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/slab.h>
  36 #include <linux/delay.h>
  37
  38 #include "iscsi_iser.h"
  39
  40 #define ISCSI_ISER_MAX_CONN     8
  41 #define ISER_MAX_RX_CQ_LEN      (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
  42 #define ISER_MAX_TX_CQ_LEN      (ISER_QP_MAX_REQ_DTOS  * 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 static void iser_event_handler(struct ib_event_handler *handler,
  58                                 struct ib_event *event)
  59 {
  60         iser_err("async event %d on device %s port %d\n", event->event,
  61                 event->device->name, event->element.port_num);
  62 }
  63
  64 /**
  65  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  66  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  67  * the adapator.
  68  *
  69  * returns 0 on success, -1 on failure
  70  */
  71 static int iser_create_device_ib_res(struct iser_device *device)
  72 {
  73         device->pd = ib_alloc_pd(device->ib_device);
  74         if (IS_ERR(device->pd))
  75                 goto pd_err;
  76
  77         device->rx_cq = ib_create_cq(device->ib_device,
  78                                   iser_cq_callback,
  79                                   iser_cq_event_callback,
  80                                   (void *)device,
  81                                   ISER_MAX_RX_CQ_LEN, 0);
  82         if (IS_ERR(device->rx_cq))
  83                 goto rx_cq_err;
  84
  85         device->tx_cq = ib_create_cq(device->ib_device,
  86                                   NULL, iser_cq_event_callback,
  87                                   (void *)device,
  88                                   ISER_MAX_TX_CQ_LEN, 0);
  89
  90         if (IS_ERR(device->tx_cq))
  91                 goto tx_cq_err;
  92
  93         if (ib_req_notify_cq(device->rx_cq, IB_CQ_NEXT_COMP))
  94                 goto cq_arm_err;
  95
  96         tasklet_init(&device->cq_tasklet,
  97                      iser_cq_tasklet_fn,
  98                      (unsigned long)device);
  99
 100         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
 101                                    IB_ACCESS_REMOTE_WRITE |
 102                                    IB_ACCESS_REMOTE_READ);
 103         if (IS_ERR(device->mr))
 104                 goto dma_mr_err;
 105
 106         INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
 107                                 iser_event_handler);
 108         if (ib_register_event_handler(&device->event_handler))
 109                 goto handler_err;
 110
 111         return 0;
 112
 113 handler_err:
 114         ib_dereg_mr(device->mr);
 115 dma_mr_err:
 116         tasklet_kill(&device->cq_tasklet);
 117 cq_arm_err:
 118         ib_destroy_cq(device->tx_cq);
 119 tx_cq_err:
 120         ib_destroy_cq(device->rx_cq);
 121 rx_cq_err:
 122         ib_dealloc_pd(device->pd);
 123 pd_err:
 124         iser_err("failed to allocate an IB resource\n");
 125         return -1;
 126 }
 127
 128 /**
 129  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 130  * CQ and PD created with the device associated with the adapator.
