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_RX_CQ_LEN      (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
  41 #define ISER_MAX_TX_CQ_LEN      (ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
  42
  43 static void iser_cq_tasklet_fn(unsigned long data);
  44 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
  45
  46 static void iser_cq_event_callback(struct ib_event *cause, void *context)
  47 {
  48         iser_err("got cq event %d \n", cause->event);
  49 }
  50
  51 static void iser_qp_event_callback(struct ib_event *cause, void *context)
  52 {
  53         iser_err("got qp event %d\n",cause->event);
  54 }
  55
  56 /**
  57  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  58  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  59  * the adapator.
  60  *
  61  * returns 0 on success, -1 on failure
  62  */
  63 static int iser_create_device_ib_res(struct iser_device *device)
  64 {
  65         device->pd = ib_alloc_pd(device->ib_device);
  66         if (IS_ERR(device->pd))
  67                 goto pd_err;
  68
  69         device->rx_cq = ib_create_cq(device->ib_device,
  70                                   iser_cq_callback,
  71                                   iser_cq_event_callback,
  72                                   (void *)device,
  73                                   ISER_MAX_RX_CQ_LEN, 0);
  74         if (IS_ERR(device->rx_cq))
  75                 goto rx_cq_err;
  76
  77         device->tx_cq = ib_create_cq(device->ib_device,
  78                                   NULL, iser_cq_event_callback,
  79                                   (void *)device,
  80                                   ISER_MAX_TX_CQ_LEN, 0);
  81
  82         if (IS_ERR(device->tx_cq))
  83                 goto tx_cq_err;
  84
  85         if (ib_req_notify_cq(device->rx_cq, IB_CQ_NEXT_COMP))
  86                 goto cq_arm_err;
  87
  88         tasklet_init(&device->cq_tasklet,
  89                      iser_cq_tasklet_fn,
  90                      (unsigned long)device);
  91
  92         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
  93                                    IB_ACCESS_REMOTE_WRITE |
  94                                    IB_ACCESS_REMOTE_READ);
  95         if (IS_ERR(device->mr))
  96                 goto dma_mr_err;
  97
  98         return 0;
  99
 100 dma_mr_err:
 101         tasklet_kill(&device->cq_tasklet);
 102 cq_arm_err:
 103         ib_destroy_cq(device->tx_cq);
 104 tx_cq_err:
 105         ib_destroy_cq(device->rx_cq);
 106 rx_cq_err:
 107         ib_dealloc_pd(device->pd);
 108 pd_err:
 109         iser_err("failed to allocate an IB resource\n");
 110         return -1;
 111 }
 112
 113 /**
 114  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 115  * CQ and PD created with the device associated with the adapator.
 116  */
 117 static void iser_free_device_ib_res(struct iser_device *device)
 118 {
 119         BUG_ON(device->mr == NULL);
 120
 121         tasklet_kill(&device->cq_tasklet);
 122
 123         (void)ib_dereg_mr(device->mr);
 124         (void)ib_destroy_cq(device->tx_cq);
 125         (void)ib_destroy_cq(device->rx_cq);
 126         (void)ib_dealloc_pd(device->pd);
 127
 128         device->mr = NULL;
 129         device->tx_cq = NULL;
 130         device->rx_cq = NULL;
 131         device->pd = NULL;
 132 }
 133
 134 /**
 135  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 136  *
 137  * returns 0 on success, -1 on failure
 138  */
 139 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 140 {
 141         struct iser_device      *device;
 142         struct ib_qp_init_attr  init_attr;
 143         int                     ret = -ENOMEM;
 144         struct ib_fmr_pool_param params;
 145
 146         BUG_ON(ib_conn->device == NULL);
 147
 148         device = ib_conn->device;
 149
 150         ib_conn->login_buf = kmalloc(ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 151         if (!ib_conn->login_buf) {
 152                 goto alloc_err;
 153                 ret = -ENOMEM;
 154         }
 155
 156         ib_conn->login_dma = ib_dma_map_single(ib_conn->device->ib_device,
 157                                 (void *)ib_conn->login_buf, ISER_RX_LOGIN_SIZE,
 158                                 DMA_FROM_DEVICE);
 159
 160         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 161                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 162                                     GFP_KERNEL);
 163         if (!ib_conn->page_vec) {
 164                 ret = -ENOMEM;
 165                 goto alloc_err;
 166         }
 167         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 168
 169         params.page_shift        = SHIFT_4K;
 170         /* when the first/last SG element are not start/end *
 171          * page aligned, the map whould be of N+1 pages     */
 172         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 173         /* make the pool size twice the max number of SCSI commands *
 174          * the ML is expected to queue, watermark for unmap at 50%  */
 175         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 176         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 177         params.cache             = 0;
 178         params.flush_function    = NULL;
 179         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 180                                     IB_ACCESS_REMOTE_WRITE |
