ctdb/ib/ibwrapper.c

   1 /*
   2  * Unix SMB/CIFS implementation.
   3  * Wrap Infiniband calls.
   4  *
   5  * Copyright (C) Sven Oehme <oehmes@de.ibm.com> 2006
   6  *
   7  * Major code contributions by Peter Somogyi <psomogyi@gamax.hu>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 3 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  21  */
  22
  23 #include <stdlib.h>
  24 #include <string.h>
  25 #include <stdio.h>
  26 #include <errno.h>
  27 #include <sys/types.h>
  28 #include <netinet/in.h>
  29 #include <sys/socket.h>
  30 #include <netdb.h>
  31 #include <arpa/inet.h>
  32 #include <malloc.h>
  33 #include <assert.h>
  34 #include <unistd.h>
  35
  36 #include "includes.h"
  37 #include "ibwrapper.h"
  38
  39 #include <infiniband/kern-abi.h>
  40 #include <rdma/rdma_cma_abi.h>
  41 #include <rdma/rdma_cma.h>
  42
  43 #include "ibwrapper_internal.h"
  44 #include "lib/util/dlinklist.h"
  45
  46 #define IBW_LASTERR_BUFSIZE 512
  47 static char ibw_lasterr[IBW_LASTERR_BUFSIZE];
  48
  49 #define IBW_MAX_SEND_WR 256
  50 #define IBW_MAX_RECV_WR 1024
  51 #define IBW_RECV_BUFSIZE 256
  52 #define IBW_RECV_THRESHOLD (1 * 1024 * 1024)
  53
  54 static void ibw_event_handler_verbs(struct event_context *ev,
  55         struct fd_event *fde, uint16_t flags, void *private_data);
  56 static int ibw_fill_cq(struct ibw_conn *conn);
  57 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc);
  58 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc);
  59 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len);
  60
  61 static void *ibw_alloc_mr(struct ibw_ctx_priv *pctx, struct ibw_conn_priv *pconn,
  62         uint32_t n, struct ibv_mr **ppmr)
  63 {
  64         void *buf;
  65
  66         DEBUG(DEBUG_DEBUG, ("ibw_alloc_mr(cmid=%p, n=%u)\n", pconn->cm_id, n));
  67         buf = memalign(pctx->pagesize, n);
  68         if (!buf) {
  69                 sprintf(ibw_lasterr, "couldn't allocate memory\n");
  70                 return NULL;
  71         }
  72
  73         *ppmr = ibv_reg_mr(pconn->pd, buf, n, IBV_ACCESS_LOCAL_WRITE);
  74         if (!*ppmr) {
  75                 sprintf(ibw_lasterr, "couldn't allocate mr\n");
  76                 free(buf);
  77                 return NULL;
  78         }
  79
  80         return buf;
  81 }
  82
  83 static void ibw_free_mr(char **ppbuf, struct ibv_mr **ppmr)
  84 {
  85         DEBUG(DEBUG_DEBUG, ("ibw_free_mr(%p %p)\n", *ppbuf, *ppmr));
  86         if (*ppmr!=NULL) {
  87                 ibv_dereg_mr(*ppmr);
  88                 *ppmr = NULL;
  89         }
  90         if (*ppbuf) {
  91                 free(*ppbuf);
  92                 *ppbuf = NULL;
  93         }
  94 }
  95
  96 static int ibw_init_memory(struct ibw_conn *conn)
  97 {
  98         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
  99         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 100         struct ibw_opts *opts = &pctx->opts;
 101         int     i;
 102         struct ibw_wr   *p;
 103
 104         DEBUG(DEBUG_DEBUG, ("ibw_init_memory(cmid: %p)\n", pconn->cm_id));
 105         pconn->buf_send = ibw_alloc_mr(pctx, pconn,
 106                 opts->max_send_wr * opts->recv_bufsize, &pconn->mr_send);
 107         if (!pconn->buf_send) {
 108                 sprintf(ibw_lasterr, "couldn't allocate work send buf\n");
 109                 return -1;
 110         }
 111
 112         pconn->buf_recv = ibw_alloc_mr(pctx, pconn,
 113                 opts->max_recv_wr * opts->recv_bufsize, &pconn->mr_recv);
 114         if (!pconn->buf_recv) {
 115                 sprintf(ibw_lasterr, "couldn't allocate work recv buf\n");
 116                 return -1;
 117         }
 118
 119         pconn->wr_index = talloc_size(pconn, opts->max_send_wr * sizeof(struct ibw_wr *));
 120         assert(pconn->wr_index!=NULL);
 121
 122         for(i=0; i<opts->max_send_wr; i++) {
 123                 p = pconn->wr_index[i] = talloc_zero(pconn, struct ibw_wr);
 124                 p->buf = pconn->buf_send + (i * opts->recv_bufsize);
 125                 p->wr_id = i;
 126
 127                 DLIST_ADD(pconn->wr_list_avail, p);
 128         }
 129
 130         return 0;
 131 }
 132
 133 static int ibw_ctx_priv_destruct(struct ibw_ctx_priv *pctx)
 134 {
 135         DEBUG(DEBUG_DEBUG, ("ibw_ctx_priv_destruct(%p)\n", pctx));
 136
 137         /*
 138          * tevent_fd must be removed before the fd is closed
 139          */
 140         TALLOC_FREE(pctx->cm_channel_event);
 141
 142         /* destroy cm */
 143         if (pctx->cm_channel) {
 144                 rdma_destroy_event_channel(pctx->cm_channel);
 145                 pctx->cm_channel = NULL;
 146         }
 147         if (pctx->cm_id) {
 148                 rdma_destroy_id(pctx->cm_id);
 149                 pctx->cm_id = NULL;
 150         }
 151
 152         return 0;
 153 }
 154
 155 static int ibw_ctx_destruct(struct ibw_ctx *ctx)
 156 {
 157         DEBUG(DEBUG_DEBUG, ("ibw_ctx_destruct(%p)\n", ctx));
 158         return 0;
 159 }
 160
 161 static int ibw_conn_priv_destruct(struct ibw_conn_priv *pconn)
 162 {
 163         DEBUG(DEBUG_DEBUG, ("ibw_conn_priv_destruct(%p, cmid: %p)\n",
 164                 pconn, pconn->cm_id));
 165
 166         /* pconn->wr_index is freed by talloc */
 167         /* pconn->wr_index[i] are freed by talloc */
 168
 169         /*
 170          * tevent_fd must be removed before the fd is closed
 171          */
 172         TALLOC_FREE(pconn->verbs_channel_event);
 173
 174         /* destroy verbs */
 175         if (pconn->cm_id!=NULL && pconn->cm_id->qp!=NULL) {
 176                 rdma_destroy_qp(pconn->cm_id);
 177                 pconn->cm_id->qp = NULL;
 178         }
 179
 180         if (pconn->cq!=NULL) {
 181                 ibv_destroy_cq(pconn->cq);
 182                 pconn->cq = NULL;
 183         }
 184
 185         if (pconn->verbs_channel!=NULL) {
 186                 ibv_destroy_comp_channel(pconn->verbs_channel);
 187                 pconn->verbs_channel = NULL;
 188         }
 189
 190         /* free memory regions */
 191         ibw_free_mr(&pconn->buf_send, &pconn->mr_send);
 192         ibw_free_mr(&pconn->buf_recv, &pconn->mr_recv);
 193
 194         if (pconn->pd) {
 195                 ibv_dealloc_pd(pconn->pd);
 196                 pconn->pd = NULL;
 197                 DEBUG(DEBUG_DEBUG, ("pconn=%p pd deallocated\n", pconn));
 198         }
 199
 200         if (pconn->cm_id) {
 201                 rdma_destroy_id(pconn->cm_id);
 202                 pconn->cm_id = NULL;
 203                 DEBUG(DEBUG_DEBUG, ("pconn=%p cm_id destroyed\n", pconn));
 204         }
 205
 206         return 0;
 207 }
 208
 209 static int ibw_wr_destruct(struct ibw_wr *wr)
 210 {
 211         if (wr->buf_large!=NULL)
 212                 ibw_free_mr(&wr->buf_large, &wr->mr_large);
 213         return 0;
 214 }
 215
 216 static int ibw_conn_destruct(struct ibw_conn *conn)
 217 {