 131  */
 132 static void iser_free_device_ib_res(struct iser_device *device)
 133 {
 134         BUG_ON(device->mr == NULL);
 135
 136         tasklet_kill(&device->cq_tasklet);
 137         (void)ib_unregister_event_handler(&device->event_handler);
 138         (void)ib_dereg_mr(device->mr);
 139         (void)ib_destroy_cq(device->tx_cq);
 140         (void)ib_destroy_cq(device->rx_cq);
 141         (void)ib_dealloc_pd(device->pd);
 142
 143         device->mr = NULL;
 144         device->tx_cq = NULL;
 145         device->rx_cq = NULL;
 146         device->pd = NULL;
 147 }
 148
 149 /**
 150  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 151  *
 152  * returns 0 on success, -1 on failure
 153  */
 154 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 155 {
 156         struct iser_device      *device;
 157         struct ib_qp_init_attr  init_attr;
 158         int                     req_err, resp_err, ret = -ENOMEM;
 159         struct ib_fmr_pool_param params;
 160
 161         BUG_ON(ib_conn->device == NULL);
 162
 163         device = ib_conn->device;
 164
 165         ib_conn->login_buf = kmalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 166                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 167         if (!ib_conn->login_buf)
 168                 goto out_err;
 169
 170         ib_conn->login_req_buf  = ib_conn->login_buf;
 171         ib_conn->login_resp_buf = ib_conn->login_buf + ISCSI_DEF_MAX_RECV_SEG_LEN;
 172
 173         ib_conn->login_req_dma = ib_dma_map_single(ib_conn->device->ib_device,
 174                                 (void *)ib_conn->login_req_buf,
 175                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
 176
 177         ib_conn->login_resp_dma = ib_dma_map_single(ib_conn->device->ib_device,
 178                                 (void *)ib_conn->login_resp_buf,
 179                                 ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
 180
 181         req_err  = ib_dma_mapping_error(device->ib_device, ib_conn->login_req_dma);
 182         resp_err = ib_dma_mapping_error(device->ib_device, ib_conn->login_resp_dma);
 183
 184         if (req_err || resp_err) {
 185                 if (req_err)
 186                         ib_conn->login_req_dma = 0;
 187                 if (resp_err)
 188                         ib_conn->login_resp_dma = 0;
 189                 goto out_err;
 190         }
 191
 192         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 193                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 194                                     GFP_KERNEL);
 195         if (!ib_conn->page_vec)
 196                 goto out_err;
 197
 198         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 199
 200         params.page_shift        = SHIFT_4K;
 201         /* when the first/last SG element are not start/end *
 202          * page aligned, the map whould be of N+1 pages     */
 203         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 204         /* make the pool size twice the max number of SCSI commands *
 205          * the ML is expected to queue, watermark for unmap at 50%  */
 206         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 207         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 208         params.cache             = 0;
 209         params.flush_function    = NULL;
 210         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 211                                     IB_ACCESS_REMOTE_WRITE |
 212                                     IB_ACCESS_REMOTE_READ);
 213
 214         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 215         if (IS_ERR(ib_conn->fmr_pool)) {
 216                 ret = PTR_ERR(ib_conn->fmr_pool);
 217                 ib_conn->fmr_pool = NULL;
 218                 goto out_err;
 219         }
 220
 221         memset(&init_attr, 0, sizeof init_attr);
 222
 223         init_attr.event_handler = iser_qp_event_callback;
 224         init_attr.qp_context    = (void *)ib_conn;
 225         init_attr.send_cq       = device->tx_cq;
 226         init_attr.recv_cq       = device->rx_cq;
 227         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 228         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 229         init_attr.cap.max_send_sge = 2;
 230         init_attr.cap.max_recv_sge = 1;
 231         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 232         init_attr.qp_type       = IB_QPT_RC;
 233
 234         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 235         if (ret)
 236                 goto out_err;
 237
 238         ib_conn->qp = ib_conn->cma_id->qp;
 239         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 240                  ib_conn, ib_conn->cma_id,
 241                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 242         return ret;
 243
 244 out_err:
 245         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 246         return ret;
 247 }
 248
 249 /**
 250  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 251  * -1 on failure
 252  */
 253 static int iser_free_ib_conn_res(struct iser_conn *ib_conn, int can_destroy_id)
 254 {
 255         BUG_ON(ib_conn == NULL);
 256
 257         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 258                  ib_conn, ib_conn->cma_id,
 259                  ib_conn->fmr_pool, ib_conn->qp);
 260
 261         /* qp is created only once both addr & route are resolved */
 262         if (ib_conn->fmr_pool != NULL)
 263                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 264
 265         if (ib_conn->qp != NULL)
 266                 rdma_destroy_qp(ib_conn->cma_id);
 267
 268         /* if cma handler context, the caller acts s.t the cma destroy the id */
 269         if (ib_conn->cma_id != NULL && can_destroy_id)
 270                 rdma_destroy_id(ib_conn->cma_id);
 271
 272         ib_conn->fmr_pool = NULL;
 273         ib_conn->qp       = NULL;
 274         ib_conn->cma_id   = NULL;
 275         kfree(ib_conn->page_vec);
 276
 277         if (ib_conn->login_buf) {
 278                 if (ib_conn->login_req_dma)
 279                         ib_dma_unmap_single(ib_conn->device->ib_device,
 280                                 ib_conn->login_req_dma,
 281                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
 282                 if (ib_conn->login_resp_dma)
 283                         ib_dma_unmap_single(ib_conn->device->ib_device,
 284                                 ib_conn->login_resp_dma,
 285                                 ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
 286                 kfree(ib_conn->login_buf);
 287         }
 288
 289         return 0;
 290 }
 291
 292 /**
 293  * based on the resolved device node GUID see if there already allocated
 294  * device for this device. If there's no such, create one.
