/* * Copyright (c) 2005 Ammasso, Inc. All rights reserved. * Copyright (c) 2006-2009 Open Grid Computing, Inc. All rights reserved. * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU * General Public License (GPL) Version 2, available from the file * COPYING in the main directory of this source tree, or the * OpenIB.org BSD license below: * * Redistribution and use in source and binary forms, with or * without modification, are permitted provided that the following * conditions are met: * * - Redistributions of source code must retain the above * copyright notice, this list of conditions and the following * disclaimer. * * - Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following * disclaimer in the documentation and/or other materials * provided with the distribution. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include "krping.h" #include "getopt.h" extern int krping_debug; #define DEBUG_LOG(cb, x...) if (krping_debug) krping_printf((cb)->cookie, x) #define PRINTF(cb, x...) krping_printf((cb)->cookie, x) MODULE_AUTHOR("Steve Wise"); MODULE_DESCRIPTION("RDMA ping client/server"); MODULE_LICENSE("Dual BSD/GPL"); MODULE_VERSION(krping, 1); MODULE_DEPEND(krping, linuxapi, 1, 1, 1); static __inline uint64_t get_cycles(void) { uint32_t low, high; __asm __volatile("rdtsc" : "=a" (low), "=d" (high)); return (low | ((u_int64_t)high << 32)); } typedef uint64_t cycles_t; enum mem_type { DMA = 1, FASTREG = 2, MW = 3, MR = 4 }; static const struct krping_option krping_opts[] = { {"count", OPT_INT, 'C'}, {"size", OPT_INT, 'S'}, {"addr", OPT_STRING, 'a'}, {"port", OPT_INT, 'p'}, {"verbose", OPT_NOPARAM, 'v'}, {"validate", OPT_NOPARAM, 'V'}, {"server", OPT_NOPARAM, 's'}, {"client", OPT_NOPARAM, 'c'}, {"mem_mode", OPT_STRING, 'm'}, {"server_inv", OPT_NOPARAM, 'I'}, {"wlat", OPT_NOPARAM, 'l'}, {"rlat", OPT_NOPARAM, 'L'}, {"bw", OPT_NOPARAM, 'B'}, {"duplex", OPT_NOPARAM, 'd'}, {"txdepth", OPT_INT, 'T'}, {"poll", OPT_NOPARAM, 'P'}, {"local_dma_lkey", OPT_NOPARAM, 'Z'}, {"read_inv", OPT_NOPARAM, 'R'}, {"fr", OPT_NOPARAM, 'f'}, {NULL, 0, 0} }; #define htonll(x) cpu_to_be64((x)) #define ntohll(x) cpu_to_be64((x)) static struct mutex krping_mutex; /* * List of running krping threads. */ static LIST_HEAD(krping_cbs); /* * krping "ping/pong" loop: * client sends source rkey/addr/len * server receives source rkey/add/len * server rdma reads "ping" data from source * server sends "go ahead" on rdma read completion * client sends sink rkey/addr/len * server receives sink rkey/addr/len * server rdma writes "pong" data to sink * server sends "go ahead" on rdma write completion * */ /* * These states are used to signal events between the completion handler * and the main client or server thread. * * Once CONNECTED, they cycle through RDMA_READ_ADV, RDMA_WRITE_ADV, * and RDMA_WRITE_COMPLETE for each ping. */ enum test_state { IDLE = 1, CONNECT_REQUEST, ADDR_RESOLVED, ROUTE_RESOLVED, CONNECTED, RDMA_READ_ADV, RDMA_READ_COMPLETE, RDMA_WRITE_ADV, RDMA_WRITE_COMPLETE, ERROR }; struct krping_rdma_info { uint64_t buf; uint32_t rkey; uint32_t size; }; /* * Default max buffer size for IO... */ #define RPING_BUFSIZE 128*1024 #define RPING_SQ_DEPTH 64 /* * Control block struct. */ struct krping_cb { void *cookie; int server; /* 0 iff client */ struct ib_cq *cq; struct ib_pd *pd; struct ib_qp *qp; enum mem_type mem; struct ib_mr *dma_mr; struct ib_fast_reg_page_list *page_list; int page_list_len; struct ib_send_wr fastreg_wr; struct ib_send_wr invalidate_wr; struct ib_mr *fastreg_mr; int server_invalidate; int read_inv; u8 key; struct ib_mw *mw; struct ib_mw_bind bind_attr; struct ib_recv_wr rq_wr; /* recv work request record */ struct ib_sge recv_sgl; /* recv single SGE */ struct krping_rdma_info recv_buf;/* malloc'd buffer */ u64 recv_dma_addr; DECLARE_PCI_UNMAP_ADDR(recv_mapping) struct ib_mr *recv_mr; struct ib_send_wr sq_wr; /* send work requrest record */ struct ib_sge send_sgl; struct krping_rdma_info send_buf;/* single send buf */ u64 send_dma_addr; DECLARE_PCI_UNMAP_ADDR(send_mapping) struct ib_mr *send_mr; struct ib_send_wr rdma_sq_wr; /* rdma work request record */ struct ib_sge rdma_sgl; /* rdma single SGE */ char *rdma_buf; /* used as rdma sink */ u64 rdma_dma_addr; DECLARE_PCI_UNMAP_ADDR(rdma_mapping) struct ib_mr *rdma_mr; uint32_t remote_rkey; /* remote guys RKEY */ uint64_t remote_addr; /* remote guys TO */ uint32_t remote_len; /* remote guys LEN */ char *start_buf; /* rdma read src */ u64 start_dma_addr; DECLARE_PCI_UNMAP_ADDR(start_mapping) struct ib_mr *start_mr; enum test_state state; /* used for cond/signalling */ wait_queue_head_t sem; struct krping_stats stats; uint16_t port; /* dst port in NBO */ struct in_addr addr; /* dst addr in NBO */ char *addr_str; /* dst addr string */ int verbose; /* verbose logging */ int count; /* ping count */ int size; /* ping data size */ int validate; /* validate ping data */ int wlat; /* run wlat test */ int rlat; /* run rlat test */ int bw; /* run bw test */ int duplex; /* run bw full duplex test */ int poll; /* poll or block for rlat test */ int txdepth; /* SQ depth */ int local_dma_lkey; /* use 0 for lkey */ int frtest; /* fastreg test */ /* CM stuff */ struct rdma_cm_id *cm_id; /* connection on client side,*/ /* listener on server side. */ struct rdma_cm_id *child_cm_id; /* connection on server side */ struct list_head list; }; static int krping_cma_event_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event) { int ret; struct krping_cb *cb = cma_id->context; DEBUG_LOG(cb, "cma_event type %d cma_id %p (%s)\n", event->event, cma_id, (cma_id == cb->cm_id) ? "parent" : "child"); switch (event->event) { case RDMA_CM_EVENT_ADDR_RESOLVED: cb->state = ADDR_RESOLVED; ret = rdma_resolve_route(cma_id, 2000); if (ret) { PRINTF(cb, "rdma_resolve_route error %d\n", ret); wake_up_interruptible(&cb->sem); } break; case RDMA_CM_EVENT_ROUTE_RESOLVED: cb->state = ROUTE_RESOLVED; wake_up_interruptible(&cb->sem); break; case RDMA_CM_EVENT_CONNECT_REQUEST: cb->state = CONNECT_REQUEST; cb->child_cm_id = cma_id; DEBUG_LOG(cb, "child cma %p\n", cb->child_cm_id); wake_up_interruptible(&cb->sem); break; case RDMA_CM_EVENT_ESTABLISHED: DEBUG_LOG(cb, "ESTABLISHED\n"); if (!cb->server) { cb->state = CONNECTED; } wake_up_interruptible(&cb->sem); break; case RDMA_CM_EVENT_ADDR_ERROR: case RDMA_CM_EVENT_ROUTE_ERROR: case RDMA_CM_EVENT_CONNECT_ERROR: case RDMA_CM_EVENT_UNREACHABLE: case RDMA_CM_EVENT_REJECTED: PRINTF(cb, "cma event %d, error %d\n", event->event, event->status); cb->state = ERROR; wake_up_interruptible(&cb->sem); break; case RDMA_CM_EVENT_DISCONNECTED: PRINTF(cb, "DISCONNECT EVENT...