1/* 2 * Copyright (c) 2005 Ammasso, Inc. All rights reserved. 3 * Copyright (c) 2006-2009 Open Grid Computing, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34#include <sys/cdefs.h> 35__FBSDID("$FreeBSD: stable/10/sys/contrib/rdma/krping/krping.c 325602 2017-11-09 17:51:55Z hselasky $"); 36 37#include <linux/module.h> 38#include <linux/moduleparam.h> 39#include <linux/slab.h> 40#include <linux/err.h> 41#include <linux/string.h> 42#include <linux/list.h> 43#include <linux/in.h> 44#include <linux/device.h> 45#include <linux/pci.h> 46#include <linux/sched.h> 47 48#include <asm/atomic.h> 49 50#include <rdma/ib_verbs.h> 51#include <rdma/rdma_cm.h> 52 53#include "krping.h" 54#include "getopt.h" 55 56extern int krping_debug; 57#define DEBUG_LOG(cb, x...) if (krping_debug) krping_printf((cb)->cookie, x) 58#define PRINTF(cb, x...) krping_printf((cb)->cookie, x) 59 60MODULE_AUTHOR("Steve Wise"); 61MODULE_DESCRIPTION("RDMA ping client/server"); 62MODULE_LICENSE("Dual BSD/GPL"); 63 64static __inline uint64_t 65get_cycles(void) 66{ 67 uint32_t low, high; 68 __asm __volatile("rdtsc" : "=a" (low), "=d" (high)); 69 return (low | ((u_int64_t)high << 32)); 70} 71 72typedef uint64_t cycles_t; 73 74enum mem_type { 75 DMA = 1, 76 FASTREG = 2, 77 MW = 3, 78 MR = 4 79}; 80 81static const struct krping_option krping_opts[] = { 82 {"count", OPT_INT, 'C'}, 83 {"size", OPT_INT, 'S'}, 84 {"addr", OPT_STRING, 'a'}, 85 {"addr6", OPT_STRING, 'A'}, 86 {"port", OPT_INT, 'p'}, 87 {"verbose", OPT_NOPARAM, 'v'}, 88 {"validate", OPT_NOPARAM, 'V'}, 89 {"server", OPT_NOPARAM, 's'}, 90 {"client", OPT_NOPARAM, 'c'}, 91 {"mem_mode", OPT_STRING, 'm'}, 92 {"server_inv", OPT_NOPARAM, 'I'}, 93 {"wlat", OPT_NOPARAM, 'l'}, 94 {"rlat", OPT_NOPARAM, 'L'}, 95 {"bw", OPT_NOPARAM, 'B'}, 96 {"duplex", OPT_NOPARAM, 'd'}, 97 {"txdepth", OPT_INT, 'T'}, 98 {"poll", OPT_NOPARAM, 'P'}, 99 {"local_dma_lkey", OPT_NOPARAM, 'Z'}, 100 {"read_inv", OPT_NOPARAM, 'R'}, 101 {"fr", OPT_NOPARAM, 'f'}, 102 {NULL, 0, 0} 103}; 104 105#define htonll(x) cpu_to_be64((x)) 106#define ntohll(x) cpu_to_be64((x)) 107 108static struct mutex krping_mutex; 109 110/* 111 * List of running krping threads. 112 */ 113static LIST_HEAD(krping_cbs); 114 115/* 116 * krping "ping/pong" loop: 117 * client sends source rkey/addr/len 118 * server receives source rkey/add/len 119 * server rdma reads "ping" data from source 120 * server sends "go ahead" on rdma read completion 121 * client sends sink rkey/addr/len 122 * server receives sink rkey/addr/len 123 * server rdma writes "pong" data to sink 124 * server sends "go ahead" on rdma write completion 125 * <repeat loop> 126 */ 127 128/* 129 * These states are used to signal events between the completion handler 130 * and the main client or server thread. 131 * 132 * Once CONNECTED, they cycle through RDMA_READ_ADV, RDMA_WRITE_ADV, 133 * and RDMA_WRITE_COMPLETE for each ping. 134 */ 135enum test_state { 136 IDLE = 1, 137 CONNECT_REQUEST, 138 ADDR_RESOLVED, 139 ROUTE_RESOLVED, 140 CONNECTED, 141 RDMA_READ_ADV, 142 RDMA_READ_COMPLETE, 143 RDMA_WRITE_ADV, 144 RDMA_WRITE_COMPLETE, 145 ERROR 146}; 147 148struct krping_rdma_info { 149 uint64_t buf; 150 uint32_t rkey; 151 uint32_t size; 152}; 153 154/* 155 * Default max buffer size for IO... 156 */ 157#define RPING_BUFSIZE 128*1024 158#define RPING_SQ_DEPTH 64 159 160/* 161 * Control block struct. 162 */ 163struct krping_cb { 164 void *cookie; 165 int server; /* 0 iff client */ 166 struct ib_cq *cq; 167 struct ib_pd *pd; 168 struct ib_qp *qp; 169 170 enum mem_type mem; 171 struct ib_mr *dma_mr; 172 173 struct ib_fast_reg_page_list *page_list; 174 int page_list_len; 175 struct ib_send_wr fastreg_wr; 176 struct ib_send_wr invalidate_wr; 177 struct ib_mr *fastreg_mr; 178 int server_invalidate; 179 int read_inv; 180 u8 key; 181 182 struct ib_mw *mw; 183 struct ib_mw_bind bind_attr; 184 185 struct ib_recv_wr rq_wr; /* recv work request record */ 186 struct ib_sge recv_sgl; /* recv single SGE */ 187 struct krping_rdma_info recv_buf;/* malloc'd buffer */ 188 u64 recv_dma_addr; 189 DECLARE_PCI_UNMAP_ADDR(recv_mapping) 190 struct ib_mr *recv_mr; 191 192 struct ib_send_wr sq_wr; /* send work requrest record */ 193 struct ib_sge send_sgl; 194 struct krping_rdma_info send_buf;/* single send buf */ 195 u64 send_dma_addr; 196 DECLARE_PCI_UNMAP_ADDR(send_mapping) 197 struct ib_mr *send_mr; 198 199 struct ib_send_wr rdma_sq_wr; /* rdma work request record */ 200 struct ib_sge rdma_sgl; /* rdma single SGE */ 201 char *rdma_buf; /* used as rdma sink */ 202 u64 rdma_dma_addr; 203 DECLARE_PCI_UNMAP_ADDR(rdma_mapping) 204 struct ib_mr *rdma_mr; 205 206 uint32_t remote_rkey; /* remote guys RKEY */ 207 uint64_t remote_addr; /* remote guys TO */ 208 uint32_t remote_len; /* remote guys LEN */ 209 210 char *start_buf; /* rdma read src */ 211 u64 start_dma_addr; 212 DECLARE_PCI_UNMAP_ADDR(start_mapping) 213 struct ib_mr *start_mr; 214 215 enum test_state state; /* used for cond/signalling */ 216 wait_queue_head_t sem; 217 struct krping_stats stats; 218 219 uint16_t port; /* dst port in NBO */ 220 union { 221 struct in_addr v4; 222 struct in6_addr v6; 223 } addr; /* dst addr in NBO */ 224 int addr_type; /* AF_INET or AF_INET6 */ 225 char *addr_str; /* dst addr string */ 226 int verbose; /* verbose logging */ 227 int count; /* ping count */ 228 int size; /* ping data size */ 229 int validate; /* validate ping data */ 230 int wlat; /* run wlat test */ 231 int rlat; /* run rlat test */ 232 int bw; /* run bw test */ 233 int duplex; /* run bw full duplex test */ 234 int poll; /* poll or block for rlat test */ 235 int txdepth; /* SQ depth */ 236 int local_dma_lkey; /* use 0 for lkey */ 237 int frtest; /* fastreg test */ 238 239 /* CM stuff */ 240 struct rdma_cm_id *cm_id; /* connection on client side,*/ 241 /* listener on server side. */ 242 struct rdma_cm_id *child_cm_id; /* connection on server side */ 243 struct list_head list; 244}; 245 246static int krping_cma_event_handler(struct rdma_cm_id *cma_id, 247 struct rdma_cm_event *event) 248{ 249 int ret; 250 struct krping_cb *cb = cma_id->context; 251 252 DEBUG_LOG(cb, "cma_event type %d cma_id %p (%s)\n", event->event, 253 cma_id, (cma_id == cb->cm_id) ? "parent" : "child"); 254 255 switch (event->event) { 256 case RDMA_CM_EVENT_ADDR_RESOLVED: 257 cb->state = ADDR_RESOLVED; 258 ret = rdma_resolve_route(cma_id, 2000); 259 if (ret) { 260 PRINTF(cb, "rdma_resolve_route error %d\n", ret); 261 wake_up_interruptible(&cb->sem); 262 } 263 break; 264 265 case RDMA_CM_EVENT_ROUTE_RESOLVED: 266 cb->state = ROUTE_RESOLVED; 267 cb->child_cm_id = cma_id; 268 wake_up_interruptible(&cb->sem); 269 break; 270 271 case RDMA_CM_EVENT_CONNECT_REQUEST: 272 cb->state = CONNECT_REQUEST; 273 cb->child_cm_id = cma_id; 274 DEBUG_LOG(cb, "child cma %p\n", cb->child_cm_id); 275 wake_up_interruptible(&cb->sem); 276 break; 277 278 case RDMA_CM_EVENT_ESTABLISHED: 279 DEBUG_LOG(cb, "ESTABLISHED\n"); 280 if (!cb->server) { 281 cb->state = CONNECTED; 282 } 283 wake_up_interruptible(&cb->sem); 284 break; 285 286 case RDMA_CM_EVENT_ADDR_ERROR: 287 case RDMA_CM_EVENT_ROUTE_ERROR: 288 case RDMA_CM_EVENT_CONNECT_ERROR: 289 case RDMA_CM_EVENT_UNREACHABLE: 290 case RDMA_CM_EVENT_REJECTED: 291 PRINTF(cb, "cma event %d, error %d\n", event->event, 292 event->status); 293 cb->state = ERROR; 294 wake_up_interruptible(&cb->sem); 295 break; 296 297 case RDMA_CM_EVENT_DISCONNECTED: 298 PRINTF(cb, "DISCONNECT EVENT...\n"); 299 cb->state = ERROR; 300 wake_up_interruptible(&cb->sem); 301 break; 302 303 case RDMA_CM_EVENT_DEVICE_REMOVAL: 304 PRINTF(cb, "cma detected device removal!!!!\n"); 305 break; 306 307 default: 308 PRINTF(cb, "oof bad type!