2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
 
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
 
   5  * This software is available to you under a choice of one of two
 
   6  * licenses.  You may choose to be licensed under the terms of the GNU
 
   7  * General Public License (GPL) Version 2, available from the file
 
   8  * COPYING in the main directory of this source tree, or the
 
   9  * OpenIB.org BSD license below:
 
  11  *     Redistribution and use in source and binary forms, with or
 
  12  *     without modification, are permitted provided that the following
 
  15  *      - Redistributions of source code must retain the above
 
  16  *        copyright notice, this list of conditions and the following
 
  19  *      - Redistributions in binary form must reproduce the above
 
  20  *        copyright notice, this list of conditions and the following
 
  21  *        disclaimer in the documentation and/or other materials
 
  22  *        provided with the distribution.
 
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 
  33 #include <linux/kernel.h>
 
  34 #include <linux/module.h>
 
  35 #include <linux/delay.h>
 
  37 #include "iscsi_iser.h"
 
  39 #define ISCSI_ISER_MAX_CONN     8
 
  40 #define ISER_MAX_CQ_LEN         ((ISER_QP_MAX_RECV_DTOS + \
 
  41                                 ISER_QP_MAX_REQ_DTOS) *   \
 
  44 static void iser_cq_tasklet_fn(unsigned long data);
 
  45 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
 
  47 static void iser_cq_event_callback(struct ib_event *cause, void *context)
 
  49         iser_err("got cq event %d \n", cause->event);
 
  52 static void iser_qp_event_callback(struct ib_event *cause, void *context)
 
  54         iser_err("got qp event %d\n",cause->event);
 
  58  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
 
  59  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
 
  62  * returns 0 on success, -1 on failure
 
  64 static int iser_create_device_ib_res(struct iser_device *device)
 
  66         device->pd = ib_alloc_pd(device->ib_device);
 
  67         if (IS_ERR(device->pd))
 
  70         device->cq = ib_create_cq(device->ib_device,
 
  72                                   iser_cq_event_callback,
 
  75         if (IS_ERR(device->cq))
 
  78         if (ib_req_notify_cq(device->cq, IB_CQ_NEXT_COMP))
 
  81         tasklet_init(&device->cq_tasklet,
 
  83                      (unsigned long)device);
 
  85         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
 
  86                                    IB_ACCESS_REMOTE_WRITE |
 
  87                                    IB_ACCESS_REMOTE_READ);
 
  88         if (IS_ERR(device->mr))
 
  94         tasklet_kill(&device->cq_tasklet);
 
  96         ib_destroy_cq(device->cq);
 
  98         ib_dealloc_pd(device->pd);
 
 100         iser_err("failed to allocate an IB resource\n");
 
 105  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 
 106  * CQ and PD created with the device associated with the adapator.
 
 108 static void iser_free_device_ib_res(struct iser_device *device)
 
 110         BUG_ON(device->mr == NULL);
 
 112         tasklet_kill(&device->cq_tasklet);
 
 114         (void)ib_dereg_mr(device->mr);
 
 115         (void)ib_destroy_cq(device->cq);
 
 116         (void)ib_dealloc_pd(device->pd);
 
 124  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 
 126  * returns 0 on success, -1 on failure
 
 128 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 
 130         struct iser_device      *device;
 
 131         struct ib_qp_init_attr  init_attr;
 
 133         struct ib_fmr_pool_param params;
 
 135         BUG_ON(ib_conn->device == NULL);
 
 137         device = ib_conn->device;
 
 139         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 
 140                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 
 142         if (!ib_conn->page_vec) {
 
 146         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 
 148         params.page_shift        = SHIFT_4K;
 
 149         /* when the first/last SG element are not start/end *
 
 150          * page aligned, the map whould be of N+1 pages     */
 
 151         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 
 152         /* make the pool size twice the max number of SCSI commands *
 
 153          * the ML is expected to queue, watermark for unmap at 50%  */
 
 154         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 
 155         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 
 157         params.flush_function    = NULL;
 
 158         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 
 159                                     IB_ACCESS_REMOTE_WRITE |
 
