// Copyright (c) Vitaliy Filippov, 2019+ // License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details) #include #include #include "msgr_rdma.h" #include "messenger.h" std::string msgr_rdma_address_t::to_string() { char msg[sizeof "0000:00000000:00000000:00000000000000000000000000000000"]; sprintf( msg, "%04x:%06x:%06x:%016jx%016jx", lid, qpn, psn, htobe64(((uint64_t*)&gid)[0]), htobe64(((uint64_t*)&gid)[1]) ); return std::string(msg); } bool msgr_rdma_address_t::from_string(const char *str, msgr_rdma_address_t *dest) { uint64_t* gid = (uint64_t*)&dest->gid; int scanned = sscanf( str, "%hx:%x:%x:%16jx%16jx", &dest->lid, &dest->qpn, &dest->psn, gid, gid+1 ); gid[0] = be64toh(gid[0]); gid[1] = be64toh(gid[1]); return scanned == 5; } msgr_rdma_context_t::~msgr_rdma_context_t() { if (cq) ibv_destroy_cq(cq); if (channel) ibv_destroy_comp_channel(channel); if (mr) ibv_dereg_mr(mr); if (pd) ibv_dealloc_pd(pd); if (context) ibv_close_device(context); } msgr_rdma_connection_t::~msgr_rdma_connection_t() { ctx->used_max_cqe -= max_send+max_recv; if (qp) ibv_destroy_qp(qp); if (recv_buffers.size()) { for (auto b: recv_buffers) { if (b.mr) ibv_dereg_mr(b.mr); free(b.buf); } recv_buffers.clear(); } if (send_out.mr) { ibv_dereg_mr(send_out.mr); send_out.mr = NULL; } if (send_out.buf) { free(send_out.buf); send_out.buf = NULL; } send_out_size = 0; } msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, bool odp, int log_level) { int res; ibv_device **dev_list = NULL; msgr_rdma_context_t *ctx = new msgr_rdma_context_t(); ctx->mtu = mtu; timespec tv; clock_gettime(CLOCK_REALTIME, &tv); srand48(tv.tv_sec*1000000000 + tv.tv_nsec); dev_list = ibv_get_device_list(NULL); if (!dev_list) { if (errno == -ENOSYS || errno == ENOSYS) { if (log_level > 0) fprintf(stderr, "No RDMA devices found (RDMA device list returned ENOSYS)\n"); } else fprintf(stderr, "Failed to get RDMA device list: %s\n", strerror(errno)); goto cleanup; } if (!ib_devname) { ctx->dev = *dev_list; if (!ctx->dev) { if (log_level > 0) fprintf(stderr, "No RDMA devices found\n"); goto cleanup; } } else { int i; for (i = 0; dev_list[i]; ++i) if (!strcmp(ibv_get_device_name(dev_list[i]), ib_devname)) break; ctx->dev = dev_list[i]; if (!ctx->dev) { fprintf(stderr, "RDMA device %s not found\n", ib_devname); goto cleanup; } } ctx->context = ibv_open_device(ctx->dev); if (!ctx->context) { fprintf(stderr, "Couldn't get RDMA context for %s\n", ibv_get_device_name(ctx->dev)); goto cleanup; } ctx->ib_port = ib_port; ctx->gid_index = gid_index; if ((res = ibv_query_port(ctx->context, ib_port, &ctx->portinfo)) != 0) { fprintf(stderr, "Couldn't get RDMA device %s port %d info: %s\n", ibv_get_device_name(ctx->dev), ib_port, strerror(res)); goto cleanup; } ctx->my_lid = ctx->portinfo.lid; if (ctx->portinfo.link_layer != IBV_LINK_LAYER_ETHERNET && !ctx->my_lid) { fprintf(stderr, "RDMA device %s must have local LID because it's not Ethernet, but LID is zero\n", ibv_get_device_name(ctx->dev)); goto cleanup; } if (ibv_query_gid(ctx->context, ib_port, gid_index, &ctx->my_gid)) { fprintf(stderr, "Couldn't read RDMA device %s GID index %d\n", ibv_get_device_name(ctx->dev), gid_index); goto cleanup; } ctx->pd = ibv_alloc_pd(ctx->context); if (!ctx->pd) { fprintf(stderr, "Couldn't allocate RDMA protection domain\n"); goto cleanup; } { if (ibv_query_device_ex(ctx->context, NULL, &ctx->attrx)) { fprintf(stderr, "Couldn't query RDMA device for its features\n"); goto cleanup; } ctx->odp = odp; if (ctx->odp && (!