\n"); 309 wake_up_interruptible(&cb->sem); 310 break; 311 } 312 return 0; 313} 314 315static int server_recv(struct krping_cb *cb, struct ib_wc *wc) 316{ 317 if (wc->byte_len != sizeof(cb->recv_buf)) { 318 PRINTF(cb, "Received bogus data, size %d\n", 319 wc->byte_len); 320 return -1; 321 } 322 323 cb->remote_rkey = ntohl(cb->recv_buf.rkey); 324 cb->remote_addr = ntohll(cb->recv_buf.buf); 325 cb->remote_len = ntohl(cb->recv_buf.size); 326 DEBUG_LOG(cb, "Received rkey %x addr %llx len %d from peer\n", 327 cb->remote_rkey, (unsigned long long)cb->remote_addr, 328 cb->remote_len); 329 330 if (cb->state <= CONNECTED || cb->state == RDMA_WRITE_COMPLETE) 331 cb->state = RDMA_READ_ADV; 332 else 333 cb->state = RDMA_WRITE_ADV; 334 335 return 0; 336} 337 338static int client_recv(struct krping_cb *cb, struct ib_wc *wc) 339{ 340 if (wc->byte_len != sizeof(cb->recv_buf)) { 341 PRINTF(cb, "Received bogus data, size %d\n", 342 wc->byte_len); 343 return -1; 344 } 345 346 if (cb->state == RDMA_READ_ADV) 347 cb->state = RDMA_WRITE_ADV; 348 else 349 cb->state = RDMA_WRITE_COMPLETE; 350 351 return 0; 352} 353 354static void krping_cq_event_handler(struct ib_cq *cq, void *ctx) 355{ 356 struct krping_cb *cb = ctx; 357 struct ib_wc wc; 358 struct ib_recv_wr *bad_wr; 359 int ret; 360 361 BUG_ON(cb->cq != cq); 362 if (cb->state == ERROR) { 363 PRINTF(cb, "cq completion in ERROR state\n"); 364 return; 365 } 366 if (cb->frtest) { 367 PRINTF(cb, "cq completion event in frtest!\n"); 368 return; 369 } 370 if (!cb->wlat && !cb->rlat && !cb->bw) 371 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); 372 while ((ret = ib_poll_cq(cb->cq, 1, &wc)) == 1) { 373 if (wc.status) { 374 if (wc.status == IB_WC_WR_FLUSH_ERR) { 375 DEBUG_LOG(cb, "cq flushed\n"); 376 continue; 377 } else { 378 PRINTF(cb, "cq completion failed with " 379 "wr_id %Lx status %d opcode %d vender_err %x\n", 380 wc.wr_id, wc.status, wc.opcode, wc.vendor_err); 381 goto error; 382 } 383 } 384 385 switch (wc.opcode) { 386 case IB_WC_SEND: 387 DEBUG_LOG(cb, "send completion\n"); 388 cb->stats.send_bytes += cb->send_sgl.length; 389 cb->stats.send_msgs++; 390 break; 391 392 case IB_WC_RDMA_WRITE: 393 DEBUG_LOG(cb, "rdma write completion\n"); 394 cb->stats.write_bytes += cb->rdma_sq_wr.sg_list->length; 395 cb->stats.write_msgs++; 396 cb->state = RDMA_WRITE_COMPLETE; 397 wake_up_interruptible(&cb->sem); 398 break; 399 400 case IB_WC_RDMA_READ: 401 DEBUG_LOG(cb, "rdma read completion\n"); 402 cb->stats.read_bytes += cb->rdma_sq_wr.sg_list->length; 403 cb->stats.read_msgs++; 404 cb->state = RDMA_READ_COMPLETE; 405 wake_up_interruptible(&cb->sem); 406 break; 407 408 case IB_WC_RECV: 409 DEBUG_LOG(cb, "recv completion\n"); 410 cb->stats.recv_bytes += sizeof(cb->recv_buf); 411 cb->stats.recv_msgs++; 412 if (cb->wlat || cb->rlat || cb->bw) 413 ret = server_recv(cb, &wc); 414 else 415 ret = cb->server ? server_recv(cb, &wc) : 416 client_recv(cb, &wc); 417 if (ret) { 418 PRINTF(cb, "recv wc error: %d\n", ret); 419 goto error; 420 } 421 422 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); 423 if (ret) { 424 PRINTF(cb, "post recv error: %d\n", 425 ret); 426 goto error; 427 } 428 wake_up_interruptible(&cb->sem); 429 break; 430 431 default: 432 PRINTF(cb, 433 "%s:%d Unexpected opcode %d, Shutting down\n", 434 __func__, __LINE__, wc.opcode); 435 goto error; 436 } 437 } 438 if (ret) { 439 PRINTF(cb, "poll error %d\n", ret); 440 goto error; 441 } 442 return; 443error: 444 cb->state = ERROR; 445 wake_up_interruptible(&cb->sem); 446} 447 448static int krping_accept(struct krping_cb *cb) 449{ 450 struct rdma_conn_param conn_param; 451 int ret; 452 453 DEBUG_LOG(cb, "accepting client connection request\n"); 454 455 memset(&conn_param, 0, sizeof conn_param); 456 conn_param.responder_resources = 1; 457 conn_param.initiator_depth = 1; 458 459 ret = rdma_accept(cb->child_cm_id, &conn_param); 460 if (ret) { 461 PRINTF(cb, "rdma_accept error: %d\n", ret); 462 return ret; 463 } 464 465 if (!cb->wlat && !cb->rlat && !cb->bw) { 466 wait_event_interruptible(cb->sem, cb->state >= CONNECTED); 467 if (cb->state == ERROR) { 468 PRINTF(cb, "wait for CONNECTED state %d\n", 469 cb->state); 470 return -1; 471 } 472 } 473 return 0; 474} 475 476static void krping_setup_wr(struct krping_cb *cb) 477{ 478 cb->recv_sgl.addr = cb->recv_dma_addr; 479 cb->recv_sgl.length = sizeof cb->recv_buf; 480 if (cb->local_dma_lkey) 481 cb->recv_sgl.lkey = cb->qp->device->local_dma_lkey; 482 else if (cb->mem == DMA) 483 cb->recv_sgl.lkey = cb->dma_mr->lkey; 484 else 485 cb->recv_sgl.lkey = cb->recv_mr->lkey; 486 cb->rq_wr.sg_list = &cb->recv_sgl; 487 cb->rq_wr.num_sge = 1; 488 489 cb->send_sgl.addr = cb->send_dma_addr; 490 cb->send_sgl.length = sizeof cb->send_buf; 491 if (cb->local_dma_lkey) 492 cb->send_sgl.lkey = cb->qp->device->local_dma_lkey; 493 else if (cb->mem == DMA) 494 cb->send_sgl.lkey = cb->dma_mr->lkey; 495 else 496 cb->send_sgl.lkey = cb->send_mr->lkey; 497 498 cb->sq_wr.opcode = IB_WR_SEND; 499 cb->sq_wr.send_flags = IB_SEND_SIGNALED; 500 cb->sq_wr.sg_list = &cb->send_sgl; 501 cb->sq_wr.num_sge = 1; 502 503 if (cb->server || cb->wlat || cb->rlat || cb->bw) { 504 cb->rdma_sgl.addr = cb->rdma_dma_addr; 505 if (cb->mem == MR) 506 cb->rdma_sgl.lkey = cb->rdma_mr->lkey; 507 cb->rdma_sq_wr.send_flags = IB_SEND_SIGNALED; 508 cb->rdma_sq_wr.sg_list = &cb->rdma_sgl; 509 cb->rdma_sq_wr.num_sge = 1; 510 } 511 512 switch(cb->mem) { 513 case FASTREG: 514 515 /* 516 * A chain of 2 WRs, INVALDATE_MR + FAST_REG_MR. 517 * both unsignaled. The client uses them to reregister 518 * the rdma buffers with a new key each iteration. 519 */ 520 cb->fastreg_wr.opcode = IB_WR_FAST_REG_MR; 521 cb->fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; 522 cb->fastreg_wr.wr.fast_reg.length = cb->size; 523 cb->fastreg_wr.wr.fast_reg.page_list = cb->page_list; 524 cb->fastreg_wr.wr.fast_reg.page_list_len = cb->page_list_len; 525 526 cb->invalidate_wr.next = &cb->fastreg_wr; 527 cb->invalidate_wr.opcode = IB_WR_LOCAL_INV; 528 break; 529 case MW: 530 cb->bind_attr.wr_id = 0xabbaabba; 531 cb->bind_attr.send_flags = 0; /* unsignaled */ 532 cb->bind_attr.length = cb->size; 533 break; 534 default: 535 break; 536 } 537} 538 539static int krping_setup_buffers(struct krping_cb *cb) 540{ 541 int ret; 542 struct ib_phys_buf buf; 543 u64 iovbase; 544 545 DEBUG_LOG(cb, "krping_setup_buffers called on cb %p\n", cb); 546 547 cb->recv_dma_addr = dma_map_single(cb->pd->device->dma_device, 548 &cb->recv_buf, 549 sizeof(cb->recv_buf), DMA_BIDIRECTIONAL); 550 pci_unmap_addr_set(cb, recv_mapping, cb->recv_dma_addr); 551 cb->send_dma_addr = dma_map_single(cb->pd->device->dma_device, 552 &cb->send_buf, sizeof(cb->send_buf), 553 DMA_BIDIRECTIONAL); 554 pci_unmap_addr_set(cb, send_mapping, cb->send_dma_addr); 555 556 if (cb->mem == DMA) { 557 cb->dma_mr = ib_get_dma_mr(cb->pd, IB_ACCESS_LOCAL_WRITE| 558 IB_ACCESS_REMOTE_READ| 559 IB_ACCESS_REMOTE_WRITE); 560 if (IS_ERR(cb->dma_mr)) { 561 DEBUG_LOG(cb, "reg_dmamr failed\n"); 562 ret = PTR_ERR(cb->dma_mr); 563 goto bail; 564 } 565 } else { 566 if (!cb->local_dma_lkey) { 567 buf.addr = cb->recv_dma_addr; 568 buf.size = sizeof cb->recv_buf; 569 DEBUG_LOG(cb, "recv buf dma_addr %llx size %d\n", buf.addr, 570 (int)buf.size); 571 iovbase = cb->recv_dma_addr; 572 cb->recv_mr = ib_reg_phys_mr(cb->pd, &buf, 1, 573 IB_ACCESS_LOCAL_WRITE, 574 &iovbase); 575 576 if (IS_ERR(cb->recv_mr)) { 577 DEBUG_LOG(cb, "recv_buf reg_mr failed\n"); 578 ret = PTR_ERR(cb->recv_mr); 579 goto bail; 580 } 581 582 buf.addr = cb->send_dma_addr; 583 buf.size = sizeof cb->send_buf; 584 DEBUG_LOG(cb, "send buf dma_addr %llx size %d\n", buf.addr, 585 (int)buf.size); 586 iovbase = cb->send_dma_addr; 587 cb->send_mr = ib_reg_phys_mr(cb->pd, &buf, 1, 588 0, &iovbase); 589 590 if (IS_ERR(cb->send_mr)) { 591 DEBUG_LOG(cb, "send_buf reg_mr failed\n"); 592 ret = PTR_ERR(cb->send_mr); 593 goto bail; 594 } 595 } 596 } 597 598 cb->rdma_buf = kmalloc(cb->size, GFP_KERNEL); 599 if (!cb->rdma_buf) { 600 DEBUG_LOG(cb, "rdma_buf malloc failed\n"); 601 ret = -ENOMEM; 602 goto bail; 603 } 604 605 cb->rdma_dma_addr = dma_map_single(cb->pd->device->dma_device, 606 cb->rdma_buf, cb->size, 607 DMA_BIDIRECTIONAL); 608 pci_unmap_addr_set(cb, rdma_mapping, cb->rdma_dma_addr); 609 if (cb->mem != DMA) { 610 switch (cb->mem) { 611 case FASTREG: 612 cb->page_list_len = (((cb->size - 1) & PAGE_MASK) + 613 PAGE_SIZE) >> PAGE_SHIFT; 614 