 295  */
 296 static
 297 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 298 {
 299         struct iser_device *device;
 300
 301         mutex_lock(&ig.device_list_mutex);
 302
 303         list_for_each_entry(device, &ig.device_list, ig_list)
 304                 /* find if there's a match using the node GUID */
 305                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 306                         goto inc_refcnt;
 307
 308         device = kzalloc(sizeof *device, GFP_KERNEL);
 309         if (device == NULL)
 310                 goto out;
 311
 312         /* assign this device to the device */
 313         device->ib_device = cma_id->device;
 314         /* init the device and link it into ig device list */
 315         if (iser_create_device_ib_res(device)) {
 316                 kfree(device);
 317                 device = NULL;
 318                 goto out;
 319         }
 320         list_add(&device->ig_list, &ig.device_list);
 321
 322 inc_refcnt:
 323         device->refcount++;
 324 out:
 325         mutex_unlock(&ig.device_list_mutex);
 326         return device;
 327 }
 328
 329 /* if there's no demand for this device, release it */
 330 static void iser_device_try_release(struct iser_device *device)
 331 {
 332         mutex_lock(&ig.device_list_mutex);
 333         device->refcount--;
 334         iser_err("device %p refcount %d\n",device,device->refcount);
 335         if (!device->refcount) {
 336                 iser_free_device_ib_res(device);
 337                 list_del(&device->ig_list);
 338                 kfree(device);
 339         }
 340         mutex_unlock(&ig.device_list_mutex);
 341 }
 342
 343 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 344                                      enum iser_ib_conn_state comp,
 345                                      enum iser_ib_conn_state exch)
 346 {
 347         int ret;
 348
 349         spin_lock_bh(&ib_conn->lock);
 350         if ((ret = (ib_conn->state == comp)))
 351                 ib_conn->state = exch;
 352         spin_unlock_bh(&ib_conn->lock);
 353         return ret;
 354 }
 355
 356 /**
 357  * Frees all conn objects and deallocs conn descriptor
 358  */
 359 static void iser_conn_release(struct iser_conn *ib_conn, int can_destroy_id)
 360 {
 361         struct iser_device  *device = ib_conn->device;
 362
 363         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 364
 365         mutex_lock(&ig.connlist_mutex);
 366         list_del(&ib_conn->conn_list);
 367         mutex_unlock(&ig.connlist_mutex);
 368         iser_free_rx_descriptors(ib_conn);
 369         iser_free_ib_conn_res(ib_conn, can_destroy_id);
 370         ib_conn->device = NULL;
 371         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 372         if (device != NULL)
 373                 iser_device_try_release(device);
 374         iscsi_destroy_endpoint(ib_conn->ep);
 375 }
 376
 377 void iser_conn_get(struct iser_conn *ib_conn)
 378 {
 379         atomic_inc(&ib_conn->refcount);
 380 }
 381
 382 int iser_conn_put(struct iser_conn *ib_conn, int can_destroy_id)
 383 {
 384         if (atomic_dec_and_test(&ib_conn->refcount)) {
 385                 iser_conn_release(ib_conn, can_destroy_id);
 386                 return 1;
 387         }
 388         return 0;
 389 }
 390
 391 /**
 392  * triggers start of the disconnect procedures and wait for them to be done
 393  */
 394 void iser_conn_terminate(struct iser_conn *ib_conn)
 395 {
 396         int err = 0;
 397
 398         /* change the ib conn state only if the conn is UP, however always call
 399          * rdma_disconnect since this is the only way to cause the CMA to change
 400          * the QP state to ERROR
 401          */
 402
 403         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 404         err = rdma_disconnect(ib_conn->cma_id);
 405         if (err)
 406                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 407                          ib_conn,err);
 408
 409         wait_event_interruptible(ib_conn->wait,
 410                                  ib_conn->state == ISER_CONN_DOWN);
 411
 412         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 413 }
 414
 415 static int iser_connect_error(struct rdma_cm_id *cma_id)
 416 {
 417         struct iser_conn *ib_conn;
 418         ib_conn = (struct iser_conn *)cma_id->context;
 419
 420         ib_conn->state = ISER_CONN_DOWN;
 421         wake_up_interruptible(&ib_conn->wait);
 422         return iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 423 }
 424
 425 static int iser_addr_handler(struct rdma_cm_id *cma_id)
