// Copyright (c) Vitaliy Filippov, 2019+ // License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details) #define _XOPEN_SOURCE #include #include #include "messenger.h" #include "msgr_iothread.h" void osd_messenger_t::outbox_push(osd_op_t *cur_op) { assert(cur_op->client_id); auto cl_it = clients.find(cur_op->client_id); if (cl_it == clients.end() || cl_it->second->peer_state == PEER_STOPPED) { delete cur_op; return; } osd_client_t *cl = cl_it->second; if (cur_op->op_type == OSD_OP_OUT) { clock_gettime(CLOCK_REALTIME, &cur_op->tv_begin); cur_op->req.hdr.id = ++cl->send_op_id; } else { // Remove the operation from received op list bool found = false; for (auto it = cl->received_ops.begin(); it != cl->received_ops.end(); it++) { if (*it == cur_op) { found = true; cl->received_ops.erase(it, it+1); break; } } // Can't be not found because client IDs are unique assert(found); } auto & to_send_list = cl->write_msg.msg_iovlen ? cl->next_send_list : cl->send_list; auto & to_outbox = cl->write_msg.msg_iovlen ? cl->next_outbox : cl->outbox; if (cur_op->op_type == OSD_OP_IN) { measure_exec(cur_op); to_send_list.push_back((iovec){ .iov_base = cur_op->reply.buf, .iov_len = OSD_PACKET_SIZE }); } else { to_send_list.push_back((iovec){ .iov_base = cur_op->req.buf, .iov_len = OSD_PACKET_SIZE }); cl->sent_ops[cur_op->req.hdr.id] = cur_op; } to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = MSGR_SENDP_HDR }); // Bitmap if (cur_op->op_type == OSD_OP_IN && cur_op->req.hdr.opcode == OSD_OP_SEC_READ && cur_op->reply.sec_rw.attr_len > 0) { to_send_list.push_back((iovec){ .iov_base = cur_op->bitmap, .iov_len = cur_op->reply.sec_rw.attr_len, }); to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 }); } else if (cur_op->op_type == OSD_OP_OUT && (cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE || cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE) && cur_op->req.sec_rw.attr_len > 0) { to_send_list.push_back((iovec){ .iov_base = cur_op->bitmap, .iov_len = cur_op->req.sec_rw.attr_len, }); to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 }); } // Operation data if ((cur_op->op_type == OSD_OP_IN ? (cur_op->req.hdr.opcode == OSD_OP_READ || cur_op->req.hdr.opcode == OSD_OP_SEC_READ || cur_op->req.hdr.opcode == OSD_OP_SEC_LIST || cur_op->req.hdr.opcode == OSD_OP_SHOW_CONFIG || cur_op->req.hdr.opcode == OSD_OP_DESCRIBE) : (cur_op->req.hdr.opcode == OSD_OP_WRITE || cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE || cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE || cur_op->req.hdr.opcode == OSD_OP_SEC_STABILIZE || cur_op->req.hdr.opcode == OSD_OP_SEC_ROLLBACK || cur_op->req.hdr.opcode == OSD_OP_SHOW_CONFIG)) && cur_op->iov.count > 0) { for (int i = 0; i < cur_op->iov.count; i++) { if (cur_op->iov.buf[i].iov_len > 0) { assert(cur_op->iov.buf[i].iov_base); to_send_list.push_back(cur_op->iov.buf[i]); to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 }); } } } if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP) { if (cur_op->op_type == OSD_OP_IN && cur_op->reply.hdr.retval > 0) { to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->reply.hdr.retval }); to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 }); } else if (cur_op->op_type == OSD_OP_OUT && cur_op->req.sec_read_bmp.len > 0) { to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->req.sec_read_bmp.len }); to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 }); } } if (cur_op->op_type == OSD_OP_IN) { to_outbox[to_outbox.size()-1].flags |= MSGR_SENDP_FREE; } #ifdef WITH_RDMA if (cl->peer_state == PEER_RDMA) { try_send_rdma(cl); return; } #endif if (!ringloop) { // FIXME: It's worse because it doesn't allow batching while (cl->outbox.size()) { try_send(cl); } } else { if ((cl->write_msg.msg_iovlen > 0 || !try_send(cl)) && (cl->write_state == 0)) { cl->write_state = CL_WRITE_READY; write_ready_clients.push_back(cur_op->client_id); } ringloop->wakeup(); } } bool osd_messenger_t::try_send(osd_client_t *cl) { if (!cl->send_list.size() || cl->write_msg.msg_iovlen > 0 || cl->peer_state == PEER_STOPPED || cl->peer_fd < 0) { return true; } assert(cl->peer_state != PEER_RDMA); if (ringloop && !use_sync_send_recv) { auto iothread = iothreads.size() ? iothreads[cl->peer_fd % iothreads.size()] : NULL; io_uring_sqe