\n"); cb->state = ERROR; wake_up_interruptible(&cb->sem); break; case RDMA_CM_EVENT_DEVICE_REMOVAL: PRINTF(cb, "cma detected device removal!!!!\n"); break; default: PRINTF(cb, "oof bad type!\n"); wake_up_interruptible(&cb->sem); break; } return 0; } static int server_recv(struct krping_cb *cb, struct ib_wc *wc) { if (wc->byte_len != sizeof(cb->recv_buf)) { PRINTF(cb, "Received bogus data, size %d\n", wc->byte_len); return -1; } cb->remote_rkey = ntohl(cb->recv_buf.rkey); cb->remote_addr = ntohll(cb->recv_buf.buf); cb->remote_len = ntohl(cb->recv_buf.size); DEBUG_LOG(cb, "Received rkey %x addr %llx len %d from peer\n", cb->remote_rkey, (unsigned long long)cb->remote_addr, cb->remote_len); if (cb->state <= CONNECTED || cb->state == RDMA_WRITE_COMPLETE) cb->state = RDMA_READ_ADV; else cb->state = RDMA_WRITE_ADV; return 0; } static int client_recv(struct krping_cb *cb, struct ib_wc *wc) { if (wc->byte_len != sizeof(cb->recv_buf)) { PRINTF(cb, "Received bogus data, size %d\n", wc->byte_len); return -1; } if (cb->state == RDMA_READ_ADV) cb->state = RDMA_WRITE_ADV; else cb->state = RDMA_WRITE_COMPLETE; return 0; } static void krping_cq_event_handler(struct ib_cq *cq, void *ctx) { struct krping_cb *cb = ctx; struct ib_wc wc; struct ib_recv_wr *bad_wr; int ret; BUG_ON(cb->cq != cq); if (cb->state == ERROR) { PRINTF(cb, "cq completion in ERROR state\n"); return; } if (cb->frtest) { PRINTF(cb, "cq completion event in frtest!\n"); return; } if (!cb->wlat && !cb->rlat && !cb->bw) ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); while ((ret = ib_poll_cq(cb->cq, 1, &wc)) == 1) { if (wc.status) { if (wc.status == IB_WC_WR_FLUSH_ERR) { DEBUG_LOG(cb, "cq flushed\n"); continue; } else { PRINTF(cb, "cq completion failed with " "wr_id %Lx status %d opcode %d vender_err %x\n", wc.wr_id, wc.status, wc.opcode, wc.vendor_err); goto error; } } switch (wc.opcode) { case IB_WC_SEND: DEBUG_LOG(cb, "send completion\n"); cb->stats.send_bytes += cb->send_sgl.length; cb->stats.send_msgs++; break; case IB_WC_RDMA_WRITE: DEBUG_LOG(cb, "rdma write completion\n"); cb->stats.write_bytes += cb->rdma_sq_wr.sg_list->length; cb->stats.write_msgs++; cb->state = RDMA_WRITE_COMPLETE; wake_up_interruptible(&cb->sem); break; case IB_WC_RDMA_READ: DEBUG_LOG(cb, "rdma read completion\n"); cb->stats.read_bytes += cb->rdma_sq_wr.sg_list->length; cb->stats.read_msgs++; cb->state = RDMA_READ_COMPLETE; wake_up_interruptible(&cb->sem); break; case IB_WC_RECV: DEBUG_LOG(cb, "recv completion\n"); cb->stats.recv_bytes += sizeof(cb->recv_buf); cb->stats.recv_msgs++; if (cb->wlat || cb->rlat || cb->bw) ret = server_recv(cb, &wc); else ret = cb->server ? server_recv(cb, &wc) : client_recv(cb, &wc); if (ret) { PRINTF(cb, "recv wc error: %d\n", ret); goto error; } ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); if (ret) { PRINTF(cb, "post recv error: %d\n", ret); goto error; } wake_up_interruptible(&cb->sem); break; default: PRINTF(cb, "%s:%d Unexpected opcode %d, Shutting down\n", __func__, __LINE__, wc.opcode); goto error; } } if (ret) { PRINTF(cb, "poll error %d\n", ret); goto error; } return; error: cb->state = ERROR; wake_up_interruptible(&cb->sem); } static int krping_accept(struct krping_cb *cb) { struct rdma_conn_param conn_param; int ret; DEBUG_LOG(cb, "accepting client connection request\n"); memset(&conn_param, 0, sizeof conn_param); conn_param.responder_resources = 1; conn_param.initiator_depth = 1; ret = rdma_accept(cb->child_cm_id, &conn_param); if (ret) { PRINTF(cb, "rdma_accept error: %d\n", ret); return ret; } if (!cb->wlat && !cb->rlat && !cb->bw) { wait_event_interruptible(cb->sem, cb->state >= CONNECTED); if (cb->state == ERROR) { PRINTF(cb, "wait for CONNECTED state %d\n", cb->state); return -1; } } return 0; } static void krping_setup_wr(struct krping_cb *cb) { cb->recv_sgl.addr = cb->recv_dma_addr; cb->recv_sgl.length = sizeof cb->recv_buf; if (cb->local_dma_lkey) cb->recv_sgl.lkey = cb->qp->device->local_dma_lkey; else if (cb->mem == DMA) cb->recv_sgl.lkey = cb->dma_mr->lkey; else cb->recv_sgl.lkey = cb->recv_mr->lkey; cb->rq_wr.sg_list = &cb->recv_sgl; cb->rq_wr.num_sge = 1; cb->send_sgl.addr = cb->send_dma_addr; cb->send_sgl.length = sizeof cb->send_buf; if (cb->local_dma_lkey) cb->send_sgl.lkey = cb->qp->device->local_dma_lkey; else if (cb->mem == DMA) cb->send_sgl.lkey = cb->dma_mr->lkey; else cb->send_sgl.lkey = cb->send_mr->lkey; cb->sq_wr.opcode = IB_WR_SEND; cb->sq_wr.send_flags = IB_SEND_SIGNALED; cb->sq_wr.sg_list = &cb->send_sgl; cb->sq_wr.num_sge = 1; if (cb->server || cb->wlat || cb->rlat || cb->bw) { cb->rdma_sgl.addr = cb->rdma_dma_addr; if (cb->mem == MR) cb->rdma_sgl.lkey = cb->rdma_mr->lkey; cb->rdma_sq_wr.send_flags = IB_SEND_SIGNALED; cb->rdma_sq_wr.sg_list = &cb->rdma_sgl; cb->rdma_sq_wr.num_sge = 1; } switch(cb->mem) { case FASTREG: /* * A chain of 2 WRs, INVALDATE_MR + FAST_REG_MR. * both unsignaled. The client uses them to reregister * the rdma buffers with a new key each iteration. */ cb->fastreg_wr.opcode = IB_WR_FAST_REG_MR; cb->fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; cb->fastreg_wr.wr.fast_reg.length = cb->size; cb->fastreg_wr.wr.fast_reg.page_list = cb->page_list; cb->fastreg_wr.wr.fast_reg.page_list_len = cb->page_list_len; cb->invalidate_wr.next = &cb->fastreg_wr; cb->invalidate_wr.opcode = IB_WR_LOCAL_INV; break; case MW: cb->bind_attr.wr_id = 0xabbaabba; cb->bind_attr.send_flags = 0; /* unsignaled */ cb->bind_attr.bind_info.length = cb->size; break; default: break; } } static int krping_setup_buffers(struct krping_cb *cb) { int ret; struct ib_phys_buf buf; u64 iovbase; DEBUG_LOG(cb, "krping_setup_buffers called on cb %p\n", cb); cb->recv_dma_addr = dma_map_single(cb->pd->device->dma_device, &cb->recv_buf, sizeof(cb->recv_buf), DMA_BIDIRECTIONAL); pci_unmap_addr_set(cb, recv_mapping, cb->recv_dma_addr); cb->send_dma_addr = dma_map_single(cb->pd->device->dma_device, &cb->send_buf, sizeof(cb->send_buf), DMA_BIDIRECTIONAL); pci_unmap_addr_set(cb, send_mapping, cb->send_dma_addr); if (cb->mem == DMA) { cb->dma_mr = ib_get_dma_mr(cb->pd, IB_ACCESS_LOCAL_WRITE| IB_ACCESS_REMOTE_READ| IB_ACCESS_REMOTE_WRITE); if (IS_ERR(cb->dma_mr)) { DEBUG_LOG(cb, "reg_dmamr failed\n"); ret = PTR_ERR(cb->dma_mr); goto bail; } } else { if (!cb->local_dma_lkey) { buf.addr = cb->recv_dma_addr; buf.size = sizeof cb->recv_buf; DEBUG_LOG(cb, "recv buf dma_addr %llx size %d\n", buf.addr, (int)buf.size); iovbase = cb->recv_dma_addr; cb->recv_mr = ib_reg_phys_mr(cb->pd, &buf, 1, IB_ACCESS_LOCAL_WRITE, &iovbase); if (IS_ERR(cb->recv_mr)) { DEBUG_LOG(cb, "recv_buf reg_mr failed\n"); ret = PTR_ERR(cb->recv_mr); goto bail; } buf.addr = cb->send_dma_addr; buf.size = sizeof cb->send_buf; DEBUG_LOG(cb, "send buf dma_addr %llx size %d\n", buf.addr, (int)buf.size); iovbase = cb->send_dma_addr; cb->send_mr = ib_reg_phys_mr(cb->pd, &buf, 1, 0, &iovbase); if (IS_ERR(cb->send_mr)) { DEBUG_LOG(cb, "send_buf reg_mr failed\n"); ret = PTR_ERR(cb->send_mr); goto bail; } } } cb->rdma_buf = kmalloc(cb->size, GFP_KERNEL); if (!cb->rdma_buf) { DEBUG_LOG(cb, "rdma_buf malloc failed\n"); ret = -ENOMEM; goto bail; } cb->rdma_dma_addr = dma_map_single(cb->pd->device->dma_device, cb->rdma_buf, cb->size, DMA_BIDIRECTIONAL); pci_unmap_addr_set(cb, rdma_mapping, cb->rdma_dma_addr); if (cb->mem != DMA) { switch (cb->mem) { case FASTREG: cb->page_list_len = (((cb->size - 1) & PAGE_MASK) + PAGE_SIZE) >> PAGE_SHIFT; cb->page_list = ib_alloc_fast_reg_page_list( cb->pd->device, cb->page_list_len); if (IS_ERR(cb->page_list)) { DEBUG_LOG(cb, "recv_buf reg_mr failed\n"); ret = PTR_ERR(cb->page_list); goto bail; } cb->fastreg_mr = ib_alloc_fast_reg_mr(cb->pd, cb->page_list->max_page_list_len); if (IS_ERR(cb->fastreg_mr)) { DEBUG_LOG(cb, "recv_buf reg_mr failed\n"); ret = PTR_ERR(cb->fastreg_mr); goto bail; } DEBUG_LOG(cb, "fastreg rkey 0x%x page_list %p" " page_list_len %u\n", cb->fastreg_mr->rkey, cb->page_list, cb->page_list_len); break; case MW: cb->mw = ib_alloc_mw(cb->pd,IB_MW_TYPE_1); if (IS_ERR(cb->mw)) { DEBUG_LOG(cb, "recv_buf alloc_mw failed\n"); ret = PTR_ERR(cb->mw); goto bail; } DEBUG_LOG(cb, "mw rkey 0x%x\n", cb->mw->rkey); /*FALLTHROUGH*/ case MR: buf.addr = cb->rdma_dma_addr; buf.size = cb->size; iovbase = cb->rdma_dma_addr; cb->rdma_mr = ib_reg_phys_mr(cb->pd, &buf, 1, IB_ACCESS_REMOTE_READ| IB_ACCESS_REMOTE_WRITE, &iovbase); if (IS_ERR(cb->rdma_mr)) { DEBUG_LOG(cb, "rdma_buf reg_mr failed\n"); ret = PTR_ERR(cb->rdma_mr); goto bail; } DEBUG_LOG(cb, "rdma buf dma_addr %llx size %d mr rkey 0x%x\n", buf.addr, (int)buf.size, cb->rdma_mr->rkey); break; default: ret = -EINVAL; goto bail; break; } } if (!cb->server || cb->wlat || cb->rlat || cb->bw) { cb->start_buf = kmalloc(cb->size, GFP_KERNEL); if (!cb->start_buf) { DEBUG_LOG(cb, "start_buf malloc failed\n"); ret = -ENOMEM; goto bail; } cb->start_dma_addr = dma_map_single(cb->pd->device->dma_device, cb->start_buf, cb->size, DMA_BIDIRECTIONAL); pci_unmap_addr_set(cb, start_mapping, cb->start_dma_addr); if (cb->mem == MR || cb->mem == MW) { unsigned flags = IB_ACCESS_REMOTE_READ; if (cb->wlat || cb->rlat || cb->bw) flags |= IB_ACCESS_REMOTE_WRITE; buf.addr = cb->start_dma_addr; buf.size = cb->size; DEBUG_LOG(cb, "start buf dma_addr %llx size %d\n", buf.addr, (int)buf.size); iovbase = cb->start_dma_addr; cb->start_mr = ib_reg_phys_mr(cb->pd, &buf, 1, flags, &iovbase); if (IS_ERR(cb->start_mr)) { DEBUG_LOG(cb, "start_buf reg_mr failed\n"); ret = PTR_ERR(cb->start_mr); goto bail; } } } krping_setup_wr(cb); DEBUG_LOG(cb, "allocated & registered buffers...\n"); return 0; bail: if (cb->fastreg_mr && !IS_ERR(cb->fastreg_mr)) ib_dereg_mr(cb->fastreg_mr); if (cb->mw && !IS_ERR(cb->mw)) ib_dealloc_mw(cb->mw); if (cb->rdma_mr && !IS_ERR(cb->rdma_mr)) ib_dereg_mr(cb->rdma_mr); if (cb->page_list && !IS_ERR(cb->page_list)) ib_free_fast_reg_page_list(cb->page_list); if (cb->dma_mr && !IS_ERR(cb->dma_mr)) ib_dereg_mr(cb->dma_mr); if (cb->recv_mr && !IS_ERR(cb->recv_mr)) ib_dereg_mr(cb->recv_mr); if (cb->send_mr && !IS_ERR(cb->send_mr)) ib_dereg_mr(cb->send_mr); if (cb->rdma_buf) kfree(cb->rdma_buf); if (cb->start_buf) kfree(cb->start_buf); return ret; } static void krping_free_buffers(struct krping_cb *cb) { DEBUG_LOG(cb, "krping_free_buffers called on cb %p\n", cb); if (cb->dma_mr) ib_dereg_mr(cb->dma_mr); if (cb->send_mr) ib_dereg_mr(cb->send_mr); if (cb->recv_mr) ib_dereg_mr(cb->recv_mr); if (cb->rdma_mr) ib_dereg_mr(cb->rdma_mr); if (cb->start_mr) ib_dereg_mr(cb->start_mr); if (cb->fastreg_mr) ib_dereg_mr(cb->fastreg_mr); if (cb->mw) ib_dealloc_mw(cb->mw); dma_unmap_single(cb->pd->device->dma_device, pci_unmap_addr(cb, recv_mapping), sizeof(cb->recv_buf), DMA_BIDIRECTIONAL); dma_unmap_single(cb->pd->device->dma_device, pci_unmap_addr(cb, send_mapping), sizeof(cb->send_buf), DMA_BIDIRECTIONAL); dma_unmap_single(cb->pd->device->dma_device, pci_unmap_addr(cb, rdma_mapping), cb->size, DMA_BIDIRECTIONAL); kfree(cb->rdma_buf); if (cb->start_buf) { dma_unmap_single(cb->pd->device->dma_device, pci_unmap_addr(cb, start_mapping), cb->size, DMA_BIDIRECTIONAL); kfree(cb->start_buf); } } static int krping_create_qp(struct krping_cb *cb) { struct ib_qp_init_attr init_attr; int ret; memset(&init_attr, 0, sizeof(init_attr)); init_attr.cap.max_send_wr = cb->txdepth; init_attr.cap.max_recv_wr = 2; init_attr.cap.max_recv_sge = 1; init_attr.cap.max_send_sge = 1; init_attr.qp_type = IB_QPT_RC; init_attr.send_cq = cb->cq; init_attr.recv_cq = cb->cq; init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; if (cb->server) { ret = rdma_create_qp(cb->child_cm_id, cb->pd, &init_attr); if (!ret) cb->qp = cb->child_cm_id->qp; } else { ret = rdma_create_qp(cb->cm_id, cb->pd, &init_attr); if (!ret) cb->qp = cb->cm_id->qp; } return ret; } static void krping_free_qp(struct krping_cb *cb) { ib_destroy_qp(cb->qp); ib_destroy_cq(cb->cq); ib_dealloc_pd(cb->pd); } static int krping_setup_qp(struct krping_cb *cb, struct rdma_cm_id *cm_id) { int ret; cb->pd = ib_alloc_pd(cm_id->device); if (IS_ERR(cb->pd)) { PRINTF(cb, "ib_alloc_pd failed\n"); return PTR_ERR(cb->pd); } DEBUG_LOG(cb, "created pd %p\n", cb->pd); strlcpy(cb->stats.name, cb->pd->device->name, sizeof(cb->stats.name)); cb->cq = ib_create_cq(cm_id->device, krping_cq_event_handler, NULL, cb, cb->txdepth * 2, 0); if (IS_ERR(cb->cq)) { PRINTF(cb, "ib_create_cq failed\n"); ret = PTR_ERR(cb->cq); goto err1; } DEBUG_LOG(cb, "created cq %p\n", cb->cq); if (!cb->wlat && !cb->rlat && !cb->bw && !cb->frtest) { ret = ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); if (ret) { PRINTF(cb, "ib_create_cq failed\n"); goto err2; } } ret = krping_create_qp(cb); if (ret) { PRINTF(cb, "krping_create_qp failed: %d\n", ret); goto err2; } DEBUG_LOG(cb, "created qp %p\n", cb->qp); return 0; err2: ib_destroy_cq(cb->cq); err1: ib_dealloc_pd(cb->pd); return ret; } /* * return the (possibly rebound) rkey for the rdma buffer. * FASTREG mode: invalidate and rebind via fastreg wr. * MW mode: rebind the MW. * other modes: just return the mr rkey. */ static u32 krping_rdma_rkey(struct krping_cb *cb, u64 buf, int post_inv) { u32 rkey = 0xffffffff; u64 p; struct ib_send_wr *bad_wr; int i; int ret; switch (cb->mem) { case FASTREG: cb->invalidate_wr.ex.invalidate_rkey = cb->fastreg_mr->rkey; /* * Update the fastreg key. */ ib_update_fast_reg_key(cb->fastreg_mr, ++cb->key); cb->fastreg_wr.wr.fast_reg.rkey = cb->fastreg_mr->rkey; /* * Update the fastreg WR with new buf info. */ if (buf == (u64)cb->start_dma_addr) cb->fastreg_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_READ; else cb->fastreg_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE; cb->fastreg_wr.wr.fast_reg.iova_start = buf; p = (u64)(buf & PAGE_MASK); for (i=0; i < cb->fastreg_wr.wr.fast_reg.page_list_len; i++, p += PAGE_SIZE) { cb->page_list->page_list[i] = p; DEBUG_LOG(cb, "page_list[%d] 0x%llx\n", i, p); } DEBUG_LOG(cb, "post_inv = %d, fastreg new rkey 0x%x shift %u len %u" " iova_start %llx page_list_len %u\n", post_inv, cb->fastreg_wr.wr.fast_reg.rkey, cb->fastreg_wr.wr.fast_reg.page_shift, cb->fastreg_wr.wr.fast_reg.length, cb->fastreg_wr.wr.fast_reg.iova_start, cb->fastreg_wr.wr.fast_reg.page_list_len); if (post_inv) ret = ib_post_send(cb->qp, &cb->invalidate_wr, &bad_wr); else ret = ib_post_send(cb->qp, &cb->fastreg_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); cb->state = ERROR; } rkey = cb->fastreg_mr->rkey; break; case MW: /* * Update the MW with new buf info. */ if (buf == (u64)cb->start_dma_addr) { cb->bind_attr.bind_info.mw_access_flags = IB_ACCESS_REMOTE_READ; cb->bind_attr.bind_info.mr = cb->start_mr; } else { cb->bind_attr.bind_info.mw_access_flags = IB_ACCESS_REMOTE_WRITE; cb->bind_attr.bind_info.mr = cb->rdma_mr; } cb->bind_attr.bind_info.addr = buf; DEBUG_LOG(cb, "binding mw rkey 0x%x to buf %llx mr rkey 0x%x\n", cb->mw->rkey, buf, cb->bind_attr.bind_info.mr->rkey); ret = ib_bind_mw(cb->qp, cb->mw, &cb->bind_attr); if (ret) { PRINTF(cb, "bind mw error %d\n", ret); cb->state = ERROR; } else rkey = cb->mw->rkey; break; case MR: if (buf == (u64)cb->start_dma_addr) rkey = cb->start_mr->rkey; else rkey = cb->rdma_mr->rkey; break; case DMA: rkey = cb->dma_mr->rkey; break; default: PRINTF(cb, "%s:%d case ERROR\n", __func__, __LINE__); cb->state = ERROR; break; } return rkey; } static void krping_format_send(struct krping_cb *cb, u64 buf) { struct krping_rdma_info *info = &cb->send_buf; u32 rkey; /* * Client side will do fastreg or mw bind before * advertising the rdma buffer. Server side * sends have no data. */ if (!cb->server || cb->wlat || cb->rlat || cb->bw) { rkey = krping_rdma_rkey(cb, buf, !cb->server_invalidate); info->buf = htonll(buf); info->rkey = htonl(rkey); info->size = htonl(cb->size); DEBUG_LOG(cb, "RDMA addr %llx rkey %x len %d\n", (unsigned long long)buf, rkey, cb->size); } } static void krping_test_server(struct krping_cb *cb) { struct ib_send_wr *bad_wr, inv; int ret; while (1) { /* Wait for client's Start STAG/TO/Len */ wait_event_interruptible(cb->sem, cb->state >= RDMA_READ_ADV); if (cb->state != RDMA_READ_ADV) { PRINTF(cb, "wait for RDMA_READ_ADV state %d\n", cb->state); break; } DEBUG_LOG(cb, "server received sink adv\n"); cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; cb->rdma_sq_wr.sg_list->length = cb->remote_len; cb->rdma_sgl.lkey = krping_rdma_rkey(cb, cb->rdma_dma_addr, 1); cb->rdma_sq_wr.next = NULL; /* Issue RDMA Read. */ if (cb->read_inv) cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ_WITH_INV; else { cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ; if (cb->mem == FASTREG) { /* * Immediately follow the read with a * fenced LOCAL_INV. */ cb->rdma_sq_wr.next = &inv; memset(&inv, 0, sizeof inv); inv.opcode = IB_WR_LOCAL_INV; inv.ex.invalidate_rkey = cb->fastreg_mr->rkey; inv.send_flags = IB_SEND_FENCE; } } ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); break; } cb->rdma_sq_wr.next = NULL; DEBUG_LOG(cb, "server posted rdma read req \n"); /* Wait for read completion */ wait_event_interruptible(cb->sem, cb->state >= RDMA_READ_COMPLETE); if (cb->state != RDMA_READ_COMPLETE) { PRINTF(cb, "wait for RDMA_READ_COMPLETE state %d\n", cb->state); break; } DEBUG_LOG(cb, "server received read complete\n"); /* Display data in recv buf */ if (cb->verbose) { if (strlen(cb->rdma_buf) > 128) { char msgbuf[128]; strlcpy(msgbuf, cb->rdma_buf, sizeof(msgbuf)); PRINTF(cb, "server ping data stripped: %s\n", msgbuf); } else PRINTF(cb, "server ping data: %s\n", cb->rdma_buf); } /* Tell client to continue */ if (cb->server && cb->server_invalidate) { cb->sq_wr.ex.invalidate_rkey = cb->remote_rkey; cb->sq_wr.opcode = IB_WR_SEND_WITH_INV; DEBUG_LOG(cb, "send-w-inv rkey 0x%x\n", cb->remote_rkey); } ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); break; } DEBUG_LOG(cb, "server posted go ahead\n"); /* Wait for client's RDMA STAG/TO/Len */ wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_ADV); if (cb->state != RDMA_WRITE_ADV) { PRINTF(cb, "wait for RDMA_WRITE_ADV state %d\n", cb->state); break; } DEBUG_LOG(cb, "server received sink adv\n"); /* RDMA Write echo data */ cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; cb->rdma_sq_wr.sg_list->length = strlen(cb->rdma_buf) + 1; if (cb->local_dma_lkey) cb->rdma_sgl.lkey = cb->qp->device->local_dma_lkey; else cb->rdma_sgl.lkey = krping_rdma_rkey(cb, cb->rdma_dma_addr, 0); DEBUG_LOG(cb, "rdma write from lkey %x laddr %llx len %d\n", cb->rdma_sq_wr.sg_list->lkey, (unsigned long long)cb->rdma_sq_wr.sg_list->addr, cb->rdma_sq_wr.sg_list->length); ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); break; } /* Wait for completion */ ret = wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_COMPLETE); if (cb->state != RDMA_WRITE_COMPLETE) { PRINTF(cb, "wait for RDMA_WRITE_COMPLETE state %d\n", cb->state); break; } DEBUG_LOG(cb, "server rdma write complete \n"); cb->state = CONNECTED; /* Tell client to begin again */ if (cb->server && cb->server_invalidate) { cb->sq_wr.ex.invalidate_rkey = cb->remote_rkey; cb->sq_wr.opcode = IB_WR_SEND_WITH_INV; DEBUG_LOG(cb, "send-w-inv rkey 