 426 {
 427         struct iser_device *device;
 428         struct iser_conn   *ib_conn;
 429         int    ret;
 430
 431         device = iser_device_find_by_ib_device(cma_id);
 432         if (!device) {
 433                 iser_err("device lookup/creation failed\n");
 434                 return iser_connect_error(cma_id);
 435         }
 436
 437         ib_conn = (struct iser_conn *)cma_id->context;
 438         ib_conn->device = device;
 439
 440         ret = rdma_resolve_route(cma_id, 1000);
 441         if (ret) {
 442                 iser_err("resolve route failed: %d\n", ret);
 443                 return iser_connect_error(cma_id);
 444         }
 445
 446         return 0;
 447 }
 448
 449 static int iser_route_handler(struct rdma_cm_id *cma_id)
 450 {
 451         struct rdma_conn_param conn_param;
 452         int    ret;
 453
 454         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 455         if (ret)
 456                 goto failure;
 457
 458         memset(&conn_param, 0, sizeof conn_param);
 459         conn_param.responder_resources = 4;
 460         conn_param.initiator_depth     = 1;
 461         conn_param.retry_count         = 7;
 462         conn_param.rnr_retry_count     = 6;
 463
 464         ret = rdma_connect(cma_id, &conn_param);
 465         if (ret) {
 466                 iser_err("failure connecting: %d\n", ret);
 467                 goto failure;
 468         }
 469
 470         return 0;
 471 failure:
 472         return iser_connect_error(cma_id);
 473 }
 474
 475 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 476 {
 477         struct iser_conn *ib_conn;
 478
 479         ib_conn = (struct iser_conn *)cma_id->context;
 480         ib_conn->state = ISER_CONN_UP;
 481         wake_up_interruptible(&ib_conn->wait);
 482 }
 483
 484 static int iser_disconnected_handler(struct rdma_cm_id *cma_id)
 485 {
 486         struct iser_conn *ib_conn;
 487         int ret;
 488
 489         ib_conn = (struct iser_conn *)cma_id->context;
 490
 491         /* getting here when the state is UP means that the conn is being *
 492          * terminated asynchronously from the iSCSI layer's perspective.  */
 493         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 494                                       ISER_CONN_TERMINATING))
 495                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 496                                    ISCSI_ERR_CONN_FAILED);
 497
 498         /* Complete the termination process if no posts are pending */
 499         if (ib_conn->post_recv_buf_count == 0 &&
 500             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 501                 ib_conn->state = ISER_CONN_DOWN;
 502                 wake_up_interruptible(&ib_conn->wait);
 503         }
 504
 505         ret = iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 506         return ret;
 507 }
 508
 509 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 510 {
 511         int ret = 0;
 512
 513         iser_err("event %d status %d conn %p id %p\n",
 514                 event->event, event->status, cma_id->context, cma_id);
 515
 516         switch (event->event) {
 517         case RDMA_CM_EVENT_ADDR_RESOLVED:
 518                 ret = iser_addr_handler(cma_id);
 519                 break;
 520         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 521                 ret = iser_route_handler(cma_id);
 522                 break;
 523         case RDMA_CM_EVENT_ESTABLISHED:
 524                 iser_connected_handler(cma_id);
 525                 break;
 526         case RDMA_CM_EVENT_ADDR_ERROR:
 527         case RDMA_CM_EVENT_ROUTE_ERROR:
 528         case RDMA_CM_EVENT_CONNECT_ERROR:
 529         case RDMA_CM_EVENT_UNREACHABLE:
 530         case RDMA_CM_EVENT_REJECTED:
 531                 ret = iser_connect_error(cma_id);
 532                 break;
 533         case RDMA_CM_EVENT_DISCONNECTED:
 534         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 535         case RDMA_CM_EVENT_ADDR_CHANGE:
 536                 ret = iser_disconnected_handler(cma_id);
 537                 break;
 538         default:
 539                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 540                 break;
 541         }
 542         return ret;
 543 }
 544
 545 void iser_conn_init(struct iser_conn *ib_conn)
 546 {
 547         ib_conn->state = ISER_CONN_INIT;
 548         init_waitqueue_head(&ib_conn->wait);
 549         ib_conn->post_recv_buf_count = 0;
 550         atomic_set(&ib_conn->post_send_buf_count, 0);
 551         atomic_set(&ib_conn->refcount, 1); /* ref ib conn allocation */