cb->page_list = ib_alloc_fast_reg_page_list( 615 cb->pd->device, 616 cb->page_list_len); 617 if (IS_ERR(cb->page_list)) { 618 DEBUG_LOG(cb, "recv_buf reg_mr failed\n"); 619 ret = PTR_ERR(cb->page_list); 620 goto bail; 621 } 622 cb->fastreg_mr = ib_alloc_fast_reg_mr(cb->pd, 623 cb->page_list->max_page_list_len); 624 if (IS_ERR(cb->fastreg_mr)) { 625 DEBUG_LOG(cb, "recv_buf reg_mr failed\n"); 626 ret = PTR_ERR(cb->fastreg_mr); 627 goto bail; 628 } 629 DEBUG_LOG(cb, "fastreg rkey 0x%x page_list %p" 630 " page_list_len %u\n", cb->fastreg_mr->rkey, 631 cb->page_list, cb->page_list_len); 632 break; 633 case MW: 634 cb->mw = ib_alloc_mw(cb->pd); 635 if (IS_ERR(cb->mw)) { 636 DEBUG_LOG(cb, "recv_buf alloc_mw failed\n"); 637 ret = PTR_ERR(cb->mw); 638 goto bail; 639 } 640 DEBUG_LOG(cb, "mw rkey 0x%x\n", cb->mw->rkey); 641 /*FALLTHROUGH*/ 642 case MR: 643 buf.addr = cb->rdma_dma_addr; 644 buf.size = cb->size; 645 iovbase = cb->rdma_dma_addr; 646 cb->rdma_mr = ib_reg_phys_mr(cb->pd, &buf, 1, 647 IB_ACCESS_REMOTE_READ| 648 IB_ACCESS_REMOTE_WRITE, 649 &iovbase); 650 if (IS_ERR(cb->rdma_mr)) { 651 DEBUG_LOG(cb, "rdma_buf reg_mr failed\n"); 652 ret = PTR_ERR(cb->rdma_mr); 653 goto bail; 654 } 655 DEBUG_LOG(cb, "rdma buf dma_addr %llx size %d mr rkey 0x%x\n", 656 buf.addr, (int)buf.size, cb->rdma_mr->rkey); 657 break; 658 default: 659 ret = -EINVAL; 660 goto bail; 661 break; 662 } 663 } 664 665 if (!cb->server || cb->wlat || cb->rlat || cb->bw) { 666 667 cb->start_buf = kmalloc(cb->size, GFP_KERNEL); 668 if (!cb->start_buf) { 669 DEBUG_LOG(cb, "start_buf malloc failed\n"); 670 ret = -ENOMEM; 671 goto bail; 672 } 673 674 cb->start_dma_addr = dma_map_single(cb->pd->device->dma_device, 675 cb->start_buf, cb->size, 676 DMA_BIDIRECTIONAL); 677 pci_unmap_addr_set(cb, start_mapping, cb->start_dma_addr); 678 679 if (cb->mem == MR || cb->mem == MW) { 680 unsigned flags = IB_ACCESS_REMOTE_READ; 681 682 if (cb->wlat || cb->rlat || cb->bw) 683 flags |= IB_ACCESS_REMOTE_WRITE; 684 685 buf.addr = cb->start_dma_addr; 686 buf.size = cb->size; 687 DEBUG_LOG(cb, "start buf dma_addr %llx size %d\n", 688 buf.addr, (int)buf.size); 689 iovbase = cb->start_dma_addr; 690 cb->start_mr = ib_reg_phys_mr(cb->pd, &buf, 1, 691 flags, 692 &iovbase); 693 694 if (IS_ERR(cb->start_mr)) { 695 DEBUG_LOG(cb, "start_buf reg_mr failed\n"); 696 ret = PTR_ERR(cb->start_mr); 697 goto bail; 698 } 699 } 700 } 701 702 krping_setup_wr(cb); 703 DEBUG_LOG(cb, "allocated & registered buffers...\n"); 704 return 0; 705bail: 706 if (cb->fastreg_mr && !IS_ERR(cb->fastreg_mr)) 707 ib_dereg_mr(cb->fastreg_mr); 708 if (cb->mw && !IS_ERR(cb->mw)) 709 ib_dealloc_mw(cb->mw); 710 if (cb->rdma_mr && !IS_ERR(cb->rdma_mr)) 711 ib_dereg_mr(cb->rdma_mr); 712 if (cb->page_list && !IS_ERR(cb->page_list)) 713 ib_free_fast_reg_page_list(cb->page_list); 714 if (cb->dma_mr && !IS_ERR(cb->dma_mr)) 715 ib_dereg_mr(cb->dma_mr); 716 if (cb->recv_mr && !IS_ERR(cb->recv_mr)) 717 ib_dereg_mr(cb->recv_mr); 718 if (cb->send_mr && !IS_ERR(cb->send_mr)) 719 ib_dereg_mr(cb->send_mr); 720 if (cb->rdma_buf) 721 kfree(cb->rdma_buf); 722 if (cb->start_buf) 723 kfree(cb->start_buf); 724 return ret; 725} 726 727static void krping_free_buffers(struct krping_cb *cb) 728{ 729 DEBUG_LOG(cb, "krping_free_buffers called on cb %p\n", cb); 730 731 if (cb->dma_mr) 732 ib_dereg_mr(cb->dma_mr); 733 if (cb->send_mr) 734 ib_dereg_mr(cb->send_mr); 735 if (cb->recv_mr) 736 ib_dereg_mr(cb->recv_mr); 737 if (cb->rdma_mr) 738 ib_dereg_mr(cb->rdma_mr); 739 if (cb->start_mr) 740 ib_dereg_mr(cb->start_mr); 741 if (cb->fastreg_mr) 742 ib_dereg_mr(cb->fastreg_mr); 743 if (cb->mw) 744 ib_dealloc_mw(cb->mw); 745 746 dma_unmap_single(cb->pd->device->dma_device, 747 pci_unmap_addr(cb, recv_mapping), 748 sizeof(cb->recv_buf), DMA_BIDIRECTIONAL); 749 dma_unmap_single(cb->pd->device->dma_device, 750 pci_unmap_addr(cb, send_mapping), 751 sizeof(cb->send_buf), DMA_BIDIRECTIONAL); 752 dma_unmap_single(cb->pd->device->dma_device, 753 pci_unmap_addr(cb, rdma_mapping), 754 cb->size, DMA_BIDIRECTIONAL); 755 kfree(cb->rdma_buf); 756 if (cb->start_buf) { 757 dma_unmap_single(cb->pd->device->dma_device, 758 pci_unmap_addr(cb, start_mapping), 759 cb->size, DMA_BIDIRECTIONAL); 760 kfree(cb->start_buf); 761 } 762} 763 764static int krping_create_qp(struct krping_cb *cb) 765{ 766 struct ib_qp_init_attr init_attr; 767 int ret; 768 769 memset(&init_attr, 0, sizeof(init_attr)); 770 init_attr.cap.max_send_wr = cb->txdepth; 771 init_attr.cap.max_recv_wr = 2; 772 init_attr.cap.max_recv_sge = 1; 773 init_attr.cap.max_send_sge = 1; 774 init_attr.qp_type = IB_QPT_RC; 775 init_attr.send_cq = cb->cq; 776 init_attr.recv_cq = cb->cq; 777 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 778 779 if (cb->server) { 780 ret = rdma_create_qp(cb->child_cm_id, cb->pd, &init_attr); 781 if (!ret) 782 cb->qp = cb->child_cm_id->qp; 783 } else { 784 ret = rdma_create_qp(cb->cm_id, cb->pd, &init_attr); 785 if (!ret) 786 cb->qp = cb->cm_id->qp; 787 } 788 789 return ret; 790} 791 792static void krping_free_qp(struct krping_cb *cb) 793{ 794 ib_destroy_qp(cb->qp); 795 ib_destroy_cq(cb->cq); 796 ib_dealloc_pd(cb->pd); 797} 798 799static int krping_setup_qp(struct krping_cb *cb, struct rdma_cm_id *cm_id) 800{ 801 int ret; 802 cb->pd = ib_alloc_pd(cm_id->device); 803 if (IS_ERR(cb->pd)) { 804 PRINTF(cb, "ib_alloc_pd failed\n"); 805 return PTR_ERR(cb->pd); 806 } 807 DEBUG_LOG(cb, "created pd %p\n", cb->pd); 808 809 strlcpy(cb->stats.name, cb->pd->device->name, sizeof(cb->stats.name)); 810 811 cb->cq = ib_create_cq(cm_id->device, krping_cq_event_handler, NULL, 812 cb, cb->txdepth * 2, 0); 813 if (IS_ERR(cb->cq)) { 814 PRINTF(cb, "ib_create_cq failed\n"); 815 ret = PTR_ERR(cb->cq); 816 goto err1; 817 } 818 DEBUG_LOG(cb, "created cq %p\n", cb->cq); 819 820 if (!cb->wlat && !cb->rlat && !cb->bw && !cb->frtest) { 821 ret = ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); 822 if (ret) { 823 PRINTF(cb, "ib_create_cq failed\n"); 824 goto err2; 825 } 826 } 827 828 ret = krping_create_qp(cb); 829 if (ret) { 830 PRINTF(cb, "krping_create_qp failed: %d\n", ret); 831 goto err2; 832 } 833 DEBUG_LOG(cb, "created qp %p\n", cb->qp); 834 return 0; 835err2: 836 ib_destroy_cq(cb->cq); 837err1: 838 ib_dealloc_pd(cb->pd); 839 return ret; 840} 841 842/* 843 * return the (possibly rebound) rkey for the rdma buffer. 844 * FASTREG mode: invalidate and rebind via fastreg wr. 845 * MW mode: rebind the MW. 846 * other modes: just return the mr rkey. 847 */ 848static u32 krping_rdma_rkey(struct krping_cb *cb, u64 buf, int post_inv) 849{ 850 u32 rkey = 0xffffffff; 851 u64 p; 852 struct ib_send_wr *bad_wr; 853 int i; 854 int ret; 855 856 switch (cb->mem) { 857 case FASTREG: 858 cb->invalidate_wr.ex.invalidate_rkey = cb->fastreg_mr->rkey; 859 860 /* 861 * Update the fastreg key. 862 */ 863 ib_update_fast_reg_key(cb->fastreg_mr, ++cb->key); 864 cb->fastreg_wr.wr.fast_reg.rkey = cb->fastreg_mr->rkey; 865 866 /* 867 * Update the fastreg WR with new buf info. 868 */ 869 if (buf == (u64)cb->start_dma_addr) 870 cb->fastreg_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_READ; 871 else 872 cb->fastreg_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE; 873 cb->fastreg_wr.wr.fast_reg.iova_start = buf; 874 p = (u64)(buf & PAGE_MASK); 875 for (i=0; i < cb->fastreg_wr.wr.fast_reg.page_list_len; 876 i++, p += PAGE_SIZE) { 877 cb->page_list->page_list[i] = p; 878 DEBUG_LOG(cb, "page_list[%d] 0x%llx\n", i, p); 879 } 880 881 DEBUG_LOG(cb, "post_inv = %d, fastreg new rkey 0x%x shift %u len %u" 882 " iova_start %llx page_list_len %u\n", 883 post_inv, 884 cb->fastreg_wr.wr.fast_reg.rkey, 885 cb->fastreg_wr.wr.fast_reg.page_shift, 886 cb->fastreg_wr.wr.fast_reg.length, 887 cb->fastreg_wr.wr.fast_reg.iova_start, 888 cb->fastreg_wr.wr.fast_reg.page_list_len); 889 890 if (post_inv) 891 ret = ib_post_send(cb->qp, &cb->invalidate_wr, &bad_wr); 892 else 893 ret = ib_post_send(cb->qp, &cb->fastreg_wr, &bad_wr); 894 if (ret) { 895 PRINTF(cb, "post send error %d\n", ret); 896 cb->state = ERROR; 897 } 898 rkey = cb->fastreg_mr->rkey; 899 break; 900 case MW: 901 /* 902 * Update the MW with new buf info. 903 */ 904 if (buf == (u64)cb->start_dma_addr) { 905 cb->bind_attr.mw_access_flags = IB_ACCESS_REMOTE_READ; 906 cb->bind_attr.mr = cb->start_mr; 907 } else { 908 cb->bind_attr.mw_access_flags = IB_ACCESS_REMOTE_WRITE; 909 cb->bind_attr.mr = cb->rdma_mr; 910 } 911 cb->bind_attr.addr = buf; 912 DEBUG_LOG(cb, "binding mw rkey 0x%x to buf %llx mr rkey 0x%x\n", 913 cb->mw->rkey, buf, cb->bind_attr.mr->rkey); 914 ret = ib_bind_mw(cb->qp, cb->mw, &cb->bind_attr); 915 if (ret) { 916 PRINTF(cb, "bind mw error %d\n", ret); 917 cb->state = ERROR; 918 } else 919 rkey = cb->mw->rkey; 920 break; 921 case MR: 922 if (buf == (u64)cb->start_dma_addr) 923 rkey = cb->start_mr->rkey; 924 else 925 rkey = cb->rdma_mr->rkey; 926 break; 927 case DMA: 928 rkey = cb->dma_mr->rkey; 929 break; 930 default: 931 PRINTF(cb, "%s:%d case ERROR\n", __func__, __LINE__); 932 cb->state = ERROR; 933 break; 934 } 935 return rkey; 936} 937 938static void krping_format_send(struct krping_cb *cb, u64 buf) 939{ 940 struct krping_rdma_info *info = &cb->send_buf; 941 u32 rkey; 942 943 /* 944 * Client side will do fastreg or mw bind before 945 * advertising the rdma buffer. Server side 946 * sends have no data. 947 */ 948 if (!cb->server || cb->wlat || cb->rlat || cb->bw) { 949 rkey = krping_rdma_rkey(cb, buf, !cb->server_invalidate); 950 info->buf = htonll(buf); 951 info->rkey = htonl(rkey); 952 info->size = htonl(cb->size); 953 DEBUG_LOG(cb, "RDMA addr %llx rkey %x len %d\n", 954 (unsigned long long)buf, rkey, cb->size); 955 } 956} 957 958static void krping_test_server(struct krping_cb *cb) 959{ 960 struct ib_send_wr *bad_wr, inv; 961 int ret; 962 963 while (1) { 964 /* Wait for client's Start STAG/TO/Len */ 965 wait_event_interruptible(cb->sem, cb->state >= RDMA_READ_ADV); 966 if (cb->state != RDMA_READ_ADV) { 967 PRINTF(cb, "wait for RDMA_READ_ADV state %d\n", 968 cb->state); 969 break; 970 } 971 972 DEBUG_LOG(cb, "server received sink adv\n"); 973 974 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; 975 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; 976 cb->rdma_sq_wr.sg_list->length = cb->remote_len; 977 cb->rdma_sgl.lkey = krping_rdma_rkey(cb, cb->rdma_dma_addr, 1); 978 cb->rdma_sq_wr.next = NULL; 979 980 /* Issue RDMA Read. */ 981 if (cb->read_inv) 982 cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ_WITH_INV; 983 else { 984 985 cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ; 986 if (cb->mem == FASTREG) { 987 /* 988 * Immediately follow the read with a 989 * fenced LOCAL_INV. 990 */ 991 cb->rdma_sq_wr.next = &inv; 992 memset(&inv, 0, sizeof inv); 993 inv.opcode = IB_WR_LOCAL_INV; 994 inv.ex.invalidate_rkey = cb->fastreg_mr->rkey; 995 inv.send_flags = IB_SEND_FENCE; 996 } 997 } 998 999 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr); 1000 if (ret) { 1001 PRINTF(cb, "post send error %d\n", ret); 1002 break; 1003 } 1004 cb->rdma_sq_wr.next = NULL; 1005 1006 DEBUG_LOG(cb, "server posted rdma read req \n"); 1007 1008 /* Wait for read completion */ 1009 wait_event_interruptible(cb->sem, 1010 cb->state >= RDMA_READ_COMPLETE); 1011 if (cb->state != RDMA_READ_COMPLETE) { 1012 PRINTF(cb, 1013 "wait for RDMA_READ_COMPLETE state %d\n", 1014 cb->state); 1015 break; 1016 } 1017 DEBUG_LOG(cb, "server received read complete\n"); 1018 1019 /* Display data in recv buf */ 1020 if (cb->verbose) 1021 PRINTF(cb, "server ping data: %s\n", 1022 cb->rdma_buf); 1023 1024 /* Tell client to continue */ 1025 if (cb->server && cb->server_invalidate) { 1026 cb->sq_wr.ex.invalidate_rkey = cb->remote_rkey; 1027 cb->sq_wr.opcode = IB_WR_SEND_WITH_INV; 1028 DEBUG_LOG(cb, "send-w-inv rkey 0x%x\n", cb->remote_rkey); 1029 } 1030 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1031 if (ret) { 1032 PRINTF(cb, "post send error %d\n", ret); 1033 break; 1034 } 1035 DEBUG_LOG(cb, "server posted go ahead\n"); 1036 1037 /* Wait for client's RDMA STAG/TO/Len */ 1038 wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_ADV); 1039 if (cb->state != RDMA_WRITE_ADV) { 1040 PRINTF(cb, 1041 "wait for RDMA_WRITE_ADV state %d\n", 1042 cb->state); 1043 break; 1044 } 1045 DEBUG_LOG(cb, "server received sink adv\n"); 1046 1047 /* RDMA Write echo data */ 1048 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; 1049 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; 1050 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; 1051 cb->rdma_sq_wr.sg_list->length = strlen(cb->rdma_buf) + 1; 1052 if (cb->local_dma_lkey) 1053 cb->rdma_sgl.lkey = cb->qp->device->local_dma_lkey; 1054 else 1055 cb->rdma_sgl.lkey = krping_rdma_rkey(cb, cb->rdma_dma_addr, 0); 1056 1057 DEBUG_LOG(cb, "rdma write from lkey %x laddr %llx len %d\n", 1058 cb->rdma_sq_wr.sg_list->lkey, 1059 (unsigned long long)cb->rdma_sq_wr.sg_list->addr, 1060 cb->rdma_sq_wr.sg_list->length); 1061 1062 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr); 1063 if (ret) { 1064 PRINTF(cb, "post send error %d\n", ret); 1065 break; 1066 } 1067 1068 /* Wait for completion */ 1069 ret = wait_event_interruptible(cb->sem, cb->state >= 1070 RDMA_WRITE_COMPLETE); 1071 if (cb->state != RDMA_WRITE_COMPLETE) { 1072 PRINTF(cb, 1073 "wait for RDMA_WRITE_COMPLETE state %d\n", 1074 cb->state); 1075 break; 1076 } 1077 DEBUG_LOG(cb, "server rdma write complete \n"); 1078 1079 cb->state = CONNECTED; 1080 1081 /* Tell client to begin again */ 1082 if (cb->server && cb->server_invalidate) { 1083 cb->sq_wr.ex.invalidate_rkey = cb->remote_rkey; 1084 cb->sq_wr.opcode = IB_WR_SEND_WITH_INV; 1085 DEBUG_LOG(cb, "send-w-inv rkey 0x%x\n", cb->remote_rkey); 1086 } 1087 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1088 if (ret) { 1089 PRINTF(cb, "post send error %d\n", ret); 1090 break; 1091 } 1092 DEBUG_LOG(cb, "server posted go ahead\n"); 1093 } 1094} 1095 1096static void rlat_test(struct krping_cb *cb) 1097{ 1098 int scnt; 1099 int iters = cb->count; 1100 struct timeval start_tv, stop_tv; 1101 int ret; 1102 struct ib_wc wc; 1103 struct ib_send_wr *bad_wr; 1104 int ne; 1105 1106 scnt = 0; 1107 cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ; 1108 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; 1109 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; 1110 cb->rdma_sq_wr.sg_list->length = cb->size; 1111 1112 microtime(&start_tv); 1113 if (!cb->poll) { 1114 cb->state = RDMA_READ_ADV; 1115 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); 1116 } 1117 while (scnt < iters) { 1118 1119 cb->state = RDMA_READ_ADV; 1120 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr); 1121 if (ret) { 1122 PRINTF(cb, 1123 "Couldn't post send: ret=%d scnt %d\n", 1124 ret, scnt); 1125 return; 1126 } 1127 1128 do { 1129 if (!cb->poll) { 1130 wait_event_interruptible(cb->sem, 1131 cb->state != RDMA_READ_ADV); 1132 if (cb->state == RDMA_READ_COMPLETE) { 1133 ne = 1; 1134 ib_req_notify_cq(cb->cq, 1135 IB_CQ_NEXT_COMP); 1136 } else { 1137 ne = -1; 1138 } 1139 } else 1140 ne = ib_poll_cq(cb->cq, 1, &wc); 1141 if (cb->state == ERROR) { 1142 PRINTF(cb, 1143 "state == ERROR...bailing scnt %d\n", 1144 scnt); 1145 return; 1146 } 1147 } while (ne == 0); 1148 1149 if (ne < 0) { 1150 PRINTF(cb, "poll CQ failed %d\n", ne); 1151 return; 1152 } 1153 if (cb->poll && wc.status != IB_WC_SUCCESS) { 1154 PRINTF(cb, "Completion wth error at %s:\n", 1155 cb->server ? "server" : "client"); 1156 PRINTF(cb, "Failed status %d: wr_id %d\n", 1157 wc.status, (int) wc.wr_id); 1158 return; 1159 } 1160 ++scnt; 1161 } 1162 microtime(&stop_tv); 1163 1164 if (stop_tv.tv_usec < start_tv.tv_usec) { 1165 stop_tv.tv_usec += 1000000; 1166 stop_tv.tv_sec -= 1; 1167 } 1168 1169 PRINTF(cb, "delta sec %lu delta usec %lu iter %d size %d\n", 1170 stop_tv.tv_sec - start_tv.tv_sec, 1171 stop_tv.tv_usec - start_tv.tv_usec, 1172 scnt, cb->size); 1173} 1174 1175static void wlat_test(struct krping_cb *cb) 1176{ 1177 int ccnt, scnt, rcnt; 1178 int iters=cb->count; 1179 volatile char *poll_buf = (char *) cb->start_buf; 1180 char *buf = (char *)cb->rdma_buf; 1181 