Pileus Git - ~andy/linux/blob - drivers/infiniband/ulp/iser/iser_verbs.c

   1 /*
   2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/module.h>
  35 #include <linux/delay.h>
  36
  37 #include "iscsi_iser.h"
  38
  39 #define ISCSI_ISER_MAX_CONN     8
  40 #define ISER_MAX_CQ_LEN         ((ISER_QP_MAX_RECV_DTOS + \
  41                                 ISER_QP_MAX_REQ_DTOS) *   \
  42                                  ISCSI_ISER_MAX_CONN)
  43
  44 static void iser_cq_tasklet_fn(unsigned long data);
  45 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
  46
  47 static void iser_cq_event_callback(struct ib_event *cause, void *context)
  48 {
  49         iser_err("got cq event %d \n", cause->event);
  50 }
  51
  52 static void iser_qp_event_callback(struct ib_event *cause, void *context)
  53 {
  54         iser_err("got qp event %d\n",cause->event);
  55 }
  56
  57 /**
  58  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  59  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  60  * the adapator.
  61  *
  62  * returns 0 on success, -1 on failure
  63  */
  64 static int iser_create_device_ib_res(struct iser_device *device)
  65 {
  66         device->pd = ib_alloc_pd(device->ib_device);
  67         if (IS_ERR(device->pd))
  68                 goto pd_err;
  69
  70         device->cq = ib_create_cq(device->ib_device,
  71                                   iser_cq_callback,
  72                                   iser_cq_event_callback,
  73                                   (void *)device,
  74                                   ISER_MAX_CQ_LEN, 0);
  75         if (IS_ERR(device->cq))
  76                 goto cq_err;
  77
  78         if (ib_req_notify_cq(device->cq, IB_CQ_NEXT_COMP))
  79                 goto cq_arm_err;
  80
  81         tasklet_init(&device->cq_tasklet,
  82                      iser_cq_tasklet_fn,
  83                      (unsigned long)device);
  84
  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))
  89                 goto dma_mr_err;
  90
  91         return 0;
  92
  93 dma_mr_err:
  94         tasklet_kill(&device->cq_tasklet);
  95 cq_arm_err:
  96         ib_destroy_cq(device->cq);
  97 cq_err:
  98         ib_dealloc_pd(device->pd);
  99 pd_err:
 100         iser_err("failed to allocate an IB resource\n");
 101         return -1;
 102 }
 103
 104 /**
 105  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 106  * CQ and PD created with the device associated with the adapator.
 107  */
 108 static void iser_free_device_ib_res(struct iser_device *device)
 109 {
 110         BUG_ON(device->mr == NULL);
 111
 112         tasklet_kill(&device->cq_tasklet);
 113
 114         (void)ib_dereg_mr(device->mr);
 115         (void)ib_destroy_cq(device->cq);
 116         (void)ib_dealloc_pd(device->pd);
 117
 118         device->mr = NULL;
 119         device->cq = NULL;
 120         device->pd = NULL;
 121 }
 122
 123 /**
 124  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 125  *
 126  * returns 0 on success, -1 on failure
 127  */
 128 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 129 {
 130         struct iser_device      *device;
 131         struct ib_qp_init_attr  init_attr;
 132         int                     ret = -ENOMEM;
 133         struct ib_fmr_pool_param params;
 134
 135         BUG_ON(ib_conn->device == NULL);
 136
 137         device = ib_conn->device;
 138
 139         ib_conn->login_buf = kmalloc(ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 140         if (!ib_conn->login_buf) {
 141                 goto alloc_err;
 142                 ret = -ENOMEM;
 143         }
 144
 145         ib_conn->login_dma = ib_dma_map_single(ib_conn->device->ib_device,
 146                                 (void *)ib_conn->login_buf, ISER_RX_LOGIN_SIZE,
 147                                 DMA_FROM_DEVICE);
 148
