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 2 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, write to the Free Software
  21  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <stdio.h>
  27 #include <errno.h>
  28 #include <sys/types.h>
  29 #include <netinet/in.h>
  30 #include <sys/socket.h>
  31 #include <netdb.h>
  32 #include <arpa/inet.h>
  33 #include <malloc.h>
  34 #include <assert.h>
  35 #include <unistd.h>
  36
  37 #include "includes.h"
  38 #include "lib/events/events.h"
  39 #include "ibwrapper.h"
  40
  41 #include <rdma/rdma_cma.h>
  42 #include "infiniband/sa-kern-abi.h"
  43
  44 #include "ibwrapper_internal.h"
  45 #include "lib/util/dlinklist.h"
  46
  47 #define IBW_LASTERR_BUFSIZE 512
  48 static char ibw_lasterr[IBW_LASTERR_BUFSIZE];
  49
  50 #define IBW_MAX_SEND_WR 256
  51 #define IBW_MAX_RECV_WR 1024
  52 #define IBW_RECV_BUFSIZE 256
  53 #define IBW_RECV_THRESHOLD (1 * 1024 * 1024)
  54
  55 static void ibw_event_handler_verbs(struct event_context *ev,
  56         struct fd_event *fde, uint16_t flags, void *private_data);
  57 static int ibw_fill_cq(struct ibw_conn *conn);
  58 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc);
  59 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc);
  60 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len);
  61
  62 static void *ibw_alloc_mr(struct ibw_ctx_priv *pctx, struct ibw_conn_priv *pconn,
  63         uint32_t n, struct ibv_mr **ppmr)
  64 {
  65         void *buf;
  66
  67         DEBUG(10, ("ibw_alloc_mr(cmid=%p, n=%u)\n", pconn->cm_id, n));
  68         buf = memalign(pctx->pagesize, n);
  69         if (!buf) {
  70                 sprintf(ibw_lasterr, "couldn't allocate memory\n");
  71                 return NULL;
  72         }
  73
  74         *ppmr = ibv_reg_mr(pconn->pd, buf, n, IBV_ACCESS_LOCAL_WRITE);
  75         if (!*ppmr) {
  76                 sprintf(ibw_lasterr, "couldn't allocate mr\n");
  77                 free(buf);
  78                 return NULL;
  79         }
  80
  81         return buf;
  82 }
  83
  84 static void ibw_free_mr(char **ppbuf, struct ibv_mr **ppmr)
  85 {
  86         DEBUG(10, ("ibw_free_mr(%u %u)\n", (uint32_t)*ppbuf, (uint32_t)*ppmr));
  87         if (*ppmr!=NULL) {
  88                 ibv_dereg_mr(*ppmr);
  89                 *ppmr = NULL;
  90         }
  91         if (*ppbuf) {
  92                 free(*ppbuf);
  93                 *ppbuf = NULL;
  94         }
  95 }
  96
  97 static int ibw_init_memory(struct ibw_conn *conn)
  98 {
  99         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 100         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 101         struct ibw_opts *opts = &pctx->opts;
 102         int     i;
 103         struct ibw_wr   *p;
 104
 105         DEBUG(10, ("ibw_init_memory(cmid: %p)\n", pconn->cm_id));
 106         pconn->buf_send = ibw_alloc_mr(pctx, pconn,
 107                 opts->max_send_wr * opts->recv_bufsize, &pconn->mr_send);
 108         if (!pconn->buf_send) {
 109                 sprintf(ibw_lasterr, "couldn't allocate work send buf\n");
 110                 return -1;
 111         }
 112
 113         pconn->buf_recv = ibw_alloc_mr(pctx, pconn,
 114                 opts->max_recv_wr * opts->recv_bufsize, &pconn->mr_recv);
 115         if (!pconn->buf_recv) {
 116                 sprintf(ibw_lasterr, "couldn't allocate work recv buf\n");
 117                 return -1;
 118         }
 119
 120         pconn->wr_index = talloc_size(pconn, opts->max_send_wr * sizeof(struct ibw_wr *));
 121         assert(pconn->wr_index!=NULL);
 122
 123         for(i=0; i<opts->max_send_wr; i++) {
 124                 p = pconn->wr_index[i] = talloc_zero(pconn, struct ibw_wr);
 125                 p->buf = pconn->buf_send + (i * opts->recv_bufsize);
 126                 p->wr_id = i;
 127
 128                 DLIST_ADD(pconn->wr_list_avail, p);
 129         }
 130
 131         return 0;
 132 }
 133
 134 static int ibw_ctx_priv_destruct(struct ibw_ctx_priv *pctx)
 135 {
 136         DEBUG(10, ("ibw_ctx_priv_destruct(%u)\n", (uint32_t)pctx));
 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_channel_event) {
 144                 /* TODO: do we have to do this here? */
 145                 talloc_free(pctx->cm_channel_event);
 146                 pctx->cm_channel_event = NULL;
 147         }
 148         if (pctx->cm_id) {
 149                 rdma_destroy_id(pctx->cm_id);
 150                 pctx->cm_id = NULL;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static int ibw_ctx_destruct(struct ibw_ctx *ctx)
 157 {
 158         DEBUG(10, ("ibw_ctx_destruct(%u)\n", (uint32_t)ctx));
 159         return 0;
 160 }
 161
 162 static int ibw_conn_priv_destruct(struct ibw_conn_priv *pconn)
 163 {
 164         DEBUG(10, ("ibw_conn_priv_destruct(%p, cmid: %p)\n",
 165                 pconn, pconn->cm_id));
 166
 167         /* pconn->wr_index is freed by talloc */
 168         /* pconn->wr_index[i] are freed by talloc */
 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         /* must be freed here because its order is important */
 187         if (pconn->verbs_channel_event) {
 188                 talloc_free(pconn->verbs_channel_event);
 189                 pconn->verbs_channel_event = NULL;
 190         }
 191
 192         /* free memory regions */
 193         ibw_free_mr(&pconn->buf_send, &pconn->mr_send);
 194         ibw_free_mr(&pconn->buf_recv, &pconn->mr_recv);
 195
 196         if (pconn->pd) {
 197                 ibv_dealloc_pd(pconn->pd);
 198                 pconn->pd = NULL;
 199                 DEBUG(10, ("pconn=%p pd deallocated\n", pconn));
 200         }
 201
 202         if (pconn->cm_id) {
 203                 rdma_destroy_id(pconn->cm_id);
 204                 pconn->cm_id = NULL;
 205                 DEBUG(10, ("pconn=%p cm_id destroyed\n", pconn));
 206         }
 207
 208         return 0;
 209 }
 210
 211 static int ibw_wr_destruct(struct ibw_wr *wr)
 212 {
 213         if (wr->buf_large!=NULL)
 214                 ibw_free_mr(&wr->buf_large, &wr->mr_large);
 215         return 0;
 216 }
 217
 218 static int ibw_conn_destruct(struct ibw_conn *conn)
 219 {
 220         DEBUG(10, ("ibw_conn_destruct(%u)\n", (uint32_t)conn));
 221
 222         /* important here: ctx is a talloc _parent_ */
 223         DLIST_REMOVE(conn->ctx->conn_list, conn);
 224         return 0;
 225 }
 226
 227 struct ibw_conn *ibw_conn_new(struct ibw_ctx *ctx, TALLOC_CTX *mem_ctx)
 228 {
 229         struct ibw_conn *conn;
 230         struct ibw_conn_priv *pconn;
 231
 232         assert(ctx!=NULL);
 233
 234         conn = talloc_zero(mem_ctx, struct ibw_conn);
 235         assert(conn!=NULL);
