET_DST(args[0].u32w0, vlane); len_sent = 0; for (i=0; i<count; i++) { if (!iov[i].iov_len) continue; /* Case 2.1: use long protocol for the last segment if it is large */ if (psmx2_env.tagged_rma && iov[i].iov_len > chunk_size && len_sent + iov[i].iov_len == total_len) { tag32 = PSMX2_TAG32(PSMX2_RMA_BIT, ep_priv->vlane, vlane); PSMX2_SET_TAG(psm2_tag, (uint64_t)req, tag32); PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE_LONG); args[0].u32w1 = iov[i].iov_len; args[1].u64 = (uint64_t)req; args[2].u64 = addr; args[3].u64 = key; nargs = 4; if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA); args[4].u64 = data; nargs++; } if (flags & FI_DELIVERY_COMPLETE) { args[0].u32w0 |= PSMX2_AM_FORCE_ACK; psm2_context = NULL; } else { psm2_context = (void *)&req->fi_context; } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, NULL, 0, am_flags, NULL, NULL); psm2_mq_isend2(ep_priv->domain->psm2_mq, psm2_epaddr, 0, &psm2_tag, iov[i].iov_base, iov[i].iov_len, psm2_context, &psm2_req); return 0; } /* Case 2.2: use short protocol all other segments */ PSMX2_AM_SET_OP(args[0].u32w0, PSMX2_AM_REQ_WRITE); nargs = 4; buf = iov[i].iov_base; len = iov[i].iov_len; while (len > chunk_size) { args[0].u32w1 = chunk_size; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, chunk_size, am_flags, NULL, NULL); buf += chunk_size; addr += chunk_size; len -= chunk_size; len_sent += chunk_size; } args[0].u32w1 = len; args[1].u64 = (uint64_t)(uintptr_t)req; args[2].u64 = addr; args[3].u64 = key; if (len_sent + len == total_len) { if (flags & FI_REMOTE_CQ_DATA) { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_DATA | PSMX2_AM_EOM); args[4].u64 = data; nargs++; } else { PSMX2_AM_SET_FLAG(args[0].u32w0, PSMX2_AM_EOM); } } psm2_am_request_short(psm2_epaddr, PSMX2_AM_RMA_HANDLER, args, nargs, (void *)buf, len, am_flags, NULL, NULL); addr += len; len_sent += len; } return 0; }