0x%x\n", cb->remote_rkey); } ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); break; } DEBUG_LOG(cb, "server posted go ahead\n"); } } static void rlat_test(struct krping_cb *cb) { int scnt; int iters = cb->count; struct timeval start_tv, stop_tv; int ret; struct ib_wc wc; struct ib_send_wr *bad_wr; int ne; scnt = 0; cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ; cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; cb->rdma_sq_wr.sg_list->length = cb->size; microtime(&start_tv); if (!cb->poll) { cb->state = RDMA_READ_ADV; ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); } while (scnt < iters) { cb->state = RDMA_READ_ADV; ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr); if (ret) { PRINTF(cb, "Couldn't post send: ret=%d scnt %d\n", ret, scnt); return; } do { if (!cb->poll) { wait_event_interruptible(cb->sem, cb->state != RDMA_READ_ADV); if (cb->state == RDMA_READ_COMPLETE) { ne = 1; ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); } else { ne = -1; } } else ne = ib_poll_cq(cb->cq, 1, &wc); if (cb->state == ERROR) { PRINTF(cb, "state == ERROR...bailing scnt %d\n", scnt); return; } } while (ne == 0); if (ne < 0) { PRINTF(cb, "poll CQ failed %d\n", ne); return; } if (cb->poll && wc.status != IB_WC_SUCCESS) { PRINTF(cb, "Completion wth error at %s:\n", cb->server ? "server" : "client"); PRINTF(cb, "Failed status %d: wr_id %d\n", wc.status, (int) wc.wr_id); return; } ++scnt; } microtime(&stop_tv); if (stop_tv.tv_usec < start_tv.tv_usec) { stop_tv.tv_usec += 1000000; stop_tv.tv_sec -= 1; } PRINTF(cb, "delta sec %lu delta usec %lu iter %d size %d\n", stop_tv.tv_sec - start_tv.tv_sec, stop_tv.tv_usec - start_tv.tv_usec, scnt, cb->size); } static void wlat_test(struct krping_cb *cb) { int ccnt, scnt, rcnt; int iters=cb->count; volatile char *poll_buf = (char *) cb->start_buf; char *buf = (char *)cb->rdma_buf; struct timeval start_tv, stop_tv; cycles_t *post_cycles_start, *post_cycles_stop; cycles_t *poll_cycles_start, *poll_cycles_stop; cycles_t *last_poll_cycles_start; cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0; int i; int cycle_iters = 1000; ccnt = 0; scnt = 0; rcnt = 0; post_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!post_cycles_start) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } post_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!post_cycles_stop) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!poll_cycles_start) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } poll_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!poll_cycles_stop) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } last_poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!last_poll_cycles_start) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; cb->rdma_sq_wr.sg_list->length = cb->size; if (cycle_iters > iters) cycle_iters = iters; microtime(&start_tv); while (scnt < iters || ccnt < iters || rcnt < iters) { /* Wait till buffer changes. */ if (rcnt < iters && !(scnt < 1 && !cb->server)) { ++rcnt; while (*poll_buf != (char)rcnt) { if (cb->state == ERROR) { PRINTF(cb, "state = ERROR, bailing\n"); return; } } } if (scnt < iters) { struct ib_send_wr *bad_wr; *buf = (char)scnt+1; if (scnt < cycle_iters) post_cycles_start[scnt] = get_cycles(); if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) { PRINTF(cb, "Couldn't post send: scnt=%d\n", scnt); return; } if (scnt < cycle_iters) post_cycles_stop[scnt] = get_cycles(); scnt++; } if (ccnt < iters) { struct ib_wc wc; int ne; if (ccnt < cycle_iters) poll_cycles_start[ccnt] = get_cycles(); do { if (ccnt < cycle_iters) last_poll_cycles_start[ccnt] = get_cycles(); ne = ib_poll_cq(cb->cq, 1, &wc); } while (ne == 0); if (ccnt < cycle_iters) poll_cycles_stop[ccnt] = get_cycles(); ++ccnt; if (ne < 0) { PRINTF(cb, "poll CQ failed %d\n", ne); return; } if (wc.status != IB_WC_SUCCESS) { PRINTF(cb, "Completion wth error at %s:\n", cb->server ? "server" : "client"); PRINTF(cb, "Failed status %d: wr_id %d\n", wc.status, (int) wc.wr_id); PRINTF(cb, "scnt=%d, rcnt=%d, ccnt=%d\n", scnt, rcnt, ccnt); return; } } } microtime(&stop_tv); if (stop_tv.tv_usec < start_tv.tv_usec) { stop_tv.tv_usec += 1000000; stop_tv.tv_sec -= 1; } for (i=0; i < cycle_iters; i++) { sum_post += post_cycles_stop[i] - post_cycles_start[i]; sum_poll += poll_cycles_stop[i] - poll_cycles_start[i]; sum_last_poll += poll_cycles_stop[i]-last_poll_cycles_start[i]; } PRINTF(cb, "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d" " sum_post %llu sum_poll %llu sum_last_poll %llu\n", stop_tv.tv_sec - start_tv.tv_sec, stop_tv.tv_usec - start_tv.tv_usec, scnt, cb->size, cycle_iters, (unsigned long long)sum_post, (unsigned long long)sum_poll, (unsigned long long)sum_last_poll); kfree(post_cycles_start); kfree(post_cycles_stop); kfree(poll_cycles_start); kfree(poll_cycles_stop); kfree(last_poll_cycles_start); } static void bw_test(struct krping_cb *cb) { int ccnt, scnt, rcnt; int iters=cb->count; struct timeval start_tv, stop_tv; cycles_t *post_cycles_start, *post_cycles_stop; cycles_t *poll_cycles_start, *poll_cycles_stop; cycles_t *last_poll_cycles_start; cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0; int i; int cycle_iters = 1000; ccnt = 0; scnt = 0; rcnt = 0; post_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!post_cycles_start) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } post_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!post_cycles_stop) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!poll_cycles_start) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } poll_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!poll_cycles_stop) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } last_poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); if (!last_poll_cycles_start) { PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); return; } cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; cb->rdma_sq_wr.sg_list->length = cb->size; if (cycle_iters > iters) cycle_iters = iters; microtime(&start_tv); while (scnt < iters || ccnt < iters) { while (scnt < iters && scnt - ccnt < cb->txdepth) { struct ib_send_wr *bad_wr; if (scnt < cycle_iters) post_cycles_start[scnt] = get_cycles(); if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) { PRINTF(cb, "Couldn't post send: scnt=%d\n", scnt); return; } if (scnt < cycle_iters) post_cycles_stop[scnt] = get_cycles(); ++scnt; } if (ccnt < iters) { int ne; struct