 236         talloc_set_destructor(conn, ibw_conn_destruct);
 237
 238         pconn = talloc_zero(conn, struct ibw_conn_priv);
 239         assert(pconn!=NULL);
 240         talloc_set_destructor(pconn, ibw_conn_priv_destruct);
 241
 242         conn->ctx = ctx;
 243         conn->internal = (void *)pconn;
 244
 245         DLIST_ADD(ctx->conn_list, conn);
 246
 247         return conn;
 248 }
 249
 250 static int ibw_setup_cq_qp(struct ibw_conn *conn)
 251 {
 252         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 253         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 254         struct ibv_qp_init_attr init_attr;
 255         struct ibv_qp_attr attr;
 256         int rc;
 257
 258         DEBUG(10, ("ibw_setup_cq_qp(cmid: %p)\n", pconn->cm_id));
 259
 260         /* init verbs */
 261         pconn->verbs_channel = ibv_create_comp_channel(pconn->cm_id->verbs);
 262         if (!pconn->verbs_channel) {
 263                 sprintf(ibw_lasterr, "ibv_create_comp_channel failed %d\n", errno);
 264                 return -1;
 265         }
 266         DEBUG(10, ("created channel %p\n", pconn->verbs_channel));
 267
 268         pconn->verbs_channel_event = event_add_fd(pctx->ectx, NULL, /* not pconn or conn */
 269                 pconn->verbs_channel->fd, EVENT_FD_READ, ibw_event_handler_verbs, conn);
 270
 271         pconn->pd = ibv_alloc_pd(pconn->cm_id->verbs);
 272         if (!pconn->pd) {
 273                 sprintf(ibw_lasterr, "ibv_alloc_pd failed %d\n", errno);
 274                 return -1;
 275         }
 276         DEBUG(10, ("created pd %p\n", pconn->pd));
 277
 278         /* init mr */
 279         if (ibw_init_memory(conn))
 280                 return -1;
 281
 282         /* init cq */
 283         pconn->cq = ibv_create_cq(pconn->cm_id->verbs,
 284                 pctx->opts.max_recv_wr + pctx->opts.max_send_wr,
 285                 conn, pconn->verbs_channel, 0);
 286         if (pconn->cq==NULL) {
 287                 sprintf(ibw_lasterr, "ibv_create_cq failed\n");
 288                 return -1;
 289         }
 290
 291         rc = ibv_req_notify_cq(pconn->cq, 0);
 292         if (rc) {
 293                 sprintf(ibw_lasterr, "ibv_req_notify_cq failed with %d\n", rc);
 294                 return rc;
 295         }
 296
 297         /* init qp */
 298         memset(&init_attr, 0, sizeof(init_attr));
 299         init_attr.cap.max_send_wr = pctx->opts.max_send_wr;
 300         init_attr.cap.max_recv_wr = pctx->opts.max_recv_wr;
 301         init_attr.cap.max_recv_sge = 1;
 302         init_attr.cap.max_send_sge = 1;
 303         init_attr.qp_type = IBV_QPT_RC;
 304         init_attr.send_cq = pconn->cq;
 305         init_attr.recv_cq = pconn->cq;
 306
 307         rc = rdma_create_qp(pconn->cm_id, pconn->pd, &init_attr);
 308         if (rc) {
 309                 sprintf(ibw_lasterr, "rdma_create_qp failed with %d\n", rc);
 310                 return rc;
 311         }
 312         /* elase result is in pconn->cm_id->qp */
 313
 314         rc = ibv_query_qp(pconn->cm_id->qp, &attr, IBV_QP_PATH_MTU, &init_attr);
 315         if (rc) {
 316                 sprintf(ibw_lasterr, "ibv_query_qp failed with %d\n", rc);
 317                 return rc;
 318         }
 319
 320         return ibw_fill_cq(conn);
 321 }
 322
 323 static int ibw_refill_cq_recv(struct ibw_conn *conn)
 324 {
 325         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 326         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 327         int     rc;
 328         struct ibv_sge list = {
 329                 .addr   = (uintptr_t) NULL, /* filled below */
 330                 .length = pctx->opts.recv_bufsize,
 331                 .lkey   = pconn->mr_recv->lkey /* always the same */
 332         };
 333         struct ibv_recv_wr wr = {
 334                 .wr_id      = 0, /* filled below */
 335                 .sg_list    = &list,
 336                 .num_sge    = 1,
 337         };
 338         struct ibv_recv_wr *bad_wr;
 339
 340         DEBUG(10, ("ibw_refill_cq_recv(cmid: %p)\n", pconn->cm_id));
 341
 342         list.addr = (uintptr_t) pconn->buf_recv + pctx->opts.recv_bufsize * pconn->recv_index;
 343         wr.wr_id = pconn->recv_index;
 344         pconn->recv_index = (pconn->recv_index + 1) % pctx->opts.max_recv_wr;
 345
 346         rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
 347         if (rc) {
 348                 sprintf(ibw_lasterr, "refill/ibv_post_recv failed with %d\n", rc);
 349                 DEBUG(0, (ibw_lasterr));
 350                 return -2;
 351         }
 352
 353         return 0;
 354 }
 355
 356 static int ibw_fill_cq(struct ibw_conn *conn)
 357 {
 358         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 359         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 360         int     i, rc;
 361         struct ibv_sge list = {
 362                 .addr   = (uintptr_t) NULL, /* filled below */
 363                 .length = pctx->opts.recv_bufsize,
 364                 .lkey   = pconn->mr_recv->lkey /* always the same */
 365         };
 366         struct ibv_recv_wr wr = {
 367                 .wr_id      = 0, /* filled below */
 368                 .sg_list    = &list,
 369                 .num_sge    = 1,
 370         };
 371         struct ibv_recv_wr *bad_wr;
 372
 373         DEBUG(10, ("ibw_fill_cq(cmid: %p)\n", pconn->cm_id));
 374
 375         for(i = pctx->opts.max_recv_wr; i!=0; i--) {
 376                 list.addr = (uintptr_t) pconn->buf_recv + pctx->opts.recv_bufsize * pconn->recv_index;
 377                 wr.wr_id = pconn->recv_index;
 378                 pconn->recv_index = (pconn->recv_index + 1) % pctx->opts.max_recv_wr;
 379
 380                 rc = ibv_post_recv(pconn->cm_id->qp, &wr, &bad_wr);
 381                 if (rc) {
 382                         sprintf(ibw_lasterr, "fill/ibv_post_recv failed with %d\n", rc);
 383                         DEBUG(0, (ibw_lasterr));
 384                         return -2;
 385                 }
 386         }
 387
 388         return 0;
 389 }
 390
 391 static int ibw_manage_connect(struct ibw_conn *conn)
 392 {
 393         struct rdma_conn_param conn_param;
 394         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 395         int     rc;
 396
 397         DEBUG(10, ("ibw_manage_connect(cmid: %p)\n", pconn->cm_id));
 398
 399         if (ibw_setup_cq_qp(conn))
 400                 return -1;
 401
 402         /* cm connect */
 403         memset(&conn_param, 0, sizeof conn_param);
 404         conn_param.responder_resources = 1;
 405         conn_param.initiator_depth = 1;
 406         conn_param.retry_count = 10;