 160                                     IB_ACCESS_REMOTE_READ);
 
 162         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, ¶ms);
 
 163         if (IS_ERR(ib_conn->fmr_pool)) {
 
 164                 ret = PTR_ERR(ib_conn->fmr_pool);
 
 168         memset(&init_attr, 0, sizeof init_attr);
 
 170         init_attr.event_handler = iser_qp_event_callback;
 
 171         init_attr.qp_context    = (void *)ib_conn;
 
 172         init_attr.send_cq       = device->cq;
 
 173         init_attr.recv_cq       = device->cq;
 
 174         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 
 175         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 
 176         init_attr.cap.max_send_sge = MAX_REGD_BUF_VECTOR_LEN;
 
 177         init_attr.cap.max_recv_sge = 2;
 
 178         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 
 179         init_attr.qp_type       = IB_QPT_RC;
 
 181         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 
 185         ib_conn->qp = ib_conn->cma_id->qp;
 
 186         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 
 187                  ib_conn, ib_conn->cma_id,
 
 188                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 
 192         (void)ib_destroy_fmr_pool(ib_conn->fmr_pool);
 
 194         kfree(ib_conn->page_vec);
 
 196         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 
 201  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 
 204 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
 
 206         BUG_ON(ib_conn == NULL);
 
 208         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 
 209                  ib_conn, ib_conn->cma_id,
 
 210                  ib_conn->fmr_pool, ib_conn->qp);
 
 212         /* qp is created only once both addr & route are resolved */
 
 213         if (ib_conn->fmr_pool != NULL)
 
 214                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 
 216         if (ib_conn->qp != NULL)
 
 217                 rdma_destroy_qp(ib_conn->cma_id);
 
 219         if (ib_conn->cma_id != NULL)
 
 220                 rdma_destroy_id(ib_conn->cma_id);
 
 222         ib_conn->fmr_pool = NULL;
 
 224         ib_conn->cma_id   = NULL;
 
 225         kfree(ib_conn->page_vec);
 
 231  * based on the resolved device node GUID see if there already allocated
 
 232  * device for this device. If there's no such, create one.
 
 235 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 
 237         struct iser_device *device;
 
 239         mutex_lock(&ig.device_list_mutex);
 
 241         list_for_each_entry(device, &ig.device_list, ig_list)
 
 242                 /* find if there's a match using the node GUID */
 
 243                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 
 246         device = kzalloc(sizeof *device, GFP_KERNEL);
 
 250         /* assign this device to the device */
 
 251         device->ib_device = cma_id->device;
 
 252         /* init the device and link it into ig device list */
 
 253         if (iser_create_device_ib_res(device)) {
 
 258         list_add(&device->ig_list, &ig.device_list);
 
 263         mutex_unlock(&ig.device_list_mutex);
 
 267 /* if there's no demand for this device, release it */
 
 268 static void iser_device_try_release(struct iser_device *device)
 
 270         mutex_lock(&ig.device_list_mutex);
 
 272         iser_err("device %p refcount %d\n",device,device->refcount);
 
 273         if (!device->refcount) {
 
 274                 iser_free_device_ib_res(device);
 
 275                 list_del(&device->ig_list);
 
 278         mutex_unlock(&ig.device_list_mutex);
 
 281 int iser_conn_state_comp(struct iser_conn *ib_conn,
 
 282                         enum iser_ib_conn_state comp)
 
 286         spin_lock_bh(&ib_conn->lock);
 
 287         ret = (ib_conn->state == comp);
 
 288         spin_unlock_bh(&ib_conn->lock);
 
 292 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 
 293                                      enum iser_ib_conn_state comp,
 
 294                                      enum iser_ib_conn_state exch)
 
 298         spin_lock_bh(&ib_conn->lock);
 
 299         if ((ret = (ib_conn->state == comp)))
 
 300                 ib_conn->state = exch;
 
 301         spin_unlock_bh(&ib_conn->lock);
 
 306  * Frees all conn objects and deallocs conn descriptor
 
 308 static void iser_conn_release(struct iser_conn *ib_conn)
 
 310         struct iser_device  *device = ib_conn->device;
 
 312         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 
 314         mutex_lock(&ig.connlist_mutex);
 