 149         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 150                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 151                                     GFP_KERNEL);
 152         if (!ib_conn->page_vec) {
 153                 ret = -ENOMEM;
 154                 goto alloc_err;
 155         }
 156         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 157
 158         params.page_shift        = SHIFT_4K;
 159         /* when the first/last SG element are not start/end *
 160          * page aligned, the map whould be of N+1 pages     */
 161         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 162         /* make the pool size twice the max number of SCSI commands *
 163          * the ML is expected to queue, watermark for unmap at 50%  */
 164         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 165         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 166         params.cache             = 0;
 167         params.flush_function    = NULL;
 168         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 169                                     IB_ACCESS_REMOTE_WRITE |
 170                                     IB_ACCESS_REMOTE_READ);
 171
 172         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 173         if (IS_ERR(ib_conn->fmr_pool)) {
 174                 ret = PTR_ERR(ib_conn->fmr_pool);
 175                 goto fmr_pool_err;
 176         }
 177
 178         memset(&init_attr, 0, sizeof init_attr);
 179
 180         init_attr.event_handler = iser_qp_event_callback;
 181         init_attr.qp_context    = (void *)ib_conn;
 182         init_attr.send_cq       = device->cq;
 183         init_attr.recv_cq       = device->cq;
 184         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 185         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 186         init_attr.cap.max_send_sge = MAX_REGD_BUF_VECTOR_LEN;
 187         init_attr.cap.max_recv_sge = 1;
 188         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 189         init_attr.qp_type       = IB_QPT_RC;
 190
 191         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 192         if (ret)
 193                 goto qp_err;
 194
 195         ib_conn->qp = ib_conn->cma_id->qp;
 196         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 197                  ib_conn, ib_conn->cma_id,
 198                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 199         return ret;
 200
 201 qp_err:
 202         (void)ib_destroy_fmr_pool(ib_conn->fmr_pool);
 203 fmr_pool_err:
 204         kfree(ib_conn->page_vec);
 205         kfree(ib_conn->login_buf);
 206 alloc_err:
 207         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 208         return ret;
 209 }
 210
 211 /**
 212  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 213  * -1 on failure
 214  */
 215 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
 216 {
 217         BUG_ON(ib_conn == NULL);
 218
 219         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 220                  ib_conn, ib_conn->cma_id,
 221                  ib_conn->fmr_pool, ib_conn->qp);
 222
 223         /* qp is created only once both addr & route are resolved */
 224         if (ib_conn->fmr_pool != NULL)
 225                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 226
 227         if (ib_conn->qp != NULL)
 228                 rdma_destroy_qp(ib_conn->cma_id);
 229
 230         if (ib_conn->cma_id != NULL)
 231                 rdma_destroy_id(ib_conn->cma_id);
 232
 233         ib_conn->fmr_pool = NULL;
 234         ib_conn->qp       = NULL;
 235         ib_conn->cma_id   = NULL;
 236         kfree(ib_conn->page_vec);
 237
 238         return 0;
 239 }
 240
 241 /**
 242  * based on the resolved device node GUID see if there already allocated
 243  * device for this device. If there's no such, create one.