sqe_local; ring_data_t data_local; io_uring_sqe* sqe = (iothread ? &sqe_local : ringloop->get_sqe()); if (iothread) { sqe_local = { .user_data = (uint64_t)&data_local }; data_local = {}; } if (!sqe) return false; cl->write_msg.msg_iov = cl->send_list.data(); cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX; cl->refs++; ring_data_t* data = ((ring_data_t*)sqe->user_data); data->callback = [this, cl](ring_data_t *data) { handle_send(data->res, data->prev, data->more, cl); }; bool use_zc = has_sendmsg_zc && min_zerocopy_send_size >= 0; if (use_zc && min_zerocopy_send_size > 0) { size_t avg_size = 0; for (size_t i = 0; i < cl->write_msg.msg_iovlen; i++) avg_size += cl->write_msg.msg_iov[i].iov_len; if (avg_size/cl->write_msg.msg_iovlen < min_zerocopy_send_size) use_zc = false; } if (use_zc) { io_uring_prep_sendmsg_zc(sqe, cl->peer_fd, &cl->write_msg, MSG_WAITALL); } else { io_uring_prep_sendmsg(sqe, cl->peer_fd, &cl->write_msg, MSG_WAITALL); } if (iothread) { iothread->add_sqe(sqe_local); } } else { cl->write_msg.msg_iov = cl->send_list.data(); cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX; cl->refs++; int result = sendmsg(cl->peer_fd, &cl->write_msg, MSG_NOSIGNAL); if (result < 0) { result = -errno; } // like set_immediate tfd->set_timer_us(0, false, [this, result, cl](int){ handle_send(result, false, false, cl); }); } return true; } void osd_messenger_t::send_replies() { for (int i = 0; i < write_ready_clients.size(); i++) { uint64_t client_id = write_ready_clients[i]; auto cl_it = clients.find(client_id); if (cl_it != clients.end() && cl_it->second->peer_state != PEER_RDMA && !try_send(cl_it->second)) { write_ready_clients.erase(write_ready_clients.begin(), write_ready_clients.begin() + i); return; } } write_ready_clients.clear(); } void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t *cl) { if (!prev) { cl->write_msg.msg_iovlen = 0; } if (!more) { cl->refs--; } if (cl->peer_state == PEER_STOPPED) { if (cl->refs <= 0) { destroy_client(cl); } return; } if (result < 0 && result != -EAGAIN && result != -EINTR) { // this is a client socket, so don't panic. just disconnect it fprintf(stderr, "Client %ju socket write error: %d (%s). Disconnecting client\n", cl->client_id, -result, strerror(-result)); stop_client(cl->client_id); return; } if (result >= 0) { if (prev) { // Second notification - only free a batch of postponed ops int i = 0; for (; i < cl->zc_free_list.size() && cl->zc_free_list[i]; i++) delete cl->zc_free_list[i]; if (i > 0) cl->zc_free_list.erase(cl->zc_free_list.begin(), cl->zc_free_list.begin()+i+1); return; } int done = 0; while (result > 0 && done < cl->send_list.size()) { iovec & iov = cl->send_list[done]; if (iov.iov_len <= result) { if (cl->outbox[done].flags & MSGR_SENDP_FREE) { // Reply fully sent if (more) cl->zc_free_list.push_back(cl->outbox[done].op); else delete cl->outbox[done].op; } result -= iov.iov_len; done++; } else { iov.iov_len -= result; iov.iov_base = (uint8_t*)iov.iov_base + result; break; } } if (more) { int expected = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX; if (done != expected) { fprintf(stderr, "Client %ju socket write error: expected to send " "%d iovecs with MSG_WAITALL but sent %d. Disconnecting client\n", cl->client_id, expected, done); stop_client(cl->client_id); return; } cl->zc_free_list.push_back(NULL); // end marker } if (done > 0) { cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+done); cl->outbox.erase(cl->outbox.begin(), cl->outbox.begin()+done); } if (cl->next_send_list.size()) { cl->send_list.insert(cl->send_list.end(), cl->next_send_list.begin(), cl->next_send_list.end()); cl->outbox.insert(cl->outbox.end(), cl->next_outbox.begin(), cl->next_outbox.end()); cl->next_send_list.clear(); cl->next_outbox.clear(); } cl->write_state = cl->outbox.size() > 0 ? CL_WRITE_READY : 0; #ifdef WITH_RDMA if (cl->rdma_conn && !cl->outbox.size() && cl->peer_state == PEER_RDMA_CONNECTING) { // FIXME: Ignore pings during RDMA state transition if (log_level > 0) { fprintf(stderr, "Successfully connected with client %ju using RDMA\n", cl->client_id); } cl->peer_state = PEER_RDMA; // Add the initial receive request init_recv_rdma(cl); } #endif } if (cl->write_state != 0) { write_ready_clients.push_back(cl->client_id); } }