 315         list_del(&ib_conn->conn_list);
 
 316         mutex_unlock(&ig.connlist_mutex);
 
 318         iser_free_ib_conn_res(ib_conn);
 
 319         ib_conn->device = NULL;
 
 320         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 
 322                 iser_device_try_release(device);
 
 323         if (ib_conn->iser_conn)
 
 324                 ib_conn->iser_conn->ib_conn = NULL;
 
 325         iscsi_destroy_endpoint(ib_conn->ep);
 
 328 void iser_conn_get(struct iser_conn *ib_conn)
 
 330         atomic_inc(&ib_conn->refcount);
 
 333 void iser_conn_put(struct iser_conn *ib_conn)
 
 335         if (atomic_dec_and_test(&ib_conn->refcount))
 
 336                 iser_conn_release(ib_conn);
 
 340  * triggers start of the disconnect procedures and wait for them to be done
 
 342 void iser_conn_terminate(struct iser_conn *ib_conn)
 
 346         /* change the ib conn state only if the conn is UP, however always call
 
 347          * rdma_disconnect since this is the only way to cause the CMA to change
 
 348          * the QP state to ERROR
 
 351         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 
 352         err = rdma_disconnect(ib_conn->cma_id);
 
 354                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 
 357         wait_event_interruptible(ib_conn->wait,
 
 358                                  ib_conn->state == ISER_CONN_DOWN);
 
 360         iser_conn_put(ib_conn);
 
 363 static void iser_connect_error(struct rdma_cm_id *cma_id)
 
 365         struct iser_conn *ib_conn;
 
 366         ib_conn = (struct iser_conn *)cma_id->context;
 
 368         ib_conn->state = ISER_CONN_DOWN;
 
 369         wake_up_interruptible(&ib_conn->wait);
 
 372 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 
 374         struct iser_device *device;
 
 375         struct iser_conn   *ib_conn;
 
 378         device = iser_device_find_by_ib_device(cma_id);
 
 380                 iser_err("device lookup/creation failed\n");
 
 381                 iser_connect_error(cma_id);
 
 385         ib_conn = (struct iser_conn *)cma_id->context;
 
 386         ib_conn->device = device;
 
 388         ret = rdma_resolve_route(cma_id, 1000);
 
 390                 iser_err("resolve route failed: %d\n", ret);
 
 391                 iser_connect_error(cma_id);
 
 395 static void iser_route_handler(struct rdma_cm_id *cma_id)
 
 397         struct rdma_conn_param conn_param;
 
 400         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 
 404         iser_dbg("path.mtu is %d setting it to %d\n",
 
 405                  cma_id->route.path_rec->mtu, IB_MTU_1024);
 
 407         /* we must set the MTU to 1024 as this is what the target is assuming */
 
 408         if (cma_id->route.path_rec->mtu > IB_MTU_1024)
 
 409                 cma_id->route.path_rec->mtu = IB_MTU_1024;
 
 411         memset(&conn_param, 0, sizeof conn_param);
 
 412         conn_param.responder_resources = 4;
 
 413         conn_param.initiator_depth     = 1;
 
 414         conn_param.retry_count         = 7;
 
 415         conn_param.rnr_retry_count     = 6;
 
 417         ret = rdma_connect(cma_id, &conn_param);
 
 419                 iser_err("failure connecting: %d\n", ret);
 
 425         iser_connect_error(cma_id);
 
 428 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 
 430         struct iser_conn *ib_conn;
 
 432         ib_conn = (struct iser_conn *)cma_id->context;
 
 433         ib_conn->state = ISER_CONN_UP;
 
 434         wake_up_interruptible(&ib_conn->wait);
 
 437 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 
 439         struct iser_conn *ib_conn;
 
 441         ib_conn = (struct iser_conn *)cma_id->context;
 
 442         ib_conn->disc_evt_flag = 1;
 
 444         /* getting here when the state is UP means that the conn is being *
 
 445          * terminated asynchronously from the iSCSI layer's perspective.  */
 
 446         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 
 447                                       ISER_CONN_TERMINATING))
 