 244  */
 245 static
 246 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 247 {
 248         struct iser_device *device;
 249
 250         mutex_lock(&ig.device_list_mutex);
 251
 252         list_for_each_entry(device, &ig.device_list, ig_list)
 253                 /* find if there's a match using the node GUID */
 254                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 255                         goto inc_refcnt;
 256
 257         device = kzalloc(sizeof *device, GFP_KERNEL);
 258         if (device == NULL)
 259                 goto out;
 260
 261         /* assign this device to the device */
 262         device->ib_device = cma_id->device;
 263         /* init the device and link it into ig device list */
 264         if (iser_create_device_ib_res(device)) {
 265                 kfree(device);
 266                 device = NULL;
 267                 goto out;
 268         }
 269         list_add(&device->ig_list, &ig.device_list);
 270
 271 inc_refcnt:
 272         device->refcount++;
 273 out:
 274         mutex_unlock(&ig.device_list_mutex);
 275         return device;
 276 }
 277
 278 /* if there's no demand for this device, release it */
 279 static void iser_device_try_release(struct iser_device *device)
 280 {
 281         mutex_lock(&ig.device_list_mutex);
 282         device->refcount--;
 283         iser_err("device %p refcount %d\n",device,device->refcount);
 284         if (!device->refcount) {
 285                 iser_free_device_ib_res(device);
 286                 list_del(&device->ig_list);
 287                 kfree(device);
 288         }
 289         mutex_unlock(&ig.device_list_mutex);
 290 }
 291
 292 int iser_conn_state_comp(struct iser_conn *ib_conn,
 293                         enum iser_ib_conn_state comp)
 294 {
 295         int ret;
 296
 297         spin_lock_bh(&ib_conn->lock);
 298         ret = (ib_conn->state == comp);
 299         spin_unlock_bh(&ib_conn->lock);
 300         return ret;
 301 }
 302
 303 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 304                                      enum iser_ib_conn_state comp,
 305                                      enum iser_ib_conn_state exch)
 306 {
 307         int ret;
 308
 309         spin_lock_bh(&ib_conn->lock);
 310         if ((ret = (ib_conn->state == comp)))
 311                 ib_conn->state = exch;
 312         spin_unlock_bh(&ib_conn->lock);
 313         return ret;
 314 }
 315
 316 /**
 317  * Frees all conn objects and deallocs conn descriptor
 318  */
 319 static void iser_conn_release(struct iser_conn *ib_conn)
 320 {
 321         struct iser_device  *device = ib_conn->device;
 322
 323         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 324
 325         mutex_lock(&ig.connlist_mutex);
 326         list_del(&ib_conn->conn_list);
 327         mutex_unlock(&ig.connlist_mutex);
 328         iser_free_rx_descriptors(ib_conn);
 329         iser_free_ib_conn_res(ib_conn);
 330         ib_conn->device = NULL;
 331         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 332         if (device != NULL)
 333                 iser_device_try_release(device);
 334         if (ib_conn->iser_conn)
 335                 ib_conn->iser_conn->ib_conn = NULL;
 336         iscsi_destroy_endpoint(ib_conn->ep);
 337 }
 338
 339 void iser_conn_get(struct iser_conn *ib_conn)
 340 {
 341         atomic_inc(&ib_conn->refcount);
 342 }
 343
 344 void iser_conn_put(struct iser_conn *ib_conn)
 345 {
 346         if (atomic_dec_and_test(&ib_conn->refcount))
 347                 iser_conn_release(ib_conn);
 348 }
 349
 350 /**
 351  * triggers start of the disconnect procedures and wait for them to be done
 352  */
 353 void iser_conn_terminate(struct iser_conn *ib_conn)
 354 {
 355         int err = 0;
 356
 357         /* change the ib conn state only if the conn is UP, however always call
 358          * rdma_disconnect since this is the only way to cause the CMA to change
 359          * the QP state to ERROR
 360          */
 361
 362         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 363         err = rdma_disconnect(ib_conn->cma_id);
 364         if (err)
 365                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 366                          ib_conn,err);
 367
 368         wait_event_interruptible(ib_conn->wait,
 369                                  ib_conn->state == ISER_CONN_DOWN);
 370
 371         iser_conn_put(ib_conn);
 372 }
 373
 374 static void iser_connect_error(struct rdma_cm_id *cma_id)
 375 {
 376         struct iser_conn *ib_conn;