 407
 408         rc = rdma_connect(pconn->cm_id, &conn_param);
 409         if (rc)
 410                 sprintf(ibw_lasterr, "rdma_connect error %d\n", rc);
 411
 412         return rc;
 413 }
 414
 415 static void ibw_event_handler_cm(struct event_context *ev,
 416         struct fd_event *fde, uint16_t flags, void *private_data)
 417 {
 418         int     rc;
 419         struct ibw_ctx  *ctx = talloc_get_type(private_data, struct ibw_ctx);
 420         struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
 421         struct ibw_conn *conn = NULL;
 422         struct ibw_conn_priv *pconn = NULL;
 423         struct rdma_cm_id *cma_id = NULL;
 424         struct rdma_cm_event *event = NULL;
 425
 426         assert(ctx!=NULL);
 427
 428         rc = rdma_get_cm_event(pctx->cm_channel, &event);
 429         if (rc) {
 430                 ctx->state = IBWS_ERROR;
 431                 event = NULL;
 432                 sprintf(ibw_lasterr, "rdma_get_cm_event error %d\n", rc);
 433                 goto error;
 434         }
 435         cma_id = event->id;
 436
 437         DEBUG(10, ("cma_event type %d cma_id %p (%s)\n", event->event, cma_id,
 438                   (cma_id == pctx->cm_id) ? "parent" : "child"));
 439
 440         switch (event->event) {
 441         case RDMA_CM_EVENT_ADDR_RESOLVED:
 442                 DEBUG(11, ("RDMA_CM_EVENT_ADDR_RESOLVED\n"));
 443                 /* continuing from ibw_connect ... */
 444                 rc = rdma_resolve_route(cma_id, 2000);
 445                 if (rc) {
 446                         sprintf(ibw_lasterr, "rdma_resolve_route error %d\n", rc);
 447                         goto error;
 448                 }
 449                 /* continued at RDMA_CM_EVENT_ROUTE_RESOLVED */
 450                 break;
 451
 452         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 453                 DEBUG(11, ("RDMA_CM_EVENT_ROUTE_RESOLVED\n"));
 454                 /* after RDMA_CM_EVENT_ADDR_RESOLVED: */
 455                 assert(cma_id->context!=NULL);
 456                 conn = talloc_get_type(cma_id->context, struct ibw_conn);
 457
 458                 rc = ibw_manage_connect(conn);
 459                 if (rc)
 460                         goto error;
 461
 462                 break;
 463
 464         case RDMA_CM_EVENT_CONNECT_REQUEST:
 465                 DEBUG(11, ("RDMA_CM_EVENT_CONNECT_REQUEST\n"));
 466                 ctx->state = IBWS_CONNECT_REQUEST;
 467                 conn = ibw_conn_new(ctx, ctx);
 468                 pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 469                 pconn->cm_id = cma_id; /* !!! event will be freed but id not */
 470                 cma_id->context = (void *)conn;
 471                 DEBUG(10, ("pconn->cm_id %p\n", pconn->cm_id));
 472
 473                 if (ibw_setup_cq_qp(conn))
 474                         goto error;
 475
 476                 conn->state = IBWC_INIT;
 477                 pctx->connstate_func(ctx, conn);
 478
 479                 /* continued at ibw_accept when invoked by the func above */
 480                 if (!pconn->is_accepted) {
 481                         rc = rdma_reject(cma_id, NULL, 0);
 482                         if (rc)
 483                                 DEBUG(0, ("rdma_reject failed with rc=%d\n", rc));
 484                         talloc_free(conn);
 485                         DEBUG(10, ("pconn->cm_id %p wasn't accepted\n", pconn->cm_id));
 486                 }
 487
 488                 /* TODO: clarify whether if it's needed by upper layer: */
 489                 ctx->state = IBWS_READY;
 490                 pctx->connstate_func(ctx, NULL);
 491
 492                 /* NOTE: more requests can arrive until RDMA_CM_EVENT_ESTABLISHED ! */
 493                 break;
 494
 495         case RDMA_CM_EVENT_ESTABLISHED:
 496                 /* expected after ibw_accept and ibw_connect[not directly] */
 497                 DEBUG(0, ("ESTABLISHED (conn: %p)\n", cma_id->context));
 498                 conn = talloc_get_type(cma_id->context, struct ibw_conn);
 499                 assert(conn!=NULL); /* important assumption */
 500
 501                 DEBUG(10, ("ibw_setup_cq_qp succeeded (cmid=%p)\n", cma_id));
 502
 503                 /* client conn is up */
 504                 conn->state = IBWC_CONNECTED;
 505
 506                 /* both ctx and conn have changed */
 507                 pctx->connstate_func(ctx, conn);
 508                 break;
 509
 510         case RDMA_CM_EVENT_ADDR_ERROR:
 511                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ADDR_ERROR, error %d\n", event->status);
 512         case RDMA_CM_EVENT_ROUTE_ERROR:
 513                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_ROUTE_ERROR, error %d\n", event->status);
 514         case RDMA_CM_EVENT_CONNECT_ERROR:
 515                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_CONNECT_ERROR, error %d\n", event->status);
 516         case RDMA_CM_EVENT_UNREACHABLE:
 517                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_UNREACHABLE, error %d\n", event->status);
 518         case RDMA_CM_EVENT_REJECTED:
 519                 sprintf(ibw_lasterr, "RDMA_CM_EVENT_REJECTED, error %d\n", event->status);
 520                 conn = talloc_get_type(cma_id->context, struct ibw_conn);
 521                 if (conn) {
 522                         if ((rc=rdma_ack_cm_event(event)))
 523                                 DEBUG(0, ("reject/rdma_ack_cm_event failed with %d\n", rc));
 524                         event = NULL; /* not to touch cma_id or conn */
 525                         conn->state = IBWC_ERROR;
 526                         /* it should free the conn */
 527                         pctx->connstate_func(NULL, conn);
 528                 }
 529                 goto error;
 530
 531         case RDMA_CM_EVENT_DISCONNECTED:
 532                 DEBUG(11, ("RDMA_CM_EVENT_DISCONNECTED\n"));
 533                 if ((rc=rdma_ack_cm_event(event)))
 534                         DEBUG(0, ("disc/rdma_ack_cm_event failed with %d\n", rc));
 535                 event = NULL; /* don't ack more */
 536
 537                 if (cma_id!=pctx->cm_id) {
 538                         DEBUG(0, ("client DISCONNECT event cm_id=%p\n", cma_id));
 539                         conn = talloc_get_type(cma_id->context, struct ibw_conn);
 540                         conn->state = IBWC_DISCONNECTED;
 541                         pctx->connstate_func(NULL, conn);
 542                 }
 543                 break;
 544
 545         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 546                 sprintf(ibw_lasterr, "cma detected device removal!\n");