(ctx->attrx.odp_caps.general_caps & IBV_ODP_SUPPORT) || !(ctx->attrx.odp_caps.general_caps & IBV_ODP_SUPPORT_IMPLICIT) || !(ctx->attrx.odp_caps.per_transport_caps.rc_odp_caps & IBV_ODP_SUPPORT_SEND) || !(ctx->attrx.odp_caps.per_transport_caps.rc_odp_caps & IBV_ODP_SUPPORT_RECV))) { ctx->odp = false; if (log_level > 0) fprintf(stderr, "The RDMA device isn't implicit ODP (On-Demand Paging) capable, disabling it\n"); } } if (ctx->odp) { ctx->mr = ibv_reg_mr(ctx->pd, NULL, SIZE_MAX, IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_ON_DEMAND); if (!ctx->mr) { fprintf(stderr, "Couldn't register RDMA memory region\n"); goto cleanup; } } ctx->channel = ibv_create_comp_channel(ctx->context); if (!ctx->channel) { fprintf(stderr, "Couldn't create RDMA completion channel\n"); goto cleanup; } ctx->max_cqe = 4096; ctx->cq = ibv_create_cq(ctx->context, ctx->max_cqe, NULL, ctx->channel, 0); if (!ctx->cq) { fprintf(stderr, "Couldn't create RDMA completion queue\n"); goto cleanup; } if (dev_list) ibv_free_device_list(dev_list); return ctx; cleanup: delete ctx; if (dev_list) ibv_free_device_list(dev_list); return NULL; } msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx, uint32_t max_send, uint32_t max_recv, uint32_t max_sge, uint32_t max_msg) { msgr_rdma_connection_t *conn = new msgr_rdma_connection_t; max_sge = max_sge > ctx->attrx.orig_attr.max_sge ? ctx->attrx.orig_attr.max_sge : max_sge; conn->ctx = ctx; conn->max_send = max_send; conn->max_recv = max_recv; conn->max_sge = max_sge; conn->max_msg = max_msg; ctx->used_max_cqe += max_send+max_recv; if (ctx->used_max_cqe > ctx->max_cqe) { // Resize CQ // Mellanox ConnectX-4 supports up to 4194303 CQEs, so it's fine to put everything into a single CQ int new_max_cqe = ctx->max_cqe; while (ctx->used_max_cqe > new_max_cqe) { new_max_cqe *= 2; } if (ibv_resize_cq(ctx->cq, new_max_cqe) != 0) { fprintf(stderr, "Couldn't resize RDMA completion queue to %d entries\n", new_max_cqe); delete conn; return NULL; } ctx->max_cqe = new_max_cqe; } ibv_qp_init_attr init_attr = { .send_cq = ctx->cq, .recv_cq = ctx->cq, .cap = { .max_send_wr = max_send, .max_recv_wr = max_recv, .max_send_sge = max_sge, .max_recv_sge = max_sge, }, .qp_type = IBV_QPT_RC, }; conn->qp = ibv_create_qp(ctx->pd, &init_attr); if (!conn->qp) { fprintf(stderr, "Couldn't create RDMA queue pair\n"); delete conn; return NULL; } conn->addr.lid = ctx->my_lid; conn->addr.gid = ctx->my_gid; conn->addr.qpn = conn->qp->qp_num; conn->addr.psn = lrand48() & 0xffffff; ibv_qp_attr attr = { .qp_state = IBV_QPS_INIT, .qp_access_flags = 0, .pkey_index = 0, .port_num = ctx->ib_port, }; if (ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT | IBV_QP_ACCESS_FLAGS)) { fprintf(stderr, "Failed to switch RDMA queue pair to INIT state\n"); delete conn; return NULL; } return conn; } static ibv_mtu mtu_to_ibv_mtu(uint32_t