ib_wc wc; if (ccnt < cycle_iters) poll_cycles_start[ccnt] = get_cycles(); do { if (ccnt < cycle_iters) last_poll_cycles_start[ccnt] = get_cycles(); ne = ib_poll_cq(cb->cq, 1, &wc); } while (ne == 0); if (ccnt < cycle_iters) poll_cycles_stop[ccnt] = get_cycles(); ccnt += 1; if (ne < 0) { PRINTF(cb, "poll CQ failed %d\n", ne); return; } if (wc.status != IB_WC_SUCCESS) { PRINTF(cb, "Completion wth error at %s:\n", cb->server ? "server" : "client"); PRINTF(cb, "Failed status %d: wr_id %d\n", wc.status, (int) wc.wr_id); return; } } } microtime(&stop_tv); if (stop_tv.tv_usec < start_tv.tv_usec) { stop_tv.tv_usec += 1000000; stop_tv.tv_sec -= 1; } for (i=0; i < cycle_iters; i++) { sum_post += post_cycles_stop[i] - post_cycles_start[i]; sum_poll += poll_cycles_stop[i] - poll_cycles_start[i]; sum_last_poll += poll_cycles_stop[i]-last_poll_cycles_start[i]; } PRINTF(cb, "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d" " sum_post %llu sum_poll %llu sum_last_poll %llu\n", stop_tv.tv_sec - start_tv.tv_sec, stop_tv.tv_usec - start_tv.tv_usec, scnt, cb->size, cycle_iters, (unsigned long long)sum_post, (unsigned long long)sum_poll, (unsigned long long)sum_last_poll); kfree(post_cycles_start); kfree(post_cycles_stop); kfree(poll_cycles_start); kfree(poll_cycles_stop); kfree(last_poll_cycles_start); } static void krping_rlat_test_server(struct krping_cb *cb) { struct ib_send_wr *bad_wr; struct ib_wc wc; int ret; /* Spin waiting for client's Start STAG/TO/Len */ while (cb->state < RDMA_READ_ADV) { krping_cq_event_handler(cb->cq, cb); } /* Send STAG/TO/Len to client */ krping_format_send(cb, cb->start_dma_addr); ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); return; } /* Spin waiting for send completion */ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); if (ret < 0) { PRINTF(cb, "poll error %d\n", ret); return; } if (wc.status) { PRINTF(cb, "send completiong error %d\n", wc.status); return; } wait_event_interruptible(cb->sem, cb->state == ERROR); } static void krping_wlat_test_server(struct krping_cb *cb) { struct ib_send_wr *bad_wr; struct ib_wc wc; int ret; /* Spin waiting for client's Start STAG/TO/Len */ while (cb->state < RDMA_READ_ADV) { krping_cq_event_handler(cb->cq, cb); } /* Send STAG/TO/Len to client */ krping_format_send(cb, cb->start_dma_addr); ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); return; } /* Spin waiting for send completion */ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); if (ret < 0) { PRINTF(cb, "poll error %d\n", ret); return; } if (wc.status) { PRINTF(cb, "send completiong error %d\n", wc.status); return; } wlat_test(cb); wait_event_interruptible(cb->sem, cb->state == ERROR); } static void krping_bw_test_server(struct krping_cb *cb) { struct ib_send_wr *bad_wr; struct ib_wc wc; int ret; /* Spin waiting for client's Start STAG/TO/Len */ while (cb->state < RDMA_READ_ADV) { krping_cq_event_handler(cb->cq, cb); } /* Send STAG/TO/Len to client */ krping_format_send(cb, cb->start_dma_addr); ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); return; } /* Spin waiting for send completion */ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); if (ret < 0) { PRINTF(cb, "poll error %d\n", ret); return; } if (wc.status) { PRINTF(cb, "send completiong error %d\n", wc.status); return; } if (cb->duplex) bw_test(cb); wait_event_interruptible(cb->sem, cb->state == ERROR); } static int fastreg_supported(struct krping_cb *cb) { struct ib_device *dev = cb->child_cm_id->device; struct ib_device_attr attr; int ret; ret = ib_query_device(dev, &attr); if (ret) { PRINTF(cb, "ib_query_device failed ret %d\n", ret); return 0; } if (!(attr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)) { PRINTF(cb, "Fastreg not supported - device_cap_flags 0x%x\n", attr.device_cap_flags); return 0; } DEBUG_LOG(cb, "Fastreg supported - device_cap_flags 0x%x\n", attr.device_cap_flags); return 1; } static int krping_bind_server(struct krping_cb *cb) { struct sockaddr_in sin; int ret; memset(&sin, 0, sizeof(sin)); sin.sin_len = sizeof sin; sin.sin_family = AF_INET; sin.sin_addr.s_addr = cb->addr.s_addr; sin.sin_port = cb->port; ret = rdma_bind_addr(cb->cm_id, (struct sockaddr *) &sin); if (ret) { PRINTF(cb, "rdma_bind_addr error %d\n", ret); return ret; } DEBUG_LOG(cb, "rdma_bind_addr successful\n"); DEBUG_LOG(cb, "rdma_listen\n"); ret = rdma_listen(cb->cm_id, 3); if (ret) { PRINTF(cb, "rdma_listen failed: %d\n", ret); return ret; } wait_event_interruptible(cb->sem, cb->state >= CONNECT_REQUEST); if (cb->state != CONNECT_REQUEST) { PRINTF(cb, "wait for CONNECT_REQUEST state %d\n", cb->state); return -1; } if (cb->mem == FASTREG && !fastreg_supported(cb)) return -EINVAL; return 0; } static void krping_run_server(struct krping_cb *cb) { struct ib_recv_wr *bad_wr; int ret; ret = krping_bind_server(cb); if (ret) return; ret = krping_setup_qp(cb, cb->child_cm_id); if (ret) { PRINTF(cb, "setup_qp failed: %d\n", ret); goto err0; } ret = krping_setup_buffers(cb); if (ret) { PRINTF(cb, "krping_setup_buffers failed: %d\n", ret); goto err1; } ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); if (ret) { PRINTF(cb, "ib_post_recv failed: %d\n", ret); goto err2; } ret = krping_accept(cb); if (ret) { PRINTF(cb, "connect error %d\n", ret); goto err2; } if (cb->wlat) krping_wlat_test_server(cb); else if (cb->rlat) krping_rlat_test_server(cb); else if (cb->bw) krping_bw_test_server(cb); else krping_test_server(cb); rdma_disconnect(cb->child_cm_id); err2: krping_free_buffers(cb); err1: krping_free_qp(cb); err0: rdma_destroy_id(cb->child_cm_id); } static void krping_test_client(struct krping_cb *cb) { int ping, start, cc, i, ret; struct ib_send_wr *bad_wr; unsigned char c; start = 65; for (ping = 0; !cb->count || ping < cb->count; ping++) { cb->state = RDMA_READ_ADV; /* Put some ascii text in the buffer. */ cc = sprintf(cb->start_buf, "rdma-ping-%d: ", ping); for (i = cc, c = start; i < cb->size; i++) { cb->start_buf[i] = c; c++; if (c > 122) c = 65; } start++; if (start > 122) start = 65; cb->start_buf[cb->size - 1] = 0; krping_format_send(cb, cb->start_dma_addr); if (cb->state == ERROR) { PRINTF(cb, "krping_format_send failed\n"); break; } ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); break; } /* Wait for server to ACK */ wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_ADV); if (cb->state != RDMA_WRITE_ADV) { PRINTF(cb, "wait for RDMA_WRITE_ADV state %d\n", cb->state); break; } krping_format_send(cb, cb->rdma_dma_addr); ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); break; } /* Wait for the server to say the RDMA Write is complete. */ wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_COMPLETE); if (cb->state != RDMA_WRITE_COMPLETE) { PRINTF(cb, "wait for RDMA_WRITE_COMPLETE state %d\n", cb->state); break; } if (cb->validate) if (memcmp(cb->start_buf, cb->rdma_buf, cb->size)) { PRINTF(cb, "data mismatch!