 552         INIT_LIST_HEAD(&ib_conn->conn_list);
 553         spin_lock_init(&ib_conn->lock);
 554 }
 555
 556  /**
 557  * starts the process of connecting to the target
 558  * sleeps until the connection is established or rejected
 559  */
 560 int iser_connect(struct iser_conn   *ib_conn,
 561                  struct sockaddr_in *src_addr,
 562                  struct sockaddr_in *dst_addr,
 563                  int                 non_blocking)
 564 {
 565         struct sockaddr *src, *dst;
 566         int err = 0;
 567
 568         sprintf(ib_conn->name, "%pI4:%d",
 569                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 570
 571         /* the device is known only --after-- address resolution */
 572         ib_conn->device = NULL;
 573
 574         iser_err("connecting to: %pI4, port 0x%x\n",
 575                  &dst_addr->sin_addr, dst_addr->sin_port);
 576
 577         ib_conn->state = ISER_CONN_PENDING;
 578
 579         iser_conn_get(ib_conn); /* ref ib conn's cma id */
 580         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 581                                              (void *)ib_conn,
 582                                              RDMA_PS_TCP, IB_QPT_RC);
 583         if (IS_ERR(ib_conn->cma_id)) {
 584                 err = PTR_ERR(ib_conn->cma_id);
 585                 iser_err("rdma_create_id failed: %d\n", err);
 586                 goto id_failure;
 587         }
 588
 589         src = (struct sockaddr *)src_addr;
 590         dst = (struct sockaddr *)dst_addr;
 591         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 592         if (err) {
 593                 iser_err("rdma_resolve_addr failed: %d\n", err);
 594                 goto addr_failure;
 595         }
 596
 597         if (!non_blocking) {
 598                 wait_event_interruptible(ib_conn->wait,
 599                                          (ib_conn->state != ISER_CONN_PENDING));
 600
 601                 if (ib_conn->state != ISER_CONN_UP) {
 602                         err =  -EIO;
 603                         goto connect_failure;
 604                 }
 605         }
 606
 607         mutex_lock(&ig.connlist_mutex);
 608         list_add(&ib_conn->conn_list, &ig.connlist);
 609         mutex_unlock(&ig.connlist_mutex);
 610         return 0;
 611
 612 id_failure:
 613         ib_conn->cma_id = NULL;
 614 addr_failure:
 615         ib_conn->state = ISER_CONN_DOWN;
 616         iser_conn_put(ib_conn, 1); /* deref ib conn's cma id */
 617 connect_failure:
 618         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 619         return err;
 620 }
 621
 622 /**
 623  * iser_reg_page_vec - Register physical memory
 624  *
 625  * returns: 0 on success, errno code on failure
 626  */
 627 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 628                       struct iser_page_vec *page_vec,
 629                       struct iser_mem_reg  *mem_reg)
 630 {
 631         struct ib_pool_fmr *mem;
 632         u64                io_addr;
 633         u64                *page_list;
 634         int                status;
 635
 636         page_list = page_vec->pages;
 637         io_addr   = page_list[0];
 638
 639         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 640                                     page_list,
 641                                     page_vec->length,
 642                                     io_addr);
 643
 644         if (IS_ERR(mem)) {
 645                 status = (int)PTR_ERR(mem);
 646                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 647                 return status;
 648         }
 649
 650         mem_reg->lkey  = mem->fmr->lkey;
 651         mem_reg->rkey  = mem->fmr->rkey;
 652         mem_reg->len   = page_vec->length * SIZE_4K;
 653         mem_reg->va    = io_addr;
 654         mem_reg->is_fmr = 1;
 655         mem_reg->mem_h = (void *)mem;
 656
 657         mem_reg->va   += page_vec->offset;
 658         mem_reg->len   = page_vec->data_size;
 659
 660         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 661                  "entry[0]: (0x%08lx,%ld)] -> "
 662                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 663                  page_vec, page_vec->length,
 664                  (unsigned long)page_vec->pages[0],
 665                  (unsigned long)page_vec->data_size,
 666                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 667                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 668         return 0;
 669 }
 670
 671 /**
 672  * Unregister (previosuly registered) memory.