 547                 goto error;
 548
 549         default:
 550                 sprintf(ibw_lasterr, "unknown event %d\n", event->event);
 551                 goto error;
 552         }
 553
 554         if (event!=NULL && (rc=rdma_ack_cm_event(event))) {
 555                 sprintf(ibw_lasterr, "rdma_ack_cm_event failed with %d\n", rc);
 556                 goto error;
 557         }
 558
 559         return;
 560 error:
 561         DEBUG(0, ("cm event handler: %s", ibw_lasterr));
 562
 563         if (event!=NULL) {
 564                 if (cma_id!=NULL && cma_id!=pctx->cm_id) {
 565                         conn = talloc_get_type(cma_id->context, struct ibw_conn);
 566                         if (conn) {
 567                                 conn->state = IBWC_ERROR;
 568                                 pctx->connstate_func(NULL, conn);
 569                         }
 570                 } else {
 571                         ctx->state = IBWS_ERROR;
 572                         pctx->connstate_func(ctx, NULL);
 573                 }
 574
 575                 if ((rc=rdma_ack_cm_event(event))!=0) {
 576                         DEBUG(0, ("rdma_ack_cm_event failed with %d\n", rc));
 577                 }
 578         }
 579
 580         return;
 581 }
 582
 583 static void ibw_event_handler_verbs(struct event_context *ev,
 584         struct fd_event *fde, uint16_t flags, void *private_data)
 585 {
 586         struct ibw_conn *conn = talloc_get_type(private_data, struct ibw_conn);
 587         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 588         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 589
 590         struct ibv_wc wc;
 591         int rc;
 592         struct ibv_cq *ev_cq;
 593         void          *ev_ctx;
 594
 595         DEBUG(10, ("ibw_event_handler_verbs(%u)\n", (uint32_t)flags));
 596
 597         /* TODO: check whether if it's good to have more channels here... */
 598         rc = ibv_get_cq_event(pconn->verbs_channel, &ev_cq, &ev_ctx);
 599         if (rc) {
 600                 sprintf(ibw_lasterr, "Failed to get cq_event with %d\n", rc);
 601                 goto error;
 602         }
 603         if (ev_cq != pconn->cq) {
 604                 sprintf(ibw_lasterr, "ev_cq(%p) != pconn->cq(%p)\n", ev_cq, pconn->cq);
 605                 goto error;
 606         }
 607         rc = ibv_req_notify_cq(pconn->cq, 0);
 608         if (rc) {
 609                 sprintf(ibw_lasterr, "Couldn't request CQ notification (%d)\n", rc);
 610                 goto error;
 611         }
 612
 613         while((rc=ibv_poll_cq(pconn->cq, 1, &wc))==1) {
 614                 if (wc.status) {
 615                         sprintf(ibw_lasterr, "cq completion failed status=%d, opcode=%d, rc=%d\n",
 616                                 wc.status, wc.opcode, rc);
 617                         goto error;
 618                 }
 619
 620                 switch(wc.opcode) {
 621                 case IBV_WC_SEND:
 622                         DEBUG(10, ("send completion\n"));
 623                         if (ibw_wc_send(conn, &wc))
 624                                 goto error;
 625                         break;
 626
 627                 case IBV_WC_RDMA_WRITE:
 628                         DEBUG(10, ("rdma write completion\n"));
 629                         break;
 630
 631                 case IBV_WC_RDMA_READ:
 632                         DEBUG(10, ("rdma read completion\n"));
 633                         break;
 634
 635                 case IBV_WC_RECV:
 636                         DEBUG(10, ("recv completion\n"));
 637                         if (ibw_wc_recv(conn, &wc))
 638                                 goto error;
 639                         break;
 640
 641                 default:
 642                         sprintf(ibw_lasterr, "unknown completion %d\n", wc.opcode);
 643                         goto error;
 644                 }
 645         }
 646         if (rc!=0) {
 647                 sprintf(ibw_lasterr, "ibv_poll_cq error %d\n", rc);
 648                 goto error;
 649         }
 650
 651         ibv_ack_cq_events(pconn->cq, 1);
 652
 653         return;
 654 error:
 655         ibv_ack_cq_events(pconn->cq, 1);
 656
 657         DEBUG(0, (ibw_lasterr));
 658
 659         if (conn->state!=IBWC_ERROR) {
 660                 conn->state = IBWC_ERROR;
 661                 pctx->connstate_func(NULL, conn);
 662         }
 663 }
 664
 665 static int ibw_process_queue(struct ibw_conn *conn)
 666 {
 667         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 668         struct ibw_ctx_priv *pctx;
 669         struct ibw_wr   *p;
 670         int     rc;
 671         uint32_t        msg_size;
 672
 673         if (pconn->queue==NULL)
 674                 return 0; /* NOP */
 675
 676         p = pconn->queue;
 677
 678         /* we must have at least 1 fragment to send */
 679         assert(p->queued_ref_cnt>0);
 680         p->queued_ref_cnt--;
 681
 682         pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 683         msg_size = (p->queued_ref_cnt) ? pctx->opts.recv_bufsize : p->queued_rlen;
 684
 685         assert(p->queued_msg!=NULL);
 686         assert(msg_size!=0);
 687
 688         DEBUG(10, ("ibw_process_queue refcnt=%d msgsize=%u\n",
 689                 p->queued_ref_cnt, msg_size));
 690
 691         rc = ibw_send_packet(conn, p->queued_msg, p, msg_size);
 692
 693         /* was this the last fragment? */
 694         if (p->queued_ref_cnt) {
 695                 p->queued_msg += pctx->opts.recv_bufsize;
 696         } else {
 697                 DLIST_REMOVE2(pconn->queue, p, qprev, qnext);
 698                 p->queued_msg = NULL;
 699         }
 700
 701         return rc;
 702 }
 703
 704 static int ibw_wc_send(struct ibw_conn *conn, struct ibv_wc *wc)
 705 {
 706         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 707         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 708         struct ibw_wr   *p;
 709         int     send_index;
 710
 711         DEBUG(10, ("ibw_wc_send(cmid: %p, wr_id: %u, bl: %u)\n",
 712                 pconn->cm_id, (uint32_t)wc->wr_id, (uint32_t)wc->byte_len));
 713
 714         assert(pconn->cm_id->qp->qp_num==wc->qp_num);
 715         assert(wc->wr_id >= pctx->opts.max_recv_wr);
 716         send_index = wc->wr_id - pctx->opts.max_recv_wr;
 717         pconn->wr_sent--;
 718
 719         if (send_index < pctx->opts.max_send_wr) {
 720                 DEBUG(10, ("ibw_wc_send#1 %u\n", (int)wc->wr_id));
 721                 p = pconn->wr_index[send_index];
 722                 if (p->buf_large!=NULL) {
 723                         if (p->ref_cnt) {