struct timeval start_tv, stop_tv; 1182 cycles_t *post_cycles_start, *post_cycles_stop; 1183 cycles_t *poll_cycles_start, *poll_cycles_stop; 1184 cycles_t *last_poll_cycles_start; 1185 cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0; 1186 int i; 1187 int cycle_iters = 1000; 1188 1189 ccnt = 0; 1190 scnt = 0; 1191 rcnt = 0; 1192 1193 post_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1194 if (!post_cycles_start) { 1195 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1196 return; 1197 } 1198 post_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1199 if (!post_cycles_stop) { 1200 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1201 return; 1202 } 1203 poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1204 if (!poll_cycles_start) { 1205 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1206 return; 1207 } 1208 poll_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1209 if (!poll_cycles_stop) { 1210 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1211 return; 1212 } 1213 last_poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), 1214 GFP_KERNEL); 1215 if (!last_poll_cycles_start) { 1216 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1217 return; 1218 } 1219 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; 1220 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; 1221 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; 1222 cb->rdma_sq_wr.sg_list->length = cb->size; 1223 1224 if (cycle_iters > iters) 1225 cycle_iters = iters; 1226 microtime(&start_tv); 1227 while (scnt < iters || ccnt < iters || rcnt < iters) { 1228 1229 /* Wait till buffer changes. */ 1230 if (rcnt < iters && !(scnt < 1 && !cb->server)) { 1231 ++rcnt; 1232 while (*poll_buf != (char)rcnt) { 1233 if (cb->state == ERROR) { 1234 PRINTF(cb, 1235 "state = ERROR, bailing\n"); 1236 return; 1237 } 1238 } 1239 } 1240 1241 if (scnt < iters) { 1242 struct ib_send_wr *bad_wr; 1243 1244 *buf = (char)scnt+1; 1245 if (scnt < cycle_iters) 1246 post_cycles_start[scnt] = get_cycles(); 1247 if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) { 1248 PRINTF(cb, 1249 "Couldn't post send: scnt=%d\n", 1250 scnt); 1251 return; 1252 } 1253 if (scnt < cycle_iters) 1254 post_cycles_stop[scnt] = get_cycles(); 1255 scnt++; 1256 } 1257 1258 if (ccnt < iters) { 1259 struct ib_wc wc; 1260 int ne; 1261 1262 if (ccnt < cycle_iters) 1263 poll_cycles_start[ccnt] = get_cycles(); 1264 do { 1265 if (ccnt < cycle_iters) 1266 last_poll_cycles_start[ccnt] = 1267 get_cycles(); 1268 ne = ib_poll_cq(cb->cq, 1, &wc); 1269 } while (ne == 0); 1270 if (ccnt < cycle_iters) 1271 poll_cycles_stop[ccnt] = get_cycles(); 1272 ++ccnt; 1273 1274 if (ne < 0) { 1275 PRINTF(cb, "poll CQ failed %d\n", ne); 1276 return; 1277 } 1278 if (wc.status != IB_WC_SUCCESS) { 1279 PRINTF(cb, 1280 "Completion wth error at %s:\n", 1281 cb->server ? "server" : "client"); 1282 PRINTF(cb, 1283 "Failed status %d: wr_id %d\n", 1284 wc.status, (int) wc.wr_id); 1285 PRINTF(cb, 1286 "scnt=%d, rcnt=%d, ccnt=%d\n", 1287 scnt, rcnt, ccnt); 1288 return; 1289 } 1290 } 1291 } 1292 microtime(&stop_tv); 1293 1294 if (stop_tv.tv_usec < start_tv.tv_usec) { 1295 stop_tv.tv_usec += 1000000; 1296 stop_tv.tv_sec -= 1; 1297 } 1298 1299 for (i=0; i < cycle_iters; i++) { 1300 sum_post += post_cycles_stop[i] - post_cycles_start[i]; 1301 sum_poll += poll_cycles_stop[i] - poll_cycles_start[i]; 1302 sum_last_poll += poll_cycles_stop[i]-last_poll_cycles_start[i]; 1303 } 1304 PRINTF(cb, 1305 "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d" 1306 " sum_post %llu sum_poll %llu sum_last_poll %llu\n", 1307 stop_tv.tv_sec - start_tv.tv_sec, 1308 stop_tv.tv_usec - start_tv.tv_usec, 1309 scnt, cb->size, cycle_iters, 1310 (unsigned long long)sum_post, (unsigned long long)sum_poll, 1311 (unsigned long long)sum_last_poll); 1312 kfree(post_cycles_start); 1313 kfree(post_cycles_stop); 1314 kfree(poll_cycles_start); 1315 kfree(poll_cycles_stop); 1316 kfree(last_poll_cycles_start); 1317} 1318 1319static void bw_test(struct krping_cb *cb) 1320{ 1321 int ccnt, scnt, rcnt; 1322 int iters=cb->count; 1323 struct timeval start_tv, stop_tv; 1324 cycles_t *post_cycles_start, *post_cycles_stop; 1325 cycles_t *poll_cycles_start, *poll_cycles_stop; 1326 cycles_t *last_poll_cycles_start; 1327 cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0; 1328 int i; 1329 int cycle_iters = 1000; 1330 1331 ccnt = 0; 1332 scnt = 0; 1333 rcnt = 0; 1334 1335 post_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1336 if (!post_cycles_start) { 1337 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1338 return; 1339 } 1340 post_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1341 if (!post_cycles_stop) { 1342 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1343 return; 1344 } 1345 poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1346 if (!poll_cycles_start) { 1347 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1348 return; 1349 } 1350 poll_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1351 if (!poll_cycles_stop) { 1352 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1353 return; 1354 } 1355 last_poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), 1356 GFP_KERNEL); 1357 if (!last_poll_cycles_start) { 1358 PRINTF(cb, "%s kmalloc failed\n", __FUNCTION__); 1359 return; 1360 } 1361 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; 1362 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; 1363 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; 1364 cb->rdma_sq_wr.sg_list->length = cb->size; 1365 1366 if (cycle_iters > iters) 1367 cycle_iters = iters; 1368 microtime(&start_tv); 1369 while (scnt < iters || ccnt < iters) { 1370 1371 while (scnt < iters && scnt - ccnt < cb->txdepth) { 1372 struct ib_send_wr *bad_wr; 1373 1374 if (scnt < cycle_iters) 1375 post_cycles_start[scnt] = get_cycles(); 1376 if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) { 1377 PRINTF(cb, 1378 "Couldn't post send: scnt=%d\n", 1379 scnt); 1380 return; 1381 } 1382 if (scnt < cycle_iters) 1383 post_cycles_stop[scnt] = get_cycles(); 1384 ++scnt; 1385 } 1386 1387 if (ccnt < iters) { 1388 int ne; 1389 struct ib_wc wc; 1390 1391 if (ccnt < cycle_iters) 1392 poll_cycles_start[ccnt] = get_cycles(); 1393 do { 1394 if (ccnt < cycle_iters) 1395 last_poll_cycles_start[ccnt] = 1396 get_cycles(); 1397 ne = ib_poll_cq(cb->cq, 1, &wc); 1398 } while (ne == 0); 1399 if (ccnt < cycle_iters) 1400 poll_cycles_stop[ccnt] = get_cycles(); 1401 ccnt += 1; 1402 1403 if (ne < 0) { 1404 PRINTF(cb, "poll CQ failed %d\n", ne); 1405 return; 1406 } 1407 if (wc.status != IB_WC_SUCCESS) { 1408 PRINTF(cb, 1409 "Completion wth error at %s:\n", 1410 cb->server ? "server" : "client"); 1411 PRINTF(cb, 1412 "Failed status %d: wr_id %d\n", 1413 wc.status, (int) wc.wr_id); 1414 return; 1415 } 1416 } 1417 } 1418 microtime(&stop_tv); 1419 1420 if (stop_tv.tv_usec < start_tv.tv_usec) { 1421 stop_tv.tv_usec += 1000000; 1422 stop_tv.tv_sec -= 1; 1423 } 1424 1425 for (i=0; i < cycle_iters; i++) { 1426 sum_post += post_cycles_stop[i] - post_cycles_start[i]; 1427 sum_poll += poll_cycles_stop[i] - poll_cycles_start[i]; 1428 sum_last_poll += poll_cycles_stop[i]-last_poll_cycles_start[i]; 1429 } 1430 PRINTF(cb, 1431 "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d" 1432 " sum_post %llu sum_poll %llu sum_last_poll %llu\n", 1433 stop_tv.tv_sec - start_tv.tv_sec, 1434 stop_tv.tv_usec - start_tv.tv_usec, 1435 scnt, cb->size, cycle_iters, 1436 (unsigned long long)sum_post, (unsigned long long)sum_poll, 1437 (unsigned long long)sum_last_poll); 1438 kfree(post_cycles_start); 1439 kfree(post_cycles_stop); 1440 kfree(poll_cycles_start); 1441 kfree(poll_cycles_stop); 1442 kfree(last_poll_cycles_start); 1443} 1444 1445static void krping_rlat_test_server(struct krping_cb *cb) 1446{ 1447 struct ib_send_wr *bad_wr; 1448 struct ib_wc wc; 1449 int ret; 1450 1451 /* Spin waiting for client's Start STAG/TO/Len */ 1452 while (cb->state < RDMA_READ_ADV) { 1453 krping_cq_event_handler(cb->cq, cb); 1454 } 1455 1456 /* Send STAG/TO/Len