 218         DEBUG(DEBUG_DEBUG, ("ibw_conn_destruct(%p)\n", conn));
 219
 220         /* important here: ctx is a talloc _parent_ */
 221         DLIST_REMOVE(conn->ctx->conn_list, conn);
 222         return 0;
 223 }
 224
 225 struct ibw_conn *ibw_conn_new(struct ibw_ctx *ctx, TALLOC_CTX *mem_ctx)
 226 {
 227         struct ibw_conn *conn;
 228         struct ibw_conn_priv *pconn;
 229
 230         assert(ctx!=NULL);
 231
 232         conn = talloc_zero(mem_ctx, struct ibw_conn);
 233         assert(conn!=NULL);
 234         talloc_set_destructor(conn, ibw_conn_destruct);
 235
 236         pconn = talloc_zero(conn, struct ibw_conn_priv);
 237         assert(pconn!=NULL);
 238         talloc_set_destructor(pconn, ibw_conn_priv_destruct);
 239
 240         conn->ctx = ctx;
 241         conn->internal = (void *)pconn;
 242
 243         DLIST_ADD(ctx->conn_list, conn);
 244
 245         return conn;
 246 }
 247
 248 static int ibw_setup_cq_qp(struct ibw_conn *conn)
 249 {
 250         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 251         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 252         struct ibv_qp_init_attr init_attr;
 253         struct ibv_qp_attr attr;
 254         int rc;
 255
 256         DEBUG(DEBUG_DEBUG, ("ibw_setup_cq_qp(cmid: %p)\n", pconn->cm_id));
 257
 258         /* init verbs */
 259         pconn->verbs_channel = ibv_create_comp_channel(pconn->cm_id->verbs);
 260         if (!pconn->verbs_channel) {
 261                 sprintf(ibw_lasterr, "ibv_create_comp_channel failed %d\n", errno);
 262                 return -1;
 263         }
 264         DEBUG(DEBUG_DEBUG, ("created channel %p\n", pconn->verbs_channel));
 265
 266         pconn->verbs_channel_event = event_add_fd(pctx->ectx, NULL, /* not pconn or conn */
 267                 pconn->verbs_channel->fd, EVENT_FD_READ, ibw_event_handler_verbs, conn);
 268
 269         pconn->pd = ibv_alloc_pd(pconn->cm_id->verbs);
 270         if (!pconn->pd) {
 271                 sprintf(ibw_lasterr, "ibv_alloc_pd failed %d\n", errno);
 272                 return -1;
 273         }
 274         DEBUG(DEBUG_DEBUG, ("created pd %p\n", pconn->pd));
 275
 276         /* init mr */
 277         if (ibw_init_memory(conn))
 278                 return -1;
 279
 280         /* init cq */
 281         pconn->cq = ibv_create_cq(pconn->cm_id->verbs,
 282                 pctx->opts.max_recv_wr + pctx->opts.max_send_wr,
 283                 conn, pconn->verbs_channel, 0);
 284         if (pconn->cq==NULL) {
 285                 sprintf(ibw_lasterr, "ibv_create_cq failed\n");
 286                 return -1;
 287         }
 288
 289         rc = ibv_req_notify_cq(pconn->cq, 0);
 290         if (rc) {
 291                 sprintf(ibw_lasterr, "ibv_req_notify_cq failed with %d\n", rc);
 292                 return rc;
 293         }
 294
 295         /* init qp */
 296         memset(&init_attr, 0, sizeof(init_attr));
 297         init_attr.cap.max_send_wr = pctx->opts.max_send_wr;
 298         init_attr.cap.max_recv_wr = pctx->opts.max_recv_wr;
 299         init_attr.cap.max_recv_sge = 1;
 300         init_attr.cap.max_send_sge = 1;
 301         init_attr.qp_type = IBV_QPT_RC;
 302         init_attr.send_cq = pconn->cq;
 303         init_attr.recv_cq = pconn->cq;
 304
 305         rc = rdma_create_qp(pconn->cm_id, pconn->pd, &init_attr);
 306         if (rc) {
 307                 sprintf(ibw_lasterr, "rdma_create_qp failed with %d\n", rc);
 308                 return rc;
 309         }
 310         /* elase result is in pconn->cm_id->qp */
 311
 312         rc = ibv_query_qp(pconn->cm_id->qp, &attr, IBV_QP_PATH_MTU, &init_attr);
 313         if (rc) {
 314                 sprintf(ibw_lasterr, "ibv_query_qp failed with %d\n", rc);
 315                 return rc;
 316         }
 317
 318         return ibw_fill_cq(conn);
 319 }
 320
 321 static int ibw_refill_cq_recv(struct ibw_conn *conn)
 322 {
 323         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 324         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 325         int     rc;
 326         struct ibv_sge list = {
 327                 .addr   = (uintptr_t) NULL, /* filled below */
 328                 .length = pctx->opts.recv_bufsize,
 329                 .lkey   = pconn->mr_recv->lkey /* always the same */
 330         };
 331         struct ibv_recv_wr wr = {
 332                 .wr_id      = 0, /* filled below */
 333                 .sg_list    = &list,
 334                 .num_sge    = 1,
 335         };
 336         struct ibv_recv_wr *bad_wr;
 337
 338         DEBUG(DEBUG_DEBUG, ("ibw_refill_cq_recv(cmid: %p)\n", pconn->cm_id));
 339
 340         list.addr = (uintptr_t) pconn->buf_recv + pctx->opts.recv_bufsize * pconn->recv_index;
 341         wr.wr_id = pconn->recv_index;
 342         pconn->recv_index = (pconn->recv_index + 1) % pctx->opts.max_recv_wr;
 343
 344         rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
 345         if (rc) {
 346                 sprintf(ibw_lasterr, "refill/ibv_post_recv failed with %d\n", rc);
 347                 DEBUG(DEBUG_ERR, (ibw_lasterr));
 348                 return -2;
 349         }
 350
 351         return 0;
 352 }
 353
 354 static int ibw_fill_cq(struct ibw_conn *conn)
 355 {
 356         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 357         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 358         int     i, rc;
 359         struct ibv_sge list = {
 360                 .addr   = (uintptr_t) NULL, /* filled below */
 361                 .length = pctx->opts.recv_bufsize,
 362                 .lkey   = pconn->mr_recv->lkey /* always the same */
 363         };
 364         struct ibv_recv_wr wr = {
 365                 .wr_id      = 0, /* filled below */
 366                 .sg_list    = &list,
 367                 .num_sge    = 1,
 368         };
 369         struct ibv_recv_wr *bad_wr;
 370
 371         DEBUG(DEBUG_DEBUG, ("ibw_fill_cq(cmid: %p)\n", pconn->cm_id));
 372
 373         for(i = pctx->opts.max_recv_wr; i!=0; i--) {
 374                 list.addr = (uintptr_t) pconn->buf_recv + pctx->opts.recv_bufsize * pconn->recv_index;
 375                 wr.wr_id = pconn->recv_index;
 376                 pconn->recv_index = (pconn->recv_index + 1) % pctx->opts.max_recv_wr;
 377
 378                 rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
 379                 if (rc) {
 380                         sprintf(ibw_lasterr, "fill/ibv_post_recv failed with %d\n", rc);
 381                         DEBUG(DEBUG_ERR, (ibw_lasterr));
 382                         return -2;
 383                 }
 384         }
 385
 386         return 0;
 387 }
 388
 389 static int ibw_manage_connect(struct ibw_conn *conn)
 390 {
 391         struct rdma_conn_param conn_param;
 392         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 393         int     rc;
 394
 395         DEBUG(DEBUG_DEBUG, ("ibw_manage_connect(cmid: %p)\n", pconn->cm_id));