 377         ib_conn = (struct iser_conn *)cma_id->context;
 378
 379         ib_conn->state = ISER_CONN_DOWN;
 380         wake_up_interruptible(&ib_conn->wait);
 381 }
 382
 383 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 384 {
 385         struct iser_device *device;
 386         struct iser_conn   *ib_conn;
 387         int    ret;
 388
 389         device = iser_device_find_by_ib_device(cma_id);
 390         if (!device) {
 391                 iser_err("device lookup/creation failed\n");
 392                 iser_connect_error(cma_id);
 393                 return;
 394         }
 395
 396         ib_conn = (struct iser_conn *)cma_id->context;
 397         ib_conn->device = device;
 398
 399         ret = rdma_resolve_route(cma_id, 1000);
 400         if (ret) {
 401                 iser_err("resolve route failed: %d\n", ret);
 402                 iser_connect_error(cma_id);
 403         }
 404 }
 405
 406 static void iser_route_handler(struct rdma_cm_id *cma_id)
 407 {
 408         struct rdma_conn_param conn_param;
 409         int    ret;
 410
 411         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 412         if (ret)
 413                 goto failure;
 414
 415         memset(&conn_param, 0, sizeof conn_param);
 416         conn_param.responder_resources = 4;
 417         conn_param.initiator_depth     = 1;
 418         conn_param.retry_count         = 7;
 419         conn_param.rnr_retry_count     = 6;
 420
 421         ret = rdma_connect(cma_id, &conn_param);
 422         if (ret) {
 423                 iser_err("failure connecting: %d\n", ret);
 424                 goto failure;
 425         }
 426
 427         return;
 428 failure:
 429         iser_connect_error(cma_id);
 430 }
 431
 432 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 433 {
 434         struct iser_conn *ib_conn;
 435
 436         ib_conn = (struct iser_conn *)cma_id->context;
 437         ib_conn->state = ISER_CONN_UP;
 438         wake_up_interruptible(&ib_conn->wait);
 439 }
 440
 441 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 442 {
 443         struct iser_conn *ib_conn;
 444
 445         ib_conn = (struct iser_conn *)cma_id->context;
 446         ib_conn->disc_evt_flag = 1;
 447
 448         /* getting here when the state is UP means that the conn is being *
 449          * terminated asynchronously from the iSCSI layer's perspective.  */
 450         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 451                                       ISER_CONN_TERMINATING))
 452                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 453                                    ISCSI_ERR_CONN_FAILED);
 454
 455         /* Complete the termination process if no posts are pending */
 456         if ((atomic_read(&ib_conn->post_recv_buf_count) == 0) &&
 457             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 458                 ib_conn->state = ISER_CONN_DOWN;
 459                 wake_up_interruptible(&ib_conn->wait);
 460         }
 461 }
 462
 463 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 464 {
 465         int ret = 0;
 466
 467         iser_err("event %d conn %p id %p\n",event->event,cma_id->context,cma_id);
 468
 469         switch (event->event) {
 470         case RDMA_CM_EVENT_ADDR_RESOLVED:
 471                 iser_addr_handler(cma_id);
 472                 break;
 473         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 474                 iser_route_handler(cma_id);
 475                 break;
 476         case RDMA_CM_EVENT_ESTABLISHED:
 477                 iser_connected_handler(cma_id);
 478                 break;
 479         case RDMA_CM_EVENT_ADDR_ERROR:
 480         case RDMA_CM_EVENT_ROUTE_ERROR:
 481         case RDMA_CM_EVENT_CONNECT_ERROR:
 482         case RDMA_CM_EVENT_UNREACHABLE:
 483         case RDMA_CM_EVENT_REJECTED:
 484                 iser_err("event: %d, error: %d\n", event->event, event->status);
 485                 iser_connect_error(cma_id);
 486                 break;
 487         case RDMA_CM_EVENT_DISCONNECTED:
 488         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 489         case RDMA_CM_EVENT_ADDR_CHANGE:
 490                 iser_disconnected_handler(cma_id);
 491                 break;
 492         default:
 493                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 494                 break;
 495         }
 496         return ret;
 497 }
 498
 499 void iser_conn_init(struct iser_conn *ib_conn)
 500 {
 501         ib_conn->state = ISER_CONN_INIT;
 502         init_waitqueue_head(&ib_conn->wait);