to client */ 1457 krping_format_send(cb, cb->start_dma_addr); 1458 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1459 if (ret) { 1460 PRINTF(cb, "post send error %d\n", ret); 1461 return; 1462 } 1463 1464 /* Spin waiting for send completion */ 1465 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1466 if (ret < 0) { 1467 PRINTF(cb, "poll error %d\n", ret); 1468 return; 1469 } 1470 if (wc.status) { 1471 PRINTF(cb, "send completiong error %d\n", wc.status); 1472 return; 1473 } 1474 1475 wait_event_interruptible(cb->sem, cb->state == ERROR); 1476} 1477 1478static void krping_wlat_test_server(struct krping_cb *cb) 1479{ 1480 struct ib_send_wr *bad_wr; 1481 struct ib_wc wc; 1482 int ret; 1483 1484 /* Spin waiting for client's Start STAG/TO/Len */ 1485 while (cb->state < RDMA_READ_ADV) { 1486 krping_cq_event_handler(cb->cq, cb); 1487 } 1488 1489 /* Send STAG/TO/Len to client */ 1490 krping_format_send(cb, cb->start_dma_addr); 1491 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1492 if (ret) { 1493 PRINTF(cb, "post send error %d\n", ret); 1494 return; 1495 } 1496 1497 /* Spin waiting for send completion */ 1498 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1499 if (ret < 0) { 1500 PRINTF(cb, "poll error %d\n", ret); 1501 return; 1502 } 1503 if (wc.status) { 1504 PRINTF(cb, "send completiong error %d\n", wc.status); 1505 return; 1506 } 1507 1508 wlat_test(cb); 1509 wait_event_interruptible(cb->sem, cb->state == ERROR); 1510} 1511 1512static void krping_bw_test_server(struct krping_cb *cb) 1513{ 1514 struct ib_send_wr *bad_wr; 1515 struct ib_wc wc; 1516 int ret; 1517 1518 /* Spin waiting for client's Start STAG/TO/Len */ 1519 while (cb->state < RDMA_READ_ADV) { 1520 krping_cq_event_handler(cb->cq, cb); 1521 } 1522 1523 /* Send STAG/TO/Len to client */ 1524 krping_format_send(cb, cb->start_dma_addr); 1525 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1526 if (ret) { 1527 PRINTF(cb, "post send error %d\n", ret); 1528 return; 1529 } 1530 1531 /* Spin waiting for send completion */ 1532 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1533 if (ret < 0) { 1534 PRINTF(cb, "poll error %d\n", ret); 1535 return; 1536 } 1537 if (wc.status) { 1538 PRINTF(cb, "send completiong error %d\n", wc.status); 1539 return; 1540 } 1541 1542 if (cb->duplex) 1543 bw_test(cb); 1544 wait_event_interruptible(cb->sem, cb->state == ERROR); 1545} 1546 1547static int fastreg_supported(struct krping_cb *cb) 1548{ 1549 struct ib_device *dev = cb->child_cm_id->device; 1550 struct ib_device_attr attr; 1551 int ret; 1552 1553 ret = ib_query_device(dev, &attr); 1554 if (ret) { 1555 PRINTF(cb, "ib_query_device failed ret %d\n", ret); 1556 return 0; 1557 } 1558 if (!(attr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)) { 1559 PRINTF(cb, "Fastreg not supported - device_cap_flags 0x%x\n", 1560 attr.device_cap_flags); 1561 return 0; 1562 } 1563 DEBUG_LOG(cb, "Fastreg supported - device_cap_flags 0x%x\n", 1564 attr.device_cap_flags); 1565 return 1; 1566} 1567 1568static int krping_bind_server(struct krping_cb *cb) 1569{ 1570 union { 1571 struct sockaddr_in v4; 1572 struct sockaddr_in6 v6; 1573 } sin; 1574 int ret; 1575 1576 memset(&sin, 0, sizeof(sin)); 1577 1578 switch (cb->addr_type) { 1579 case AF_INET: 1580 sin.v4.sin_len = sizeof sin.v4; 1581 sin.v4.sin_family = AF_INET; 1582 sin.v4.sin_addr = cb->addr.v4; 1583 sin.v4.sin_port = cb->port; 1584 break; 1585 case AF_INET6: 1586 sin.v6.sin6_len = sizeof sin.v6; 1587 sin.v6.sin6_family = AF_INET6; 1588 sin.v6.sin6_addr = cb->addr.v6; 1589 sin.v6.sin6_port = cb->port; 1590 break; 1591 default: 1592 return (-EINVAL); 1593 } 1594 1595 ret = rdma_bind_addr(cb->cm_id, (struct sockaddr *) &sin); 1596 if (ret) { 1597 PRINTF(cb, "rdma_bind_addr error %d\n", ret); 1598 return ret; 1599 } 1600 DEBUG_LOG(cb, "rdma_bind_addr successful\n"); 1601 1602 DEBUG_LOG(cb, "rdma_listen\n"); 1603 ret = rdma_listen(cb->cm_id, 3); 1604 if (ret) { 1605 PRINTF(cb, "rdma_listen failed: %d\n", ret); 1606 return ret; 1607 } 1608 1609 wait_event_interruptible(cb->sem, cb->state >= CONNECT_REQUEST); 1610 if (cb->state != CONNECT_REQUEST) { 1611 PRINTF(cb, "wait for CONNECT_REQUEST state %d\n", 1612 cb->state); 1613 return -1; 1614 } 1615 1616 if (cb->mem == FASTREG && !fastreg_supported(cb)) 1617 return -EINVAL; 1618 1619 return 0; 1620} 1621 1622static void krping_run_server(struct krping_cb *cb) 1623{ 1624 struct ib_recv_wr *bad_wr; 1625 int ret; 1626 1627 ret = krping_bind_server(cb); 1628 if (ret) 1629 return; 1630 1631 ret = krping_setup_qp(cb, cb->child_cm_id); 1632 if (ret) { 1633 PRINTF(cb, "setup_qp failed: %d\n", ret); 1634 goto err0; 1635 } 1636 1637 ret = krping_setup_buffers(cb); 1638 if (ret) { 1639 PRINTF(cb, "krping_setup_buffers failed: %d\n", ret); 1640 goto err1; 1641 } 1642 1643 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); 1644 if (ret) { 1645 PRINTF(cb, "ib_post_recv failed: %d\n", ret); 1646 goto err2; 1647 } 1648 1649 ret = krping_accept(cb); 1650 if (ret) { 1651 PRINTF(cb, "connect error %d\n", ret); 1652 goto err2; 1653 } 1654 1655 if (cb->wlat) 1656 krping_wlat_test_server(cb); 1657 else if (cb->rlat) 1658 krping_rlat_test_server(cb); 1659 else if (cb->bw) 1660 krping_bw_test_server(cb); 1661 else 1662 krping_test_server(cb); 1663 rdma_disconnect(cb->child_cm_id); 1664err2: 1665 krping_free_buffers(cb); 1666err1: 1667 krping_free_qp(cb); 1668err0: 1669 rdma_destroy_id(cb->child_cm_id); 1670} 1671 1672static void krping_test_client(struct krping_cb *cb) 1673{ 1674 int ping, start, cc, i, ret; 1675 struct ib_send_wr *bad_wr; 1676 unsigned char c; 1677 1678 start = 65; 1679 for (ping = 0; !cb->count || ping < cb->count; ping++) { 1680 cb->state = RDMA_READ_ADV; 1681 1682 /* Put some ascii text in the buffer. */ 1683 cc = sprintf(cb->start_buf, "rdma-ping-%d: ", ping); 1684 for (i = cc, c = start; i < cb->size; i++) { 1685 cb->start_buf[i] = c; 1686 c++; 1687 if (c > 122) 1688 c = 65; 1689 } 1690 start++; 1691 if (start > 122) 1692 start = 65; 1693 cb->start_buf[cb->size - 1] = 0; 1694 1695 krping_format_send(cb, cb->start_dma_addr); 1696 if (cb->state == ERROR) { 1697 PRINTF(cb, "krping_format_send failed\n"); 1698 break; 1699 } 1700 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1701 if (ret) { 1702 PRINTF(cb, "post send error %d\n", ret); 1703 break; 1704 } 1705 1706 /* Wait for server to ACK */ 1707 wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_ADV); 1708 if (cb->state != RDMA_WRITE_ADV) { 1709 PRINTF(cb, 1710 "wait for RDMA_WRITE_ADV state %d\n", 1711 cb->state); 1712 break; 1713 } 1714 1715 krping_format_send(cb, cb->rdma_dma_addr); 1716 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1717 if (ret) { 1718 PRINTF(cb, "post send error %d\n", ret); 1719 break; 1720 } 1721 1722 /* Wait for the server to say the RDMA Write is complete. */ 1723 wait_event_interruptible(cb->sem, 1724 cb->state >= RDMA_WRITE_COMPLETE); 1725 if (cb->state != RDMA_WRITE_COMPLETE) { 1726 PRINTF(cb, 1727 "wait for RDMA_WRITE_COMPLETE state %d\n", 1728 cb->state); 1729 break; 1730 } 1731 1732 if (cb->validate) 1733 if (memcmp(cb->start_buf, cb->rdma_buf, cb->size)) { 1734 PRINTF(cb, "data mismatch!