 181                                     IB_ACCESS_REMOTE_READ);
 182
 183         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 184         if (IS_ERR(ib_conn->fmr_pool)) {
 185                 ret = PTR_ERR(ib_conn->fmr_pool);
 186                 goto fmr_pool_err;
 187         }
 188
 189         memset(&init_attr, 0, sizeof init_attr);
 190
 191         init_attr.event_handler = iser_qp_event_callback;
 192         init_attr.qp_context    = (void *)ib_conn;
 193         init_attr.send_cq       = device->tx_cq;
 194         init_attr.recv_cq       = device->rx_cq;
 195         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 196         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 197         init_attr.cap.max_send_sge = 2;
 198         init_attr.cap.max_recv_sge = 1;
 199         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 200         init_attr.qp_type       = IB_QPT_RC;
 201
 202         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 203         if (ret)
 204                 goto qp_err;
 205
 206         ib_conn->qp = ib_conn->cma_id->qp;
 207         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 208                  ib_conn, ib_conn->cma_id,
 209                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 210         return ret;
 211
 212 qp_err:
 213         (void)ib_destroy_fmr_pool(ib_conn->fmr_pool);
 214 fmr_pool_err:
 215         kfree(ib_conn->page_vec);
 216         kfree(ib_conn->login_buf);
 217 alloc_err:
 218         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 219         return ret;
 220 }
 221
 222 /**
 223  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 224  * -1 on failure
 225  */
 226 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
 227 {
 228         BUG_ON(ib_conn == NULL);
 229
 230         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 231                  ib_conn, ib_conn->cma_id,
 232                  ib_conn->fmr_pool, ib_conn->qp);
 233
 234         /* qp is created only once both addr & route are resolved */
 235         if (ib_conn->fmr_pool != NULL)
 236                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 237
 238         if (ib_conn->qp != NULL)
 239                 rdma_destroy_qp(ib_conn->cma_id);
 240
 241         if (ib_conn->cma_id != NULL)
 242                 rdma_destroy_id(ib_conn->cma_id);
 243
 244         ib_conn->fmr_pool = NULL;
 245         ib_conn->qp       = NULL;
 246         ib_conn->cma_id   = NULL;
 247         kfree(ib_conn->page_vec);
 248
 249         return 0;
 250 }
 251
 252 /**
 253  * based on the resolved device node GUID see if there already allocated
 254  * device for this device. If there's no such, create one.
 255  */
 256 static
 257 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 258 {
 259         struct iser_device *device;
 260
 261         mutex_lock(&ig.device_list_mutex);
 262
 263         list_for_each_entry(device, &ig.device_list, ig_list)
 264                 /* find if there's a match using the node GUID */
 265                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 266                         goto inc_refcnt;
 267
 268         device = kzalloc(sizeof *device, GFP_KERNEL);
 269         if (device == NULL)
 270                 goto out;
 271
 272         /* assign this device to the device */
 273         device->ib_device = cma_id->device;
 274         /* init the device and link it into ig device list */
 275         if (iser_create_device_ib_res(device)) {
 276                 kfree(device);
 277                 device = NULL;
 278                 goto out;
 279         }
 280         list_add(&device->ig_list, &ig.device_list);
 281
 282 inc_refcnt:
 283         device->refcount++;
 284 out:
 285         mutex_unlock(&ig.device_list_mutex);
 286         return device;
 287 }
 288
 289 /* if there's no demand for this device, release it */
 290 static void iser_device_try_release(struct iser_device *device)
 291 {
 292         mutex_lock(&ig.device_list_mutex);
 293         device->refcount--;
 294         iser_err("device %p refcount %d\n",device,device->refcount);
 295         if (!device->refcount) {
 296                 iser_free_device_ib_res(device);
 297                 list_del(&device->ig_list);
 298                 kfree(device);
 299         }
 300         mutex_unlock(&ig.device_list_mutex);
 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 (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         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         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                 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         ib_conn->post_recv_buf_count += 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                 ib_conn->post_recv_buf_count -= count;
 687         } else
 688                 ib_conn->rx_desc_head = my_rx_head;
 689         return ib_ret;
 690 }
 691
 692
 693 /**
 694  * iser_start_send - Initiate a Send DTO operation
 695  *
 696  * returns 0 on success, -1 on failure
 697  */
 698 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 699 {
 700         int               ib_ret;
 701         struct ib_send_wr send_wr, *send_wr_failed;
 702
 703         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 704                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 705
 706         send_wr.next       = NULL;
