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("$FreeBSD$"); 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 #include <linux/wait.h> 48 49 #include <asm/atomic.h> 50 51 #include <rdma/ib_verbs.h> 52 #include <rdma/rdma_cm.h> 53 54 #include "krping.h" 55 #include "getopt.h" 56 57 #define PFX "krping: " 58 59 extern int krping_debug; 60 #define DEBUG_LOG(...) do { if (krping_debug) log(LOG_INFO, __VA_ARGS__); } while (0) 61 #define BIND_INFO 1 62 63 MODULE_AUTHOR("Steve Wise"); 64 MODULE_DESCRIPTION("RDMA ping server"); 65 MODULE_LICENSE("Dual BSD/GPL"); 66 MODULE_VERSION(krping, 1); 67 MODULE_DEPEND(krping, linuxkpi, 1, 1, 1); 68 69 static __inline uint64_t 70 get_cycles(void) 71 { 72 #if defined(__amd64__) || defined(__i386__) 73 uint32_t low, high; 74 __asm __volatile("rdtsc" : "=a" (low), "=d" (high)); 75 return (low | ((u_int64_t)high << 32)); 76 #elif defined(__powerpc64__) 77 return __builtin_readcyclecounter(); 78 #endif 79 } 80 81 typedef uint64_t cycles_t; 82 83 enum mem_type { 84 DMA = 1, 85 REG = 2, 86 }; 87 88 static const struct krping_option krping_opts[] = { 89 {"count", OPT_INT, 'C'}, 90 {"size", OPT_INT, 'S'}, 91 {"addr", OPT_STRING, 'a'}, 92 {"addr6", OPT_STRING, 'A'}, 93 {"port", OPT_INT, 'p'}, 94 {"verbose", OPT_NOPARAM, 'v'}, 95 {"validate", OPT_NOPARAM, 'V'}, 96 {"server", OPT_NOPARAM, 's'}, 97 {"client", OPT_NOPARAM, 'c'}, 98 {"server_inv", OPT_NOPARAM, 'I'}, 99 {"wlat", OPT_NOPARAM, 'l'}, 100 {"rlat", OPT_NOPARAM, 'L'}, 101 {"bw", OPT_NOPARAM, 'B'}, 102 {"duplex", OPT_NOPARAM, 'd'}, 103 {"tos", OPT_INT, 't'}, 104 {"txdepth", OPT_INT, 'T'}, 105 {"poll", OPT_NOPARAM, 'P'}, 106 {"local_dma_lkey", OPT_NOPARAM, 'Z'}, 107 {"read_inv", OPT_NOPARAM, 'R'}, 108 {"fr", OPT_NOPARAM, 'f'}, 109 {NULL, 0, 0} 110 }; 111 112 #define htonll(x) cpu_to_be64((x)) 113 #define ntohll(x) cpu_to_be64((x)) 114 115 static DEFINE_MUTEX(krping_mutex); 116 117 /* 118 * List of running krping threads. 119 */ 120 static LIST_HEAD(krping_cbs); 121 122 /* 123 * Invoke like this, one on each side, using the server's address on 124 * the RDMA device (iw%d): 125 * 126 * /bin/echo server,port=9999,addr=192.168.69.142,validate > /proc/krping 127 * /bin/echo client,port=9999,addr=192.168.69.142,validate > /proc/krping 128 * /bin/echo client,port=9999,addr6=2001:db8:0:f101::1,validate > /proc/krping 129 * 130 * krping "ping/pong" loop: 131 * client sends source rkey/addr/len 132 * server receives source rkey/add/len 133 * server rdma reads "ping" data from source 134 * server sends "go ahead" on rdma read completion 135 * client sends sink rkey/addr/len 136 * server receives sink rkey/addr/len 137 * server rdma writes "pong" data to sink 138 * server sends "go ahead" on rdma write completion 139 * <repeat loop> 140 */ 141 142 /* 143 * These states are used to signal events between the completion handler 144 * and the main client or server thread. 145 * 146 * Once CONNECTED, they cycle through RDMA_READ_ADV, RDMA_WRITE_ADV, 147 * and RDMA_WRITE_COMPLETE for each ping. 148 */ 149 enum test_state { 150 IDLE = 1, 151 CONNECT_REQUEST, 152 ADDR_RESOLVED, 153 ROUTE_RESOLVED, 154 CONNECTED, 155 RDMA_READ_ADV, 156 RDMA_READ_COMPLETE, 157 RDMA_WRITE_ADV, 158 RDMA_WRITE_COMPLETE, 159 ERROR 160 }; 161 162 struct krping_rdma_info { 163 uint64_t buf; 164 uint32_t rkey; 165 uint32_t size; 166 }; 167 168 /* 169 * Default max buffer size for IO... 170 */ 171 #define RPING_BUFSIZE 128*1024 172 #define RPING_SQ_DEPTH 64 173 174 /* 175 * Control block struct. 176 */ 177 struct krping_cb { 178 int server; /* 0 iff client */ 179 struct ib_cq *cq; 180 struct ib_pd *pd; 181 struct ib_qp *qp; 182 183 struct ib_mr *dma_mr; 184 185 struct ib_fast_reg_page_list *page_list; 186 int page_list_len; 187 struct ib_reg_wr reg_mr_wr; 188 struct ib_send_wr invalidate_wr; 189 struct ib_mr *reg_mr; 190 int server_invalidate; 191 int read_inv; 192 u8 key; 193 194 struct ib_recv_wr rq_wr; /* recv work request record */ 195 struct ib_sge recv_sgl; /* recv single SGE */ 196 struct krping_rdma_info recv_buf __aligned(16); /* malloc'd buffer */ 197 u64 recv_dma_addr; 198 DECLARE_PCI_UNMAP_ADDR(recv_mapping) 199 200 struct ib_send_wr sq_wr; /* send work requrest record */ 201 struct ib_sge send_sgl; 202 struct krping_rdma_info send_buf __aligned(16); /* single send buf */ 203 u64 send_dma_addr; 204 DECLARE_PCI_UNMAP_ADDR(send_mapping) 205 206 struct ib_rdma_wr rdma_sq_wr; /* rdma work request record */ 207 struct ib_sge rdma_sgl; /* rdma single SGE */ 208 char *rdma_buf; /* used as rdma sink */ 209 u64 rdma_dma_addr; 210 DECLARE_PCI_UNMAP_ADDR(rdma_mapping) 211 struct ib_mr *rdma_mr; 212 213 uint32_t remote_rkey; /* remote guys RKEY */ 214 uint64_t remote_addr; /* remote guys TO */ 215 uint32_t remote_len; /* remote guys LEN */ 216 217 char *start_buf; /* rdma read src */ 218 u64 start_dma_addr; 219 DECLARE_PCI_UNMAP_ADDR(start_mapping) 220 struct ib_mr *start_mr; 221 222 enum test_state state; /* used for cond/signalling */ 223 wait_queue_head_t sem; 224 struct krping_stats stats; 225 226 uint16_t port; /* dst port in NBO */ 227 u8 addr[16] __aligned(8); /* dst addr in NBO */ 228 char *addr_str; /* dst addr string */ 229 uint8_t addr_type; /* ADDR_FAMILY - IPv4/V6 */ 230 int verbose; /* verbose logging */ 231 int count; /* ping count */ 232 int size; /* ping data size */ 233 int validate; /* validate ping data */ 234 int wlat; /* run wlat test */ 235 int rlat; /* run rlat test */ 236 int bw; /* run bw test */ 237 int duplex; /* run bw full duplex test */ 238 int poll; /* poll or block for rlat test */ 239 int txdepth; /* SQ depth */ 240 int local_dma_lkey; /* use 0 for lkey */ 241 int frtest; /* reg test */ 242 int tos; /* type of service */ 243 244 /* CM stuff */ 245 struct rdma_cm_id *cm_id; /* connection on client side,*/ 246 /* listener on server side. */ 247 struct rdma_cm_id *child_cm_id; /* connection on server side */ 248 struct list_head list; 249 }; 250 251 static int krping_cma_event_handler(struct rdma_cm_id *cma_id, 252 struct rdma_cm_event *event) 253 { 254 int ret; 255 struct krping_cb *cb = cma_id->context; 256 257 DEBUG_LOG("cma_event type %d cma_id %p (%s)\n", event->event, cma_id, 258 (cma_id == cb->cm_id) ? "parent" : "child"); 259 260 switch (event->event) { 261 case RDMA_CM_EVENT_ADDR_RESOLVED: 262 cb->state = ADDR_RESOLVED; 263 ret = rdma_resolve_route(cma_id, 2000); 264 if (ret) { 265 printk(KERN_ERR PFX "rdma_resolve_route error %d\n", 266 ret); 267 wake_up_interruptible(&cb->sem); 268 } 269 break; 270 271 case RDMA_CM_EVENT_ROUTE_RESOLVED: 272 cb->state = ROUTE_RESOLVED; 273 wake_up_interruptible(&cb->sem); 274 break; 275 276 case RDMA_CM_EVENT_CONNECT_REQUEST: 277 cb->state = CONNECT_REQUEST; 278 cb->child_cm_id = cma_id; 279 DEBUG_LOG("child cma %p\n", cb->child_cm_id); 280 wake_up_interruptible(&cb->sem); 281 break; 282 283 case RDMA_CM_EVENT_ESTABLISHED: 284 DEBUG_LOG("ESTABLISHED\n"); 285 if (!cb->server) { 286 cb->state = CONNECTED; 287 } 288 wake_up_interruptible(&cb->sem); 289 break; 290 291 case RDMA_CM_EVENT_ADDR_ERROR: 292 case RDMA_CM_EVENT_ROUTE_ERROR: 293 case RDMA_CM_EVENT_CONNECT_ERROR: 294 case RDMA_CM_EVENT_UNREACHABLE: 295 case RDMA_CM_EVENT_REJECTED: 296 printk(KERN_ERR PFX "cma event %d, error %d\n", event->event, 297 event->status); 298 cb->state = ERROR; 299 wake_up_interruptible(&cb->sem); 300 break; 301 302 case RDMA_CM_EVENT_DISCONNECTED: 303 printk(KERN_ERR PFX "DISCONNECT EVENT...