\n"); 1735 break; 1736 } 1737 1738 if (cb->verbose) 1739 PRINTF(cb, "ping data: %s\n", cb->rdma_buf); 1740#ifdef SLOW_KRPING 1741 wait_event_interruptible_timeout(cb->sem, cb->state == ERROR, HZ); 1742#endif 1743 } 1744} 1745 1746static void krping_rlat_test_client(struct krping_cb *cb) 1747{ 1748 struct ib_send_wr *bad_wr; 1749 struct ib_wc wc; 1750 int ret; 1751 1752 cb->state = RDMA_READ_ADV; 1753 1754 /* Send STAG/TO/Len to client */ 1755 krping_format_send(cb, cb->start_dma_addr); 1756 if (cb->state == ERROR) { 1757 PRINTF(cb, "krping_format_send failed\n"); 1758 return; 1759 } 1760 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1761 if (ret) { 1762 PRINTF(cb, "post send error %d\n", ret); 1763 return; 1764 } 1765 1766 /* Spin waiting for send completion */ 1767 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1768 if (ret < 0) { 1769 PRINTF(cb, "poll error %d\n", ret); 1770 return; 1771 } 1772 if (wc.status) { 1773 PRINTF(cb, "send completion error %d\n", wc.status); 1774 return; 1775 } 1776 1777 /* Spin waiting for server's Start STAG/TO/Len */ 1778 while (cb->state < RDMA_WRITE_ADV) { 1779 krping_cq_event_handler(cb->cq, cb); 1780 } 1781 1782#if 0 1783{ 1784 int i; 1785 struct timeval start, stop; 1786 time_t sec; 1787 suseconds_t usec; 1788 unsigned long long elapsed; 1789 struct ib_wc wc; 1790 struct ib_send_wr *bad_wr; 1791 int ne; 1792 1793 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE; 1794 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey; 1795 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr; 1796 cb->rdma_sq_wr.sg_list->length = 0; 1797 cb->rdma_sq_wr.num_sge = 0; 1798 1799 microtime(&start); 1800 for (i=0; i < 100000; i++) { 1801 if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) { 1802 PRINTF(cb, "Couldn't post send\n"); 1803 return; 1804 } 1805 do { 1806 ne = ib_poll_cq(cb->cq, 1, &wc); 1807 } while (ne == 0); 1808 if (ne < 0) { 1809 PRINTF(cb, "poll CQ failed %d\n", ne); 1810 return; 1811 } 1812 if (wc.status != IB_WC_SUCCESS) { 1813 PRINTF(cb, "Completion wth error at %s:\n", 1814 cb->server ? "server" : "client"); 1815 PRINTF(cb, "Failed status %d: wr_id %d\n", 1816 wc.status, (int) wc.wr_id); 1817 return; 1818 } 1819 } 1820 microtime(&stop); 1821 1822 if (stop.tv_usec < start.tv_usec) { 1823 stop.tv_usec += 1000000; 1824 stop.tv_sec -= 1; 1825 } 1826 sec = stop.tv_sec - start.tv_sec; 1827 usec = stop.tv_usec - start.tv_usec; 1828 elapsed = sec * 1000000 + usec; 1829 PRINTF(cb, "0B-write-lat iters 100000 usec %llu\n", elapsed); 1830} 1831#endif 1832 1833 rlat_test(cb); 1834} 1835 1836static void krping_wlat_test_client(struct krping_cb *cb) 1837{ 1838 struct ib_send_wr *bad_wr; 1839 struct ib_wc wc; 1840 int ret; 1841 1842 cb->state = RDMA_READ_ADV; 1843 1844 /* Send STAG/TO/Len to client */ 1845 krping_format_send(cb, cb->start_dma_addr); 1846 if (cb->state == ERROR) { 1847 PRINTF(cb, "krping_format_send failed\n"); 1848 return; 1849 } 1850 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1851 if (ret) { 1852 PRINTF(cb, "post send error %d\n", ret); 1853 return; 1854 } 1855 1856 /* Spin waiting for send completion */ 1857 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1858 if (ret < 0) { 1859 PRINTF(cb, "poll error %d\n", ret); 1860 return; 1861 } 1862 if (wc.status) { 1863 PRINTF(cb, "send completion error %d\n", wc.status); 1864 return; 1865 } 1866 1867 /* Spin waiting for server's Start STAG/TO/Len */ 1868 while (cb->state < RDMA_WRITE_ADV) { 1869 krping_cq_event_handler(cb->cq, cb); 1870 } 1871 1872 wlat_test(cb); 1873} 1874 1875static void krping_bw_test_client(struct krping_cb *cb) 1876{ 1877 struct ib_send_wr *bad_wr; 1878 struct ib_wc wc; 1879 int ret; 1880 1881 cb->state = RDMA_READ_ADV; 1882 1883 /* Send STAG/TO/Len to client */ 1884 krping_format_send(cb, cb->start_dma_addr); 1885 if (cb->state == ERROR) { 1886 PRINTF(cb, "krping_format_send failed\n"); 1887 return; 1888 } 1889 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1890 if (ret) { 1891 PRINTF(cb, "post send error %d\n", ret); 1892 return; 1893 } 1894 1895 /* Spin waiting for send completion */ 1896 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1897 if (ret < 0) { 1898 PRINTF(cb, "poll error %d\n", ret); 1899 return; 1900 } 1901 if (wc.status) { 1902 PRINTF(cb, "send completion error %d\n", wc.status); 1903 return; 1904 } 1905 1906 /* Spin waiting for server's Start STAG/TO/Len */ 1907 while (cb->state < RDMA_WRITE_ADV) { 1908 krping_cq_event_handler(cb->cq, cb); 1909 } 1910 1911 bw_test(cb); 1912} 1913 1914static void krping_fr_test(struct krping_cb *cb) 1915{ 1916 struct ib_fast_reg_page_list *pl; 1917 struct ib_send_wr fr, inv, *bad; 1918 struct ib_wc wc; 1919 u8 key = 0; 1920 struct ib_mr *mr; 1921 int i; 1922 int ret; 1923 int size = cb->size; 1924 int plen = (((size - 1) & PAGE_MASK) + PAGE_SIZE) >> PAGE_SHIFT; 1925 time_t start; 1926 int count = 0; 1927 int scnt = 0; 1928 1929 pl = ib_alloc_fast_reg_page_list(cb->qp->device, plen); 1930 if (IS_ERR(pl)) { 1931 PRINTF(cb, "ib_alloc_fast_reg_page_list failed %ld\n", PTR_ERR(pl)); 1932 return; 1933 } 1934 1935 mr = ib_alloc_fast_reg_mr(cb->pd, plen); 1936 if (IS_ERR(mr)) { 1937 PRINTF(cb, "ib_alloc_fast_reg_mr failed %ld\n", PTR_ERR(pl)); 1938 goto err1; 1939 } 1940 1941 for (i=0; i<plen; i++) 1942 pl->page_list[i] = 0xcafebabe | i; 1943 1944 memset(&fr, 0, sizeof fr); 1945 fr.opcode = IB_WR_FAST_REG_MR; 1946 fr.wr.fast_reg.page_shift = PAGE_SHIFT; 1947 fr.wr.fast_reg.length = size; 1948 fr.wr.fast_reg.page_list = pl; 1949 fr.wr.fast_reg.page_list_len = plen; 1950 fr.wr.fast_reg.iova_start = 0; 1951 fr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE; 1952 fr.next = &inv; 1953 memset(&inv, 0, sizeof inv); 1954 inv.opcode = IB_WR_LOCAL_INV; 1955 inv.send_flags = IB_SEND_SIGNALED; 1956 1957 DEBUG_LOG(cb, "fr_test: stag index 0x%x plen %u size %u depth %u\n", mr->rkey >> 8, plen, cb->size, cb->txdepth); 1958 start = time_uptime; 1959 while (1) { 1960 if ((time_uptime - start) >= 9) { 1961 DEBUG_LOG(cb, "fr_test: pausing 1 second! count %u latest size %u plen %u\n", count, size, plen); 1962 wait_event_interruptible(cb->sem, cb->state == ERROR); 1963 if (cb->state == ERROR) 1964 break; 1965 start = time_uptime; 1966 } 1967 while (scnt < (cb->txdepth>>1)) { 1968 ib_update_fast_reg_key(mr, ++key); 1969 fr.wr.fast_reg.rkey = mr->rkey; 1970 inv.ex.invalidate_rkey = mr->rkey; 1971 size = arc4random() % cb->size; 1972 if (size == 0) 1973 size = cb->size; 1974 plen = (((size - 1) & PAGE_MASK) + PAGE_SIZE) >> PAGE_SHIFT; 1975 fr.wr.fast_reg.length = size; 1976 fr.wr.fast_reg.page_list_len = plen; 1977 ret = ib_post_send(cb->qp, &fr, &bad); 1978 if (ret) { 1979 PRINTF(cb, "ib_post_send failed %d\n", ret); 1980 goto err2; 1981 } 1982 scnt++; 1983 } 1984 1985 do { 1986 ret = ib_poll_cq(cb->cq, 1, &wc); 1987 if (ret < 0) { 1988 PRINTF(cb, "ib_poll_cq failed %d\n", ret); 1989 goto err2; 1990 } 1991 if (ret == 1) { 1992 if (wc.status) { 1993 PRINTF(cb, "completion error %u\n", wc.status); 1994 goto err2; 1995 } 1996 count++; 1997 scnt--; 1998 } 1999 else if (krping_sigpending()) { 2000 PRINTF(cb, "signal!\n"); 2001 goto err2; 2002 } 2003 } while (ret == 1); 2004 } 2005err2: 2006#if 0 2007 DEBUG_LOG(cb, "sleeping 1 second\n"); 2008 wait_event_interruptible_timeout(cb->sem, cb->state == ERROR, HZ); 2009#endif 2010 DEBUG_LOG(cb, "draining the cq...\n"); 2011 do { 2012 ret = ib_poll_cq(cb->cq, 1, &wc); 2013 if (ret < 0) { 2014 PRINTF(cb, "ib_poll_cq failed %d\n", ret); 2015 break; 2016 } 2017 if (ret == 1) { 2018 if (wc.status) { 2019 PRINTF(cb, "completion error %u opcode %u\n", wc.status, wc.opcode); 2020 } 2021 } 2022 } while (ret == 1); 2023 DEBUG_LOG(cb, "fr_test: done!