 396
 397         if (ibw_setup_cq_qp(conn))
 398                 return -1;
 399
 400         /* cm connect */
 401         memset(&conn_param, 0, sizeof conn_param);
 402         conn_param.responder_resources = 1;
 403         conn_param.initiator_depth = 1;
 404         conn_param.retry_count = 10;
 405
 406         rc = rdma_connect(pconn->cm_id, &conn_param);
 407         if (rc)
 408                 sprintf(ibw_lasterr, "rdma_connect error %d\n", rc);
 409
 410         return rc;
 411 }
 412
 413 static void ibw_event_handler_cm(struct event_context *ev,
 414         struct fd_event *fde, uint16_t flags, void *private_data)
 415 {
 416         int     rc;
 417         struct ibw_ctx  *ctx = talloc_get_type(private_data, struct ibw_ctx);
 418         struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
 419         struct ibw_conn *conn = NULL;
 420         struct ibw_conn_priv *pconn = NULL;
 421         struct rdma_cm_id *cma_id = NULL;
 422         struct rdma_cm_event *event = NULL;
 423
 424         assert(ctx!=NULL);
 425
 426         rc = rdma_get_cm_event(pctx->cm_channel, &event);
 427         if (rc) {
 428                 ctx->state = IBWS_ERROR;
 429                 event = NULL;
 430                 sprintf(ibw_lasterr, "rdma_get_cm_event error %d\n", rc);
 431                 goto error;
 432         }
 433         cma_id = event->id;
 434
 435         DEBUG(DEBUG_DEBUG, ("cma_event type %d cma_id %p (%s)\n", event->event, cma_id,
 436                   (cma_id == pctx->cm_id) ? "parent" : "child"));
 437
 438         switch (event->event) {
 439         case RDMA_CM_EVENT_ADDR_RESOLVED:
 440                 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_ADDR_RESOLVED\n"));
 441                 /* continuing from ibw_connect ... */
 442                 rc = rdma_resolve_route(cma_id, 2000);
 443                 if (rc) {
 444                         sprintf(ibw_lasterr, "rdma_resolve_route error %d\n", rc);
 445                         goto error;
 446                 }
 447                 /* continued at RDMA_CM_EVENT_ROUTE_RESOLVED */
 448                 break;
 449
 450         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 451                 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_ROUTE_RESOLVED\n"));
 452                 /* after RDMA_CM_EVENT_ADDR_RESOLVED: */
 453                 assert(cma_id->context!=NULL);
 454                 conn = talloc_get_type(cma_id->context, struct ibw_conn);
 455
 456                 rc = ibw_manage_connect(conn);
 457                 if (rc)
 458                         goto error;
 459
 460                 break;
 461
 462         case RDMA_CM_EVENT_CONNECT_REQUEST:
 463                 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_CONNECT_REQUEST\n"));
 464                 ctx->state = IBWS_CONNECT_REQUEST;
 465                 conn = ibw_conn_new(ctx, ctx);
 466                 pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 467                 pconn->cm_id = cma_id; /* !!! event will be freed but id not */
 468                 cma_id->context = (void *)conn;
 469                 DEBUG(DEBUG_DEBUG, ("pconn->cm_id %p\n", pconn->cm_id));
 470
 471                 if (ibw_setup_cq_qp(conn))
 472                         goto error;
 473
 474                 conn->state = IBWC_INIT;
 475                 pctx->connstate_func(ctx, conn);
 476
 477                 /* continued at ibw_accept when invoked by the func above */
 478                 if (!pconn->is_accepted) {
 479                         rc = rdma_reject(cma_id, NULL, 0);
 480                         if (rc)
 481                                 DEBUG(DEBUG_ERR, ("rdma_reject failed with rc=%d\n", rc));
 482                         talloc_free(conn);
 483                         DEBUG(DEBUG_DEBUG, ("pconn->cm_id %p wasn't accepted\n", pconn->cm_id));
 484                 }
 485
 486                 /* TODO: clarify whether if it's needed by upper layer: */
 487                 ctx->state = IBWS_READY;
 488                 pctx->connstate_func(ctx, NULL);
 489
 490                 /* NOTE: more requests can arrive until RDMA_CM_EVENT_ESTABLISHED ! */
 491                 break;
 492
 493         case RDMA_CM_EVENT_ESTABLISHED:
 494                 /* expected after ibw_accept and ibw_connect[not directly] */
 495                 DEBUG(DEBUG_INFO, ("ESTABLISHED (conn: %p)\n", cma_id->context));
 496                 conn = talloc_get_type(cma_id->context, struct ibw_conn);
 497                 assert(conn!=NULL); /* important assumption */
 498
 499                 DEBUG(DEBUG_DEBUG, ("ibw_setup_cq_qp succeeded (cmid=%p)\n", cma_id));
 500
 501                 /* client conn is up */
 502                 conn->state = IBWC_CONNECTED;
 503
 504                 /* both ctx and conn have changed */
 505                 pctx->connstate_func(ctx, conn);
 506                 break;
 507
 508         case RDMA_CM_EVENT_ADDR_ERROR:
 509                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ADDR_ERROR, error %d\n", event->status);
 510         case RDMA_CM_EVENT_ROUTE_ERROR:
 511                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ROUTE_ERROR, error %d\n", event->status);
 512         case RDMA_CM_EVENT_CONNECT_ERROR:
 513                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_CONNECT_ERROR, error %d\n", event->status);
 514         case RDMA_CM_EVENT_UNREACHABLE:
 515                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_UNREACHABLE, error %d\n", event->status);
 516                 goto error;
 517         case RDMA_CM_EVENT_REJECTED:
 518                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_REJECTED, error %d\n", event->status);
 519                 DEBUG(DEBUG_INFO, ("cm event handler: %s", ibw_lasterr));
 520                 conn = talloc_get_type(cma_id->context, struct ibw_conn);
 521                 if (conn) {
 522                         /* must be done BEFORE connstate */
 523                         if ((rc=rdma_ack_cm_event(event)))
 524                                 DEBUG(DEBUG_ERR, ("reject/rdma_ack_cm_event failed with %d\n", rc));
 525                         event = NULL; /* not to touch cma_id or conn */
 526                         conn->state = IBWC_ERROR;
 527                         /* it should free the conn */
 528                         pctx->connstate_func(NULL, conn);
 529                 }
 530                 break; /* this is not strictly an error */
 531
 532         case RDMA_CM_EVENT_DISCONNECTED:
 533                 DEBUG(DEBUG_DEBUG, ("RDMA_CM_EVENT_DISCONNECTED\n"));
 534                 if ((rc=rdma_ack_cm_event(event)))
 535                         DEBUG(DEBUG_ERR, ("disc/rdma_ack_cm_event failed with %d\n", rc));
 536                 event = NULL; /* don't ack more */
 537
 538                 if (cma_id!=pctx->cm_id) {