\n"); 304 cb->state = ERROR; 305 wake_up_interruptible(&cb->sem); 306 break; 307 308 case RDMA_CM_EVENT_DEVICE_REMOVAL: 309 printk(KERN_ERR PFX "cma detected device removal!!!!\n"); 310 cb->state = ERROR; 311 wake_up_interruptible(&cb->sem); 312 break; 313 314 default: 315 printk(KERN_ERR PFX "oof bad type!\n"); 316 wake_up_interruptible(&cb->sem); 317 break; 318 } 319 return 0; 320 } 321 322 static int server_recv(struct krping_cb *cb, struct ib_wc *wc) 323 { 324 if (wc->byte_len != sizeof(cb->recv_buf)) { 325 printk(KERN_ERR PFX "Received bogus data, size %d\n", 326 wc->byte_len); 327 return -1; 328 } 329 330 cb->remote_rkey = ntohl(cb->recv_buf.rkey); 331 cb->remote_addr = ntohll(cb->recv_buf.buf); 332 cb->remote_len = ntohl(cb->recv_buf.size); 333 DEBUG_LOG("Received rkey %x addr %llx len %d from peer\n", 334 cb->remote_rkey, (unsigned long long)cb->remote_addr, 335 cb->remote_len); 336 337 if (cb->state <= CONNECTED || cb->state == RDMA_WRITE_COMPLETE) 338 cb->state = RDMA_READ_ADV; 339 else 340 cb->state = RDMA_WRITE_ADV; 341 342 return 0; 343 } 344 345 static int client_recv(struct krping_cb *cb, struct ib_wc *wc) 346 { 347 if (wc->byte_len != sizeof(cb->recv_buf)) { 348 printk(KERN_ERR PFX "Received bogus data, size %d\n", 349 wc->byte_len); 350 return -1; 351 } 352 353 if (cb->state == RDMA_READ_ADV) 354 cb->state = RDMA_WRITE_ADV; 355 else 356 cb->state = RDMA_WRITE_COMPLETE; 357 358 return 0; 359 } 360 361 static void krping_cq_event_handler(struct ib_cq *cq, void *ctx) 362 { 363 struct krping_cb *cb = ctx; 364 struct ib_wc wc; 365 const struct ib_recv_wr *bad_wr; 366 int ret; 367 368 BUG_ON(cb->cq != cq); 369 if (cb->frtest) { 370 printk(KERN_ERR PFX "cq completion event in frtest!\n"); 371 return; 372 } 373 if (!cb->wlat && !cb->rlat && !cb->bw) 374 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); 375 while ((ret = ib_poll_cq(cb->cq, 1, &wc)) == 1) { 376 if (wc.status) { 377 if (wc.status == IB_WC_WR_FLUSH_ERR) { 378 DEBUG_LOG("cq flushed\n"); 379 continue; 380 } else { 381 printk(KERN_ERR PFX "cq completion failed with " 382 "wr_id %jx status %d opcode %d vender_err %x\n", 383 (uintmax_t)wc.wr_id, wc.status, wc.opcode, wc.vendor_err); 384 goto error; 385 } 386 } 387 if (cb->state == ERROR) { 388 printk(KERN_ERR PFX "cq completion in ERROR state\n"); 389 return; 390 } 391 switch (wc.opcode) { 392 case IB_WC_SEND: 393 DEBUG_LOG("send completion\n"); 394 cb->stats.send_bytes += cb->send_sgl.length; 395 cb->stats.send_msgs++; 396 break; 397 398 case IB_WC_RDMA_WRITE: 399 DEBUG_LOG("rdma write completion\n"); 400 cb->stats.write_bytes += cb->rdma_sq_wr.wr.sg_list->length; 401 cb->stats.write_msgs++; 402 cb->state = RDMA_WRITE_COMPLETE; 403 wake_up_interruptible(&cb->sem); 404 break; 405 406 case IB_WC_RDMA_READ: 407 DEBUG_LOG("rdma read completion\n"); 408 cb->stats.read_bytes += cb->rdma_sq_wr.wr.sg_list->length; 409 cb->stats.read_msgs++; 410 cb->state = RDMA_READ_COMPLETE; 411 wake_up_interruptible(&cb->sem); 412 break; 413 414 case IB_WC_RECV: 415 DEBUG_LOG("recv completion\n"); 416 cb->stats.recv_bytes += sizeof(cb->recv_buf); 417 cb->stats.recv_msgs++; 418 if (cb->wlat || cb->rlat || cb->bw) 419 ret = server_recv(cb, &wc); 420 else 421 ret = cb->server ? server_recv(cb, &wc) : 422 client_recv(cb, &wc); 423 if (ret) { 424 printk(KERN_ERR PFX "recv wc error: %d\n", ret); 425 goto error; 426 } 427 428 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); 429 if (ret) { 430 printk(KERN_ERR PFX "post recv error: %d\n", 431 ret); 432 goto error; 433 } 434 wake_up_interruptible(&cb->sem); 435 break; 436 437 default: 438 printk(KERN_ERR PFX 439 "%s:%d Unexpected opcode %d, Shutting down\n", 440 __func__, __LINE__, wc.opcode); 441 goto error; 442 } 443 } 444 if (ret) { 445 printk(KERN_ERR PFX "poll error %d\n", ret); 446 goto error; 447 } 448 return; 449 error: 450 cb->state = ERROR; 451 wake_up_interruptible(&cb->sem); 452 } 453 454 static int krping_accept(struct krping_cb *cb) 455 { 456 struct rdma_conn_param conn_param; 457 int ret; 458 459 DEBUG_LOG("accepting client connection request\n"); 460 461 memset(&conn_param, 0, sizeof conn_param); 462 conn_param.responder_resources = 1; 463 conn_param.initiator_depth = 1; 464 465 ret = rdma_accept(cb->child_cm_id, &conn_param); 466 if (ret) { 467 printk(KERN_ERR PFX "rdma_accept error: %d\n", ret); 468 return ret; 469 } 470 471 if (!cb->wlat && !cb->rlat && !cb->bw) { 472 wait_event_interruptible(cb->sem, cb->state >= CONNECTED); 473 if (cb->state == ERROR) { 474 printk(KERN_ERR PFX "wait for CONNECTED state %d\n", 475 cb->state); 476 return -1; 477 } 478 } 479 return 0; 480 } 481 482 static void krping_setup_wr(struct krping_cb *cb) 483 { 484 cb->recv_sgl.addr = cb->recv_dma_addr; 485 cb->recv_sgl.length = sizeof cb->recv_buf; 486 cb->recv_sgl.lkey = cb->pd->local_dma_lkey; 487 cb->rq_wr.sg_list = &cb->recv_sgl; 488 cb->rq_wr.num_sge = 1; 489 490 cb->send_sgl.addr = cb->send_dma_addr; 491 cb->send_sgl.length = sizeof cb->send_buf; 492 cb->send_sgl.lkey = cb->pd->local_dma_lkey; 493 494 cb->sq_wr.opcode = IB_WR_SEND; 495 cb->sq_wr.send_flags = IB_SEND_SIGNALED; 496 cb->sq_wr.sg_list = &cb->send_sgl; 497 cb->sq_wr.num_sge = 1; 498 499 if (cb->server || cb->wlat || cb->rlat || cb->bw) { 500 cb->rdma_sgl.addr = cb->rdma_dma_addr; 501 cb->rdma_sq_wr.wr.send_flags = IB_SEND_SIGNALED; 502 cb->rdma_sq_wr.wr.sg_list = &cb->rdma_sgl; 503 cb->rdma_sq_wr.wr.num_sge = 1; 504 } 505 506 /* 507 * A chain of 2 WRs, INVALDATE_MR + REG_MR. 508 * both unsignaled. The client uses them to reregister 509 * the rdma buffers with a new key each iteration. 510 */ 511 cb->reg_mr_wr.wr.opcode = IB_WR_REG_MR; 512 cb->reg_mr_wr.mr = cb->reg_mr; 513 514 cb->invalidate_wr.next = &cb->reg_mr_wr.wr; 515 cb->invalidate_wr.opcode = IB_WR_LOCAL_INV; 516 } 517 518 static int krping_setup_buffers(struct krping_cb *cb) 519 { 520 int ret; 521 522 DEBUG_LOG(PFX "krping_setup_buffers called on cb %p\n", cb); 523 524 cb->recv_dma_addr = ib_dma_map_single(cb->pd->device, 525 &cb->recv_buf, 526 sizeof(cb->recv_buf), DMA_BIDIRECTIONAL); 527 pci_unmap_addr_set(cb, recv_mapping, cb->recv_dma_addr); 528 cb->send_dma_addr = ib_dma_map_single(cb->pd->device, 529 &cb->send_buf, sizeof(cb->send_buf), 530 DMA_BIDIRECTIONAL); 531 pci_unmap_addr_set(cb, send_mapping, cb->send_dma_addr); 532 533 cb->rdma_buf = ib_dma_alloc_coherent(cb->pd->device, cb->size, 534 &cb->rdma_dma_addr, 535 GFP_KERNEL); 536 if (!cb->rdma_buf) { 537 DEBUG_LOG(PFX "rdma_buf allocation failed\n"); 538 ret = -ENOMEM; 539 goto bail; 540 } 541 pci_unmap_addr_set(cb, rdma_mapping, cb->rdma_dma_addr); 542 cb->page_list_len = (((cb->size - 1) & PAGE_MASK) + PAGE_SIZE) 543 >> PAGE_SHIFT; 544 cb->reg_mr = ib_alloc_mr(cb->pd, IB_MR_TYPE_MEM_REG, 545 cb->page_list_len); 546 if (IS_ERR(cb->reg_mr)) { 547 ret = PTR_ERR(cb->reg_mr); 548 DEBUG_LOG(PFX "recv_buf reg_mr failed %d\n", ret); 549 goto bail; 550 } 551 DEBUG_LOG(PFX "reg rkey 0x%x page_list_len %u\n", 552 cb->reg_mr->rkey, cb->page_list_len); 553 554 if (!cb->server || cb->wlat || cb->rlat || cb->bw) { 555 556 cb->start_buf = ib_dma_alloc_coherent(cb->pd->device, cb->size, 557 &cb->start_dma_addr, 558 GFP_KERNEL); 559 if (!cb->start_buf) { 560 DEBUG_LOG(PFX "start_buf malloc failed\n"); 561 ret = -ENOMEM; 562 goto bail; 563 } 564 pci_unmap_addr_set(cb, start_mapping, cb->start_dma_addr); 565 } 566 567 krping_setup_wr(cb); 568 DEBUG_LOG(PFX "allocated & registered buffers...