 673  */
 674 void iser_unreg_mem(struct iser_mem_reg *reg)
 675 {
 676         int ret;
 677
 678         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 679
 680         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 681         if (ret)
 682                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 683
 684         reg->mem_h = NULL;
 685 }
 686
 687 int iser_post_recvl(struct iser_conn *ib_conn)
 688 {
 689         struct ib_recv_wr rx_wr, *rx_wr_failed;
 690         struct ib_sge     sge;
 691         int ib_ret;
 692
 693         sge.addr   = ib_conn->login_resp_dma;
 694         sge.length = ISER_RX_LOGIN_SIZE;
 695         sge.lkey   = ib_conn->device->mr->lkey;
 696
 697         rx_wr.wr_id   = (unsigned long)ib_conn->login_resp_buf;
 698         rx_wr.sg_list = &sge;
 699         rx_wr.num_sge = 1;
 700         rx_wr.next    = NULL;
 701
 702         ib_conn->post_recv_buf_count++;
 703         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 704         if (ib_ret) {
 705                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 706                 ib_conn->post_recv_buf_count--;
 707         }
 708         return ib_ret;
 709 }
 710
 711 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 712 {
 713         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 714         int i, ib_ret;
 715         unsigned int my_rx_head = ib_conn->rx_desc_head;
 716         struct iser_rx_desc *rx_desc;
 717
 718         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 719                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 720                 rx_wr->wr_id    = (unsigned long)rx_desc;
 721                 rx_wr->sg_list  = &rx_desc->rx_sg;
 722                 rx_wr->num_sge  = 1;
 723                 rx_wr->next     = rx_wr + 1;
 724                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 725         }
 726
 727         rx_wr--;
 728         rx_wr->next = NULL; /* mark end of work requests list */
 729
 730         ib_conn->post_recv_buf_count += count;
 731         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 732         if (ib_ret) {
 733                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 734                 ib_conn->post_recv_buf_count -= count;
 735         } else
 736                 ib_conn->rx_desc_head = my_rx_head;
 737         return ib_ret;
 738 }
 739
 740
 741 /**
 742  * iser_start_send - Initiate a Send DTO operation
 743  *
 744  * returns 0 on success, -1 on failure
 745  */
 746 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 747 {
 748         int               ib_ret;
 749         struct ib_send_wr send_wr, *send_wr_failed;
 750
 751         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 752                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 753
 754         send_wr.next       = NULL;
 755         send_wr.wr_id      = (unsigned long)tx_desc;
 756         send_wr.sg_list    = tx_desc->tx_sg;
 757         send_wr.num_sge    = tx_desc->num_sge;
 758         send_wr.opcode     = IB_WR_SEND;
 759         send_wr.send_flags = IB_SEND_SIGNALED;
 760
 761         atomic_inc(&ib_conn->post_send_buf_count);
 762
 763         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 764         if (ib_ret) {
 765                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 766                 atomic_dec(&ib_conn->post_send_buf_count);
 767         }
 768         return ib_ret;
 769 }
 770
 771 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 772                                 struct iser_conn *ib_conn)
 773 {
 774         if (desc && desc->type == ISCSI_TX_DATAOUT)
 775                 kmem_cache_free(ig.desc_cache, desc);
 776
 777         if (ib_conn->post_recv_buf_count == 0 &&
 778             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 779                 /* getting here when the state is UP means that the conn is *
 780                  * being terminated asynchronously from the iSCSI layer's   *
 781                  * perspective.                                             */
