static inline int rxm_finish_rma(struct rxm_ep *rxm_ep, struct rxm_rma_buf *rma_buf, uint64_t comp_flags) { int ret = rxm_cq_tx_comp_write(rxm_ep, comp_flags, rma_buf->app_context, rma_buf->flags); assert(((comp_flags & FI_WRITE) && !(comp_flags & FI_READ)) || ((comp_flags & FI_READ) && !(comp_flags & FI_WRITE))); if (comp_flags & FI_WRITE) ofi_ep_wr_cntr_inc(&rxm_ep->util_ep); else ofi_ep_rd_cntr_inc(&rxm_ep->util_ep); if (!(rma_buf->flags & FI_INJECT) && !rxm_ep->rxm_mr_local && rxm_ep->msg_mr_local) { rxm_ep_msg_mr_closev(rma_buf->mr.mr, rma_buf->mr.count); } ofi_buf_free(rma_buf); return ret; }
static ssize_t rxm_ep_readv(struct fid_ep *ep_fid, const struct iovec *iov, void **desc, size_t count, fi_addr_t src_addr, uint64_t addr, uint64_t key, void *context) { struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); struct fi_rma_iov rma_iov = { .addr = addr, .len = ofi_total_iov_len(iov, count), .key = key, }; struct fi_msg_rma msg = { .msg_iov = iov, .desc = desc, .iov_count = count, .addr = src_addr, .rma_iov = &rma_iov, .rma_iov_count = 1, .context = context, .data = 0, }; return rxm_ep_rma_common(rxm_ep, &msg, rxm_ep_tx_flags(rxm_ep), fi_readmsg, FI_READ); } static ssize_t rxm_ep_read(struct fid_ep *ep_fid, void *buf, size_t len, void *desc, fi_addr_t src_addr, uint64_t addr, uint64_t key, void *context) { struct iovec iov = { .iov_base = (void*)buf, .iov_len = len, }; struct fi_rma_iov rma_iov = { .addr = addr, .len = len, .key = key, }; struct fi_msg_rma msg = { .msg_iov = &iov, .desc = &desc, .iov_count = 1, .addr = src_addr, .rma_iov = &rma_iov, .rma_iov_count = 1, .context = context, .data = 0, }; struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); return rxm_ep_rma_common(rxm_ep, &msg, rxm_ep_tx_flags(rxm_ep), fi_readmsg, FI_READ); } static inline void rxm_ep_format_rma_msg(struct rxm_rma_buf *rma_buf, const struct fi_msg_rma *orig_msg, struct iovec *rxm_iov, struct fi_msg_rma *rxm_msg) { rxm_msg->context = rma_buf; rxm_msg->addr = orig_msg->addr; rxm_msg->data = orig_msg->data; ofi_copy_from_iov(rma_buf->pkt.data, rma_buf->pkt.hdr.size, orig_msg->msg_iov, orig_msg->iov_count, 0); rxm_iov->iov_base = &rma_buf->pkt.data; rxm_iov->iov_len = rma_buf->pkt.hdr.size; rxm_msg->msg_iov = rxm_iov; rxm_msg->desc = &rma_buf->hdr.desc; rxm_msg->iov_count = 1; rxm_msg->rma_iov = orig_msg->rma_iov; rxm_msg->rma_iov_count = orig_msg->rma_iov_count; } static inline ssize_t rxm_ep_rma_emulate_inject_msg(struct rxm_ep *rxm_ep, struct rxm_conn *rxm_conn, size_t total_size, const struct fi_msg_rma *msg, uint64_t flags) { struct rxm_rma_buf *rma_buf; ssize_t ret; struct iovec rxm_msg_iov = { 0 }; struct fi_msg_rma rxm_rma_msg = { 0 }; assert(msg->rma_iov_count <= rxm_ep->rxm_info->tx_attr->rma_iov_limit); ofi_ep_lock_acquire(&rxm_ep->util_ep); rma_buf = rxm_rma_buf_alloc(rxm_ep); if (OFI_UNLIKELY(!rma_buf)) { FI_WARN(&rxm_prov, FI_LOG_EP_DATA, "Ran out of buffers from RMA buffer pool\n"); ret = -FI_ENOMEM; goto unlock; } rma_buf->pkt.hdr.size = total_size; rma_buf->app_context = msg->context; rma_buf->flags = flags; rxm_ep_format_rma_msg(rma_buf, msg, &rxm_msg_iov, &rxm_rma_msg); flags = (flags & ~FI_INJECT) | FI_COMPLETION; ret = fi_writemsg(rxm_conn->msg_ep, &rxm_rma_msg, flags); if (OFI_UNLIKELY(ret)) { if (ret == -FI_EAGAIN) rxm_ep_do_progress(&rxm_ep->util_ep); ofi_buf_free(rma_buf); } unlock: ofi_ep_lock_release(&rxm_ep->util_ep); return ret; } static inline ssize_t rxm_ep_rma_emulate_inject(struct rxm_ep *rxm_ep, struct rxm_conn *rxm_conn, const void *buf, size_t len, uint64_t data, fi_addr_t dest_addr, uint64_t addr, uint64_t key, uint64_t flags) { struct fi_rma_iov rma_iov = { .addr = addr, .len = len, .key = key, }; struct iovec iov = { .iov_base = (void*)buf, .iov_len = len, }; struct fi_msg_rma msg = { .msg_iov = &iov, .desc = NULL, .iov_count = 1, .addr = dest_addr, .rma_iov = &rma_iov, .rma_iov_count = 1, .context = NULL, .data = data, }; return rxm_ep_rma_emulate_inject_msg(rxm_ep, rxm_conn, len, &msg, flags); } static inline ssize_t rxm_ep_rma_inject_common(struct rxm_ep *rxm_ep, const struct fi_msg_rma *msg, uint64_t flags) { struct rxm_conn *rxm_conn; size_t total_size = ofi_total_iov_len(msg->msg_iov, msg->iov_count); ssize_t ret; assert(total_size <= rxm_ep->rxm_info->tx_attr->inject_size); ret = rxm_ep_prepare_tx(rxm_ep, msg->addr, &rxm_conn); if (OFI_UNLIKELY(ret)) return ret; if ((total_size <= rxm_ep->msg_info->tx_attr->inject_size) && !(flags & FI_COMPLETION) && (msg->iov_count == 1) && (msg->rma_iov_count == 1)) { if (flags & FI_REMOTE_CQ_DATA) { ret = fi_inject_writedata(rxm_conn->msg_ep, msg->msg_iov->iov_base, msg->msg_iov->iov_len, msg->data, msg->addr, msg->rma_iov->addr, msg->rma_iov->key); } else { ret = fi_inject_write(rxm_conn->msg_ep, msg->msg_iov->iov_base, msg->msg_iov->iov_len, msg->addr, msg->rma_iov->addr, msg->rma_iov->key); } if (OFI_LIKELY(!ret)) { ofi_ep_wr_cntr_inc(&rxm_ep->util_ep); } else { FI_DBG(&rxm_prov, FI_LOG_EP_DATA, "fi_inject_write* for MSG provider failed with ret - %" PRId64"\n", ret); if (OFI_LIKELY(ret == -FI_EAGAIN)) rxm_ep_progress(&rxm_ep->util_ep); } return ret; } else { return rxm_ep_rma_emulate_inject_msg(rxm_ep, rxm_conn, total_size, msg, flags); } } static inline ssize_t rxm_ep_generic_writemsg(struct fid_ep *ep_fid, const struct fi_msg_rma *msg, uint64_t flags) { struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); if (flags & FI_INJECT) return rxm_ep_rma_inject_common(rxm_ep, msg, flags); else return rxm_ep_rma_common(rxm_ep, msg, flags, fi_writemsg, FI_WRITE); } static inline ssize_t rxm_ep_writemsg(struct fid_ep *ep_fid, const struct fi_msg_rma *msg, uint64_t flags) { struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); return rxm_ep_generic_writemsg(ep_fid, msg, flags | rxm_ep->util_ep.tx_msg_flags); } static ssize_t rxm_ep_writev(struct fid_ep *ep_fid, const struct iovec *iov, void **desc, size_t count, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context) { struct fi_rma_iov rma_iov = { .addr = addr, .len = ofi_total_iov_len(iov, count), .key = key, }; struct fi_msg_rma msg = { .msg_iov = iov, .desc = desc, .iov_count = count, .addr = dest_addr, .rma_iov = &rma_iov, .rma_iov_count = 1, .context = context, .data = 0, }; struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); return rxm_ep_generic_writemsg(ep_fid, &msg, rxm_ep_tx_flags(rxm_ep)); } static ssize_t rxm_ep_writedata(struct fid_ep *ep_fid, const void *buf, size_t len, void *desc, uint64_t data, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context) { struct fi_rma_iov rma_iov = { .addr = addr, .len = len, .key = key, }; struct iovec iov = { .iov_base = (void*)buf, .iov_len = len, }; struct fi_msg_rma msg = { .msg_iov = &iov, .desc = &desc, .iov_count = 1, .addr = dest_addr, .rma_iov = &rma_iov, .rma_iov_count = 1, .context = context, .data = data, }; struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); return rxm_ep_generic_writemsg(ep_fid, &msg, rxm_ep_tx_flags(rxm_ep) | FI_REMOTE_CQ_DATA); } static ssize_t rxm_ep_write(struct fid_ep *ep_fid, const void *buf, size_t len, void *desc, fi_addr_t dest_addr, uint64_t addr, uint64_t key, void *context) { struct fi_rma_iov rma_iov = { .addr = addr, .len = len, .key = key, }; struct iovec iov = { .iov_base = (void*)buf, .iov_len = len, }; struct fi_msg_rma msg = { .msg_iov = &iov, .desc = &desc, .iov_count = 1, .addr = dest_addr, .rma_iov = &rma_iov, .rma_iov_count = 1, .context = context, .data = 0, }; struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); return rxm_ep_generic_writemsg(ep_fid, &msg, rxm_ep_tx_flags(rxm_ep)); } static ssize_t rxm_ep_inject_write(struct fid_ep *ep_fid, const void *buf, size_t len, fi_addr_t dest_addr, uint64_t addr, uint64_t key) { ssize_t ret; struct rxm_conn *rxm_conn; struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); ret = rxm_ep_prepare_tx(rxm_ep, dest_addr, &rxm_conn); if (OFI_UNLIKELY(ret)) return ret; if (len <= rxm_ep->msg_info->tx_attr->inject_size) { ret = fi_inject_write(rxm_conn->msg_ep, buf, len, dest_addr, addr, key); if (OFI_LIKELY(!ret)) { ofi_ep_wr_cntr_inc(&rxm_ep->util_ep); } else { FI_DBG(&rxm_prov, FI_LOG_EP_DATA, "fi_inject_write for MSG provider failed with ret - %" PRId64"\n", ret); if (OFI_LIKELY(ret == -FI_EAGAIN)) rxm_ep_progress(&rxm_ep->util_ep); } return ret; } else { return rxm_ep_rma_emulate_inject(rxm_ep, rxm_conn, buf, len, 0, dest_addr, addr, key, FI_INJECT); } } static ssize_t rxm_ep_inject_writedata(struct fid_ep *ep_fid, const void *buf, size_t len, uint64_t data, fi_addr_t dest_addr, uint64_t addr, uint64_t key) { ssize_t ret; struct rxm_conn *rxm_conn; struct rxm_ep *rxm_ep = container_of(ep_fid, struct rxm_ep, util_ep.ep_fid.fid); ret = rxm_ep_prepare_tx(rxm_ep, dest_addr, &rxm_conn); if (OFI_UNLIKELY(ret)) return ret; if (len <= rxm_ep->msg_info->tx_attr->inject_size) { ret = fi_inject_writedata(rxm_conn->msg_ep, buf, len, data, dest_addr, addr, key); if (OFI_LIKELY(!ret)) { ofi_ep_wr_cntr_inc(&rxm_ep->util_ep); } else { FI_DBG(&rxm_prov, FI_LOG_EP_DATA, "fi_inject_writedata for MSG provider failed with ret - %" PRId64"\n", ret); if (OFI_LIKELY(ret == -FI_EAGAIN)) rxm_ep_progress(&rxm_ep->util_ep); } return ret; } else { return rxm_ep_rma_emulate_inject(rxm_ep, rxm_conn, buf, len, data, dest_addr, addr, key, FI_REMOTE_CQ_DATA | FI_INJECT); } } struct fi_ops_rma rxm_ops_rma = { .size = sizeof (struct fi_ops_rma), .read = rxm_ep_read, .readv = rxm_ep_readv, .readmsg = rxm_ep_readmsg, .write = rxm_ep_write, .writev = rxm_ep_writev, .writemsg = rxm_ep_writemsg, .inject = rxm_ep_inject_write, .writedata = rxm_ep_writedata, .injectdata = rxm_ep_inject_writedata, };