 724                                 /* awaiting more of it... */
 725                                 p->ref_cnt--;
 726                         } else {
 727                                 ibw_free_mr(&p->buf_large, &p->mr_large);
 728                                 DLIST_REMOVE(pconn->wr_list_used, p);
 729                                 DLIST_ADD(pconn->wr_list_avail, p);
 730                         }
 731                 } else { /* nasty - but necessary */
 732                         DLIST_REMOVE(pconn->wr_list_used, p);
 733                         DLIST_ADD(pconn->wr_list_avail, p);
 734                 }
 735         } else { /* "extra" request - not optimized */
 736                 DEBUG(10, ("ibw_wc_send#2 %u\n", (int)wc->wr_id));
 737                 for(p=pconn->extra_sent; p!=NULL; p=p->next)
 738                         if ((p->wr_id + pctx->opts.max_recv_wr)==(int)wc->wr_id)
 739                                 break;
 740                 if (p==NULL) {
 741                         sprintf(ibw_lasterr, "failed to find wr_id %d\n", (int)wc->wr_id);
 742                                 return -1;
 743                 }
 744                 if (p->ref_cnt) {
 745                         p->ref_cnt--;
 746                 } else {
 747                         ibw_free_mr(&p->buf_large, &p->mr_large);
 748                         DLIST_REMOVE(pconn->extra_sent, p);
 749                         DLIST_ADD(pconn->extra_avail, p);
 750                 }
 751         }
 752
 753         return ibw_process_queue(conn);
 754 }
 755
 756 static int ibw_append_to_part(struct ibw_conn_priv *pconn,
 757         struct ibw_part *part, char **pp, uint32_t add_len, int info)
 758 {
 759         DEBUG(10, ("ibw_append_to_part: cmid=%p, (bs=%u, len=%u, tr=%u), al=%u, i=%u\n",
 760                 pconn->cm_id, part->bufsize, part->len, part->to_read, add_len, info));
 761
 762         /* allocate more if necessary - it's an "evergrowing" buffer... */
 763         if (part->len + add_len > part->bufsize) {
 764                 if (part->buf==NULL) {
 765                         assert(part->len==0);
 766                         part->buf = talloc_size(pconn, add_len);
 767                         if (part->buf==NULL) {
 768                                 sprintf(ibw_lasterr, "recv talloc_size error (%u) #%d\n",
 769                                         add_len, info);
 770                                 return -1;
 771                         }
 772                         part->bufsize = add_len;
 773                 } else {
 774                         part->buf = talloc_realloc_size(pconn,
 775                                 part->buf, part->len + add_len);
 776                         if (part->buf==NULL) {
 777                                 sprintf(ibw_lasterr, "recv realloc error (%u + %u) #%d\n",
 778                                         part->len, add_len, info);
 779                                 return -1;
 780                         }
 781                 }
 782                 part->bufsize = part->len + add_len;
 783         }
 784
 785         /* consume pp */
 786         memcpy(part->buf + part->len, *pp, add_len);
 787         *pp += add_len;
 788         part->len += add_len;
 789         part->to_read -= add_len;
 790
 791         return 0;
 792 }
 793
 794 static int ibw_wc_mem_threshold(struct ibw_conn_priv *pconn,
 795         struct ibw_part *part, uint32_t threshold)
 796 {
 797         DEBUG(10, ("ibw_wc_mem_threshold: cmid=%p, (bs=%u, len=%u, tr=%u), thr=%u\n",
 798                 pconn->cm_id, part->bufsize, part->len, part->to_read, threshold));
 799
 800         if (part->bufsize > threshold) {
 801                 DEBUG(3, ("ibw_wc_mem_threshold: cmid=%p, %u > %u\n",
 802                         pconn->cm_id, part->bufsize, threshold));
 803                 talloc_free(part->buf);
 804                 part->buf = talloc_size(pconn, threshold);
 805                 if (part->buf==NULL) {
 806                         sprintf(ibw_lasterr, "talloc_size failed\n");
 807                         return -1;
 808                 }
 809                 part->bufsize = threshold;
 810         }
 811         return 0;
 812 }
 813
 814 static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc)
 815 {
 816         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
 817         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
 818         struct ibw_part *part = &pconn->part;
 819         char    *p;
 820         uint32_t        remain = wc->byte_len;
 821
 822         DEBUG(10, ("ibw_wc_recv: cmid=%p, wr_id: %u, bl: %u\n",
 823                 pconn->cm_id, (uint32_t)wc->wr_id, remain));
 824
 825         assert(pconn->cm_id->qp->qp_num==wc->qp_num);
 826         assert((int)wc->wr_id < pctx->opts.max_recv_wr);
 827         assert(wc->byte_len <= pctx->opts.recv_bufsize);
 828
 829         p = pconn->buf_recv + ((int)wc->wr_id * pctx->opts.recv_bufsize);
 830
 831         while(remain) {
 832                 /* here always true: (part->len!=0 && part->to_read!=0) ||
 833                         (part->len==0 && part->to_read==0) */
 834                 if (part->len) { /* is there a partial msg to be continued? */
 835                         int read_len = (part->to_read<=remain) ? part->to_read : remain;
 836                         if (ibw_append_to_part(pconn, part, &p, read_len, 421))
 837                                 goto error;
 838                         remain -= read_len;
 839
 840                         if (part->len<=sizeof(uint32_t) && part->to_read==0) {
 841                                 assert(part->len==sizeof(uint32_t));
 842                                 /* set it again now... */
 843                                 part->to_read = *((uint32_t *)(part->buf)); /* TODO: ntohl */
 844                                 if (part->to_read<sizeof(uint32_t)) {
 845                                         sprintf(ibw_lasterr, "got msglen=%u #2\n", part->to_read);
 846                                         goto error;
 847                                 }
 848                                 part->to_read -= sizeof(uint32_t); /* it's already read */
 849                         }
 850
 851                         if (part->to_read==0) {
 852                                 pctx->receive_func(conn, part->buf, part->len);
 853                                 part->len = 0; /* tells not having partial data (any more) */
 854                                 if (ibw_wc_mem_threshold(pconn, part, pctx->opts.recv_threshold))
 855                                         goto error;
 856                         }
 857                 } else {