 782                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 783                     ISER_CONN_TERMINATING))
 784                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 785                                            ISCSI_ERR_CONN_FAILED);
 786
 787                 /* no more non completed posts to the QP, complete the
 788                  * termination process w.o worrying on disconnect event */
 789                 ib_conn->state = ISER_CONN_DOWN;
 790                 wake_up_interruptible(&ib_conn->wait);
 791         }
 792 }
 793
 794 static int iser_drain_tx_cq(struct iser_device  *device)
 795 {
 796         struct ib_cq  *cq = device->tx_cq;
 797         struct ib_wc  wc;
 798         struct iser_tx_desc *tx_desc;
 799         struct iser_conn *ib_conn;
 800         int completed_tx = 0;
 801
 802         while (ib_poll_cq(cq, 1, &wc) == 1) {
 803                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 804                 ib_conn = wc.qp->qp_context;
 805                 if (wc.status == IB_WC_SUCCESS) {
 806                         if (wc.opcode == IB_WC_SEND)
 807                                 iser_snd_completion(tx_desc, ib_conn);
 808                         else
 809                                 iser_err("expected opcode %d got %d\n",
 810                                         IB_WC_SEND, wc.opcode);
 811                 } else {
 812                         iser_err("tx id %llx status %d vend_err %x\n",
 813                                 wc.wr_id, wc.status, wc.vendor_err);
 814                         atomic_dec(&ib_conn->post_send_buf_count);
 815                         iser_handle_comp_error(tx_desc, ib_conn);
 816                 }
 817                 completed_tx++;
 818         }
 819         return completed_tx;
 820 }
 821
 822
 823 static void iser_cq_tasklet_fn(unsigned long data)
 824 {
 825          struct iser_device  *device = (struct iser_device *)data;
 826          struct ib_cq        *cq = device->rx_cq;
 827          struct ib_wc        wc;
 828          struct iser_rx_desc *desc;
 829          unsigned long       xfer_len;
 830         struct iser_conn *ib_conn;
 831         int completed_tx, completed_rx;
 832         completed_tx = completed_rx = 0;
 833
 834         while (ib_poll_cq(cq, 1, &wc) == 1) {
 835                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 836                 BUG_ON(desc == NULL);
 837                 ib_conn = wc.qp->qp_context;
 838                 if (wc.status == IB_WC_SUCCESS) {
 839                         if (wc.opcode == IB_WC_RECV) {
 840                                 xfer_len = (unsigned long)wc.byte_len;
 841                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 842                         } else
 843                                 iser_err("expected opcode %d got %d\n",
 844                                         IB_WC_RECV, wc.opcode);
 845                 } else {
 846                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 847                                 iser_err("rx id %llx status %d vend_err %x\n",
 848                                         wc.wr_id, wc.status, wc.vendor_err);
 849                         ib_conn->post_recv_buf_count--;
 850                         iser_handle_comp_error(NULL, ib_conn);
 851                 }
 852                 completed_rx++;
 853                 if (!(completed_rx & 63))
 854                         completed_tx += iser_drain_tx_cq(device);
 855         }
 856         /* #warning "it is assumed here that arming CQ only once its empty" *
 857          * " would not cause interrupts to be missed"                       */
 858         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 859
 860         completed_tx += iser_drain_tx_cq(device);
 861         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 862 }
 863
 864 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 865 {
 866         struct iser_device  *device = (struct iser_device *)cq_context;
 867
 868         tasklet_schedule(&device->cq_tasklet);
 869 }