 503         atomic_set(&ib_conn->post_recv_buf_count, 0);
 504         atomic_set(&ib_conn->post_send_buf_count, 0);
 505         atomic_set(&ib_conn->refcount, 1);
 506         INIT_LIST_HEAD(&ib_conn->conn_list);
 507         spin_lock_init(&ib_conn->lock);
 508 }
 509
 510  /**
 511  * starts the process of connecting to the target
 512  * sleeps until the connection is established or rejected
 513  */
 514 int iser_connect(struct iser_conn   *ib_conn,
 515                  struct sockaddr_in *src_addr,
 516                  struct sockaddr_in *dst_addr,
 517                  int                 non_blocking)
 518 {
 519         struct sockaddr *src, *dst;
 520         int err = 0;
 521
 522         sprintf(ib_conn->name, "%pI4:%d",
 523                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 524
 525         /* the device is known only --after-- address resolution */
 526         ib_conn->device = NULL;
 527
 528         iser_err("connecting to: %pI4, port 0x%x\n",
 529                  &dst_addr->sin_addr, dst_addr->sin_port);
 530
 531         ib_conn->state = ISER_CONN_PENDING;
 532
 533         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 534                                              (void *)ib_conn,
 535                                              RDMA_PS_TCP);
 536         if (IS_ERR(ib_conn->cma_id)) {
 537                 err = PTR_ERR(ib_conn->cma_id);
 538                 iser_err("rdma_create_id failed: %d\n", err);
 539                 goto id_failure;
 540         }
 541
 542         src = (struct sockaddr *)src_addr;
 543         dst = (struct sockaddr *)dst_addr;
 544         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 545         if (err) {
 546                 iser_err("rdma_resolve_addr failed: %d\n", err);
 547                 goto addr_failure;
 548         }
 549
 550         if (!non_blocking) {
 551                 wait_event_interruptible(ib_conn->wait,
 552                                          (ib_conn->state != ISER_CONN_PENDING));
 553
 554                 if (ib_conn->state != ISER_CONN_UP) {
 555                         err =  -EIO;
 556                         goto connect_failure;
 557                 }
 558         }
 559
 560         mutex_lock(&ig.connlist_mutex);
 561         list_add(&ib_conn->conn_list, &ig.connlist);
 562         mutex_unlock(&ig.connlist_mutex);
 563         return 0;
 564
 565 id_failure:
 566         ib_conn->cma_id = NULL;
 567 addr_failure:
 568         ib_conn->state = ISER_CONN_DOWN;
 569 connect_failure:
 570         iser_conn_release(ib_conn);
 571         return err;
 572 }
 573
 574 /**
 575  * iser_reg_page_vec - Register physical memory
 576  *
 577  * returns: 0 on success, errno code on failure
 578  */
 579 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 580                       struct iser_page_vec *page_vec,
 581                       struct iser_mem_reg  *mem_reg)
 582 {
 583         struct ib_pool_fmr *mem;
 584         u64                io_addr;
 585         u64                *page_list;
 586         int                status;
 587
 588         page_list = page_vec->pages;
 589         io_addr   = page_list[0];
 590
 591         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 592                                     page_list,
 593                                     page_vec->length,
 594                                     io_addr);
 595
 596         if (IS_ERR(mem)) {
 597                 status = (int)PTR_ERR(mem);
 598                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 599                 return status;
 600         }
 601
 602         mem_reg->lkey  = mem->fmr->lkey;
 603         mem_reg->rkey  = mem->fmr->rkey;
 604         mem_reg->len   = page_vec->length * SIZE_4K;
 605         mem_reg->va    = io_addr;
 606         mem_reg->is_fmr = 1;
 607         mem_reg->mem_h = (void *)mem;
 608
 609         mem_reg->va   += page_vec->offset;
 610         mem_reg->len   = page_vec->data_size;
 611
 612         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 613                  "entry[0]: (0x%08lx,%ld)] -> "
 614                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 615                  page_vec, page_vec->length,
 616                  (unsigned long)page_vec->pages[0],
 617                  (unsigned long)page_vec->data_size,
 618                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 619                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 620         return 0;
 621 }
 622
 623 /**
 624  * Unregister (previosuly registered) memory.