 539                         DEBUG(DEBUG_ERR, ("client DISCONNECT event cm_id=%p\n", cma_id));
 540                         conn = talloc_get_type(cma_id->context, struct ibw_conn);
 541                         conn->state = IBWC_DISCONNECTED;
 542                         pctx->connstate_func(NULL, conn);
 543                 }
 544                 break;
 545
 546         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 547                 sprintf(ibw_lasterr, "cma detected device removal!\n");
 548                 goto error;
 549
 550         default:
 551                 sprintf(ibw_lasterr, "unknown event %d\n", event->event);
 552                 goto error;
 553         }
 554
 555         if (event!=NULL && (rc=rdma_ack_cm_event(event))) {
 556                 sprintf(ibw_lasterr, "rdma_ack_cm_event failed with %d\n", rc);
 557                 goto error;
 558         }
 559
 560         return;
 561 error:
 562         DEBUG(DEBUG_ERR, ("cm event handler: %s", ibw_lasterr));
 563
 564         if (event!=NULL) {
 565                 if (cma_id!=NULL && cma_id!=pctx->cm_id) {
 566                         conn = talloc_get_type(cma_id->context, struct ibw_conn);
 567                         if (conn) {
 568                                 conn->state = IBWC_ERROR;
 569                                 pctx->connstate_func(NULL, conn);
 570                         }
 571                 } else {
 572                         ctx->state = IBWS_ERROR;
 573                         pctx->connstate_func(ctx, NULL);
 574                 }
 575
 576                 if ((rc=rdma_ack_cm_event(event))!=0) {
 577                         DEBUG(DEBUG_ERR, ("rdma_ack_cm_event failed with %d\n", rc));
 578                 }
 579         }
 580
 581         return;
 582 }
 583
 584 static void ibw_event_handler_verbs(struct event_context *ev,
 585         struct fd_event *fde, uint16_t flags, void *private_data)
 586 {
 587         struct ibw_conn *conn = talloc_get_type(private_data, struct ibw_conn);
 588         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 589         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 590
 591         struct ibv_wc wc;
 592         int rc;
 593         struct ibv_cq *ev_cq;
 594         void          *ev_ctx;
 595
 596         DEBUG(DEBUG_DEBUG, ("ibw_event_handler_verbs(%u)\n", (uint32_t)flags));
 597
 598         /* TODO: check whether if it's good to have more channels here... */
 599         rc = ibv_get_cq_event(pconn->verbs_channel, &ev_cq, &ev_ctx);
 600         if (rc) {
 601                 sprintf(ibw_lasterr, "Failed to get cq_event with %d\n", rc);
 602                 goto error;
 603         }
 604         if (ev_cq != pconn->cq) {
 605                 sprintf(ibw_lasterr, "ev_cq(%p) != pconn->cq(%p)\n", ev_cq, pconn->cq);
 606                 goto error;
 607         }
 608         rc = ibv_req_notify_cq(pconn->cq, 0);
 609         if (rc) {
 610                 sprintf(ibw_lasterr, "Couldn't request CQ notification (%d)\n", rc);
 611                 goto error;
 612         }
 613
 614         while((rc=ibv_poll_cq(pconn->cq, 1, &wc))==1) {
 615                 if (wc.status) {
 616                         sprintf(ibw_lasterr, "cq completion failed status=%d, opcode=%d, rc=%d\n",
 617                                 wc.status, wc.opcode, rc);
 618                         goto error;
 619                 }
 620
 621                 switch(wc.opcode) {
 622                 case IBV_WC_SEND:
 623                         DEBUG(DEBUG_DEBUG, ("send completion\n"));
 624                         if (ibw_wc_send(conn, &wc))
 625                                 goto error;
 626                         break;
 627
 628                 case IBV_WC_RDMA_WRITE:
 629                         DEBUG(DEBUG_DEBUG, ("rdma write completion\n"));
 630                         break;
 631
 632                 case IBV_WC_RDMA_READ:
 633                         DEBUG(DEBUG_DEBUG, ("rdma read completion\n"));
 634                         break;
 635
 636                 case IBV_WC_RECV:
 637                         DEBUG(DEBUG_DEBUG, ("recv completion\n"));
 638                         if (ibw_wc_recv(conn, &wc))
 639                                 goto error;
 640                         break;
 641
 642                 default:
 643                         sprintf(ibw_lasterr, "unknown completion %d\n", wc.opcode);
 644                         goto error;
 645                 }
 646         }
 647         if (rc!=0) {
 648                 sprintf(ibw_lasterr, "ibv_poll_cq error %d\n", rc);
 649                 goto error;
 650         }
 651
 652         ibv_ack_cq_events(pconn->cq, 1);
 653
 654         return;
 655 error:
 656         ibv_ack_cq_events(pconn->cq, 1);
 657
 658         DEBUG(DEBUG_ERR, (ibw_lasterr));
 659
 660         if (conn->state!=IBWC_ERROR) {
 661                 conn->state = IBWC_ERROR;
 662                 pctx->connstate_func(NULL, conn);
 663         }
 664 }
 665
 666 static int ibw_process_queue(struct ibw_conn *conn)
 667 {
 668         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 669         struct ibw_ctx_priv *pctx;
 670         struct ibw_wr   *p;
 671         int     rc;
 672         uint32_t        msg_size;
 673
 674         if (pconn->queue==NULL)
 675                 return 0; /* NOP */
 676
 677         p = pconn->queue;
 678
 679         /* we must have at least 1 fragment to send */
 680         assert(p->queued_ref_cnt>0);
 681         p->queued_ref_cnt--;
 682
 683         pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 684         msg_size = (p->queued_ref_cnt) ? pctx->opts.recv_bufsize : p->queued_rlen;
 685
 686         assert(p->queued_msg!=NULL);
 687         assert(msg_size!=0);
 688
 689         DEBUG(DEBUG_DEBUG, ("ibw_process_queue refcnt=%d msgsize=%u\n",
 690                 p->queued_ref_cnt, msg_size));
 691
 692         rc = ibw_send_packet(conn, p->queued_msg, p, msg_size);
 693
 694         /* was this the last fragment? */
 695         if (p->queued_ref_cnt) {
 696                 p->queued_msg += pctx->opts.recv_bufsize;
 697         } else {
 698                 DLIST_REMOVE2(pconn->queue, p, qprev, qnext);
 699                 p->queued_msg = NULL;
 700         }
 701
 702         return rc;
 703 }
 704
 705 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc)
 706 {
 707         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 708         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 709         struct ibw_wr   *p;
 710         int     send_index;
 711
 712         DEBUG(DEBUG_DEBUG, ("ibw_wc_send(cmid: %p, wr_id: %u, bl: %u)\n",
 713                 pconn->cm_id, (uint32_t)wc->wr_id, (uint32_t)wc->byte_len));
 714
 715         assert(pconn->cm_id->qp->qp_num==wc->qp_num);
 716         assert(wc->wr_id >= pctx->opts.max_recv_wr);
 717         send_index = wc->wr_id - pctx->opts.max_recv_wr;
 718         pconn->wr_sent--;