 448                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 
 449                                    ISCSI_ERR_CONN_FAILED);
 
 451         /* Complete the termination process if no posts are pending */
 
 452         if ((atomic_read(&ib_conn->post_recv_buf_count) == 0) &&
 
 453             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 
 454                 ib_conn->state = ISER_CONN_DOWN;
 
 455                 wake_up_interruptible(&ib_conn->wait);
 
 459 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 
 463         iser_err("event %d conn %p id %p\n",event->event,cma_id->context,cma_id);
 
 465         switch (event->event) {
 
 466         case RDMA_CM_EVENT_ADDR_RESOLVED:
 
 467                 iser_addr_handler(cma_id);
 
 469         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 
 470                 iser_route_handler(cma_id);
 
 472         case RDMA_CM_EVENT_ESTABLISHED:
 
 473                 iser_connected_handler(cma_id);
 
 475         case RDMA_CM_EVENT_ADDR_ERROR:
 
 476         case RDMA_CM_EVENT_ROUTE_ERROR:
 
 477         case RDMA_CM_EVENT_CONNECT_ERROR:
 
 478         case RDMA_CM_EVENT_UNREACHABLE:
 
 479         case RDMA_CM_EVENT_REJECTED:
 
 480                 iser_err("event: %d, error: %d\n", event->event, event->status);
 
 481                 iser_connect_error(cma_id);
 
 483         case RDMA_CM_EVENT_DISCONNECTED:
 
 484         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 
 485         case RDMA_CM_EVENT_ADDR_CHANGE:
 
 486                 iser_disconnected_handler(cma_id);
 
 489                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 
 495 void iser_conn_init(struct iser_conn *ib_conn)
 
 497         ib_conn->state = ISER_CONN_INIT;
 
 498         init_waitqueue_head(&ib_conn->wait);
 
 499         atomic_set(&ib_conn->post_recv_buf_count, 0);
 
 500         atomic_set(&ib_conn->post_send_buf_count, 0);
 
 501         atomic_set(&ib_conn->unexpected_pdu_count, 0);
 
 502         atomic_set(&ib_conn->refcount, 1);
 
 503         INIT_LIST_HEAD(&ib_conn->conn_list);
 
 504         spin_lock_init(&ib_conn->lock);
 
 508  * starts the process of connecting to the target
 
 509  * sleeps untill the connection is established or rejected
 
 511 int iser_connect(struct iser_conn   *ib_conn,
 
 512                  struct sockaddr_in *src_addr,
 
 513                  struct sockaddr_in *dst_addr,
 
 516         struct sockaddr *src, *dst;
 
 519         sprintf(ib_conn->name, "%pI4:%d",
 
 520                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 
 522         /* the device is known only --after-- address resolution */
 
 523         ib_conn->device = NULL;
 
 525         iser_err("connecting to: %pI4, port 0x%x\n",
 
 526                  &dst_addr->sin_addr, dst_addr->sin_port);
 
 528         ib_conn->state = ISER_CONN_PENDING;
 
 530         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 
 533         if (IS_ERR(ib_conn->cma_id)) {
 
 534                 err = PTR_ERR(ib_conn->cma_id);
 
 535                 iser_err("rdma_create_id failed: %d\n", err);
 
 539         src = (struct sockaddr *)src_addr;
 
 540         dst = (struct sockaddr *)dst_addr;
 
 541         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 
 543                 iser_err("rdma_resolve_addr failed: %d\n", err);
 
 548                 wait_event_interruptible(ib_conn->wait,
 
 549                                          (ib_conn->state != ISER_CONN_PENDING));
 
 551                 if (ib_conn->state != ISER_CONN_UP) {
 
 553                         goto connect_failure;
 
 557         mutex_lock(&ig.connlist_mutex);
 
 558         list_add(&ib_conn->conn_list, &ig.connlist);
 
 559         mutex_unlock(&ig.connlist_mutex);
 
 563         ib_conn->cma_id = NULL;
 
 565         ib_conn->state = ISER_CONN_DOWN;
 
 567         iser_conn_release(ib_conn);
 