\n"); break; } if (cb->verbose) { if (strlen(cb->rdma_buf) > 128) { char msgbuf[128]; strlcpy(msgbuf, cb->rdma_buf, sizeof(msgbuf)); PRINTF(cb, "ping data stripped: %s\n", msgbuf); } else PRINTF(cb, "ping data: %s\n", cb->rdma_buf); } #ifdef SLOW_KRPING wait_event_interruptible_timeout(cb->sem, cb->state == ERROR, HZ); #endif } } static void krping_rlat_test_client(struct krping_cb *cb) { struct ib_send_wr *bad_wr; struct ib_wc wc; int ret; cb->state = RDMA_READ_ADV; /* Send STAG/TO/Len to client */ krping_format_send(cb, cb->start_dma_addr); if (cb->state == ERROR) { PRINTF(cb, "krping_format_send failed\n"); return; } ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); return; } /* Spin waiting for send completion */ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); if (ret < 0) { PRINTF(cb, "poll error %d\n", ret); return; } if (wc.status) { PRINTF(cb, "send completion error %d\n", wc.status); return; } /* Spin waiting for server's Start STAG/TO/Len */ while (cb->state < RDMA_WRITE_ADV) { krping_cq_event_handler(cb->cq, cb); } #if 0 { int i; struct timeval start, stop; time_t sec; suseconds_t usec; unsigned long long elapsed; struct ib_wc wc; struct ib_send_wr *bad_wr; int ne; cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; cb->rdma_sq_wr.sg_list->length = 0; cb->rdma_sq_wr.num_sge = 0; microtime(&start); for (i=0; i < 100000; i++) { if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) { PRINTF(cb, "Couldn't post send\n"); return; } do { ne = ib_poll_cq(cb->cq, 1, &wc); } while (ne == 0); if (ne < 0) { PRINTF(cb, "poll CQ failed %d\n", ne); return; } if (wc.status != IB_WC_SUCCESS) { PRINTF(cb, "Completion wth error at %s:\n", cb->server ? "server" : "client"); PRINTF(cb, "Failed status %d: wr_id %d\n", wc.status, (int) wc.wr_id); return; } } microtime(&stop); if (stop.tv_usec < start.tv_usec) { stop.tv_usec += 1000000; stop.tv_sec -= 1; } sec = stop.tv_sec - start.tv_sec; usec = stop.tv_usec - start.tv_usec; elapsed = sec * 1000000 + usec; PRINTF(cb, "0B-write-lat iters 100000 usec %llu\n", elapsed); } #endif rlat_test(cb); } static void krping_wlat_test_client(struct krping_cb *cb) { struct ib_send_wr *bad_wr; struct ib_wc wc; int ret; cb->state = RDMA_READ_ADV; /* Send STAG/TO/Len to client */ krping_format_send(cb, cb->start_dma_addr); if (cb->state == ERROR) { PRINTF(cb, "krping_format_send failed\n"); return; } ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); return; } /* Spin waiting for send completion */ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); if (ret < 0) { PRINTF(cb, "poll error %d\n", ret); return; } if (wc.status) { PRINTF(cb, "send completion error %d\n", wc.status); return; } /* Spin waiting for server's Start STAG/TO/Len */ while (cb->state < RDMA_WRITE_ADV) { krping_cq_event_handler(cb->cq, cb); } wlat_test(cb); } static void krping_bw_test_client(struct krping_cb *cb) { struct ib_send_wr *bad_wr; struct ib_wc wc; int ret; cb->state = RDMA_READ_ADV; /* Send STAG/TO/Len to client */ krping_format_send(cb, cb->start_dma_addr); if (cb->state == ERROR) { PRINTF(cb, "krping_format_send failed\n"); return; } ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); if (ret) { PRINTF(cb, "post send error %d\n", ret); return; } /* Spin waiting for send completion */ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); if (ret < 0) { PRINTF(cb, "poll error %d\n", ret); return; } if (wc.status) { PRINTF(cb, "send completion error %d\n", wc.status); return; } /* Spin waiting for server's Start STAG/TO/Len */ while (cb->state < RDMA_WRITE_ADV) { krping_cq_event_handler(cb->cq, cb); } bw_test(cb); } static void krping_fr_test(struct krping_cb *cb) { struct ib_fast_reg_page_list *pl; struct ib_send_wr fr, inv, *bad; struct ib_wc wc; u8 key = 0; struct ib_mr *mr; int i; int ret; int size = cb->size; int plen = (((size - 1) & PAGE_MASK) + PAGE_SIZE) >> PAGE_SHIFT; time_t start; int count = 0; int scnt = 0; pl = ib_alloc_fast_reg_page_list(cb->qp->device, plen); if (IS_ERR(pl)) { PRINTF(cb, "ib_alloc_fast_reg_page_list failed %ld\n", PTR_ERR(pl)); return; } mr = ib_alloc_fast_reg_mr(cb->pd, plen); if (IS_ERR(mr)) { PRINTF(cb, "ib_alloc_fast_reg_mr failed %ld\n", PTR_ERR(pl)); goto err1; } for (i=0; ipage_list[i] = 0xcafebabe | i; memset(&fr, 0, sizeof fr); fr.opcode = IB_WR_FAST_REG_MR; fr.wr.fast_reg.page_shift = PAGE_SHIFT; fr.wr.fast_reg.length = size; fr.wr.fast_reg.page_list = pl; fr.wr.fast_reg.page_list_len = plen; fr.wr.fast_reg.iova_start = 0; fr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE; fr.next = &inv; memset(&inv, 0, sizeof inv); inv.opcode = IB_WR_LOCAL_INV; inv.send_flags = IB_SEND_SIGNALED; DEBUG_LOG(cb, "fr_test: stag index 0x%x plen %u size %u depth %u\n", mr->rkey >> 8, plen, cb->size, cb->txdepth); start = time_uptime; while (1) { if ((time_uptime - start) >= 9) { DEBUG_LOG(cb, "fr_test: pausing 1 second! count %u latest size %u plen %u\n", count, size, plen); wait_event_interruptible(cb->sem, cb->state == ERROR); if (cb->state == ERROR) break; start = time_uptime; } while (scnt < (cb->txdepth>>1)) { ib_update_fast_reg_key(mr, ++key); fr.wr.fast_reg.rkey = mr->rkey; inv.ex.invalidate_rkey = mr->rkey; size = arc4random() % cb->size; if (size == 0) size = cb->size; plen = (((size - 1) & PAGE_MASK) + PAGE_SIZE) >> PAGE_SHIFT; fr.wr.fast_reg.length = size; fr.wr.fast_reg.page_list_len = plen; ret = ib_post_send(cb->qp, &fr, &bad); if (ret) { PRINTF(cb, "ib_post_send failed %d\n", ret); goto err2; } scnt++; } do { ret = ib_poll_cq(cb->cq, 1, &wc); if (ret < 0) { PRINTF(cb, "ib_poll_cq failed %d\n", ret); goto err2; } if (ret == 1) { if (wc.status) { PRINTF(cb, "completion error %u\n", wc.status); goto err2; } count++; scnt--; } else if (krping_sigpending()) { PRINTF(cb, "signal!\n"); goto err2; } } while (ret == 1); } err2: #if 0 DEBUG_LOG(cb, "sleeping 1 second\n"); wait_event_interruptible_timeout(cb->sem, cb->state == ERROR, HZ); #endif DEBUG_LOG(cb, "draining the cq...