 858                         if (remain>=sizeof(uint32_t)) {
 859                                 uint32_t msglen = *(uint32_t *)p; /* TODO: ntohl */
 860                                 if (msglen<sizeof(uint32_t)) {
 861                                         sprintf(ibw_lasterr, "got msglen=%u\n", msglen);
 862                                         goto error;
 863                                 }
 864
 865                                 /* mostly awaited case: */
 866                                 if (msglen<=remain) {
 867                                         pctx->receive_func(conn, p, msglen);
 868                                         p += msglen;
 869                                         remain -= msglen;
 870                                 } else {
 871                                         part->to_read = msglen;
 872                                         /* part->len is already 0 */
 873                                         if (ibw_append_to_part(pconn, part, &p, remain, 422))
 874                                                 goto error;
 875                                         remain = 0; /* to be continued ... */
 876                                         /* part->to_read > 0 here */
 877                                 }
 878                         } else { /* edge case: */
 879                                 part->to_read = sizeof(uint32_t);
 880                                 /* part->len is already 0 */
 881                                 if (ibw_append_to_part(pconn, part, &p, remain, 423))
 882                                         goto error;
 883                                 remain = 0;
 884                                 /* part->to_read > 0 here */
 885                         }
 886                 }
 887         } /* <remain> is always decreased at least by 1 */
 888
 889         if (ibw_refill_cq_recv(conn))
 890                 goto error;
 891
 892         return 0;
 893
 894 error:
 895         DEBUG(0, ("ibw_wc_recv error: %s", ibw_lasterr));
 896         return -1;
 897 }
 898
 899 static int ibw_process_init_attrs(struct ibw_initattr *attr, int nattr, struct ibw_opts *opts)
 900 {
 901         int     i;
 902         const char *name, *value;
 903
 904         DEBUG(10, ("ibw_process_init_attrs: nattr: %d\n", nattr));
 905
 906         opts->max_send_wr = IBW_MAX_SEND_WR;
 907         opts->max_recv_wr = IBW_MAX_RECV_WR;
 908         opts->recv_bufsize = IBW_RECV_BUFSIZE;
 909         opts->recv_threshold = IBW_RECV_THRESHOLD;
 910
 911         for(i=0; i<nattr; i++) {
 912                 name = attr[i].name;
 913                 value = attr[i].value;
 914
 915                 assert(name!=NULL && value!=NULL);
 916                 if (strcmp(name, "max_send_wr")==0)
 917                         opts->max_send_wr = atoi(value);
 918                 else if (strcmp(name, "max_recv_wr")==0)
 919                         opts->max_recv_wr = atoi(value);
 920                 else if (strcmp(name, "recv_bufsize")==0)
 921                         opts->recv_bufsize = atoi(value);
 922                 else if (strcmp(name, "recv_threshold")==0)
 923                         opts->recv_threshold = atoi(value);
 924                 else {
 925                         sprintf(ibw_lasterr, "ibw_init: unknown name %s\n", name);
 926                         return -1;
 927                 }
 928         }
 929         return 0;
 930 }
 931
 932 struct ibw_ctx *ibw_init(struct ibw_initattr *attr, int nattr,
 933         void *ctx_userdata,
 934         ibw_connstate_fn_t ibw_connstate,
 935         ibw_receive_fn_t ibw_receive,
 936         struct event_context *ectx)
 937 {
 938         struct ibw_ctx *ctx = talloc_zero(NULL, struct ibw_ctx);
 939         struct ibw_ctx_priv *pctx;
 940         int     rc;
 941
 942         DEBUG(10, ("ibw_init(ctx_userdata: %p, ectx: %p)\n", ctx_userdata, ectx));
 943
 944         /* initialize basic data structures */
 945         memset(ibw_lasterr, 0, IBW_LASTERR_BUFSIZE);
 946
 947         assert(ctx!=NULL);
 948         ibw_lasterr[0] = '\0';
 949         talloc_set_destructor(ctx, ibw_ctx_destruct);
 950         ctx->ctx_userdata = ctx_userdata;
 951
 952         pctx = talloc_zero(ctx, struct ibw_ctx_priv);
 953         talloc_set_destructor(pctx, ibw_ctx_priv_destruct);
 954         ctx->internal = (void *)pctx;
 955         assert(pctx!=NULL);
 956
 957         pctx->connstate_func = ibw_connstate;
 958         pctx->receive_func = ibw_receive;
 959
 960         pctx->ectx = ectx;
 961
 962         /* process attributes */
 963         if (ibw_process_init_attrs(attr, nattr, &pctx->opts))
 964                 goto cleanup;
 965
 966         /* init cm */
 967         pctx->cm_channel = rdma_create_event_channel();
 968         if (!pctx->cm_channel) {
 969                 sprintf(ibw_lasterr, "rdma_create_event_channel error %d\n", errno);
 970                 goto cleanup;
 971         }
 972
 973         pctx->cm_channel_event = event_add_fd(pctx->ectx, pctx,
 974                 pctx->cm_channel->fd, EVENT_FD_READ, ibw_event_handler_cm, ctx);
 975
 976         rc = rdma_create_id(pctx->cm_channel, &pctx->cm_id, ctx, RDMA_PS_TCP);
 977         if (rc) {
 978                 rc = errno;
 979                 sprintf(ibw_lasterr, "rdma_create_id error %d\n", rc);
 980                 goto cleanup;
 981         }
 982         DEBUG(10, ("created cm_id %p\n", pctx->cm_id));
 983
 984         pctx->pagesize = sysconf(_SC_PAGESIZE);
 985
 986         return ctx;
 987         /* don't put code here */
 988 cleanup:
 989         DEBUG(0, (ibw_lasterr));
 990
 991         if (ctx)
 992                 talloc_free(ctx);
 993
 994         return NULL;
 995 }
 996
 997 int ibw_stop(struct ibw_ctx *ctx)
 998 {
 999         struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1000         struct ibw_conn *p;
1001
1002         DEBUG(10, ("ibw_stop\n"));
1003
1004         for(p=ctx->conn_list; p!=NULL; p=p->next) {
1005                 if (ctx->state==IBWC_ERROR || ctx->state==IBWC_CONNECTED) {
1006                         if (ibw_disconnect(p))
1007                                 return -1;
1008                 }
1009         }
1010
1011         ctx->state = IBWS_STOPPED;
1012         pctx->connstate_func(ctx, NULL);
1013
1014         return 0;
1015 }
1016
1017 int ibw_bind(struct ibw_ctx *ctx, struct sockaddr_in *my_addr)
1018 {
1019         struct ibw_ctx_priv *pctx = (struct ibw_ctx_priv *)ctx->internal;
1020         int     rc;
1021
1022         DEBUG(10, ("ibw_bind: addr=%s, port=%u\n",
1023                 inet_ntoa(my_addr->sin_addr), ntohs(my_addr->sin_port)));