 719
 720         if (send_index < pctx->opts.max_send_wr) {
 721                 DEBUG(DEBUG_DEBUG, ("ibw_wc_send#1 %u\n", (int)wc->wr_id));
 722                 p = pconn->wr_index[send_index];
 723                 if (p->buf_large!=NULL) {
 724                         if (p->ref_cnt) {
 725                                 /* awaiting more of it... */
 726                                 p->ref_cnt--;
 727                         } else {
 728                                 ibw_free_mr(&p->buf_large, &p->mr_large);
 729                                 DLIST_REMOVE(pconn->wr_list_used, p);
 730                                 DLIST_ADD(pconn->wr_list_avail, p);
 731                         }
 732                 } else { /* nasty - but necessary */
 733                         DLIST_REMOVE(pconn->wr_list_used, p);
 734                         DLIST_ADD(pconn->wr_list_avail, p);
 735                 }
 736         } else { /* "extra" request - not optimized */
 737                 DEBUG(DEBUG_DEBUG, ("ibw_wc_send#2 %u\n", (int)wc->wr_id));
 738                 for(p=pconn->extra_sent; p!=NULL; p=p->next)
 739                         if ((p->wr_id + pctx->opts.max_recv_wr)==(int)wc->wr_id)
 740                                 break;
 741                 if (p==NULL) {
 742                         sprintf(ibw_lasterr, "failed to find wr_id %d\n", (int)wc->wr_id);
 743                                 return -1;
 744                 }
 745                 if (p->ref_cnt) {
 746                         p->ref_cnt--;
 747                 } else {
 748                         ibw_free_mr(&p->buf_large, &p->mr_large);
 749                         DLIST_REMOVE(pconn->extra_sent, p);
 750                         DLIST_ADD(pconn->extra_avail, p);
 751                 }
 752         }
 753
 754         return ibw_process_queue(conn);
 755 }
 756
 757 static int ibw_append_to_part(struct ibw_conn_priv *pconn,
 758         struct ibw_part *part, char **pp, uint32_t add_len, int info)
 759 {
 760         DEBUG(DEBUG_DEBUG, ("ibw_append_to_part: cmid=%p, (bs=%u, len=%u, tr=%u), al=%u, i=%u\n",
 761                 pconn->cm_id, part->bufsize, part->len, part->to_read, add_len, info));
 762
 763         /* allocate more if necessary - it's an "evergrowing" buffer... */
 764         if (part->len + add_len > part->bufsize) {
 765                 if (part->buf==NULL) {
 766                         assert(part->len==0);
 767                         part->buf = talloc_size(pconn, add_len);
 768                         if (part->buf==NULL) {
 769                                 sprintf(ibw_lasterr, "recv talloc_size error (%u) #%d\n",
 770                                         add_len, info);
 771                                 return -1;
 772                         }
 773                         part->bufsize = add_len;
 774                 } else {
 775                         part->buf = talloc_realloc_size(pconn,
 776                                 part->buf, part->len + add_len);
 777                         if (part->buf==NULL) {
 778                                 sprintf(ibw_lasterr, "recv realloc error (%u + %u) #%d\n",
 779                                         part->len, add_len, info);
 780                                 return -1;
 781                         }
 782                 }
 783                 part->bufsize = part->len + add_len;
 784         }
 785
 786         /* consume pp */
 787         memcpy(part->buf + part->len, *pp, add_len);
 788         *pp += add_len;
 789         part->len += add_len;
 790         part->to_read -= add_len;
 791
 792         return 0;
 793 }
 794
 795 static int ibw_wc_mem_threshold(struct ibw_conn_priv *pconn,
 796         struct ibw_part *part, uint32_t threshold)
 797 {
 798         DEBUG(DEBUG_DEBUG, ("ibw_wc_mem_threshold: cmid=%p, (bs=%u, len=%u, tr=%u), thr=%u\n",
 799                 pconn->cm_id, part->bufsize, part->len, part->to_read, threshold));
 800
 801         if (part->bufsize > threshold) {
 802                 DEBUG(DEBUG_DEBUG, ("ibw_wc_mem_threshold: cmid=%p, %u > %u\n",
 803                         pconn->cm_id, part->bufsize, threshold));
 804                 talloc_free(part->buf);
 805                 part->buf = talloc_size(pconn, threshold);
 806                 if (part->buf==NULL) {
 807                         sprintf(ibw_lasterr, "talloc_size failed\n");
 808                         return -1;
 809                 }
 810                 part->bufsize = threshold;
 811         }
 812         return 0;
 813 }
 814
 815 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc)
 816 {
 817         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 818         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 819         struct ibw_part *part = &pconn->part;
 820         char    *p;
 821         uint32_t        remain = wc->byte_len;
 822
 823         DEBUG(DEBUG_DEBUG, ("ibw_wc_recv: cmid=%p, wr_id: %u, bl: %u\n",
 824                 pconn->cm_id, (uint32_t)wc->wr_id, remain));
 825
 826         assert(pconn->cm_id->qp->qp_num==wc->qp_num);
 827         assert((int)wc->wr_id < pctx->opts.max_recv_wr);
 828         assert(wc->byte_len <= pctx->opts.recv_bufsize);
 829
 830         p = pconn->buf_recv + ((int)wc->wr_id * pctx->opts.recv_bufsize);
 831
 832         while(remain) {
 833                 /* here always true: (part->len!=0 && part->to_read!=0) ||
 834                         (part->len==0 && part->to_read==0) */
 835                 if (part->len) { /* is there a partial msg to be continued? */
 836                         int read_len = (part->to_read<=remain) ? part->to_read : remain;
 837                         if (ibw_append_to_part(pconn, part, &p, read_len, 421))
 838                                 goto error;
 839                         remain -= read_len;
 840
 841                         if (part->len<=sizeof(uint32_t) && part->to_read==0) {
 842                                 assert(part->len==sizeof(uint32_t));
 843                                 /* set it again now... */
 844                                 part->to_read = *((uint32_t *)(part->buf)); /* TODO: ntohl */
 845                                 if (part->to_read<sizeof(uint32_t)) {
 846                                         sprintf(ibw_lasterr, "got msglen=%u #2\n", part->to_read);
 847                                         goto error;
 848                                 }
 849                                 part->to_read -= sizeof(uint32_t); /* it's already read */
 850                         }
 851
 852                         if (part->to_read==0) {
 853                                 if (pctx->receive_func(conn, part->buf, part->len) != 0) {
 854                                         goto error;
 855                                 }
 856                                 part->len = 0; /* tells not having partial data (any more) */