\n"); 2024 ib_dereg_mr(mr); 2025err1: 2026 ib_free_fast_reg_page_list(pl); 2027} 2028 2029static int krping_connect_client(struct krping_cb *cb) 2030{ 2031 struct rdma_conn_param conn_param; 2032 int ret; 2033 2034 memset(&conn_param, 0, sizeof conn_param); 2035 conn_param.responder_resources = 1; 2036 conn_param.initiator_depth = 1; 2037 conn_param.retry_count = 10; 2038 2039 ret = rdma_connect(cb->cm_id, &conn_param); 2040 if (ret) { 2041 PRINTF(cb, "rdma_connect error %d\n", ret); 2042 return ret; 2043 } 2044 2045 wait_event_interruptible(cb->sem, cb->state >= CONNECTED); 2046 if (cb->state == ERROR) { 2047 PRINTF(cb, "wait for CONNECTED state %d\n", cb->state); 2048 return -1; 2049 } 2050 2051 DEBUG_LOG(cb, "rdma_connect successful\n"); 2052 return 0; 2053} 2054 2055static int krping_bind_client(struct krping_cb *cb) 2056{ 2057 union { 2058 struct sockaddr_in v4; 2059 struct sockaddr_in6 v6; 2060 } sin; 2061 int ret; 2062 2063 memset(&sin, 0, sizeof(sin)); 2064 2065 switch (cb->addr_type) { 2066 case AF_INET: 2067 sin.v4.sin_len = sizeof sin.v4; 2068 sin.v4.sin_family = AF_INET; 2069 sin.v4.sin_addr = cb->addr.v4; 2070 sin.v4.sin_port = cb->port; 2071 break; 2072 case AF_INET6: 2073 sin.v6.sin6_len = sizeof sin.v6; 2074 sin.v6.sin6_family = AF_INET6; 2075 sin.v6.sin6_addr = cb->addr.v6; 2076 sin.v6.sin6_port = cb->port; 2077 break; 2078 default: 2079 return (-EINVAL); 2080 } 2081 2082 ret = rdma_resolve_addr(cb->cm_id, NULL, (struct sockaddr *) &sin, 2083 2000); 2084 if (ret) { 2085 PRINTF(cb, "rdma_resolve_addr error %d\n", ret); 2086 return ret; 2087 } 2088 2089 wait_event_interruptible(cb->sem, cb->state >= ROUTE_RESOLVED); 2090 if (cb->state != ROUTE_RESOLVED) { 2091 PRINTF(cb, 2092 "addr/route resolution did not resolve: state %d\n", 2093 cb->state); 2094 return -EINTR; 2095 } 2096 2097 if (cb->mem == FASTREG && !fastreg_supported(cb)) 2098 return -EINVAL; 2099 2100 DEBUG_LOG(cb, "rdma_resolve_addr - rdma_resolve_route successful\n"); 2101 return 0; 2102} 2103 2104static void krping_run_client(struct krping_cb *cb) 2105{ 2106 struct ib_recv_wr *bad_wr; 2107 int ret; 2108 2109 ret = krping_bind_client(cb); 2110 if (ret) 2111 return; 2112 2113 ret = krping_setup_qp(cb, cb->cm_id); 2114 if (ret) { 2115 PRINTF(cb, "setup_qp failed: %d\n", ret); 2116 return; 2117 } 2118 2119 ret = krping_setup_buffers(cb); 2120 if (ret) { 2121 PRINTF(cb, "krping_setup_buffers failed: %d\n", ret); 2122 goto err1; 2123 } 2124 2125 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); 2126 if (ret) { 2127 PRINTF(cb, "ib_post_recv failed: %d\n", ret); 2128 goto err2; 2129 } 2130 2131 ret = krping_connect_client(cb); 2132 if (ret) { 2133 PRINTF(cb, "connect error %d\n", ret); 2134 goto err2; 2135 } 2136 2137 if (cb->wlat) 2138 krping_wlat_test_client(cb); 2139 else if (cb->rlat) 2140 krping_rlat_test_client(cb); 2141 else if (cb->bw) 2142 krping_bw_test_client(cb); 2143 else if (cb->frtest) 2144 krping_fr_test(cb); 2145 else 2146 krping_test_client(cb); 2147 rdma_disconnect(cb->cm_id); 2148err2: 2149 krping_free_buffers(cb); 2150err1: 2151 krping_free_qp(cb); 2152} 2153 2154static uint16_t 2155krping_get_ipv6_scope_id(char *name) 2156{ 2157 struct ifnet *ifp; 2158 uint16_t retval; 2159 2160 if (name == NULL) 2161 return (0); 2162 ifp = ifunit_ref(name); 2163 if (ifp == NULL) 2164 return (0); 2165 retval = ifp->if_index; 2166 if_rele(ifp); 2167 return (retval); 2168} 2169 2170int krping_doit(char *cmd, void *cookie) 2171{ 2172 struct krping_cb *cb; 2173 int op; 2174 int ret = 0; 2175 char *optarg; 2176 char *scope; 2177 unsigned long optint; 2178 2179 cb = kzalloc(sizeof(*cb), GFP_KERNEL); 2180 if (!cb) 2181 return -ENOMEM; 2182 2183 mutex_lock(&krping_mutex); 2184 list_add_tail(&cb->list, &krping_cbs); 2185 mutex_unlock(&krping_mutex); 2186 2187 cb->cookie = cookie; 2188 cb->server = -1; 2189 cb->state = IDLE; 2190 cb->size = 64; 2191 cb->txdepth = RPING_SQ_DEPTH; 2192 cb->mem = DMA; 2193 cb->addr_type = AF_INET; 2194 init_waitqueue_head(&cb->sem); 2195 2196 while ((op = krping_getopt("krping", &cmd, krping_opts, NULL, &optarg, 2197 &optint)) != 0) { 2198 switch (op) { 2199 case 'a': 2200 cb->addr_str = optarg; 2201 cb->addr_type = AF_INET; 2202 DEBUG_LOG(cb, "ipv4addr (%s)\n", optarg); 2203 if (inet_pton(AF_INET, optarg, &cb->addr) != 1) { 2204 PRINTF(cb, "bad addr string %s\n", 2205 optarg); 2206 ret = EINVAL; 2207 } 2208 break; 2209 case 'A': 2210 cb->addr_str = optarg; 2211 cb->addr_type = AF_INET6; 2212 DEBUG_LOG(cb, "ipv6addr (%s)\n", optarg); 2213 scope = strstr(optarg, "%"); 2214 /* extract scope ID, if any */ 2215 if (scope != NULL) 2216 *scope++ = 0; 2217 /* extract IPv6 network address */ 2218 if (inet_pton(AF_INET6, optarg, &cb->addr) != 1) { 2219 PRINTF(cb, "bad addr string %s\n", 2220 optarg); 2221 ret = EINVAL; 2222 } else if (IN6_IS_SCOPE_LINKLOCAL(&cb->addr.v6) || 2223 IN6_IS_ADDR_MC_INTFACELOCAL(&cb->addr.v6)) { 2224 uint16_t scope_id = krping_get_ipv6_scope_id(scope); 2225 DEBUG_LOG(cb, "ipv6 scope ID = %d\n", scope_id); 2226 cb->addr.v6.s6_addr[2] = scope_id >> 8; 2227 cb->addr.v6.s6_addr[3] = scope_id & 0xFF; 2228 } 2229 break; 2230 case 'p': 2231 cb->port = htons(optint); 2232 DEBUG_LOG(cb, "port %d\n", (int)optint); 2233 break; 2234 case 'P': 2235 cb->poll = 1; 2236 DEBUG_LOG(cb, "server\n"); 2237 break; 2238 case 's': 2239 cb->server = 1; 2240 DEBUG_LOG(cb, "server\n"); 2241 break; 2242 case 'c': 2243 cb->server = 0; 2244 DEBUG_LOG(cb, "client\n"); 2245 break; 2246 case 'S': 2247 cb->size = optint; 2248 if ((cb->size < 1) || 2249 (cb->size > RPING_BUFSIZE)) { 2250 PRINTF(cb, "Invalid size %d " 2251 "(valid range is 1 to %d)\n", 2252 cb->size, RPING_BUFSIZE); 2253 ret = EINVAL; 2254 } else 2255 DEBUG_LOG(cb, "size %d\n", (int)optint); 2256 break; 2257 case 'C': 2258 cb->count = optint; 2259 if (cb->count < 0) { 2260 PRINTF(cb, "Invalid count %d\n", 2261 cb->count); 2262 ret = EINVAL; 2263 } else 2264 DEBUG_LOG(cb, "count %d\n", (int) cb->count); 2265 break; 2266 case 'v': 2267 cb->verbose++; 2268 DEBUG_LOG(cb, "verbose\n"); 2269 break; 2270 case 'V': 2271 cb->validate++; 2272 DEBUG_LOG(cb, "validate data\n"); 2273 break; 2274 case 'l': 2275 cb->wlat++; 2276 break; 2277 case 'L': 2278 cb->rlat++; 2279 break; 2280 case 'B': 2281 cb->bw++; 2282 break; 2283 case 'd': 2284 cb->duplex++; 2285 break; 2286 case 'm': 2287 if (!strncmp(optarg, "dma", 3)) 2288 cb->mem = DMA; 2289 else if (!strncmp(optarg, "fastreg", 7)) 2290 cb->mem = FASTREG; 2291 else if (!strncmp(optarg, "mw", 2)) 2292 cb->mem = MW; 2293 else if (!strncmp(optarg, "mr", 2)) 2294 cb->mem = MR; 2295 else { 2296 PRINTF(cb, "unknown mem mode %s. " 2297 "Must be dma, fastreg, mw, or mr\n", 2298 optarg); 2299 ret = -EINVAL; 2300 break; 2301 } 2302 break; 2303 case 'I': 2304 cb->server_invalidate = 1; 2305 break; 2306 case 'T': 2307 cb->txdepth = optint; 2308 DEBUG_LOG(cb, "txdepth %d\n", (int) cb->txdepth); 2309 break; 2310 case 'Z': 2311 cb->local_dma_lkey = 1; 2312 DEBUG_LOG(cb, "using local dma lkey\n"); 2313 break; 2314 case 'R': 2315 cb->read_inv = 1; 2316 DEBUG_LOG(cb, "using read-with-inv\n"); 2317 break; 2318 case 'f': 2319 cb->frtest = 1; 2320 DEBUG_LOG(cb, "fast-reg test!\n"); 2321 break; 2322 default: 2323 PRINTF(cb, "unknown opt %s\n", optarg); 2324 ret = -EINVAL; 2325 break; 2326 } 2327 } 2328 if (ret) 2329 goto out; 2330 2331 if (cb->server == -1) { 2332 PRINTF(cb, "must be either client or server\n"); 2333 ret = -EINVAL; 2334 goto out; 2335 } 2336 2337 if (cb->server && cb->frtest) { 2338 PRINTF(cb, "must be client to run frtest\n"); 2339 ret = -EINVAL; 2340 goto out; 2341 } 2342 2343 if ((cb->frtest + cb->bw + cb->rlat + cb->wlat) > 1) { 2344 PRINTF(cb, "Pick only one test: fr, bw, rlat, wlat\n"); 2345 ret = -EINVAL; 2346 goto out; 2347 } 2348 2349 if (cb->server_invalidate && cb->mem != FASTREG) { 2350 PRINTF(cb, "server_invalidate only valid with fastreg mem_mode\n"); 2351 ret = -EINVAL; 2352 goto out; 2353 } 2354 2355 if (cb->read_inv && cb->mem != FASTREG) { 2356 PRINTF(cb, "read_inv only valid with fastreg mem_mode\n"); 2357 ret = -EINVAL; 2358 goto out; 2359 } 2360 2361 if (cb->mem != MR && (cb->wlat || cb->rlat || cb->bw)) { 2362 PRINTF(cb, "wlat, rlat, and bw tests only support mem_mode MR\n"); 2363 ret = -EINVAL; 2364 goto out; 2365 } 2366 2367 cb->cm_id = rdma_create_id(krping_cma_event_handler, cb, RDMA_PS_TCP); 2368 if (IS_ERR(cb->cm_id)) { 2369 ret = PTR_ERR(cb->cm_id); 2370 PRINTF(cb, "rdma_create_id error %d\n", ret); 2371 goto out; 2372 } 2373 DEBUG_LOG(cb, "created cm_id %p\n", cb->cm_id); 2374 2375 if (cb->server) 2376 krping_run_server(cb); 2377 else 2378 krping_run_client(cb); 2379 2380 DEBUG_LOG(cb, "destroy cm_id %p\n", cb->cm_id); 2381 rdma_destroy_id(cb->cm_id); 2382out: 2383 mutex_lock(&krping_mutex); 2384 list_del(&cb->list); 2385 mutex_unlock(&krping_mutex); 2386 kfree(cb); 2387 return ret; 2388} 2389 2390void 2391krping_walk_cb_list(void (*f)(struct krping_stats *, void *), void *arg) 2392{ 2393 struct krping_cb *cb; 2394 2395 mutex_lock(&krping_mutex); 2396 list_for_each_entry(cb, &krping_cbs, list) 2397 (*f)(cb->pd ? &cb->stats : NULL, arg); 2398 mutex_unlock(&krping_mutex); 2399} 2400 2401void krping_init(void) 2402{ 2403 2404 mutex_init(&krping_mutex); 2405} 2406