1024         rc = rdma_bind_addr(pctx->cm_id, (struct sockaddr *) my_addr);
1025         if (rc) {
1026                 sprintf(ibw_lasterr, "rdma_bind_addr error %d\n", rc);
1027                 DEBUG(0, (ibw_lasterr));
1028                 return rc;
1029         }
1030         DEBUG(10, ("rdma_bind_addr successful\n"));
1031
1032         return 0;
1033 }
1034
1035 int ibw_listen(struct ibw_ctx *ctx, int backlog)
1036 {
1037         struct ibw_ctx_priv *pctx = talloc_get_type(ctx->internal, struct ibw_ctx_priv);
1038         int     rc;
1039
1040         DEBUG(10, ("ibw_listen\n"));
1041         rc = rdma_listen(pctx->cm_id, backlog);
1042         if (rc) {
1043                 sprintf(ibw_lasterr, "rdma_listen failed: %d\n", rc);
1044                 DEBUG(0, (ibw_lasterr));
1045                 return rc;
1046         }
1047
1048         return 0;
1049 }
1050
1051 int ibw_accept(struct ibw_ctx *ctx, struct ibw_conn *conn, void *conn_userdata)
1052 {
1053         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1054         struct rdma_conn_param  conn_param;
1055         int     rc;
1056
1057         DEBUG(10, ("ibw_accept: cmid=%p\n", pconn->cm_id));
1058         conn->conn_userdata = conn_userdata;
1059
1060         memset(&conn_param, 0, sizeof(struct rdma_conn_param));
1061         conn_param.responder_resources = 1;
1062         conn_param.initiator_depth = 1;
1063         rc = rdma_accept(pconn->cm_id, &conn_param);
1064         if (rc) {
1065                 sprintf(ibw_lasterr, "rdma_accept failed %d\n", rc);
1066                 DEBUG(0, (ibw_lasterr));
1067                 return -1;;
1068         }
1069
1070         pconn->is_accepted = 1;
1071
1072         /* continued at RDMA_CM_EVENT_ESTABLISHED */
1073
1074         return 0;
1075 }
1076
1077 int ibw_connect(struct ibw_conn *conn, struct sockaddr_in *serv_addr, void *conn_userdata)
1078 {
1079         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1080         struct ibw_conn_priv *pconn = NULL;
1081         int     rc;
1082
1083         assert(conn!=NULL);
1084
1085         conn->conn_userdata = conn_userdata;
1086         pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1087         DEBUG(10, ("ibw_connect: addr=%s, port=%u\n", inet_ntoa(serv_addr->sin_addr),
1088                 ntohs(serv_addr->sin_port)));
1089
1090         /* clean previous - probably half - initialization */
1091         if (ibw_conn_priv_destruct(pconn)) {
1092                 DEBUG(0, ("ibw_connect/ibw_pconn_destruct failed for cm_id=%p\n", pconn->cm_id));
1093                 return -1;
1094         }
1095
1096         /* init cm */
1097         rc = rdma_create_id(pctx->cm_channel, &pconn->cm_id, conn, RDMA_PS_TCP);
1098         if (rc) {
1099                 rc = errno;
1100                 sprintf(ibw_lasterr, "ibw_connect/rdma_create_id error %d\n", rc);
1101                 talloc_free(conn);
1102                 return -1;
1103         }
1104         DEBUG(10, ("ibw_connect: rdma_create_id succeeded, cm_id=%p\n", pconn->cm_id));
1105
1106         rc = rdma_resolve_addr(pconn->cm_id, NULL, (struct sockaddr *) serv_addr, 2000);
1107         if (rc) {
1108                 sprintf(ibw_lasterr, "rdma_resolve_addr error %d\n", rc);
1109                 DEBUG(0, (ibw_lasterr));
1110                 talloc_free(conn);
1111                 return -1;
1112         }
1113
1114         /* continued at RDMA_CM_EVENT_ADDR_RESOLVED */
1115
1116         return 0;
1117 }
1118
1119 int ibw_disconnect(struct ibw_conn *conn)
1120 {
1121         int     rc;
1122         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1123
1124         DEBUG(10, ("ibw_disconnect: cmid=%p\n", pconn->cm_id));
1125
1126         assert(pconn!=NULL);
1127
1128         switch(conn->state) {
1129         case IBWC_ERROR:
1130                 ibw_conn_priv_destruct(pconn); /* do this here right now */
1131                 break;
1132         case IBWC_CONNECTED:
1133                 rc = rdma_disconnect(pconn->cm_id);
1134                 if (rc) {
1135                         sprintf(ibw_lasterr, "ibw_disconnect failed with %d\n", rc);
1136                         DEBUG(0, (ibw_lasterr));
1137                         return rc;
1138                 }
1139                 break;
1140         default:
1141                 DEBUG(9, ("invalid state for disconnect: %d\n", conn->state));
1142                 break;
1143         }
1144
1145         return 0;
1146 }
1147
1148 int ibw_alloc_send_buf(struct ibw_conn *conn, void **buf, void **key, uint32_t len)
1149 {
1150         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1151         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1152         struct ibw_wr *p = pconn->wr_list_avail;
1153
1154         if (p!=NULL) {
1155                 DEBUG(10, ("ibw_alloc_send_buf#1: cmid=%p, len=%d\n", pconn->cm_id, len));
1156
1157                 DLIST_REMOVE(pconn->wr_list_avail, p);
1158                 DLIST_ADD(pconn->wr_list_used, p);
1159
1160                 if (len <= pctx->opts.recv_bufsize) {
1161                         *buf = (void *)p->buf;
1162                 } else {
1163                         p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1164                         if (p->buf_large==NULL) {
1165                                 sprintf(ibw_lasterr, "ibw_alloc_mr#1 failed\n");
1166                                 goto error;
1167                         }
1168                         *buf = (void *)p->buf_large;
1169                 }
1170                 /* p->wr_id is already filled in ibw_init_memory */
1171         } else {
1172                 DEBUG(10, ("ibw_alloc_send_buf#2: cmid=%p, len=%d\n", pconn->cm_id, len));
1173                 /* not optimized */
1174                 p = pconn->extra_avail;
1175                 if (!p) {
1176                         p = pconn->extra_avail = talloc_zero(pconn, struct ibw_wr);
1177                         talloc_set_destructor(p, ibw_wr_destruct);
1178                         if (p==NULL) {
1179                                 sprintf(ibw_lasterr, "talloc_zero failed (emax: %u)\n", pconn->extra_max);
1180                                 goto error;
1181                         }
1182                         p->wr_id = pctx->opts.max_send_wr + pconn->extra_max;
1183                         pconn->extra_max++;
1184                         switch(pconn->extra_max) {
1185                                 case 1: DEBUG(2, ("warning: queue performed\n")); break;
1186                                 case 10: DEBUG(0, ("warning: queue reached 10\n")); break;
1187                                 case 100: DEBUG(0, ("warning: queue reached 100\n")); break;