\n"); 569 return 0; 570 bail: 571 if (cb->reg_mr && !IS_ERR(cb->reg_mr)) 572 ib_dereg_mr(cb->reg_mr); 573 if (cb->rdma_mr && !IS_ERR(cb->rdma_mr)) 574 ib_dereg_mr(cb->rdma_mr); 575 if (cb->dma_mr && !IS_ERR(cb->dma_mr)) 576 ib_dereg_mr(cb->dma_mr); 577 if (cb->rdma_buf) { 578 ib_dma_free_coherent(cb->pd->device, cb->size, cb->rdma_buf, 579 cb->rdma_dma_addr); 580 } 581 if (cb->start_buf) { 582 ib_dma_free_coherent(cb->pd->device, cb->size, cb->start_buf, 583 cb->start_dma_addr); 584 } 585 return ret; 586 } 587 588 static void krping_free_buffers(struct krping_cb *cb) 589 { 590 DEBUG_LOG("krping_free_buffers called on cb %p\n", cb); 591 592 if (cb->dma_mr) 593 ib_dereg_mr(cb->dma_mr); 594 if (cb->rdma_mr) 595 ib_dereg_mr(cb->rdma_mr); 596 if (cb->start_mr) 597 ib_dereg_mr(cb->start_mr); 598 if (cb->reg_mr) 599 ib_dereg_mr(cb->reg_mr); 600 601 dma_unmap_single(cb->pd->device->dma_device, 602 pci_unmap_addr(cb, recv_mapping), 603 sizeof(cb->recv_buf), DMA_BIDIRECTIONAL); 604 dma_unmap_single(cb->pd->device->dma_device, 605 pci_unmap_addr(cb, send_mapping), 606 sizeof(cb->send_buf), DMA_BIDIRECTIONAL); 607 608 ib_dma_free_coherent(cb->pd->device, cb->size, cb->rdma_buf, 609 cb->rdma_dma_addr); 610 611 if (cb->start_buf) { 612 ib_dma_free_coherent(cb->pd->device, cb->size, cb->start_buf, 613 cb->start_dma_addr); 614 } 615 } 616 617 static int krping_create_qp(struct krping_cb *cb) 618 { 619 struct ib_qp_init_attr init_attr; 620 int ret; 621 622 memset(&init_attr, 0, sizeof(init_attr)); 623 init_attr.cap.max_send_wr = cb->txdepth; 624 init_attr.cap.max_recv_wr = 2; 625 626 /* For flush_qp() */ 627 init_attr.cap.max_send_wr++; 628 init_attr.cap.max_recv_wr++; 629 630 init_attr.cap.max_recv_sge = 1; 631 init_attr.cap.max_send_sge = 1; 632 init_attr.qp_type = IB_QPT_RC; 633 init_attr.send_cq = cb->cq; 634 init_attr.recv_cq = cb->cq; 635 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 636 637 if (cb->server) { 638 ret = rdma_create_qp(cb->child_cm_id, cb->pd, &init_attr); 639 if (!ret) 640 cb->qp = cb->child_cm_id->qp; 641 } else { 642 ret = rdma_create_qp(cb->cm_id, cb->pd, &init_attr); 643 if (!ret) 644 cb->qp = cb->cm_id->qp; 645 } 646 647 return ret; 648 } 649 650 static void krping_free_qp(struct krping_cb *cb) 651 { 652 ib_destroy_qp(cb->qp); 653 ib_destroy_cq(cb->cq); 654 ib_dealloc_pd(cb->pd); 655 } 656 657 static int krping_setup_qp(struct krping_cb *cb, struct rdma_cm_id *cm_id) 658 { 659 int ret; 660 struct ib_cq_init_attr attr = {0}; 661 662 cb->pd = ib_alloc_pd(cm_id->device, 0); 663 if (IS_ERR(cb->pd)) { 664 printk(KERN_ERR PFX "ib_alloc_pd failed\n"); 665 return PTR_ERR(cb->pd); 666 } 667 DEBUG_LOG("created pd %p\n", cb->pd); 668 669 strlcpy(cb->stats.name, cb->pd->device->name, sizeof(cb->stats.name)); 670 671 attr.cqe = cb->txdepth * 2; 672 attr.comp_vector = 0; 673 cb->cq = ib_create_cq(cm_id->device, krping_cq_event_handler, NULL, 674 cb, &attr); 675 if (IS_ERR(cb->cq)) { 676 printk(KERN_ERR PFX "ib_create_cq failed\n"); 677 ret = PTR_ERR(cb->cq); 678 goto err1; 679 } 680 DEBUG_LOG("created cq %p\n", cb->cq); 681 682 if (!cb->wlat && !cb->rlat && !cb->bw && !cb->frtest) { 683 ret = ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); 684 if (ret) { 685 printk(KERN_ERR PFX "ib_create_cq failed\n"); 686 goto err2; 687 } 688 } 689 690 ret = krping_create_qp(cb); 691 if (ret) { 692 printk(KERN_ERR PFX "krping_create_qp failed: %d\n", ret); 693 goto err2; 694 } 695 DEBUG_LOG("created qp %p\n", cb->qp); 696 return 0; 697 err2: 698 ib_destroy_cq(cb->cq); 699 err1: 700 ib_dealloc_pd(cb->pd); 701 return ret; 702 } 703 704 /* 705 * return the (possibly rebound) rkey for the rdma buffer. 706 * REG mode: invalidate and rebind via reg wr. 707 * other modes: just return the mr rkey. 708 */ 709 static u32 krping_rdma_rkey(struct krping_cb *cb, u64 buf, int post_inv) 710 { 711 u32 rkey; 712 const struct ib_send_wr *bad_wr; 713 int ret; 714 struct scatterlist sg = {0}; 715 716 cb->invalidate_wr.ex.invalidate_rkey = cb->reg_mr->rkey; 717 718 /* 719 * Update the reg key. 720 */ 721 ib_update_fast_reg_key(cb->reg_mr, ++cb->key); 722 cb->reg_mr_wr.key = cb->reg_mr->rkey; 723 724 /* 725 * Update the reg WR with new buf info. 726 */ 727 if (buf == (u64)cb->start_dma_addr) 728 cb->reg_mr_wr.access = IB_ACCESS_REMOTE_READ; 729 else 730 cb->reg_mr_wr.access = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE; 731 sg_dma_address(&sg) = buf; 732 sg_dma_len(&sg) = cb->size; 733 734 ret = ib_map_mr_sg(cb->reg_mr, &sg, 1, NULL, PAGE_SIZE); 735 BUG_ON(ret <= 0 || ret > cb->page_list_len); 736 737 DEBUG_LOG(PFX "post_inv = %d, reg_mr new rkey 0x%x pgsz %u len %u" 738 " iova_start %llx\n", 739 post_inv, 740 cb->reg_mr_wr.key, 741 cb->reg_mr->page_size, 742 (unsigned)cb->reg_mr->length, 743 (unsigned long long)cb->reg_mr->iova); 744 745 if (post_inv) 746 ret = ib_post_send(cb->qp, &cb->invalidate_wr, &bad_wr); 747 else 748 ret = ib_post_send(cb->qp, &cb->reg_mr_wr.wr, &bad_wr); 749 if (ret) { 750 printk(KERN_ERR PFX "post send error %d\n", ret); 751 cb->state = ERROR; 752 } 753 rkey = cb->reg_mr->rkey; 754 return rkey; 755 } 756 757 static void krping_format_send(struct krping_cb *cb, u64 buf) 758 { 759 struct krping_rdma_info *info = &cb->send_buf; 760 u32 rkey; 761 762 /* 763 * Client side will do reg or mw bind before 764 * advertising the rdma buffer. Server side 765 * sends have no data. 766 */ 767 if (!cb->server || cb->wlat || cb->rlat || cb->bw) { 768 rkey = krping_rdma_rkey(cb, buf, !cb->server_invalidate); 769 info->buf = htonll(buf); 770 info->rkey = htonl(rkey); 771 info->size = htonl(cb->size); 772 DEBUG_LOG("RDMA addr %llx rkey %x len %d\n", 773 (unsigned long long)buf, rkey, cb->size); 774 } 775 } 776 777 static void krping_test_server(struct krping_cb *cb) 778 { 779 const struct ib_send_wr *bad_wr; 780 struct ib_send_wr inv; 781 int ret; 782 783 while (1) { 784 /* Wait for client's Start STAG/TO/Len */ 785 wait_event_interruptible(cb->sem, cb->state >= RDMA_READ_ADV); 786 if (cb->state != RDMA_READ_ADV) { 787 printk(KERN_ERR PFX "wait for RDMA_READ_ADV state %d\n", 788 cb->state); 789 break; 790 } 791 792 DEBUG_LOG("server received sink adv\n"); 793 794 cb->rdma_sq_wr.rkey = cb->remote_rkey; 795 cb->rdma_sq_wr.remote_addr = cb->remote_addr; 796 cb->rdma_sq_wr.wr.sg_list->length = cb->remote_len; 797 cb->rdma_sgl.lkey = krping_rdma_rkey(cb, cb->rdma_dma_addr, !cb->read_inv); 798 cb->rdma_sq_wr.wr.next = NULL; 799 800 /* Issue RDMA Read. */ 801 if (cb->read_inv) 802 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV; 803 else { 804 805 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_READ; 806 /* 807 * Immediately follow the read with a 808 * fenced LOCAL_INV. 809 */ 810 cb->rdma_sq_wr.wr.next = &inv; 811 memset(&inv, 0, sizeof inv); 812 inv.opcode = IB_WR_LOCAL_INV; 813 inv.ex.invalidate_rkey = cb->reg_mr->rkey; 814 inv.send_flags = IB_SEND_FENCE; 815 } 816 817 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr.wr, &bad_wr); 818 if (ret) { 819 printk(KERN_ERR PFX "post send error %d\n", ret); 820 break; 821 } 822 cb->rdma_sq_wr.wr.next = NULL; 823 824 DEBUG_LOG("server posted rdma read req \n"); 825 826 /* Wait for read completion */ 827 wait_event_interruptible(cb->sem, 828 cb->state >= RDMA_READ_COMPLETE); 829 if (cb->state != RDMA_READ_COMPLETE) { 830 printk(KERN_ERR PFX 831 "wait for RDMA_READ_COMPLETE state %d\n", 832 cb->state); 833 break; 834 } 835 DEBUG_LOG("server received read complete\n"); 836 837 /* Display data in recv buf */ 838 if (cb->verbose) 839 printk(KERN_INFO PFX "server ping data: %s\n", 840 cb->rdma_buf); 841 842 /* Tell client to continue */ 843 if (cb->server && cb->server_invalidate) { 844 cb->sq_wr.ex.invalidate_rkey = cb->remote_rkey; 845 cb->sq_wr.opcode = IB_WR_SEND_WITH_INV; 846 DEBUG_LOG("send-w-inv rkey 0x%x\n", cb->remote_rkey); 847 } 848 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 849 if (ret) { 850 printk(KERN_ERR PFX "post send error %d\n", ret); 851 break; 852 } 853 DEBUG_LOG("server posted go ahead\n"); 854 855 /* Wait for client's RDMA STAG/TO/Len */ 856 wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_ADV); 857 if (cb->state != RDMA_WRITE_ADV) { 858 printk(KERN_ERR PFX 859 "wait for RDMA_WRITE_ADV state %d\n", 860 cb->state); 861 break; 862 } 863 DEBUG_LOG("server received sink adv\n"); 864 865 /* RDMA Write echo data */ 866 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_WRITE; 867 cb->rdma_sq_wr.rkey = cb->remote_rkey; 868 cb->rdma_sq_wr.remote_addr = cb->remote_addr; 869 cb->rdma_sq_wr.wr.sg_list->length = strlen(cb->rdma_buf) + 1; 870 if (cb->local_dma_lkey) 871 cb->rdma_sgl.lkey = cb->pd->local_dma_lkey; 872 else 873 cb->rdma_sgl.lkey = krping_rdma_rkey(cb, cb->rdma_dma_addr, 0); 874 875 DEBUG_LOG("rdma write from lkey %x laddr %llx len %d\n", 876 cb->rdma_sq_wr.wr.sg_list->lkey, 877 (unsigned long long)cb->rdma_sq_wr.wr.sg_list->addr, 878 cb->rdma_sq_wr.wr.sg_list->length); 879 880 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr.wr, &bad_wr); 881 if (ret) { 882 printk(KERN_ERR PFX "post send error %d\n", ret); 883 break; 884 } 885 886 /* Wait for completion */ 887 ret = wait_event_interruptible(cb->sem, cb->state >= 888 RDMA_WRITE_COMPLETE); 889 if (cb->state != RDMA_WRITE_COMPLETE) { 890 printk(KERN_ERR PFX 891 "wait for RDMA_WRITE_COMPLETE state %d\n", 892 cb->state); 893 break; 894 } 895 DEBUG_LOG("server rdma write complete \n"); 896 897 cb->state = CONNECTED; 898 899 /* Tell client to begin again */ 900 if (cb->server && cb->server_invalidate) { 901 cb->sq_wr.ex.invalidate_rkey = cb->remote_rkey; 902 cb->sq_wr.opcode = IB_WR_SEND_WITH_INV; 903 DEBUG_LOG("send-w-inv rkey 0x%x\n", cb->remote_rkey); 904 } 905 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 906 if (ret) { 907 printk(KERN_ERR PFX "post send error %d\n", ret); 908 break; 909 } 910 DEBUG_LOG("server posted go ahead\n"); 911 } 912 } 913 914 static void rlat_test(struct krping_cb *cb) 915 { 916 int scnt; 917 int iters = cb->count; 918 struct timeval start_tv, stop_tv; 919 int ret; 920 struct ib_wc wc; 921 const struct ib_send_wr *bad_wr; 922 int ne; 923 924 scnt = 0; 925 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_READ; 926 cb->rdma_sq_wr.rkey = cb->remote_rkey; 927 cb->rdma_sq_wr.remote_addr = cb->remote_addr; 928 cb->rdma_sq_wr.wr.sg_list->length = cb->size; 929 930 microtime(&start_tv); 931 if (!cb->poll) { 932 cb->state = RDMA_READ_ADV; 933 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP); 934 } 935 while (scnt < iters) { 936 937 cb->state = RDMA_READ_ADV; 938 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr.wr, &bad_wr); 939 if (ret) { 940 printk(KERN_ERR PFX 941 "Couldn't post send: ret=%d scnt %d\n", 942 ret, scnt); 943 return; 944 } 945 946 do { 947 if (!cb->poll) { 948 wait_event_interruptible(cb->sem, 949 cb->state != RDMA_READ_ADV); 950 if (cb->state == RDMA_READ_COMPLETE) { 951 ne = 1; 952 ib_req_notify_cq(cb->cq, 953 IB_CQ_NEXT_COMP); 954 } else { 955 ne = -1; 956 } 957 } else 958 ne = ib_poll_cq(cb->cq, 1, &wc); 959 if (cb->state == ERROR) { 960 printk(KERN_ERR PFX 961 "state == ERROR...bailing scnt %d\n", 962 scnt); 963 return; 964 } 965 } while (ne == 0); 966 967 if (ne < 0) { 968 printk(KERN_ERR PFX "poll CQ failed %d\n", ne); 969 return; 970 } 971 if (cb->poll && wc.status != IB_WC_SUCCESS) { 972 printk(KERN_ERR PFX "Completion wth error at %s:\n", 973 cb->server ? "server" : "client"); 974 printk(KERN_ERR PFX "Failed status %d: wr_id %d\n", 975 wc.status, (int) wc.wr_id); 976 return; 977 } 978 ++scnt; 979 } 980 microtime(&stop_tv); 981 982 if (stop_tv.tv_usec < start_tv.tv_usec) { 983 stop_tv.tv_usec += 1000000; 984 stop_tv.tv_sec -= 1; 985 } 986 987 printk(KERN_ERR PFX "delta sec %lu delta usec %lu iter %d size %d\n", 988 (unsigned long)(stop_tv.tv_sec - start_tv.tv_sec), 989 (unsigned long)(stop_tv.tv_usec - start_tv.tv_usec), 990 scnt, cb->size); 991 } 992 993 static void wlat_test(struct krping_cb *cb) 994 { 995 int ccnt, scnt, rcnt; 996 int iters=cb->count; 997 volatile char *poll_buf = (char *) cb->start_buf; 998 char *buf = (char *)cb->rdma_buf; 999 struct timeval start_tv, stop_tv; 1000 cycles_t *post_cycles_start = NULL; 1001 cycles_t *post_cycles_stop = NULL; 1002 cycles_t *poll_cycles_start = NULL; 1003 cycles_t *poll_cycles_stop = NULL; 1004 cycles_t *last_poll_cycles_start = NULL; 1005 cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0; 1006 int i; 1007 int cycle_iters = 1000; 1008 1009 ccnt = 0; 1010 scnt = 0; 1011 rcnt = 0; 1012 1013 post_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1014 if (!post_cycles_start) { 1015 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1016 goto done; 1017 } 1018 post_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1019 if (!post_cycles_stop) { 1020 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1021 goto done; 1022 } 1023 poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1024 if (!poll_cycles_start) { 1025 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1026 goto done; 1027 } 1028 poll_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1029 if (!poll_cycles_stop) { 1030 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1031 goto done; 1032 } 1033 last_poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), 1034 GFP_KERNEL); 1035 if (!last_poll_cycles_start) { 1036 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1037 goto done; 1038 } 1039 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_WRITE; 1040 cb->rdma_sq_wr.rkey = cb->remote_rkey; 1041 cb->rdma_sq_wr.remote_addr = cb->remote_addr; 1042 cb->rdma_sq_wr.wr.sg_list->length = cb->size; 1043 1044 if (cycle_iters > iters) 1045 cycle_iters = iters; 1046 microtime(&start_tv); 1047 while (scnt < iters || ccnt < iters || rcnt < iters) { 1048 1049 /* Wait till buffer changes. */ 1050 if (rcnt < iters && !(scnt < 1 && !cb->server)) { 1051 ++rcnt; 1052 while (*poll_buf != (char)rcnt) { 1053 if (cb->state == ERROR) { 1054 printk(KERN_ERR PFX 1055 "state = ERROR, bailing\n"); 1056 goto done; 1057 } 1058 } 1059 } 1060 1061 if (scnt < iters) { 1062 const struct ib_send_wr *bad_wr; 1063 1064 *buf = (char)scnt+1; 1065 if (scnt < cycle_iters) 1066 post_cycles_start[scnt] = get_cycles(); 1067 if (ib_post_send(cb->qp, &cb->rdma_sq_wr.wr, &bad_wr)) { 1068 printk(KERN_ERR PFX 1069 "Couldn't post send: scnt=%d\n", 1070 scnt); 1071 goto done; 1072 } 1073 if (scnt < cycle_iters) 1074 post_cycles_stop[scnt] = get_cycles(); 1075 scnt++; 1076 } 1077 1078 if (ccnt < iters) { 1079 struct ib_wc wc; 1080 int ne; 1081 1082 if (ccnt < cycle_iters) 1083 poll_cycles_start[ccnt] = get_cycles(); 1084 do { 1085 if (ccnt < cycle_iters) 1086 last_poll_cycles_start[ccnt] = 1087 get_cycles(); 1088 ne = ib_poll_cq(cb->cq, 1, &wc); 1089 } while (ne == 0); 1090 if (ccnt < cycle_iters) 1091 poll_cycles_stop[ccnt] = get_cycles(); 1092 ++ccnt; 1093 1094 if (ne < 0) { 1095 printk(KERN_ERR PFX "poll CQ failed %d\n", ne); 1096 goto done; 1097 } 1098 if (wc.status != IB_WC_SUCCESS) { 1099 printk(KERN_ERR PFX 1100 "Completion wth error at %s:\n", 1101 cb->server ? "server" : "client"); 1102 printk(KERN_ERR PFX 1103 "Failed status %d: wr_id %d\n", 