 572  * iser_reg_page_vec - Register physical memory
 
 574  * returns: 0 on success, errno code on failure
 
 576 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 
 577                       struct iser_page_vec *page_vec,
 
 578                       struct iser_mem_reg  *mem_reg)
 
 580         struct ib_pool_fmr *mem;
 
 585         page_list = page_vec->pages;
 
 586         io_addr   = page_list[0];
 
 588         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 
 594                 status = (int)PTR_ERR(mem);
 
 595                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 
 599         mem_reg->lkey  = mem->fmr->lkey;
 
 600         mem_reg->rkey  = mem->fmr->rkey;
 
 601         mem_reg->len   = page_vec->length * SIZE_4K;
 
 602         mem_reg->va    = io_addr;
 
 604         mem_reg->mem_h = (void *)mem;
 
 606         mem_reg->va   += page_vec->offset;
 
 607         mem_reg->len   = page_vec->data_size;
 
 609         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 
 610                  "entry[0]: (0x%08lx,%ld)] -> "
 
 611                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 
 612                  page_vec, page_vec->length,
 
 613                  (unsigned long)page_vec->pages[0],
 
 614                  (unsigned long)page_vec->data_size,
 
 615                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 
 616                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 
 621  * Unregister (previosuly registered) memory.
 
 623 void iser_unreg_mem(struct iser_mem_reg *reg)
 
 627         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 
 629         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 
 631                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 
 637  * iser_dto_to_iov - builds IOV from a dto descriptor
 
 639 static void iser_dto_to_iov(struct iser_dto *dto, struct ib_sge *iov, int iov_len)
 
 643         struct iser_regd_buf *regd_buf;
 
 645         if (dto->regd_vector_len > iov_len) {
 
 646                 iser_err("iov size %d too small for posting dto of len %d\n",
 
 647                          iov_len, dto->regd_vector_len);
 
 651         for (i = 0; i < dto->regd_vector_len; i++) {
 
 653                 regd_buf  = dto->regd[i];
 
 655                 sge->addr   = regd_buf->reg.va;
 
 656                 sge->length = regd_buf->reg.len;
 
 657                 sge->lkey   = regd_buf->reg.lkey;
 
 659                 if (dto->used_sz[i] > 0)  /* Adjust size */
 
 660                         sge->length = dto->used_sz[i];
 
 662                 /* offset and length should not exceed the regd buf length */
 
 663                 if (sge->length + dto->offset[i] > regd_buf->reg.len) {
 
 664                         iser_err("Used len:%ld + offset:%d, exceed reg.buf.len:"
 
 665                                  "%ld in dto:0x%p [%d], va:0x%08lX\n",
 
 666                                  (unsigned long)sge->length, dto->offset[i],
 
 667                                  (unsigned long)regd_buf->reg.len, dto, i,
 
 668                                  (unsigned long)sge->addr);
 
 672                 sge->addr += dto->offset[i]; /* Adjust offset */
 
 677  * iser_post_recv - Posts a receive buffer.
 
 679  * returns 0 on success, -1 on failure
 
 681 int iser_post_recv(struct iser_desc *rx_desc)
 
 683         int               ib_ret, ret_val = 0;
 
 684         struct ib_recv_wr recv_wr, *recv_wr_failed;
 
 685         struct ib_sge     iov[2];
 
 686         struct iser_conn  *ib_conn;
 
 687         struct iser_dto   *recv_dto = &rx_desc->dto;
 
 690         ib_conn = recv_dto->ib_conn;
 
 692         iser_dto_to_iov(recv_dto, iov, 2);
 
 695         recv_wr.sg_list = iov;
 
 696         recv_wr.num_sge = recv_dto->regd_vector_len;
 
 697         recv_wr.wr_id   = (unsigned long)rx_desc;
 
 699         atomic_inc(&ib_conn->post_recv_buf_count);
 
 700         ib_ret  = ib_post_recv(ib_conn->qp, &recv_wr, &recv_wr_failed);
 
 702                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 
 703                 atomic_dec(&ib_conn->post_recv_buf_count);
 
 711  * iser_start_send - Initiate a Send DTO operation
 
 713  * returns 0 on success, -1 on failure
 
 715 int iser_post_send(struct iser_desc *tx_desc)
 