 625  */
 626 void iser_unreg_mem(struct iser_mem_reg *reg)
 627 {
 628         int ret;
 629
 630         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 631
 632         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 633         if (ret)
 634                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 635
 636         reg->mem_h = NULL;
 637 }
 638
 639 int iser_post_recvl(struct iser_conn *ib_conn)
 640 {
 641         struct ib_recv_wr rx_wr, *rx_wr_failed;
 642         struct ib_sge     sge;
 643         int ib_ret;
 644
 645         sge.addr   = ib_conn->login_dma;
 646         sge.length = ISER_RX_LOGIN_SIZE;
 647         sge.lkey   = ib_conn->device->mr->lkey;
 648
 649         rx_wr.wr_id   = (unsigned long)ib_conn->login_buf;
 650         rx_wr.sg_list = &sge;
 651         rx_wr.num_sge = 1;
 652         rx_wr.next    = NULL;
 653
 654         atomic_inc(&ib_conn->post_recv_buf_count);
 655         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 656         if (ib_ret) {
 657                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 658                 atomic_dec(&ib_conn->post_recv_buf_count);
 659         }
 660         return ib_ret;
 661 }
 662
 663 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 664 {
 665         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 666         int i, ib_ret;
 667         unsigned int my_rx_head = ib_conn->rx_desc_head;
 668         struct iser_rx_desc *rx_desc;
 669
 670         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 671                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 672                 rx_wr->wr_id    = (unsigned long)rx_desc;
 673                 rx_wr->sg_list  = &rx_desc->rx_sg;
 674                 rx_wr->num_sge  = 1;
 675                 rx_wr->next     = rx_wr + 1;
 676                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 677         }
 678
 679         rx_wr--;
 680         rx_wr->next = NULL; /* mark end of work requests list */
 681
 682         atomic_add(count, &ib_conn->post_recv_buf_count);
 683         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 684         if (ib_ret) {
 685                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 686                 atomic_sub(count, &ib_conn->post_recv_buf_count);
 687         } else
 688                 ib_conn->rx_desc_head = my_rx_head;
 689         return ib_ret;
 690 }
 691
 692
 693 /**
 694  * iser_dto_to_iov - builds IOV from a dto descriptor
 695  */
 696 static void iser_dto_to_iov(struct iser_dto *dto, struct ib_sge *iov, int iov_len)
 697 {
 698         int                  i;
 699         struct ib_sge        *sge;
 700         struct iser_regd_buf *regd_buf;
 701
 702         if (dto->regd_vector_len > iov_len) {
 703                 iser_err("iov size %d too small for posting dto of len %d\n",
 704                          iov_len, dto->regd_vector_len);
 705                 BUG();
 706         }
 707
 708         for (i = 0; i < dto->regd_vector_len; i++) {
 709                 sge         = &iov[i];
 710                 regd_buf  = dto->regd[i];
 711
 712                 sge->addr   = regd_buf->reg.va;
 713                 sge->length = regd_buf->reg.len;
 714                 sge->lkey   = regd_buf->reg.lkey;
 715
 716                 if (dto->used_sz[i] > 0)  /* Adjust size */
 717                         sge->length = dto->used_sz[i];
 718
 719                 /* offset and length should not exceed the regd buf length */
 720                 if (sge->length + dto->offset[i] > regd_buf->reg.len) {
 721                         iser_err("Used len:%ld + offset:%d, exceed reg.buf.len:"
 722                                  "%ld in dto:0x%p [%d], va:0x%08lX\n",
 723                                  (unsigned long)sge->length, dto->offset[i],
 724                                  (unsigned long)regd_buf->reg.len, dto, i,
 725                                  (unsigned long)sge->addr);
 726                         BUG();
 727                 }
 728
 729                 sge->addr += dto->offset[i]; /* Adjust offset */
 730         }
 731 }
 732
 733
 734 /**
 735  * iser_start_send - Initiate a Send DTO operation
 736  *
 737  * returns 0 on success, -1 on failure
 738  */
 739 int iser_post_send(struct iser_desc *tx_desc)
 740 {
 741         int               ib_ret, ret_val = 0;
 742         struct ib_send_wr send_wr, *send_wr_failed;
 743         struct ib_sge     iov[MAX_REGD_BUF_VECTOR_LEN];
 744         struct iser_conn  *ib_conn;
 745         struct iser_dto   *dto = &tx_desc->dto;
 746
 747         ib_conn = dto->ib_conn;
 748
 749         iser_dto_to_iov(dto, iov, MAX_REGD_BUF_VECTOR_LEN);