1104 wc.status, (int) wc.wr_id); 1105 printk(KERN_ERR PFX 1106 "scnt=%d, rcnt=%d, ccnt=%d\n", 1107 scnt, rcnt, ccnt); 1108 goto done; 1109 } 1110 } 1111 } 1112 microtime(&stop_tv); 1113 1114 if (stop_tv.tv_usec < start_tv.tv_usec) { 1115 stop_tv.tv_usec += 1000000; 1116 stop_tv.tv_sec -= 1; 1117 } 1118 1119 for (i=0; i < cycle_iters; i++) { 1120 sum_post += post_cycles_stop[i] - post_cycles_start[i]; 1121 sum_poll += poll_cycles_stop[i] - poll_cycles_start[i]; 1122 sum_last_poll += poll_cycles_stop[i]-last_poll_cycles_start[i]; 1123 } 1124 printk(KERN_ERR PFX 1125 "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d" 1126 " sum_post %llu sum_poll %llu sum_last_poll %llu\n", 1127 (unsigned long)(stop_tv.tv_sec - start_tv.tv_sec), 1128 (unsigned long)(stop_tv.tv_usec - start_tv.tv_usec), 1129 scnt, cb->size, cycle_iters, 1130 (unsigned long long)sum_post, (unsigned long long)sum_poll, 1131 (unsigned long long)sum_last_poll); 1132 done: 1133 kfree(post_cycles_start); 1134 kfree(post_cycles_stop); 1135 kfree(poll_cycles_start); 1136 kfree(poll_cycles_stop); 1137 kfree(last_poll_cycles_start); 1138 } 1139 1140 static void bw_test(struct krping_cb *cb) 1141 { 1142 int ccnt, scnt; 1143 int iters=cb->count; 1144 struct timeval start_tv, stop_tv; 1145 cycles_t *post_cycles_start = NULL; 1146 cycles_t *post_cycles_stop = NULL; 1147 cycles_t *poll_cycles_start = NULL; 1148 cycles_t *poll_cycles_stop = NULL; 1149 cycles_t *last_poll_cycles_start = NULL; 1150 cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0; 1151 int i; 1152 int cycle_iters = 1000; 1153 1154 ccnt = 0; 1155 scnt = 0; 1156 1157 post_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1158 if (!post_cycles_start) { 1159 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1160 goto done; 1161 } 1162 post_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1163 if (!post_cycles_stop) { 1164 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1165 goto done; 1166 } 1167 poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1168 if (!poll_cycles_start) { 1169 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1170 goto done; 1171 } 1172 poll_cycles_stop = kmalloc(cycle_iters * sizeof(cycles_t), GFP_KERNEL); 1173 if (!poll_cycles_stop) { 1174 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1175 goto done; 1176 } 1177 last_poll_cycles_start = kmalloc(cycle_iters * sizeof(cycles_t), 1178 GFP_KERNEL); 1179 if (!last_poll_cycles_start) { 1180 printk(KERN_ERR PFX "%s kmalloc failed\n", __FUNCTION__); 1181 goto done; 1182 } 1183 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_WRITE; 1184 cb->rdma_sq_wr.rkey = cb->remote_rkey; 1185 cb->rdma_sq_wr.remote_addr = cb->remote_addr; 1186 cb->rdma_sq_wr.wr.sg_list->length = cb->size; 1187 1188 if (cycle_iters > iters) 1189 cycle_iters = iters; 1190 microtime(&start_tv); 1191 while (scnt < iters || ccnt < iters) { 1192 1193 while (scnt < iters && scnt - ccnt < cb->txdepth) { 1194 const struct ib_send_wr *bad_wr; 1195 1196 if (scnt < cycle_iters) 1197 post_cycles_start[scnt] = get_cycles(); 1198 if (ib_post_send(cb->qp, &cb->rdma_sq_wr.wr, &bad_wr)) { 1199 printk(KERN_ERR PFX 1200 "Couldn't post send: scnt=%d\n", 1201 scnt); 1202 goto done; 1203 } 1204 if (scnt < cycle_iters) 1205 post_cycles_stop[scnt] = get_cycles(); 1206 ++scnt; 1207 } 1208 1209 if (ccnt < iters) { 1210 int ne; 1211 struct ib_wc wc; 1212 1213 if (ccnt < cycle_iters) 1214 poll_cycles_start[ccnt] = get_cycles(); 1215 do { 1216 if (ccnt < cycle_iters) 1217 last_poll_cycles_start[ccnt] = 1218 get_cycles(); 1219 ne = ib_poll_cq(cb->cq, 1, &wc); 1220 } while (ne == 0); 1221 if (ccnt < cycle_iters) 1222 poll_cycles_stop[ccnt] = get_cycles(); 1223 ccnt += 1; 1224 1225 if (ne < 0) { 1226 printk(KERN_ERR PFX "poll CQ failed %d\n", ne); 1227 goto done; 1228 } 1229 if (wc.status != IB_WC_SUCCESS) { 1230 printk(KERN_ERR PFX 1231 "Completion wth error at %s:\n", 1232 cb->server ? "server" : "client"); 1233 printk(KERN_ERR PFX 1234 "Failed status %d: wr_id %d\n", 1235 wc.status, (int) wc.wr_id); 1236 goto done; 1237 } 1238 } 1239 } 1240 microtime(&stop_tv); 1241 1242 if (stop_tv.tv_usec < start_tv.tv_usec) { 1243 stop_tv.tv_usec += 1000000; 1244 stop_tv.tv_sec -= 1; 1245 } 1246 1247 for (i=0; i < cycle_iters; i++) { 1248 sum_post += post_cycles_stop[i] - post_cycles_start[i]; 1249 sum_poll += poll_cycles_stop[i] - poll_cycles_start[i]; 1250 sum_last_poll += poll_cycles_stop[i]-last_poll_cycles_start[i]; 1251 } 1252 printk(KERN_ERR PFX 1253 "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d" 1254 " sum_post %llu sum_poll %llu sum_last_poll %llu\n", 1255 (unsigned long)(stop_tv.tv_sec - start_tv.tv_sec), 1256 (unsigned long)(stop_tv.tv_usec - start_tv.tv_usec), 1257 scnt, cb->size, cycle_iters, 1258 (unsigned long long)sum_post, (unsigned long long)sum_poll, 1259 (unsigned long long)sum_last_poll); 1260 done: 1261 kfree(post_cycles_start); 1262 kfree(post_cycles_stop); 1263 kfree(poll_cycles_start); 1264 kfree(poll_cycles_stop); 1265 kfree(last_poll_cycles_start); 1266 } 1267 1268 static void krping_rlat_test_server(struct krping_cb *cb) 1269 { 1270 const struct ib_send_wr *bad_wr; 1271 struct ib_wc wc; 1272 int ret; 1273 1274 /* Spin waiting for client's Start STAG/TO/Len */ 1275 while (cb->state < RDMA_READ_ADV) { 1276 krping_cq_event_handler(cb->cq, cb); 1277 } 1278 1279 /* Send STAG/TO/Len to client */ 1280 krping_format_send(cb, cb->start_dma_addr); 1281 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1282 if (ret) { 1283 printk(KERN_ERR PFX "post send error %d\n", ret); 1284 return; 1285 } 1286 1287 /* Spin waiting for send completion */ 1288 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1289 if (ret < 0) { 1290 printk(KERN_ERR PFX "poll error %d\n", ret); 1291 return; 1292 } 1293 if (wc.status) { 1294 printk(KERN_ERR PFX "send completiong error %d\n", wc.status); 1295 return; 1296 } 1297 1298 wait_event_interruptible(cb->sem, cb->state == ERROR); 1299 } 1300 1301 static void krping_wlat_test_server(struct krping_cb *cb) 1302 { 1303 const struct ib_send_wr *bad_wr; 1304 struct ib_wc wc; 1305 int ret; 1306 1307 /* Spin waiting for client's Start STAG/TO/Len */ 1308 while (cb->state < RDMA_READ_ADV) { 1309 krping_cq_event_handler(cb->cq, cb); 1310 } 1311 1312 /* Send STAG/TO/Len to client */ 1313 krping_format_send(cb, cb->start_dma_addr); 1314 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1315 if (ret) { 1316 printk(KERN_ERR PFX "post send error %d\n", ret); 1317 return; 1318 } 1319 1320 /* Spin waiting for send completion */ 1321 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1322 if (ret < 0) { 1323 printk(KERN_ERR PFX "poll error %d\n", ret); 1324 return; 1325 } 1326 if (wc.status) { 1327 printk(KERN_ERR PFX "send completiong error %d\n", wc.status); 1328 return; 1329 } 1330 1331 wlat_test(cb); 1332 wait_event_interruptible(cb->sem, cb->state == ERROR); 1333 } 1334 1335 static void krping_bw_test_server(struct krping_cb *cb) 1336 { 1337 const struct ib_send_wr *bad_wr; 1338 struct ib_wc wc; 1339 int ret; 1340 1341 /* Spin waiting for client's Start STAG/TO/Len */ 1342 while (cb->state < RDMA_READ_ADV) { 1343 krping_cq_event_handler(cb->cq, cb); 1344 } 1345 1346 /* Send STAG/TO/Len to client */ 1347 krping_format_send(cb, cb->start_dma_addr); 1348 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1349 if (ret) { 1350 printk(KERN_ERR PFX "post send error %d\n", ret); 1351 return; 1352 } 1353 1354 /* Spin waiting for send completion */ 1355 