 717         int               ib_ret, ret_val = 0;
 
 718         struct ib_send_wr send_wr, *send_wr_failed;
 
 719         struct ib_sge     iov[MAX_REGD_BUF_VECTOR_LEN];
 
 720         struct iser_conn  *ib_conn;
 
 721         struct iser_dto   *dto = &tx_desc->dto;
 
 723         ib_conn = dto->ib_conn;
 
 725         iser_dto_to_iov(dto, iov, MAX_REGD_BUF_VECTOR_LEN);
 
 728         send_wr.wr_id      = (unsigned long)tx_desc;
 
 729         send_wr.sg_list    = iov;
 
 730         send_wr.num_sge    = dto->regd_vector_len;
 
 731         send_wr.opcode     = IB_WR_SEND;
 
 732         send_wr.send_flags = dto->notify_enable ? IB_SEND_SIGNALED : 0;
 
 734         atomic_inc(&ib_conn->post_send_buf_count);
 
 736         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 
 738                 iser_err("Failed to start SEND DTO, dto: 0x%p, IOV len: %d\n",
 
 739                          dto, dto->regd_vector_len);
 
 740                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 
 741                 atomic_dec(&ib_conn->post_send_buf_count);
 
 748 static void iser_handle_comp_error(struct iser_desc *desc)
 
 750         struct iser_dto  *dto     = &desc->dto;
 
 751         struct iser_conn *ib_conn = dto->ib_conn;
 
 753         iser_dto_buffs_release(dto);
 
 755         if (desc->type == ISCSI_RX) {
 
 757                 kmem_cache_free(ig.desc_cache, desc);
 
 758                 atomic_dec(&ib_conn->post_recv_buf_count);
 
 759         } else { /* type is TX control/command/dataout */
 
 760                 if (desc->type == ISCSI_TX_DATAOUT)
 
 761                         kmem_cache_free(ig.desc_cache, desc);
 
 762                 atomic_dec(&ib_conn->post_send_buf_count);
 
 765         if (atomic_read(&ib_conn->post_recv_buf_count) == 0 &&
 
 766             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 
 767                 /* getting here when the state is UP means that the conn is *
 
 768                  * being terminated asynchronously from the iSCSI layer's   *
 
 770                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 
 771                     ISER_CONN_TERMINATING))
 
 772                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 
 773                                            ISCSI_ERR_CONN_FAILED);
 
 775                 /* complete the termination process if disconnect event was delivered *
 
 776                  * note there are no more non completed posts to the QP               */
 
 777                 if (ib_conn->disc_evt_flag) {
 
 778                         ib_conn->state = ISER_CONN_DOWN;
 
 779                         wake_up_interruptible(&ib_conn->wait);
 
 784 static void iser_cq_tasklet_fn(unsigned long data)
 
 786          struct iser_device  *device = (struct iser_device *)data;
 
 787          struct ib_cq        *cq = device->cq;
 
 789          struct iser_desc    *desc;
 
 790          unsigned long       xfer_len;
 
 792         while (ib_poll_cq(cq, 1, &wc) == 1) {
 
 793                 desc     = (struct iser_desc *) (unsigned long) wc.wr_id;
 
 794                 BUG_ON(desc == NULL);
 
 796                 if (wc.status == IB_WC_SUCCESS) {
 
 797                         if (desc->type == ISCSI_RX) {
 
 798                                 xfer_len = (unsigned long)wc.byte_len;
 
 799                                 iser_rcv_completion(desc, xfer_len);
 
 800                         } else /* type == ISCSI_TX_CONTROL/SCSI_CMD/DOUT */
 
 801                                 iser_snd_completion(desc);
 
 803                         iser_err("comp w. error op %d status %d\n",desc->type,wc.status);
 
 804                         iser_handle_comp_error(desc);
 
 807         /* #warning "it is assumed here that arming CQ only once its empty" *
 
 808          * " would not cause interrupts to be missed"                       */
 
 809         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 
 812 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 
 814         struct iser_device  *device = (struct iser_device *)cq_context;
 
 816         tasklet_schedule(&device->cq_tasklet);