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