 750
 751         send_wr.next       = NULL;
 752         send_wr.wr_id      = (unsigned long)tx_desc;
 753         send_wr.sg_list    = iov;
 754         send_wr.num_sge    = dto->regd_vector_len;
 755         send_wr.opcode     = IB_WR_SEND;
 756         send_wr.send_flags = dto->notify_enable ? IB_SEND_SIGNALED : 0;
 757
 758         atomic_inc(&ib_conn->post_send_buf_count);
 759
 760         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 761         if (ib_ret) {
 762                 iser_err("Failed to start SEND DTO, dto: 0x%p, IOV len: %d\n",
 763                          dto, dto->regd_vector_len);
 764                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 765                 atomic_dec(&ib_conn->post_send_buf_count);
 766                 ret_val = -1;
 767         }
 768
 769         return ret_val;
 770 }
 771
 772 static void iser_handle_comp_error(struct iser_desc *desc,
 773                                 struct iser_conn *ib_conn)
 774 {
 775         struct iser_rx_desc *rx       = (struct iser_rx_desc *)desc;
 776         struct iser_rx_desc *rx_first = ib_conn->rx_descs;
 777         struct iser_rx_desc *rx_last  = rx_first + (ISER_QP_MAX_RECV_DTOS - 1);
 778
 779         if ((char *)desc == ib_conn->login_buf ||
 780                         (rx_first <= rx && rx <= rx_last))
 781                 atomic_dec(&ib_conn->post_recv_buf_count);
 782          else { /* type is TX control/command/dataout */
 783                 if (desc->type == ISCSI_TX_DATAOUT)
 784                         kmem_cache_free(ig.desc_cache, desc);
 785                 atomic_dec(&ib_conn->post_send_buf_count);
 786         }
 787
 788         if (atomic_read(&ib_conn->post_recv_buf_count) == 0 &&
 789             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 790                 /* getting here when the state is UP means that the conn is *
 791                  * being terminated asynchronously from the iSCSI layer's   *
 792                  * perspective.                                             */
 793                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 794                     ISER_CONN_TERMINATING))
 795                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 796                                            ISCSI_ERR_CONN_FAILED);
 797
 798                 /* complete the termination process if disconnect event was delivered *
 799                  * note there are no more non completed posts to the QP               */
 800                 if (ib_conn->disc_evt_flag) {
 801                         ib_conn->state = ISER_CONN_DOWN;
 802                         wake_up_interruptible(&ib_conn->wait);
 803                 }
 804         }
 805 }
 806
 807 static void iser_cq_tasklet_fn(unsigned long data)
 808 {
 809          struct iser_device  *device = (struct iser_device *)data;
 810          struct ib_cq        *cq = device->cq;
 811          struct ib_wc        wc;
 812          struct iser_desc    *desc;
 813          unsigned long       xfer_len;
 814         struct iser_conn *ib_conn;
 815
 816         while (ib_poll_cq(cq, 1, &wc) == 1) {
 817                 desc     = (struct iser_desc *) (unsigned long) wc.wr_id;
 818                 BUG_ON(desc == NULL);
 819                 ib_conn = wc.qp->qp_context;
 820
 821                 if (wc.status == IB_WC_SUCCESS) {
 822                         if (wc.opcode == IB_WC_RECV) {
 823                                 xfer_len = (unsigned long)wc.byte_len;
 824                                 iser_rcv_completion((struct iser_rx_desc *)desc,
 825                                                         xfer_len, ib_conn);
 826                         } else /* type == ISCSI_TX_CONTROL/SCSI_CMD/DOUT */
 827                                 iser_snd_completion(desc);
 828                 } else {
 829                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 830                                 iser_err("id %llx status %d vend_err %x\n",
 831                                         wc.wr_id, wc.status, wc.vendor_err);
 832                         iser_handle_comp_error(desc, ib_conn);
 833                 }
 834         }
 835         /* #warning "it is assumed here that arming CQ only once its empty" *
 836          * " would not cause interrupts to be missed"                       */
 837         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 838 }
 839
 840 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 841 {
 842         struct iser_device  *device = (struct iser_device *)cq_context;
 843
 844         tasklet_schedule(&device->cq_tasklet);
 845 }