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1356 if (ret < 0) { 1357 printk(KERN_ERR PFX "poll error %d\n", ret); 1358 return; 1359 } 1360 if (wc.status) { 1361 printk(KERN_ERR PFX "send completiong error %d\n", wc.status); 1362 return; 1363 } 1364 1365 if (cb->duplex) 1366 bw_test(cb); 1367 wait_event_interruptible(cb->sem, cb->state == ERROR); 1368 } 1369 1370 static int reg_supported(struct ib_device *dev) 1371 { 1372 u64 needed_flags = IB_DEVICE_MEM_MGT_EXTENSIONS; 1373 1374 if ((dev->attrs.device_cap_flags & needed_flags) != needed_flags) { 1375 printk(KERN_ERR PFX 1376 "Fastreg not supported - device_cap_flags 0x%llx\n", 1377 (unsigned long long)dev->attrs.device_cap_flags); 1378 return 0; 1379 } 1380 DEBUG_LOG("Fastreg supported - device_cap_flags 0x%llx\n", 1381 (unsigned long long)dev->attrs.device_cap_flags); 1382 return 1; 1383 } 1384 1385 static void fill_sockaddr(struct sockaddr_storage *sin, struct krping_cb *cb) 1386 { 1387 memset(sin, 0, sizeof(*sin)); 1388 1389 if (cb->addr_type == AF_INET) { 1390 struct sockaddr_in *sin4 = (struct sockaddr_in *)sin; 1391 sin4->sin_len = sizeof(*sin4); 1392 sin4->sin_family = AF_INET; 1393 memcpy((void *)&sin4->sin_addr.s_addr, cb->addr, 4); 1394 sin4->sin_port = cb->port; 1395 } else if (cb->addr_type == AF_INET6) { 1396 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sin; 1397 sin6->sin6_len = sizeof(*sin6); 1398 sin6->sin6_family = AF_INET6; 1399 memcpy((void *)&sin6->sin6_addr, cb->addr, 16); 1400 sin6->sin6_port = cb->port; 1401 } 1402 } 1403 1404 static int krping_bind_server(struct krping_cb *cb) 1405 { 1406 struct sockaddr_storage sin; 1407 int ret; 1408 1409 1410 fill_sockaddr(&sin, cb); 1411 1412 ret = rdma_bind_addr(cb->cm_id, (struct sockaddr *)&sin); 1413 if (ret) { 1414 printk(KERN_ERR PFX "rdma_bind_addr error %d\n", ret); 1415 return ret; 1416 } 1417 DEBUG_LOG("rdma_bind_addr successful\n"); 1418 1419 DEBUG_LOG("rdma_listen\n"); 1420 ret = rdma_listen(cb->cm_id, 3); 1421 if (ret) { 1422 printk(KERN_ERR PFX "rdma_listen failed: %d\n", ret); 1423 return ret; 1424 } 1425 1426 wait_event_interruptible(cb->sem, cb->state >= CONNECT_REQUEST); 1427 if (cb->state != CONNECT_REQUEST) { 1428 printk(KERN_ERR PFX "wait for CONNECT_REQUEST state %d\n", 1429 cb->state); 1430 return -1; 1431 } 1432 1433 if (!reg_supported(cb->child_cm_id->device)) 1434 return -EINVAL; 1435 1436 return 0; 1437 } 1438 1439 static void krping_run_server(struct krping_cb *cb) 1440 { 1441 const struct ib_recv_wr *bad_wr; 1442 int ret; 1443 1444 ret = krping_bind_server(cb); 1445 if (ret) 1446 return; 1447 1448 ret = krping_setup_qp(cb, cb->child_cm_id); 1449 if (ret) { 1450 printk(KERN_ERR PFX "setup_qp failed: %d\n", ret); 1451 goto err0; 1452 } 1453 1454 ret = krping_setup_buffers(cb); 1455 if (ret) { 1456 printk(KERN_ERR PFX "krping_setup_buffers failed: %d\n", ret); 1457 goto err1; 1458 } 1459 1460 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); 1461 if (ret) { 1462 printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret); 1463 goto err2; 1464 } 1465 1466 ret = krping_accept(cb); 1467 if (ret) { 1468 printk(KERN_ERR PFX "connect error %d\n", ret); 1469 goto err2; 1470 } 1471 1472 if (cb->wlat) 1473 krping_wlat_test_server(cb); 1474 else if (cb->rlat) 1475 krping_rlat_test_server(cb); 1476 else if (cb->bw) 1477 krping_bw_test_server(cb); 1478 else 1479 krping_test_server(cb); 1480 rdma_disconnect(cb->child_cm_id); 1481 err2: 1482 krping_free_buffers(cb); 1483 err1: 1484 krping_free_qp(cb); 1485 err0: 1486 rdma_destroy_id(cb->child_cm_id); 1487 } 1488 1489 static void krping_test_client(struct krping_cb *cb) 1490 { 1491 int ping, start, cc, i, ret; 1492 const struct ib_send_wr *bad_wr; 1493 unsigned char c; 1494 1495 start = 65; 1496 for (ping = 0; !cb->count || ping < cb->count; ping++) { 1497 cb->state = RDMA_READ_ADV; 1498 1499 /* Put some ascii text in the buffer. */ 1500 cc = sprintf(cb->start_buf, "rdma-ping-%d: ", ping); 1501 for (i = cc, c = start; i < cb->size; i++) { 1502 cb->start_buf[i] = c; 1503 c++; 1504 if (c > 122) 1505 c = 65; 1506 } 1507 start++; 1508 if (start > 122) 1509 start = 65; 1510 cb->start_buf[cb->size - 1] = 0; 1511 1512 krping_format_send(cb, cb->start_dma_addr); 1513 if (cb->state == ERROR) { 1514 printk(KERN_ERR PFX "krping_format_send failed\n"); 1515 break; 1516 } 1517 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1518 if (ret) { 1519 printk(KERN_ERR PFX "post send error %d\n", ret); 1520 break; 1521 } 1522 1523 /* Wait for server to ACK */ 1524 wait_event_interruptible(cb->sem, cb->state >= RDMA_WRITE_ADV); 1525 if (cb->state != RDMA_WRITE_ADV) { 1526 printk(KERN_ERR PFX 1527 "wait for RDMA_WRITE_ADV state %d\n", 1528 cb->state); 1529 break; 1530 } 1531 1532 krping_format_send(cb, cb->rdma_dma_addr); 1533 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1534 if (ret) { 1535 printk(KERN_ERR PFX "post send error %d\n", ret); 1536 break; 1537 } 1538 1539 /* Wait for the server to say the RDMA Write is complete. */ 1540 wait_event_interruptible(cb->sem, 1541 cb->state >= RDMA_WRITE_COMPLETE); 1542 if (cb->state != RDMA_WRITE_COMPLETE) { 1543 printk(KERN_ERR PFX 1544 "wait for RDMA_WRITE_COMPLETE state %d\n", 1545 cb->state); 1546 break; 1547 } 1548 1549 if (cb->validate) 1550 if (memcmp(cb->start_buf, cb->rdma_buf, cb->size)) { 1551 printk(KERN_ERR PFX "data mismatch!\n"); 1552 break; 1553 } 1554 1555 if (cb->verbose) 1556 printk(KERN_INFO PFX "ping data: %s\n", cb->rdma_buf); 1557 #ifdef SLOW_KRPING 1558 wait_event_interruptible_timeout(cb->sem, cb->state == ERROR, HZ); 1559 #endif 1560 } 1561 } 1562 1563 static void krping_rlat_test_client(struct krping_cb *cb) 1564 { 1565 const struct ib_send_wr *bad_wr; 1566 struct ib_wc wc; 1567 int ret; 1568 1569 cb->state = RDMA_READ_ADV; 1570 1571 /* Send STAG/TO/Len to client */ 1572 krping_format_send(cb, cb->start_dma_addr); 1573 if (cb->state == ERROR) { 1574 printk(KERN_ERR PFX "krping_format_send failed\n"); 1575 return; 1576 } 1577 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1578 if (ret) { 1579 printk(KERN_ERR PFX "post send error %d\n", ret); 1580 return; 1581 } 1582 1583 /* Spin waiting for send completion */ 1584 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1585 if (ret < 0) { 1586 printk(KERN_ERR PFX "poll error %d\n", ret); 1587 return; 1588 } 1589 if (wc.status) { 1590 printk(KERN_ERR PFX "send completion error %d\n", wc.status); 1591 return; 1592 } 1593 1594 /* Spin waiting for server's Start STAG/TO/Len */ 1595 while (cb->state < RDMA_WRITE_ADV) { 1596 krping_cq_event_handler(cb->cq, cb); 1597 } 1598 1599 #if 0 1600 { 1601 int i; 1602 struct timeval start, stop; 1603 time_t sec; 1604 suseconds_t usec; 1605 unsigned long long elapsed; 1606 struct ib_wc wc; 1607 const struct ib_send_wr *bad_wr; 1608 int ne; 1609 1610 cb->rdma_sq_wr.wr.opcode = IB_WR_RDMA_WRITE; 1611 cb->rdma_sq_wr.rkey = cb->remote_rkey; 1612 cb->rdma_sq_wr.remote_addr = cb->remote_addr; 1613 cb->rdma_sq_wr.wr.sg_list->length = 0; 1614 cb->rdma_sq_wr.wr.num_sge = 0; 1615 1616 microtime(&start); 1617 for (i=0; i < 100000; i++) { 1618 if (ib_post_send(cb->qp, &cb->rdma_sq_wr.wr, &bad_wr)) { 1619 printk(KERN_ERR PFX "Couldn't post send\n"); 1620 return; 1621 } 1622 do { 1623 ne = ib_poll_cq(cb->cq, 1, &wc); 1624 } while (ne == 0); 1625 if (ne < 0) { 1626 printk(KERN_ERR PFX "poll CQ failed %d\n", ne); 1627 return; 1628 } 1629 if (wc.status != IB_WC_SUCCESS) { 1630 printk(KERN_ERR PFX "Completion wth error at %s:\n", 1631 cb->server ? "server" : "client"); 1632 printk(KERN_ERR PFX "Failed status %d: wr_id %d\n", 1633 wc.status, (int) wc.wr_id); 1634 return; 1635 } 1636 } 1637 microtime(&stop); 1638 1639 if (stop.tv_usec < start.tv_usec) { 1640 stop.tv_usec += 1000000; 1641 stop.tv_sec -= 1; 1642 } 1643 sec = stop.tv_sec - start.tv_sec; 1644 usec = stop.tv_usec - start.tv_usec; 1645 elapsed = sec * 1000000 + usec; 1646 printk(KERN_ERR PFX "0B-write-lat iters 100000 usec %llu\n", elapsed); 1647 } 1648 #endif 1649 1650 rlat_test(cb); 1651 } 1652 1653 static void krping_wlat_test_client(struct krping_cb *cb) 1654 { 1655 const struct ib_send_wr *bad_wr; 1656 struct ib_wc wc; 1657 int ret; 1658 1659 cb->state = RDMA_READ_ADV; 1660 1661 /* Send STAG/TO/Len to client */ 1662 krping_format_send(cb, cb->start_dma_addr); 1663 if (cb->state == ERROR) { 1664 printk(KERN_ERR PFX "krping_format_send failed\n"); 1665 return; 1666 } 1667 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1668 if (ret) { 1669 printk(KERN_ERR PFX "post send error %d\n", ret); 1670 return; 1671 } 1672 1673 /* Spin waiting for send completion */ 1674 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1675 if (ret < 0) { 1676 printk(KERN_ERR PFX "poll error %d\n", ret); 1677 return; 1678 } 1679 if (wc.status) { 1680 printk(KERN_ERR PFX "send completion error %d\n", wc.status); 1681 return; 1682 } 1683 1684 /* Spin waiting for server's Start STAG/TO/Len */ 1685 while (cb->state < RDMA_WRITE_ADV) { 1686 krping_cq_event_handler(cb->cq, cb); 1687 } 1688 1689 wlat_test(cb); 1690 } 1691 1692 static void krping_bw_test_client(struct krping_cb *cb) 1693 { 1694 const struct ib_send_wr *bad_wr; 1695 struct ib_wc wc; 1696 int ret; 1697 1698 cb->state = RDMA_READ_ADV; 1699 1700 /* Send STAG/TO/Len to client */ 1701 krping_format_send(cb, cb->start_dma_addr); 1702 if (cb->state == ERROR) { 1703 printk(KERN_ERR PFX "krping_format_send failed\n"); 1704 return; 1705 } 1706 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr); 1707 if (ret) { 1708 printk(KERN_ERR PFX "post send error %d\n", ret); 1709 return; 1710 } 1711 1712 /* Spin waiting for send completion */ 1713 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0)); 1714 if (ret < 0) { 1715 printk(KERN_ERR PFX "poll error %d\n", ret); 1716 return; 1717 } 1718 if (wc.status) { 1719 printk(KERN_ERR PFX "send completion error %d\n", wc.status); 1720 return; 1721 } 1722 1723 /* Spin waiting for server's Start STAG/TO/Len */ 1724 while (cb->state < RDMA_WRITE_ADV) { 1725 krping_cq_event_handler(cb->cq, cb); 1726 } 1727 1728 bw_test(cb); 1729 } 1730 1731 /* 1732 * Manual qp flush test 1733 */ 1734 static void flush_qp(struct krping_cb *cb) 1735 { 1736 struct ib_send_wr wr = { 0 }; 1737 const struct ib_send_wr *bad; 1738 struct ib_recv_wr recv_wr = { 0 }; 1739 const struct ib_recv_wr *recv_bad; 1740 struct ib_wc wc; 1741 int ret; 1742 int flushed = 0; 1743 int ccnt = 0; 1744 1745 rdma_disconnect(cb->cm_id); 1746 DEBUG_LOG("disconnected!\n"); 1747 1748 wr.opcode = IB_WR_SEND; 1749 wr.wr_id = 0xdeadbeefcafebabe; 1750 ret = ib_post_send(cb->qp, &wr, &bad); 1751 if (ret) { 1752 printk(KERN_ERR PFX "%s post_send failed ret %d\n", __func__, ret); 1753 return; 1754 } 1755 1756 recv_wr.wr_id = 0xcafebabedeadbeef; 1757 ret = ib_post_recv(cb->qp, &recv_wr, &recv_bad); 1758 if (ret) { 1759 printk(KERN_ERR PFX "%s post_recv failed ret %d\n", __func__, ret); 1760 return; 1761 } 1762 1763 /* poll until the flush WRs complete */ 1764 do { 1765 ret = ib_poll_cq(cb->cq, 1, &wc); 1766 if (ret < 0) { 1767 printk(KERN_ERR PFX "ib_poll_cq failed %d\n", ret); 1768 return; 1769 } 1770 if (ret == 0) 1771 continue; 1772 ccnt++; 1773 if (wc.wr_id == 0xdeadbeefcafebabe || 1774 wc.wr_id == 0xcafebabedeadbeef) 1775 flushed++; 1776 } while (flushed != 2); 1777 DEBUG_LOG("qp_flushed! ccnt %u\n", ccnt); 1778 } 1779 1780 static void krping_fr_test(struct krping_cb *cb) 1781 { 1782 struct ib_send_wr inv; 1783 const struct ib_send_wr *bad; 1784 struct ib_reg_wr fr; 1785 struct ib_wc wc; 1786 u8 key = 0; 1787 struct ib_mr *mr; 1788 int ret; 1789 int size = cb->size; 1790 int plen = (((size - 1) & PAGE_MASK) + PAGE_SIZE) >> PAGE_SHIFT; 1791 unsigned long start; 1792 int count = 0; 1793 int scnt = 0; 1794 struct scatterlist sg = {0}; 1795 1796 mr = ib_alloc_mr(cb->pd, IB_MR_TYPE_MEM_REG, plen); 1797 if (IS_ERR(mr)) { 1798 printk(KERN_ERR PFX "ib_alloc_mr failed %ld\n", PTR_ERR(mr)); 1799 return; 1800 } 1801 1802 sg_dma_address(&sg) = (dma_addr_t)0xcafebabe0000ULL; 1803 sg_dma_len(&sg) = size; 1804 ret = ib_map_mr_sg(mr, &sg, 1, NULL, PAGE_SIZE); 1805 if (ret <= 0) { 1806 printk(KERN_ERR PFX "ib_map_mr_sge err %d\n", ret); 1807 goto err2; 1808 } 1809 1810 memset(&fr, 0, sizeof fr); 1811 fr.wr.opcode = IB_WR_REG_MR; 1812 fr.access = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE; 1813 fr.mr = mr; 1814 fr.wr.next = &inv; 1815 1816 memset(&inv, 0, sizeof inv); 1817 inv.opcode = IB_WR_LOCAL_INV; 1818 inv.send_flags = IB_SEND_SIGNALED; 1819 1820 DEBUG_LOG("fr_test: stag index 0x%x plen %u size %u depth %u\n", mr->rkey >> 8, plen, cb->size, cb->txdepth); 1821 start = time_uptime; 1822 while (!cb->count || count <= cb->count) { 1823 if (SIGPENDING(curthread)) { 1824 printk(KERN_ERR PFX "signal!\n"); 1825 break; 1826 } 1827 if ((time_uptime - start) >= 9) { 1828 DEBUG_LOG("fr_test: pausing 1 second! count %u latest size %u plen %u\n", count, size, plen); 1829 wait_event_interruptible_timeout(cb->sem, cb->state == ERROR, HZ); 1830 if (cb->state == ERROR) 1831 break; 1832 start = time_uptime; 1833 } 1834 while (scnt < (cb->txdepth>>1)) { 1835 ib_update_fast_reg_key(mr, ++key); 1836 fr.key = mr->rkey; 1837 inv.ex.invalidate_rkey = mr->rkey; 1838 1839 size = arc4random() % cb->size; 1840 if (size == 0) 1841 size = cb->size; 1842 sg_dma_len(&sg) = size; 1843 ret = ib_map_mr_sg(mr, &sg, 1, NULL, PAGE_SIZE); 1844 if (ret <= 0) { 1845 printk(KERN_ERR PFX "ib_map_mr_sge err %d\n", ret); 1846 goto err2; 1847 } 1848 ret = ib_post_send(cb->qp, &fr.wr, &bad); 1849 if (ret) { 1850 printk(KERN_ERR PFX "ib_post_send failed %d\n", ret); 1851 goto err2; 1852 } 1853 scnt++; 1854 } 1855 1856 ret = ib_poll_cq(cb->cq, 1, &wc); 1857 if (ret < 0) { 1858 printk(KERN_ERR PFX "ib_poll_cq failed %d\n", ret); 1859 goto err2; 1860 } 1861 if (ret == 1) { 1862 if (wc.status) { 1863 printk(KERN_ERR PFX "completion error %u\n", wc.status); 1864 goto err2; 1865 } 1866 count++; 1867 scnt--; 1868 } 1869 } 1870 err2: 1871 flush_qp(cb); 1872 DEBUG_LOG("fr_test: done!