 857                                 if (ibw_wc_mem_threshold(pconn, part, pctx->opts.recv_threshold))
 858                                         goto error;
 859                         }
 860                 } else {
 861                         if (remain>=sizeof(uint32_t)) {
 862                                 uint32_t msglen = *(uint32_t *)p; /* TODO: ntohl */
 863                                 if (msglen<sizeof(uint32_t)) {
 864                                         sprintf(ibw_lasterr, "got msglen=%u\n", msglen);
 865                                         goto error;
 866                                 }
 867
 868                                 /* mostly awaited case: */
 869                                 if (msglen<=remain) {
 870                                         if (pctx->receive_func(conn, p, msglen) != 0) {
 871                                                 goto error;
 872                                         }
 873                                         p += msglen;
 874                                         remain -= msglen;
 875                                 } else {
 876                                         part->to_read = msglen;
 877                                         /* part->len is already 0 */
 878                                         if (ibw_append_to_part(pconn, part, &p, remain, 422))
 879                                                 goto error;
 880                                         remain = 0; /* to be continued ... */
 881                                         /* part->to_read > 0 here */
 882                                 }
 883                         } else { /* edge case: */
 884                                 part->to_read = sizeof(uint32_t);
 885                                 /* part->len is already 0 */
 886                                 if (ibw_append_to_part(pconn, part, &p, remain, 423))
 887                                         goto error;
 888                                 remain = 0;
 889                                 /* part->to_read > 0 here */
 890                         }
 891                 }
 892         } /* <remain> is always decreased at least by 1 */
 893
 894         if (ibw_refill_cq_recv(conn))
 895                 goto error;
 896
 897         return 0;
 898
 899 error:
 900         DEBUG(DEBUG_ERR, ("ibw_wc_recv error: %s", ibw_lasterr));
 901         return -1;
 902 }
 903
 904 static int ibw_process_init_attrs(struct ibw_initattr *attr, int nattr, struct ibw_opts *opts)
 905 {
 906         int     i;
 907         const char *name, *value;
 908
 909         DEBUG(DEBUG_DEBUG, ("ibw_process_init_attrs: nattr: %d\n", nattr));
 910
 911         opts->max_send_wr = IBW_MAX_SEND_WR;
 912         opts->max_recv_wr = IBW_MAX_RECV_WR;
 913         opts->recv_bufsize = IBW_RECV_BUFSIZE;
 914         opts->recv_threshold = IBW_RECV_THRESHOLD;
 915
 916         for(i=0; i<nattr; i++) {
 917                 name = attr[i].name;
 918                 value = attr[i].value;
 919
 920                 assert(name!=NULL && value!=NULL);
 921                 if (strcmp(name, "max_send_wr")==0)
 922                         opts->max_send_wr = atoi(value);
 923                 else if (strcmp(name, "max_recv_wr")==0)
 924                         opts->max_recv_wr = atoi(value);
 925                 else if (strcmp(name, "recv_bufsize")==0)
 926                         opts->recv_bufsize = atoi(value);
 927                 else if (strcmp(name, "recv_threshold")==0)
 928                         opts->recv_threshold = atoi(value);
 929                 else {
 930                         sprintf(ibw_lasterr, "ibw_init: unknown name %s\n", name);
 931                         return -1;
 932                 }
 933         }
 934         return 0;
 935 }
 936
 937 struct ibw_ctx *ibw_init(struct ibw_initattr *attr, int nattr,
 938         void *ctx_userdata,
 939         ibw_connstate_fn_t ibw_connstate,
 940         ibw_receive_fn_t ibw_receive,
 941         struct event_context *ectx)
 942 {
 943         struct ibw_ctx *ctx = talloc_zero(NULL, struct ibw_ctx);
 944         struct ibw_ctx_priv *pctx;
 945         int     rc;
 946
 947         DEBUG(DEBUG_DEBUG, ("ibw_init(ctx_userdata: %p, ectx: %p)\n", ctx_userdata, ectx));
 948
 949         /* initialize basic data structures */
 950         memset(ibw_lasterr, 0, IBW_LASTERR_BUFSIZE);
 951
 952         assert(ctx!=NULL);
 953         ibw_lasterr[0] = '\0';
 954         talloc_set_destructor(ctx, ibw_ctx_destruct);
 955         ctx->ctx_userdata = ctx_userdata;
 956
 957         pctx = talloc_zero(ctx, struct ibw_ctx_priv);
 958         talloc_set_destructor(pctx, ibw_ctx_priv_destruct);
 959         ctx->internal = (void *)pctx;
 960         assert(pctx!=NULL);
 961
 962         pctx->connstate_func = ibw_connstate;
 963         pctx->receive_func = ibw_receive;
 964
 965         pctx->ectx = ectx;
 966
 967         /* process attributes */
 968         if (ibw_process_init_attrs(attr, nattr, &pctx->opts))
 969                 goto cleanup;
 970
 971         /* init cm */
 972         pctx->cm_channel = rdma_create_event_channel();
 973         if (!pctx->cm_channel) {
 974                 sprintf(ibw_lasterr, "rdma_create_event_channel error %d\n", errno);
 975                 goto cleanup;
 976         }
 977
 978         pctx->cm_channel_event = event_add_fd(pctx->ectx, pctx,
 979                 pctx->cm_channel->fd, EVENT_FD_READ, ibw_event_handler_cm, ctx);
 980
 981 #if RDMA_USER_CM_MAX_ABI_VERSION >= 2
 982         rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx, RDMA_PS_TCP);
 983 #else
 984         rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx);
 985 #endif
 986         if (rc) {
 987                 rc = errno;
 988                 sprintf(ibw_lasterr, "rdma_create_id error %d\n", rc);
 989                 goto cleanup;
 990         }
 991         DEBUG(DEBUG_DEBUG, ("created cm_id %p\n", pctx->cm_id));
 992
 993         pctx->pagesize = sysconf(_SC_PAGESIZE);
 994
 995         return ctx;
 996         /* don't put code here */
 997 cleanup:
 998         DEBUG(DEBUG_ERR, (ibw_lasterr));
 999
1000         if (ctx)
1001                 talloc_free(ctx);
1002
1003         return NULL;
1004 }
1005
1006 int ibw_stop(struct ibw_ctx *ctx)
1007 {
1008         struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1009         struct ibw_conn *p;
1010
1011         DEBUG(DEBUG_DEBUG, ("ibw_stop\n"));
1012
1013         for(p=ctx->conn_list; p!=NULL; p=p->next) {
1014                 if (p->state==IBWC_ERROR || p->state==IBWC_CONNECTED) {
1015                         if (ibw_disconnect(p))
1016                                 return -1;
1017                 }
1018         }
1019
1020         ctx->state = IBWS_STOPPED;
1021         pctx->connstate_func(ctx, NULL);
1022
1023         return 0;
1024 }
1025
1026 int ibw_bind(struct ibw_ctx *ctx, struct sockaddr_in *my_addr)
1027 {