 707         send_wr.wr_id      = (unsigned long)tx_desc;
 708         send_wr.sg_list    = tx_desc->tx_sg;
 709         send_wr.num_sge    = tx_desc->num_sge;
 710         send_wr.opcode     = IB_WR_SEND;
 711         send_wr.send_flags = IB_SEND_SIGNALED;
 712
 713         atomic_inc(&ib_conn->post_send_buf_count);
 714
 715         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 716         if (ib_ret) {
 717                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 718                 atomic_dec(&ib_conn->post_send_buf_count);
 719         }
 720         return ib_ret;
 721 }
 722
 723 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 724                                 struct iser_conn *ib_conn)
 725 {
 726         if (desc && desc->type == ISCSI_TX_DATAOUT)
 727                 kmem_cache_free(ig.desc_cache, desc);
 728
 729         if (ib_conn->post_recv_buf_count == 0 &&
 730             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 731                 /* getting here when the state is UP means that the conn is *
 732                  * being terminated asynchronously from the iSCSI layer's   *
 733                  * perspective.                                             */
 734                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 735                     ISER_CONN_TERMINATING))
 736                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 737                                            ISCSI_ERR_CONN_FAILED);
 738
 739                 /* complete the termination process if disconnect event was delivered *
 740                  * note there are no more non completed posts to the QP               */
 741                 if (ib_conn->disc_evt_flag) {
 742                         ib_conn->state = ISER_CONN_DOWN;
 743                         wake_up_interruptible(&ib_conn->wait);
 744                 }
 745         }
 746 }
 747
 748 static int iser_drain_tx_cq(struct iser_device  *device)
 749 {
 750         struct ib_cq  *cq = device->tx_cq;
 751         struct ib_wc  wc;
 752         struct iser_tx_desc *tx_desc;
 753         struct iser_conn *ib_conn;
 754         int completed_tx = 0;
 755
 756         while (ib_poll_cq(cq, 1, &wc) == 1) {
 757                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 758                 ib_conn = wc.qp->qp_context;
 759                 if (wc.status == IB_WC_SUCCESS) {
 760                         if (wc.opcode == IB_WC_SEND)
 761                                 iser_snd_completion(tx_desc, ib_conn);
 762                         else
 763                                 iser_err("expected opcode %d got %d\n",
 764                                         IB_WC_SEND, wc.opcode);
 765                 } else {
 766                         iser_err("tx id %llx status %d vend_err %x\n",
 767                                 wc.wr_id, wc.status, wc.vendor_err);
 768                         atomic_dec(&ib_conn->post_send_buf_count);
 769                         iser_handle_comp_error(tx_desc, ib_conn);
 770                 }
 771                 completed_tx++;
 772         }
 773         return completed_tx;
 774 }
 775
 776
 777 static void iser_cq_tasklet_fn(unsigned long data)
 778 {
 779          struct iser_device  *device = (struct iser_device *)data;
 780          struct ib_cq        *cq = device->rx_cq;
 781          struct ib_wc        wc;
 782          struct iser_rx_desc *desc;
 783          unsigned long       xfer_len;
 784         struct iser_conn *ib_conn;
 785         int completed_tx, completed_rx;
 786         completed_tx = completed_rx = 0;
 787
 788         while (ib_poll_cq(cq, 1, &wc) == 1) {
 789                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 790                 BUG_ON(desc == NULL);
 791                 ib_conn = wc.qp->qp_context;
 792                 if (wc.status == IB_WC_SUCCESS) {
 793                         if (wc.opcode == IB_WC_RECV) {
 794                                 xfer_len = (unsigned long)wc.byte_len;
 795                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 796                         } else
 797                                 iser_err("expected opcode %d got %d\n",
 798                                         IB_WC_RECV, wc.opcode);
 799                 } else {
 800                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 801                                 iser_err("rx id %llx status %d vend_err %x\n",
 802                                         wc.wr_id, wc.status, wc.vendor_err);
 803                         ib_conn->post_recv_buf_count--;
 804                         iser_handle_comp_error(NULL, ib_conn);
 805                 }
 806                 completed_rx++;
 807                 if (!(completed_rx & 63))
 808                         completed_tx += iser_drain_tx_cq(device);
 809         }
 810         /* #warning "it is assumed here that arming CQ only once its empty" *
 811          * " would not cause interrupts to be missed"                       */
 812         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 813
 814         completed_tx += iser_drain_tx_cq(device);
 815         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 816 }
 817
 818 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 819 {
 820         struct iser_device  *device = (struct iser_device *)cq_context;
 821
 822         tasklet_schedule(&device->cq_tasklet);
 823 }