\n"); 1873 ib_dereg_mr(mr); 1874 } 1875 1876 static int krping_connect_client(struct krping_cb *cb) 1877 { 1878 struct rdma_conn_param conn_param; 1879 int ret; 1880 1881 memset(&conn_param, 0, sizeof conn_param); 1882 conn_param.responder_resources = 1; 1883 conn_param.initiator_depth = 1; 1884 conn_param.retry_count = 10; 1885 1886 ret = rdma_connect(cb->cm_id, &conn_param); 1887 if (ret) { 1888 printk(KERN_ERR PFX "rdma_connect error %d\n", ret); 1889 return ret; 1890 } 1891 1892 wait_event_interruptible(cb->sem, cb->state >= CONNECTED); 1893 if (cb->state == ERROR) { 1894 printk(KERN_ERR PFX "wait for CONNECTED state %d\n", cb->state); 1895 return -1; 1896 } 1897 1898 DEBUG_LOG("rdma_connect successful\n"); 1899 return 0; 1900 } 1901 1902 static int krping_bind_client(struct krping_cb *cb) 1903 { 1904 struct sockaddr_storage sin; 1905 int ret; 1906 1907 fill_sockaddr(&sin, cb); 1908 1909 ret = rdma_resolve_addr(cb->cm_id, NULL, (struct sockaddr *)&sin, 2000); 1910 if (ret) { 1911 printk(KERN_ERR PFX "rdma_resolve_addr error %d\n", ret); 1912 return ret; 1913 } 1914 1915 wait_event_interruptible(cb->sem, cb->state >= ROUTE_RESOLVED); 1916 if (cb->state != ROUTE_RESOLVED) { 1917 printk(KERN_ERR PFX 1918 "addr/route resolution did not resolve: state %d\n", 1919 cb->state); 1920 return -EINTR; 1921 } 1922 1923 if (!reg_supported(cb->cm_id->device)) 1924 return -EINVAL; 1925 1926 DEBUG_LOG("rdma_resolve_addr - rdma_resolve_route successful\n"); 1927 return 0; 1928 } 1929 1930 static void krping_run_client(struct krping_cb *cb) 1931 { 1932 const struct ib_recv_wr *bad_wr; 1933 int ret; 1934 1935 /* set type of service, if any */ 1936 if (cb->tos != 0) 1937 rdma_set_service_type(cb->cm_id, cb->tos); 1938 1939 ret = krping_bind_client(cb); 1940 if (ret) 1941 return; 1942 1943 ret = krping_setup_qp(cb, cb->cm_id); 1944 if (ret) { 1945 printk(KERN_ERR PFX "setup_qp failed: %d\n", ret); 1946 return; 1947 } 1948 1949 ret = krping_setup_buffers(cb); 1950 if (ret) { 1951 printk(KERN_ERR PFX "krping_setup_buffers failed: %d\n", ret); 1952 goto err1; 1953 } 1954 1955 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr); 1956 if (ret) { 1957 printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret); 1958 goto err2; 1959 } 1960 1961 ret = krping_connect_client(cb); 1962 if (ret) { 1963 printk(KERN_ERR PFX "connect error %d\n", ret); 1964 goto err2; 1965 } 1966 1967 if (cb->wlat) 1968 krping_wlat_test_client(cb); 1969 else if (cb->rlat) 1970 krping_rlat_test_client(cb); 1971 else if (cb->bw) 1972 krping_bw_test_client(cb); 1973 else if (cb->frtest) 1974 krping_fr_test(cb); 1975 else 1976 krping_test_client(cb); 1977 rdma_disconnect(cb->cm_id); 1978 err2: 1979 krping_free_buffers(cb); 1980 err1: 1981 krping_free_qp(cb); 1982 } 1983 1984 static uint16_t 1985 krping_get_ipv6_scope_id(char *name) 1986 { 1987 struct ifnet *ifp; 1988 uint16_t retval; 1989 1990 if (name == NULL) 1991 return (0); 1992 CURVNET_SET_QUIET(TD_TO_VNET(curthread)); 1993 ifp = ifunit_ref(name); 1994 CURVNET_RESTORE(); 1995 if (ifp == NULL) 1996 return (0); 1997 retval = ifp->if_index; 1998 if_rele(ifp); 1999 return (retval); 2000 } 2001 2002 int krping_doit(char *cmd) 2003 { 2004 struct krping_cb *cb; 2005 int op; 2006 int ret = 0; 2007 char *optarg; 2008 char *scope; 2009 unsigned long optint; 2010 2011 cb = kzalloc(sizeof(*cb), GFP_KERNEL); 2012 if (!cb) 2013 return -ENOMEM; 2014 2015 mutex_lock(&krping_mutex); 2016 list_add_tail(&cb->list, &krping_cbs); 2017 mutex_unlock(&krping_mutex); 2018 2019 cb->server = -1; 2020 cb->state = IDLE; 2021 cb->size = 64; 2022 cb->txdepth = RPING_SQ_DEPTH; 2023 init_waitqueue_head(&cb->sem); 2024 2025 while ((op = krping_getopt("krping", &cmd, krping_opts, NULL, &optarg, 2026 &optint)) != 0) { 2027 switch (op) { 2028 case 'a': 2029 cb->addr_str = optarg; 2030 cb->addr_type = AF_INET; 2031 DEBUG_LOG("ipaddr (%s)\n", optarg); 2032 if (inet_pton(AF_INET, optarg, cb->addr) != 1) { 2033 printk(KERN_ERR PFX "bad addr string %s\n", 2034 optarg); 2035 ret = EINVAL; 2036 } 2037 break; 2038 case 'A': 2039 cb->addr_str = optarg; 2040 cb->addr_type = AF_INET6; 2041 DEBUG_LOG("ipv6addr (%s)\n", optarg); 2042 scope = strstr(optarg, "%"); 2043 /* extract scope ID, if any */ 2044 if (scope != NULL) 2045 *scope++ = 0; 2046 /* extract IPv6 network address */ 2047 if (inet_pton(AF_INET6, optarg, cb->addr) != 1) { 2048 printk(KERN_ERR PFX "bad addr string %s\n", 2049 optarg); 2050 ret = EINVAL; 2051 } else if (IN6_IS_SCOPE_LINKLOCAL((struct in6_addr *)cb->addr) || 2052 IN6_IS_ADDR_MC_INTFACELOCAL((struct in6_addr *)cb->addr)) { 2053 uint16_t scope_id = krping_get_ipv6_scope_id(scope); 2054 DEBUG_LOG("ipv6 scope ID = %d\n", scope_id); 2055 cb->addr[2] = scope_id >> 8; 2056 cb->addr[3] = scope_id & 0xFF; 2057 } 2058 break; 2059 case 'p': 2060 cb->port = htons(optint); 2061 DEBUG_LOG("port %d\n", (int)optint); 2062 break; 2063 case 'P': 2064 cb->poll = 1; 2065 DEBUG_LOG("server\n"); 2066 break; 2067 case 's': 2068 cb->server = 1; 2069 DEBUG_LOG("server\n"); 2070 break; 2071 case 'c': 2072 cb->server = 0; 2073 DEBUG_LOG("client\n"); 2074 break; 2075 case 'S': 2076 cb->size = optint; 2077 if ((cb->size < 1) || 2078 (cb->size > RPING_BUFSIZE)) { 2079 printk(KERN_ERR PFX "Invalid size %d " 2080 "(valid range is 1 to %d)\n", 2081 cb->size, RPING_BUFSIZE); 2082 ret = EINVAL; 2083 } else 2084 DEBUG_LOG("size %d\n", (int)optint); 2085 break; 2086 case 'C': 2087 cb->count = optint; 2088 if (cb->count < 0) { 2089 printk(KERN_ERR PFX "Invalid count %d\n", 2090 cb->count); 2091 ret = EINVAL; 2092 } else 2093 DEBUG_LOG("count %d\n", (int) cb->count); 2094 break; 2095 case 'v': 2096 cb->verbose++; 2097 DEBUG_LOG("verbose\n"); 2098 break; 2099 case 'V': 2100 cb->validate++; 2101 DEBUG_LOG("validate data\n"); 2102 break; 2103 case 'l': 2104 cb->wlat++; 2105 break; 2106 case 'L': 2107 cb->rlat++; 2108 break; 2109 case 'B': 2110 cb->bw++; 2111 break; 2112 case 'd': 2113 cb->duplex++; 2114 break; 2115 case 'I': 2116 cb->server_invalidate = 1; 2117 break; 2118 case 't': 2119 cb->tos = optint; 2120 DEBUG_LOG("type of service, tos=%d\n", (int) cb->tos); 2121 break; 2122 case 'T': 2123 cb->txdepth = optint; 2124 DEBUG_LOG("txdepth %d\n", (int) cb->txdepth); 2125 break; 2126 case 'Z': 2127 cb->local_dma_lkey = 1; 2128 DEBUG_LOG("using local dma lkey\n"); 2129 break; 2130 case 'R': 2131 cb->read_inv = 1; 2132 DEBUG_LOG("using read-with-inv\n"); 2133 break; 2134 case 'f': 2135 cb->frtest = 1; 2136 DEBUG_LOG("fast-reg test!\n"); 2137 break; 2138 default: 2139 printk(KERN_ERR PFX "unknown opt %s\n", optarg); 2140 ret = -EINVAL; 2141 break; 2142 } 2143 } 2144 if (ret) 2145 goto out; 2146 2147 if (cb->server == -1) { 2148 printk(KERN_ERR PFX "must be either client or server\n"); 2149 ret = -EINVAL; 2150 goto out; 2151 } 2152 2153 if (cb->server && cb->frtest) { 2154 printk(KERN_ERR PFX "must be client to run frtest\n"); 2155 ret = -EINVAL; 2156 goto out; 2157 } 2158 2159 if ((cb->frtest + cb->bw + cb->rlat + cb->wlat) > 1) { 2160 printk(KERN_ERR PFX "Pick only one test: fr, bw, rlat, wlat\n"); 2161 ret = -EINVAL; 2162 goto out; 2163 } 2164 2165 if (cb->wlat || cb->rlat || cb->bw) { 2166 printk(KERN_ERR PFX "wlat, rlat, and bw tests only support mem_mode MR - which is no longer supported\n"); 2167 ret = -EINVAL; 2168 goto out; 2169 } 2170 2171 cb->cm_id = rdma_create_id(TD_TO_VNET(curthread), krping_cma_event_handler, cb, RDMA_PS_TCP, IB_QPT_RC); 2172 if (IS_ERR(cb->cm_id)) { 2173 ret = PTR_ERR(cb->cm_id); 2174 printk(KERN_ERR PFX "rdma_create_id error %d\n", ret); 2175 goto out; 2176 } 2177 DEBUG_LOG("created cm_id %p\n", cb->cm_id); 2178 2179 if (cb->server) 2180 krping_run_server(cb); 2181 else 2182 krping_run_client(cb); 2183 2184 DEBUG_LOG("destroy cm_id %p\n", cb->cm_id); 2185 rdma_destroy_id(cb->cm_id); 2186 out: 2187 mutex_lock(&krping_mutex); 2188 list_del(&cb->list); 2189 mutex_unlock(&krping_mutex); 2190 kfree(cb); 2191 return ret; 2192 } 2193 2194 void 2195 krping_walk_cb_list(void (*f)(struct krping_stats *, void *), void *arg) 2196 { 2197 struct krping_cb *cb; 2198 2199 mutex_lock(&krping_mutex); 2200 list_for_each_entry(cb, &krping_cbs, list) 2201 (*f)(cb->pd ? &cb->stats : NULL, arg); 2202 mutex_unlock(&krping_mutex); 2203 } 2204 2205 void 2206 krping_cancel_all(void) 2207 { 2208 struct krping_cb *cb; 2209 2210 mutex_lock(&krping_mutex); 2211 list_for_each_entry(cb, &krping_cbs, list) { 2212 cb->state = ERROR; 2213 wake_up_interruptible(&cb->sem); 2214 } 2215 mutex_unlock(&krping_mutex); 2216 } 2217 2218