\n"); do { ret = ib_poll_cq(cb->cq, 1, &wc); if (ret < 0) { PRINTF(cb, "ib_poll_cq failed %d\n", ret); break; } if (ret == 1) { if (wc.status) { PRINTF(cb, "completion error %u opcode %u\n", wc.status, wc.opcode); } } } while (ret == 1); DEBUG_LOG(cb, "fr_test: done!\n"); ib_dereg_mr(mr); err1: ib_free_fast_reg_page_list(pl); } static int krping_connect_client(struct krping_cb *cb) { struct rdma_conn_param conn_param; int ret; memset(&conn_param, 0, sizeof conn_param); conn_param.responder_resources = 1; conn_param.initiator_depth = 1; conn_param.retry_count = 10; ret = rdma_connect(cb->cm_id, &conn_param); if (ret) { PRINTF(cb, "rdma_connect error %d\n", ret); return ret; } wait_event_interruptible(cb->sem, cb->state >= CONNECTED); if (cb->state == ERROR) { PRINTF(cb, "wait for CONNECTED state %d\n", cb->state); return -1; } DEBUG_LOG(cb, "rdma_connect successful\n"); return 0; } static int krping_bind_client(struct krping_cb *cb) { struct sockaddr_in sin; int ret; memset(&sin, 0, sizeof(sin)); sin.sin_len = sizeof sin; sin.sin_family = AF_INET; sin.sin_addr.s_addr = cb->addr.s_addr; sin.sin_port = cb->port; ret = rdma_resolve_addr(cb->cm_id, NULL, (struct sockaddr *) &sin, 2000); if (ret) { PRINTF(cb, "rdma_resolve_addr error %d\n", ret); return ret; } wait_event_interruptible(cb->sem, cb->state >= ROUTE_RESOLVED); if (cb->state != ROUTE_RESOLVED) { PRINTF(cb, "addr/route resolution did not resolve: state %d\n", cb->state); return -EINTR; } if (cb->mem == FASTREG && !fastreg_supported(cb)) return -EINVAL; DEBUG_LOG(cb, "rdma_resolve_addr - rdma_resolve_route successful\n"); return 0; } static void krping_run_client(struct krping_cb *cb) { struct ib_recv_wr *bad_wr; int ret; ret = krping_bind_client(cb); if (ret) return; ret = krping_setup_qp(cb, cb->cm_id); if (ret) { PRINTF(cb, "setup_qp failed: %d\n", ret); return; } ret = krping_setup_buffers(cb); if (ret) { PRINTF(cb, "krping_setup_buffers failed: %d\n", ret); goto err1; } ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); if (ret) { PRINTF(cb, "ib_post_recv failed: %d\n", ret); goto err2; } ret = krping_connect_client(cb); if (ret) { PRINTF(cb, "connect error %d\n", ret); goto err2; } if (cb->wlat) krping_wlat_test_client(cb); else if (cb->rlat) krping_rlat_test_client(cb); else if (cb->bw) krping_bw_test_client(cb); else if (cb->frtest) krping_fr_test(cb); else krping_test_client(cb); rdma_disconnect(cb->cm_id); err2: krping_free_buffers(cb); err1: krping_free_qp(cb); } int krping_doit(char *cmd, void *cookie) { struct krping_cb *cb; int op; int ret = 0; char *optarg; unsigned long optint; cb = kzalloc(sizeof(*cb), GFP_KERNEL); if (!cb) return -ENOMEM; mutex_lock(&krping_mutex); list_add_tail(&cb->list, &krping_cbs); mutex_unlock(&krping_mutex); cb->cookie = cookie; cb->server = -1; cb->state = IDLE; cb->size = 64; cb->txdepth = RPING_SQ_DEPTH; cb->mem = DMA; init_waitqueue_head(&cb->sem); while ((op = krping_getopt("krping", &cmd, krping_opts, NULL, &optarg, &optint)) != 0) { switch (op) { case 'a': cb->addr_str = optarg; DEBUG_LOG(cb, "ipaddr (%s)\n", optarg); if (!inet_aton(optarg, &cb->addr)) { PRINTF(cb, "bad addr string %s\n", optarg); ret = EINVAL; } break; case 'p': cb->port = htons(optint); DEBUG_LOG(cb, "port %d\n", (int)optint); break; case 'P': cb->poll = 1; DEBUG_LOG(cb, "server\n"); break; case 's': cb->server = 1; DEBUG_LOG(cb, "server\n"); break; case 'c': cb->server = 0; DEBUG_LOG(cb, "client\n"); break; case 'S': cb->size = optint; if ((cb->size < 1) || (cb->size > RPING_BUFSIZE)) { PRINTF(cb, "Invalid size %d " "(valid range is 1 to %d)\n", cb->size, RPING_BUFSIZE); ret = EINVAL; } else DEBUG_LOG(cb, "size %d\n", (int)optint); break; case 'C': cb->count = optint; if (cb->count < 0) { PRINTF(cb, "Invalid count %d\n", cb->count); ret = EINVAL; } else DEBUG_LOG(cb, "count %d\n", (int) cb->count); break; case 'v': cb->verbose++; DEBUG_LOG(cb, "verbose\n"); break; case 'V': cb->validate++; DEBUG_LOG(cb, "validate data\n"); break; case 'l': cb->wlat++; break; case 'L': cb->rlat++; break; case 'B': cb->bw++; break; case 'd': cb->duplex++; break; case 'm': if (!strncmp(optarg, "dma", 3)) cb->mem = DMA; else if (!strncmp(optarg, "fastreg", 7)) cb->mem = FASTREG; else if (!strncmp(optarg, "mw", 2)) cb->mem = MW; else if (!strncmp(optarg, "mr", 2)) cb->mem = MR; else { PRINTF(cb, "unknown mem mode %s. " "Must be dma, fastreg, mw, or mr\n", optarg); ret = -EINVAL; break; } break; case 'I': cb->server_invalidate = 1; break; case 'T': cb->txdepth = optint; DEBUG_LOG(cb, "txdepth %d\n", (int) cb->txdepth); break; case 'Z': cb->local_dma_lkey = 1; DEBUG_LOG(cb, "using local dma lkey\n"); break; case 'R': cb->read_inv = 1; DEBUG_LOG(cb, "using read-with-inv\n"); break; case 'f': cb->frtest = 1; DEBUG_LOG(cb, "fast-reg test!\n"); break; default: PRINTF(cb, "unknown opt %s\n", optarg); ret = -EINVAL; break; } } if (ret) goto out; if (cb->server == -1) { PRINTF(cb, "must be either client or server\n"); ret = -EINVAL; goto out; } if (cb->server && cb->frtest) { PRINTF(cb, "must be client to run frtest\n"); ret = -EINVAL; goto out; } if ((cb->frtest + cb->bw + cb->rlat + cb->wlat) > 1) { PRINTF(cb, "Pick only one test: fr, bw, rlat, wlat\n"); ret = -EINVAL; goto out; } if (cb->server_invalidate && cb->mem != FASTREG) { PRINTF(cb, "server_invalidate only valid with fastreg mem_mode\n"); ret = -EINVAL; goto out; } if (cb->read_inv && cb->mem != FASTREG) { PRINTF(cb, "read_inv only valid with fastreg mem_mode\n"); ret = -EINVAL; goto out; } if (cb->mem != MR && (cb->wlat || cb->rlat || cb->bw)) { PRINTF(cb, "wlat, rlat, and bw tests only support mem_mode MR\n"); ret = -EINVAL; goto out; } cb->cm_id = rdma_create_id(krping_cma_event_handler, cb, RDMA_PS_TCP, IB_QPT_RC); if (IS_ERR(cb->cm_id)) { ret = PTR_ERR(cb->cm_id); PRINTF(cb, "rdma_create_id error %d\n", ret); goto out; } DEBUG_LOG(cb, "created cm_id %p\n", cb->cm_id); if (cb->server) krping_run_server(cb); else krping_run_client(cb); DEBUG_LOG(cb, "destroy cm_id %p\n", cb->cm_id); rdma_destroy_id(cb->cm_id); out: mutex_lock(&krping_mutex); list_del(&cb->list); mutex_unlock(&krping_mutex); kfree(cb); return ret; } void krping_walk_cb_list(void (*f)(struct krping_stats *, void *), void *arg) { struct krping_cb *cb; mutex_lock(&krping_mutex); list_for_each_entry(cb, &krping_cbs, list) (*f)(cb->pd ? &cb->stats : NULL, arg); mutex_unlock(&krping_mutex); } void krping_init(void) { mutex_init(&krping_mutex); }