mtu) { switch (mtu) { case 256: return IBV_MTU_256; case 512: return IBV_MTU_512; case 1024: return IBV_MTU_1024; case 2048: return IBV_MTU_2048; case 4096: return IBV_MTU_4096; } return IBV_MTU_4096; } int msgr_rdma_connection_t::connect(msgr_rdma_address_t *dest) { auto conn = this; ibv_qp_attr attr = { .qp_state = IBV_QPS_RTR, .path_mtu = mtu_to_ibv_mtu(conn->ctx->mtu), .rq_psn = dest->psn, .sq_psn = conn->addr.psn, .dest_qp_num = dest->qpn, .ah_attr = { .grh = { .dgid = dest->gid, .sgid_index = conn->ctx->gid_index, .hop_limit = 1, // FIXME can it vary? }, .dlid = dest->lid, .sl = 0, // service level .src_path_bits = 0, .is_global = (uint8_t)(dest->gid.global.interface_id ? 1 : 0), .port_num = conn->ctx->ib_port, }, .max_rd_atomic = 1, .max_dest_rd_atomic = 1, // Timeout and min_rnr_timer actual values seem to be 4.096us*2^(timeout+1) .min_rnr_timer = 1, .timeout = 14, .retry_cnt = 7, .rnr_retry = 7, }; // FIXME No idea if ibv_modify_qp is a blocking operation or not. No idea if it has a timeout and what it is. if (ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU | IBV_QP_DEST_QPN | IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC | IBV_QP_MIN_RNR_TIMER)) { fprintf(stderr, "Failed to switch RDMA queue pair to RTR (ready-to-receive) state\n"); return 1; } attr.qp_state = IBV_QPS_RTS; if (ibv_modify_qp(conn->qp, &attr, IBV_QP_STATE | IBV_QP_TIMEOUT | IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY | IBV_QP_SQ_PSN | IBV_QP_MAX_QP_RD_ATOMIC)) { fprintf(stderr, "Failed to switch RDMA queue pair to RTS (ready-to-send) state\n"); return 1; } return 0; } bool osd_messenger_t::connect_rdma(int peer_fd, std::string rdma_address, uint64_t client_max_msg) { // Try to connect to the peer using RDMA msgr_rdma_address_t addr; if (msgr_rdma_address_t::from_string(rdma_address.c_str(), &addr)) { if (client_max_msg > rdma_max_msg) { client_max_msg = rdma_max_msg; } auto rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, client_max_msg); if (rdma_conn) { int r = rdma_conn->connect(&addr); if (r != 0) { delete rdma_conn; fprintf( stderr, "Failed to connect RDMA queue pair to %s (client %d)\n", addr.to_string().c_str(), peer_fd ); } else { // Remember connection, but switch to RDMA only after sending the configuration response auto cl = clients.at(peer_fd); cl->rdma_conn = rdma_conn; cl->peer_state = PEER_RDMA_CONNECTING; return true; } } } return false; } static void try_send_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge) { ibv_send_wr *bad_wr = NULL; ibv_send_wr wr = { .wr_id = (uint64_t)(cl->peer_fd*2+1), .sg_list = sge, .num_sge = op_sge, .opcode = IBV_WR_SEND, .send_flags = IBV_SEND_SIGNALED, }; int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr); if (err || bad_wr) { fprintf(stderr, "RDMA send failed: %s\n", strerror(err)); exit(1); } cl->rdma_conn->cur_send++; } static int try_send_rdma_copy(osd_client_t *cl, uint8_t *dst, int dst_len) { auto rc = cl->rdma_conn; int total_dst_len = dst_len; while (dst_len > 0 && rc->send_pos < cl->send_list.size()) { iovec & iov = cl->send_list[rc->send_pos]; uint32_t len = (uint32_t)(iov.iov_len-rc->send_buf_pos < dst_len ? iov.iov_len-rc->send_buf_pos : dst_len); memcpy(dst, iov.iov_base+rc->send_buf_pos, len); dst += len; dst_len -= len; rc->send_buf_pos += len; if (rc->send_buf_pos >= iov.iov_len) { rc->send_pos++; rc->send_buf_pos = 0; } } return total_dst_len-dst_len; } void osd_messenger_t::try_send_rdma_odp(osd_client_t *cl) { auto rc = cl->rdma_conn; if (!cl->send_list.size() || rc->cur_send >= rc->max_send) { return; } uint64_t op_size = 0, op_sge = 0; ibv_sge sge[rc->max_sge]; while (rc->send_pos < cl->send_list.size()) { iovec & iov = cl->send_list[rc->send_pos]; if (op_size >= rc->max_msg || op_sge >= rc->max_sge) { rc->send_sizes.push_back(op_size); try_send_rdma_wr(cl, sge, op_sge); op_sge = 0; op_size = 0; if (rc->cur_send >= rc->max_send) { break; } } uint32_t len = (uint32_t)(op_size+iov.iov_len-rc->send_buf_pos < rc->max_msg ? iov.iov_len-rc->send_buf_pos : rc->max_msg-op_size); sge[op_sge++] = { .addr = (uintptr_t)((uint8_t*)iov.iov_base+rc->send_buf_pos), .length = len, .lkey = rc->ctx->mr->lkey, }; op_size += len; rc->send_buf_pos += len; if (rc->send_buf_pos >= iov.iov_len) { rc->send_pos++; rc->send_buf_pos = 0; } } if (op_sge > 0) { rc->send_sizes.push_back(op_size); try_send_rdma_wr(cl, sge, op_sge); } } void osd_messenger_t::try_send_rdma_nodp(osd_client_t *cl) { auto rc = cl->rdma_conn; if (!rc->send_out_size) { // Allocate send ring buffer, if not yet rc->send_out_size = rc->max_msg*rdma_max_send; rc->send_out.buf = malloc_or_die(rc->send_out_size); if (!rdma_context->odp) { rc->send_out.mr = ibv_reg_mr(rdma_context->pd, rc->send_out.buf, rc->send_out_size, 0); if (!rc->send_out.mr) { fprintf(stderr, "Failed to register RDMA memory region: %s\n", strerror(errno)); exit(1); } } } // Copy data into the buffer and send it uint8_t *dst = NULL; int dst_len = 0; int copied = 1; while (!rc->send_out_full && copied > 0 && rc->cur_send < rc->max_send) { dst = (uint8_t*)rc->send_out.buf + rc->send_out_pos; dst_len = (rc->send_out_pos < rc->send_out_size ? rc->send_out_size-rc->send_out_pos : rc->send_done_pos-rc->send_out_pos); if (dst_len > rc->max_msg) dst_len = rc->max_msg; copied = try_send_rdma_copy(cl, dst, dst_len); if (copied > 0) { rc->send_out_pos += copied; if (rc->send_out_pos == rc->send_out_size) rc->send_out_pos = 0; assert(rc->send_out_pos < rc->send_out_size); if (rc->send_out_pos >= rc->send_done_pos) rc->send_out_full = true; ibv_sge sge = { .addr = (uintptr_t)dst, .length = (uint32_t)copied, .lkey = rdma_context->odp ? rdma_context->mr->lkey : rc->send_out.mr->lkey, }; try_send_rdma_wr(cl, &sge, 1); rc->send_sizes.push_back(copied); } } } void osd_messenger_t::try_send_rdma(osd_client_t *cl) { if (rdma_context->odp) try_send_rdma_odp(cl); else try_send_rdma_nodp(cl); } static void try_recv_rdma_wr(osd_client_t *cl, msgr_rdma_buf_t b) { ibv_sge sge = { .addr = (uintptr_t)b.buf, .length = (uint32_t)cl->rdma_conn->max_msg, .lkey = cl->rdma_conn->ctx->odp ? cl->rdma_conn->ctx->mr->lkey : b.mr->lkey, }; ibv_recv_wr *bad_wr = NULL; ibv_recv_wr wr = { .wr_id = (uint64_t)(cl->peer_fd*2), .sg_list = &sge, .num_sge = 1, }; int err = ibv_post_recv(cl->rdma_conn->qp, &wr, &bad_wr); if (err || bad_wr) { fprintf(stderr, "RDMA receive failed: %s\n", strerror(err)); exit(1); } cl->rdma_conn->cur_recv++; } bool osd_messenger_t::try_recv_rdma(osd_client_t *cl) { auto rc = cl->rdma_conn; while (rc->cur_recv < rc->max_recv) { msgr_rdma_buf_t b; b.buf = malloc_or_die(rc->max_msg); if (!rdma_context->odp) { b.mr = ibv_reg_mr(rdma_context->pd, b.buf, rc->max_msg, IBV_ACCESS_LOCAL_WRITE); if (!b.mr) { fprintf(stderr, "Failed to register RDMA memory region: %s\n", strerror(errno)); exit(1); } } rc->recv_buffers.push_back(b); try_recv_rdma_wr(cl, b); } return true; } #define RDMA_EVENTS_AT_ONCE 32 void osd_messenger_t::handle_rdma_events() { // Request next notification ibv_cq *ev_cq; void *ev_ctx; // FIXME: This is inefficient as it calls read()... if (ibv_get_cq_event(rdma_context->channel, &ev_cq, &ev_ctx) == 0) { ibv_ack_cq_events(rdma_context->cq, 1); } if (ibv_req_notify_cq(rdma_context->cq, 0) != 0) { fprintf(stderr, "Failed to request RDMA completion notification, exiting\n"); exit(1); } ibv_wc wc[RDMA_EVENTS_AT_ONCE]; int event_count; do { event_count = ibv_poll_cq(rdma_context->cq, RDMA_EVENTS_AT_ONCE, wc); for (int i = 0; i < event_count; i++) { int client_id = wc[i].wr_id >> 1; bool is_send = wc[i].wr_id & 1; auto cl_it = clients.find(client_id); if (cl_it == clients.end()) { continue; } osd_client_t *cl = cl_it->second; auto rc = cl->rdma_conn; if (wc[i].status != IBV_WC_SUCCESS) { fprintf(stderr, "RDMA work request failed for client %d", client_id); if (cl->osd_num) { fprintf(stderr, " (OSD %ju)", cl->osd_num); } fprintf(stderr, " with status: %s, stopping client\n", ibv_wc_status_str(wc[i].status)); stop_client(client_id); continue; } if (!is_send) { rc->cur_recv--; if (!handle_read_buffer(cl, rc->recv_buffers[rc->next_recv_buf].buf, wc[i].byte_len)) { // handle_read_buffer may stop the client continue; } try_recv_rdma_wr(cl, rc->recv_buffers[rc->next_recv_buf]); rc->next_recv_buf = (rc->next_recv_buf+1) % rc->recv_buffers.size(); } else { rc->cur_send--; uint64_t sent_size = rc->send_sizes.at(0); rc->send_sizes.erase(rc->send_sizes.begin(), rc->send_sizes.begin()+1); if (!rdma_context->odp) { rc->send_done_pos += sent_size; rc->send_out_full = false; if (rc->send_done_pos == rc->send_out_size) rc->send_done_pos = 0; assert(rc->send_done_pos < rc->send_out_size); } int send_pos = 0, send_buf_pos = 0; while (sent_size > 0) { if (sent_size >= cl->send_list.at(send_pos).iov_len) { sent_size -= cl->send_list[send_pos].iov_len; send_pos++; } else { send_buf_pos = sent_size; sent_size = 0; } } assert(rc->send_pos >= send_pos); if (rc->send_pos == send_pos) { rc->send_buf_pos -= send_buf_pos; } rc->send_pos -= send_pos; for (int i = 0; i < send_pos; i++) { if (cl->outbox[i].flags & MSGR_SENDP_FREE) { // Reply fully sent delete cl->outbox[i].op; } } if (send_pos > 0) { cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+send_pos); cl->outbox.erase(cl->outbox.begin(), cl->outbox.begin()+send_pos); } if (send_buf_pos > 0) { cl->send_list[0].iov_base = (uint8_t*)cl->send_list[0].iov_base + send_buf_pos; cl->send_list[0].iov_len -= send_buf_pos; } try_send_rdma(cl); } } } while (event_count > 0); for (auto cb: set_immediate) { cb(); } set_immediate.clear(); }