1188                                 case 1000: DEBUG(0, ("warning: queue reached 1000\n")); break;
1189                                 default: break;
1190                         }
1191                 }
1192
1193                 p->buf_large = ibw_alloc_mr(pctx, pconn, len, &p->mr_large);
1194                 if (p->buf_large==NULL) {
1195                         sprintf(ibw_lasterr, "ibw_alloc_mr#2 failed\n");
1196                         goto error;
1197                 }
1198                 *buf = (void *)p->buf_large;
1199
1200                 DLIST_REMOVE(pconn->extra_avail, p);
1201                 /* we don't have prepared index for this, so that
1202                  * we will have to find this by wr_id later on */
1203                 DLIST_ADD(pconn->extra_sent, p);
1204         }
1205
1206         *key = (void *)p;
1207
1208         return 0;
1209 error:
1210         DEBUG(0, ("ibw_alloc_send_buf error: %s", ibw_lasterr));
1211         return -1;
1212 }
1213
1214
1215 static int ibw_send_packet(struct ibw_conn *conn, void *buf, struct ibw_wr *p, uint32_t len)
1216 {
1217         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1218         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1219         int     rc;
1220
1221         /* can we send it right now? */
1222         if (pconn->wr_sent<pctx->opts.max_send_wr) {
1223                 struct ibv_send_wr *bad_wr;
1224                 struct ibv_sge list = {
1225                         .addr   = (uintptr_t)buf,
1226                         .length = len,
1227                         .lkey   = pconn->mr_send->lkey
1228                 };
1229                 struct ibv_send_wr wr = {
1230                         .wr_id      = p->wr_id + pctx->opts.max_recv_wr,
1231                         .sg_list    = &list,
1232                         .num_sge    = 1,
1233                         .opcode     = IBV_WR_SEND,
1234                         .send_flags = IBV_SEND_SIGNALED,
1235                 };
1236
1237                 if (p->buf_large==NULL) {
1238                         DEBUG(10, ("ibw_send#normal(cmid: %p, wrid: %u, n: %d)\n",
1239                                 pconn->cm_id, (uint32_t)wr.wr_id, len));
1240                 } else {
1241                         DEBUG(10, ("ibw_send#large(cmid: %p, wrid: %u, n: %d)\n",
1242                                 pconn->cm_id, (uint32_t)wr.wr_id, len));
1243                         list.lkey = p->mr_large->lkey;
1244                 }
1245
1246                 rc = ibv_post_send(pconn->cm_id->qp, &wr, &bad_wr);
1247                 if (rc) {
1248                         sprintf(ibw_lasterr, "ibv_post_send error %d (%d)\n",
1249                                 rc, pconn->wr_sent);
1250                         goto error;
1251                 }
1252
1253                 pconn->wr_sent++;
1254
1255                 return rc;
1256         } /* else put the request into our own queue: */
1257
1258         DEBUG(10, ("ibw_send#queued(cmid: %p, len: %u)\n", pconn->cm_id, len));
1259
1260         /* TODO: clarify how to continue when state==IBWC_STOPPED */
1261
1262         /* to be sent by ibw_wc_send */
1263         /* regardless "normal" or [a part of] "large" packet */
1264         if (!p->queued_ref_cnt) {
1265                 DLIST_ADD_END2(pconn->queue, p, struct ibw_wr *,
1266                         qprev, qnext); /* TODO: optimize */
1267                 p->queued_msg = buf;
1268         }
1269         p->queued_ref_cnt++;
1270         p->queued_rlen = len; /* last wins; see ibw_wc_send */
1271
1272         return 0;
1273 error:
1274         DEBUG(0, (ibw_lasterr));
1275         return -1;
1276 }
1277
1278 int ibw_send(struct ibw_conn *conn, void *buf, void *key, uint32_t len)
1279 {
1280         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1281         struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1282         int     rc;
1283
1284         assert(len>=sizeof(uint32_t));
1285         assert((*((uint32_t *)buf)==len)); /* TODO: htonl */
1286
1287         if (len > pctx->opts.recv_bufsize) {
1288                 struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1289                 int     rlen = len;
1290                 char    *packet = (char *)buf;
1291                 uint32_t        recv_bufsize = pctx->opts.recv_bufsize;
1292
1293                 DEBUG(10, ("ibw_send#frag(cmid: %p, buf: %p, len: %u)\n",
1294                         pconn->cm_id, buf, len));
1295
1296                 /* single threaded => no race here: */
1297                 assert(p->ref_cnt==0);
1298                 while(rlen > recv_bufsize) {
1299                         rc = ibw_send_packet(conn, packet, p, recv_bufsize);
1300                         if (rc)
1301                                 return rc;
1302                         packet += recv_bufsize;
1303                         rlen -= recv_bufsize;
1304                         p->ref_cnt++; /* not good to have it in ibw_send_packet */
1305                 }
1306                 if (rlen) {
1307                         rc = ibw_send_packet(conn, packet, p, rlen);
1308                         p->ref_cnt++; /* not good to have it in ibw_send_packet */
1309                 }
1310                 p->ref_cnt--; /* for the same handling */
1311         } else {
1312                 assert(p->ref_cnt==0);
1313                 assert(p->queued_ref_cnt==0);
1314
1315                 rc = ibw_send_packet(conn, buf, p, len);
1316         }
1317         return rc;
1318 }
1319
1320 int ibw_cancel_send_buf(struct ibw_conn *conn, void *buf, void *key)
1321 {
1322         struct ibw_ctx_priv *pctx = talloc_get_type(conn->ctx->internal, struct ibw_ctx_priv);
1323         struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
1324         struct ibw_wr *p = talloc_get_type(key, struct ibw_wr);
1325
1326         assert(p!=NULL);
1327         assert(buf!=NULL);
1328         assert(conn!=NULL);
1329
1330         if (p->buf_large!=NULL)
1331                 ibw_free_mr(&p->buf_large, &p->mr_large);
1332
1333         /* parallel case */
1334         if (p->wr_id < pctx->opts.max_send_wr) {
1335                 DEBUG(10, ("ibw_cancel_send_buf#1 %u", (int)p->wr_id));
1336                 DLIST_REMOVE(pconn->wr_list_used, p);
1337                 DLIST_ADD(pconn->wr_list_avail, p);
1338         } else { /* "extra" packet */
1339                 DEBUG(10, ("ibw_cancel_send_buf#2 %u", (int)p->wr_id));
1340                 DLIST_REMOVE(pconn->extra_sent, p);
1341                 DLIST_ADD(pconn->extra_avail, p);
1342         }
1343
1344         return 0;
1345 }
1346
1347 const char *ibw_getLastError(void)
1348 {
1349         return ibw_lasterr;
1350 }