1028         struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1029         int     rc;
1030
1031         DEBUG(DEBUG_DEBUG, ("ibw_bind: addr=%s, port=%u\n",
1032                 inet_ntoa(my_addr->sin_addr), ntohs(my_addr->sin_port)));
1033         rc = rdma_bind_addr(pctx->cm_id, (struct sockaddr *) my_addr);
1034         if (rc) {
1035                 sprintf(ibw_lasterr, "rdma_bind_addr error %d\n", rc);
1036                 DEBUG(DEBUG_ERR, (ibw_lasterr));
1037                 return rc;
1038         }
1039         DEBUG(DEBUG_DEBUG, ("rdma_bind_addr successful\n"));
1040
1041         return 0;
1042 }
1043
1044 int ibw_listen(struct ibw_ctx *ctx, int backlog)
1045 {
1046         struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
1047         int     rc;
1048
1049         DEBUG(DEBUG_DEBUG, ("ibw_listen\n"));
1050         rc = rdma_listen(pctx->cm_id, backlog);
1051         if (rc) {
1052                 sprintf(ibw_lasterr, "rdma_listen failed: %d\n", rc);
1053                 DEBUG(DEBUG_ERR, (ibw_lasterr));
1054                 return rc;
1055         }
1056
1057         return 0;
1058 }
1059
1060 int ibw_accept(struct ibw_ctx *ctx, struct ibw_conn *conn, void *conn_userdata)
1061 {
1062         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1063         struct rdma_conn_param  conn_param;
1064         int     rc;
1065
1066         DEBUG(DEBUG_DEBUG, ("ibw_accept: cmid=%p\n", pconn->cm_id));
1067         conn->conn_userdata = conn_userdata;
1068
1069         memset(&conn_param, 0, sizeof(struct rdma_conn_param));
1070         conn_param.responder_resources = 1;
1071         conn_param.initiator_depth = 1;
1072         rc = rdma_accept(pconn->cm_id, &conn_param);
1073         if (rc) {
1074                 sprintf(ibw_lasterr, "rdma_accept failed %d\n", rc);
1075                 DEBUG(DEBUG_ERR, (ibw_lasterr));
1076                 return -1;;
1077         }
1078
1079         pconn->is_accepted = 1;
1080
1081         /* continued at RDMA_CM_EVENT_ESTABLISHED */
1082
1083         return 0;
1084 }
1085
1086 int ibw_connect(struct ibw_conn *conn, struct sockaddr_in *serv_addr, void *conn_userdata)
1087 {
1088         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1089         struct ibw_conn_priv *pconn = NULL;
1090         int     rc;
1091
1092         assert(conn!=NULL);
1093
1094         conn->conn_userdata = conn_userdata;
1095         pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1096         DEBUG(DEBUG_DEBUG, ("ibw_connect: addr=%s, port=%u\n", inet_ntoa(serv_addr->sin_addr),
1097                 ntohs(serv_addr->sin_port)));
1098
1099         /* clean previous - probably half - initialization */
1100         if (ibw_conn_priv_destruct(pconn)) {
1101                 DEBUG(DEBUG_ERR, ("ibw_connect/ibw_pconn_destruct failed for cm_id=%p\n", pconn->cm_id));
1102                 return -1;
1103         }
1104
1105         /* init cm */
1106 #if RDMA_USER_CM_MAX_ABI_VERSION >= 2
1107         rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn, RDMA_PS_TCP);
1108 #else
1109         rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn);
1110 #endif
1111         if (rc) {
1112                 rc = errno;
1113                 sprintf(ibw_lasterr, "ibw_connect/rdma_create_id error %d\n", rc);
1114                 talloc_free(conn);
1115                 return -1;
1116         }
1117         DEBUG(DEBUG_DEBUG, ("ibw_connect: rdma_create_id succeeded, cm_id=%p\n", pconn->cm_id));
1118
1119         rc = rdma_resolve_addr(pconn->cm_id, NULL, (struct sockaddr *) serv_addr, 2000);
1120         if (rc) {
1121                 sprintf(ibw_lasterr, "rdma_resolve_addr error %d\n", rc);
1122                 DEBUG(DEBUG_ERR, (ibw_lasterr));
1123                 talloc_free(conn);
1124                 return -1;
1125         }
1126
1127         /* continued at RDMA_CM_EVENT_ADDR_RESOLVED */
1128
1129         return 0;
1130 }
1131
1132 int ibw_disconnect(struct ibw_conn *conn)
1133 {
1134         int     rc;
1135         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1136
1137         DEBUG(DEBUG_DEBUG, ("ibw_disconnect: cmid=%p\n", pconn->cm_id));
1138
1139         assert(pconn!=NULL);
1140
1141         switch(conn->state) {
1142         case IBWC_ERROR:
1143                 ibw_conn_priv_destruct(pconn); /* do this here right now */
1144                 break;
1145         case IBWC_CONNECTED:
1146                 rc = rdma_disconnect(pconn->cm_id);
1147                 if (rc) {
1148                         sprintf(ibw_lasterr, "ibw_disconnect failed with %d\n", rc);
1149                         DEBUG(DEBUG_ERR, (ibw_lasterr));
1150                         return rc;
1151                 }
1152                 break;
1153         default:
1154                 DEBUG(DEBUG_DEBUG, ("invalid state for disconnect: %d\n", conn->state));
1155                 break;
1156         }
1157
1158         return 0;
1159 }
1160
1161 int ibw_alloc_send_buf(struct ibw_conn *conn, void **buf, void **key, uint32_t len)
1162 {
1163         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1164         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1165         struct ibw_wr *p = pconn->wr_list_avail;
1166
1167         if (p!=NULL) {
1168                 DEBUG(DEBUG_DEBUG, ("ibw_alloc_send_buf#1: cmid=%p, len=%d\n", pconn->cm_id, len));
1169
1170                 DLIST_REMOVE(pconn->wr_list_avail, p);
1171                 DLIST_ADD(pconn->wr_list_used, p);
1172
1173                 if (len <= pctx->opts.recv_bufsize) {
1174                         *buf = (void *)p->buf;
1175                 } else {
1176                         p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1177                         if (p->buf_large==NULL) {
1178                                 sprintf(ibw_lasterr, "ibw_alloc_mr#1 failed\n");
1179                                 goto error;
1180                         }
1181                         *buf = (void *)p->buf_large;
1182                 }
1183                 /* p->wr_id is already filled in ibw_init_memory */
1184         } else {
1185                 DEBUG(DEBUG_DEBUG, ("ibw_alloc_send_buf#2: cmid=%p, len=%d\n", pconn->cm_id, len));
1186                 /* not optimized */
1187                 p = pconn->extra_avail;
1188                 if (!p) {
1189                         p = pconn->extra_avail = talloc_zero(pconn, struct ibw_wr);
1190                         talloc_set_destructor(p, ibw_wr_destruct);
1191                         if (p==NULL) {
1192                                 sprintf(ibw_lasterr, "talloc_zero failed (emax: %u)\n", pconn->extra_max);
1193                                 goto error;
1194                         }
1195                         p->wr_id = pctx->opts.max_send_wr + pconn->extra_max;
1196                         pconn->extra_max++;
1197                         switch(pconn->extra_max) {
1198                                 case 1: DEBUG(DEBUG_INFO, ("warning: queue performed\n")); break;
