sys/contrib/rdma/krping/krping.c

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