1199                                 case 10: DEBUG(DEBUG_INFO, ("warning: queue reached 10\n")); break;
1200                                 case 100: DEBUG(DEBUG_INFO, ("warning: queue reached 100\n")); break;
1201                                 case 1000: DEBUG(DEBUG_INFO, ("warning: queue reached 1000\n")); break;
1202                                 default: break;
1203                         }
1204                 }
1205
1206                 p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1207                 if (p->buf_large==NULL) {
1208                         sprintf(ibw_lasterr, "ibw_alloc_mr#2 failed\n");
1209                         goto error;
1210                 }
1211                 *buf = (void *)p->buf_large;
1212
1213                 DLIST_REMOVE(pconn->extra_avail, p);
1214                 /* we don't have prepared index for this, so that
1215                  * we will have to find this by wr_id later on */
1216                 DLIST_ADD(pconn->extra_sent, p);
1217         }
1218
1219         *key = (void *)p;
1220
1221         return 0;
1222 error:
1223         DEBUG(DEBUG_ERR, ("ibw_alloc_send_buf error: %s", ibw_lasterr));
1224         return -1;
1225 }
1226
1227
1228 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len)
1229 {
1230         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1231         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1232         int     rc;
1233
1234         /* can we send it right now? */
1235         if (pconn->wr_sent<pctx->opts.max_send_wr) {
1236                 struct ibv_send_wr *bad_wr;
1237                 struct ibv_sge list = {
1238                         .addr   = (uintptr_t)buf,
1239                         .length = len,
1240                         .lkey   = pconn->mr_send->lkey
1241                 };
1242                 struct ibv_send_wr wr = {
1243                         .wr_id      = p->wr_id + pctx->opts.max_recv_wr,
1244                         .sg_list    = &list,
1245                         .num_sge    = 1,
1246                         .opcode     = IBV_WR_SEND,
1247                         .send_flags = IBV_SEND_SIGNALED,
1248                 };
1249
1250                 if (p->buf_large==NULL) {
1251                         DEBUG(DEBUG_DEBUG, ("ibw_send#normal(cmid: %p, wrid: %u, n: %d)\n",
1252                                 pconn->cm_id, (uint32_t)wr.wr_id, len));
1253                 } else {
1254                         DEBUG(DEBUG_DEBUG, ("ibw_send#large(cmid: %p, wrid: %u, n: %d)\n",
1255                                 pconn->cm_id, (uint32_t)wr.wr_id, len));
1256                         list.lkey = p->mr_large->lkey;
1257                 }
1258
1259                 rc = ibv_post_send(pconn->cm_id->qp, &wr, &bad_wr);
1260                 if (rc) {
1261                         sprintf(ibw_lasterr, "ibv_post_send error %d (%d)\n",
1262                                 rc, pconn->wr_sent);
1263                         goto error;
1264                 }
1265
1266                 pconn->wr_sent++;
1267
1268                 return rc;
1269         } /* else put the request into our own queue: */
1270
1271         DEBUG(DEBUG_DEBUG, ("ibw_send#queued(cmid: %p, len: %u)\n", pconn->cm_id, len));
1272
1273         /* TODO: clarify how to continue when state==IBWC_STOPPED */
1274
1275         /* to be sent by ibw_wc_send */
1276         /* regardless "normal" or [a part of] "large" packet */
1277         if (!p->queued_ref_cnt) {
1278                 DLIST_ADD_END2(pconn->queue, p, struct ibw_wr *,
1279                         qprev, qnext); /* TODO: optimize */
1280                 p->queued_msg = buf;
1281         }
1282         p->queued_ref_cnt++;
1283         p->queued_rlen = len; /* last wins; see ibw_wc_send */
1284
1285         return 0;
1286 error:
1287         DEBUG(DEBUG_ERR, (ibw_lasterr));
1288         return -1;
1289 }
1290
1291 int ibw_send(struct ibw_conn *conn, void *buf, void *key, uint32_t len)
1292 {
1293         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1294         struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1295         int     rc;
1296
1297         assert(len>=sizeof(uint32_t));
1298         assert((*((uint32_t *)buf)==len)); /* TODO: htonl */
1299
1300         if (len > pctx->opts.recv_bufsize) {
1301                 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1302                 int     rlen = len;
1303                 char    *packet = (char *)buf;
1304                 uint32_t        recv_bufsize = pctx->opts.recv_bufsize;
1305
1306                 DEBUG(DEBUG_DEBUG, ("ibw_send#frag(cmid: %p, buf: %p, len: %u)\n",
1307                         pconn->cm_id, buf, len));
1308
1309                 /* single threaded => no race here: */
1310                 assert(p->ref_cnt==0);
1311                 while(rlen > recv_bufsize) {
1312                         rc = ibw_send_packet(conn, packet, p, recv_bufsize);
1313                         if (rc)
1314                                 return rc;
1315                         packet += recv_bufsize;
1316                         rlen -= recv_bufsize;
1317                         p->ref_cnt++; /* not good to have it in ibw_send_packet */
1318                 }
1319                 if (rlen) {
1320                         rc = ibw_send_packet(conn, packet, p, rlen);
1321                         p->ref_cnt++; /* not good to have it in ibw_send_packet */
1322                 }
1323                 p->ref_cnt--; /* for the same handling */
1324         } else {
1325                 assert(p->ref_cnt==0);
1326                 assert(p->queued_ref_cnt==0);
1327
1328                 rc = ibw_send_packet(conn, buf, p, len);
1329         }
1330         return rc;
1331 }
1332
1333 int ibw_cancel_send_buf(struct ibw_conn *conn, void *buf, void *key)
1334 {
1335         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1336         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1337         struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1338
1339         assert(p!=NULL);
1340         assert(buf!=NULL);
1341         assert(conn!=NULL);
1342
1343         if (p->buf_large!=NULL)
1344                 ibw_free_mr(&p->buf_large, &p->mr_large);
1345
1346         /* parallel case */
1347         if (p->wr_id < pctx->opts.max_send_wr) {
1348                 DEBUG(DEBUG_DEBUG, ("ibw_cancel_send_buf#1 %u", (int)p->wr_id));
1349                 DLIST_REMOVE(pconn->wr_list_used, p);
1350                 DLIST_ADD(pconn->wr_list_avail, p);
1351         } else { /* "extra" packet */
1352                 DEBUG(DEBUG_DEBUG, ("ibw_cancel_send_buf#2 %u", (int)p->wr_id));
1353                 DLIST_REMOVE(pconn->extra_sent, p);
1354                 DLIST_ADD(pconn->extra_avail, p);
1355         }
1356
1357         return 0;
1358 }
1359
1360 const char *ibw_getLastError(void)
1361 {
1362         return ibw_lasterr;
1363 }