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         int i, j;
  74         struct iser_cq_desc *cq_desc;
  75
  76         device->cqs_used = min(ISER_MAX_CQ, device->ib_device->num_comp_vectors);
  77         iser_err("using %d CQs, device %s supports %d vectors\n", device->cqs_used,
  78                  device->ib_device->name, device->ib_device->num_comp_vectors);
  79
  80         device->cq_desc = kmalloc(sizeof(struct iser_cq_desc) * device->cqs_used,
  81                                   GFP_KERNEL);
  82         if (device->cq_desc == NULL)
  83                 goto cq_desc_err;
  84         cq_desc = device->cq_desc;
  85
  86         device->pd = ib_alloc_pd(device->ib_device);
  87         if (IS_ERR(device->pd))
  88                 goto pd_err;
  89
  90         for (i = 0; i < device->cqs_used; i++) {
  91                 cq_desc[i].device   = device;
  92                 cq_desc[i].cq_index = i;
  93
  94                 device->rx_cq[i] = ib_create_cq(device->ib_device,
  95                                           iser_cq_callback,
  96                                           iser_cq_event_callback,
  97                                           (void *)&cq_desc[i],
  98                                           ISER_MAX_RX_CQ_LEN, i);
  99                 if (IS_ERR(device->rx_cq[i]))
 100                         goto cq_err;
 101
 102                 device->tx_cq[i] = ib_create_cq(device->ib_device,
 103                                           NULL, iser_cq_event_callback,
 104                                           (void *)&cq_desc[i],
 105                                           ISER_MAX_TX_CQ_LEN, i);
 106
 107                 if (IS_ERR(device->tx_cq[i]))
 108                         goto cq_err;
 109
 110                 if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
 111                         goto cq_err;
 112
 113                 tasklet_init(&device->cq_tasklet[i],
 114                              iser_cq_tasklet_fn,
 115                         (unsigned long)&cq_desc[i]);
 116         }
 117
 118         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
 119                                    IB_ACCESS_REMOTE_WRITE |
 120                                    IB_ACCESS_REMOTE_READ);
 121         if (IS_ERR(device->mr))
 122                 goto dma_mr_err;
 123
 124         INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
 125                                 iser_event_handler);
 126         if (ib_register_event_handler(&device->event_handler))
 127                 goto handler_err;
 128
 129         return 0;
 130
 131 handler_err:
 132         ib_dereg_mr(device->mr);
 133 dma_mr_err:
 134         for (j = 0; j < device->cqs_used; j++)
 135                 tasklet_kill(&device->cq_tasklet[j]);
 136 cq_err:
 137         for (j = 0; j < i; j++) {
 138                 if (device->tx_cq[j])
 139                         ib_destroy_cq(device->tx_cq[j]);
 140                 if (device->rx_cq[j])
 141                         ib_destroy_cq(device->rx_cq[j]);
 142         }
 143         ib_dealloc_pd(device->pd);
 144 pd_err:
 145         kfree(device->cq_desc);
 146 cq_desc_err:
 147         iser_err("failed to allocate an IB resource\n");
 148         return -1;
 149 }
 150
 151 /**
 152  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 153  * CQ and PD created with the device associated with the adapator.
 154  */
 155 static void iser_free_device_ib_res(struct iser_device *device)
 156 {
 157         int i;
 158         BUG_ON(device->mr == NULL);
 159
 160         for (i = 0; i < device->cqs_used; i++) {
 161                 tasklet_kill(&device->cq_tasklet[i]);
 162                 (void)ib_destroy_cq(device->tx_cq[i]);
 163                 (void)ib_destroy_cq(device->rx_cq[i]);
 164                 device->tx_cq[i] = NULL;
 165                 device->rx_cq[i] = NULL;
 166         }
 167
 168         (void)ib_unregister_event_handler(&device->event_handler);
 169         (void)ib_dereg_mr(device->mr);
 170         (void)ib_dealloc_pd(device->pd);
 171
 172         kfree(device->cq_desc);
 173
 174         device->mr = NULL;
 175         device->pd = NULL;
 176 }
 177
 178 /**
 179  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 180  *
 181  * returns 0 on success, -1 on failure
 182  */
 183 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 184 {
 185         struct iser_device      *device;
 186         struct ib_qp_init_attr  init_attr;
 187         int                     req_err, resp_err, ret = -ENOMEM;
 188         struct ib_fmr_pool_param params;
 189         int index, min_index = 0;
 190
 191         BUG_ON(ib_conn->device == NULL);
 192
 193         device = ib_conn->device;
 194
 195         ib_conn->login_buf = kmalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 196                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 197         if (!ib_conn->login_buf)
 198                 goto out_err;
 199
 200         ib_conn->login_req_buf  = ib_conn->login_buf;
 201         ib_conn->login_resp_buf = ib_conn->login_buf + ISCSI_DEF_MAX_RECV_SEG_LEN;
 202
 203         ib_conn->login_req_dma = ib_dma_map_single(ib_conn->device->ib_device,
 204                                 (void *)ib_conn->login_req_buf,
 205                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
 206
 207         ib_conn->login_resp_dma = ib_dma_map_single(ib_conn->device->ib_device,
 208                                 (void *)ib_conn->login_resp_buf,
 209                                 ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
 210
 211         req_err  = ib_dma_mapping_error(device->ib_device, ib_conn->login_req_dma);
 212         resp_err = ib_dma_mapping_error(device->ib_device, ib_conn->login_resp_dma);
 213
 214         if (req_err || resp_err) {
 215                 if (req_err)
 216                         ib_conn->login_req_dma = 0;
 217                 if (resp_err)
 218                         ib_conn->login_resp_dma = 0;
 219                 goto out_err;
 220         }
 221
 222         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 223                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 224                                     GFP_KERNEL);
 225         if (!ib_conn->page_vec)
 226                 goto out_err;
 227
 228         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 229
 230         params.page_shift        = SHIFT_4K;
 231         /* when the first/last SG element are not start/end *
 232          * page aligned, the map whould be of N+1 pages     */
 233         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 234         /* make the pool size twice the max number of SCSI commands *
 235          * the ML is expected to queue, watermark for unmap at 50%  */
 236         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 237         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 238         params.cache             = 0;
 239         params.flush_function    = NULL;
 240         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 241                                     IB_ACCESS_REMOTE_WRITE |
 242                                     IB_ACCESS_REMOTE_READ);
 243
 244         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 245         ret = PTR_ERR(ib_conn->fmr_pool);
 246         if (IS_ERR(ib_conn->fmr_pool) && ret != -ENOSYS) {
 247                 ib_conn->fmr_pool = NULL;
 248                 goto out_err;
 249         } else if (ret == -ENOSYS) {
 250                 ib_conn->fmr_pool = NULL;
 251                 iser_warn("FMRs are not supported, using unaligned mode\n");
 252                 ret = 0;
 253         }
 254
 255         memset(&init_attr, 0, sizeof init_attr);
 256
 257         mutex_lock(&ig.connlist_mutex);
 258         /* select the CQ with the minimal number of usages */
 259         for (index = 0; index < device->cqs_used; index++)
 260                 if (device->cq_active_qps[index] <
 261                     device->cq_active_qps[min_index])
 262                         min_index = index;
 263         device->cq_active_qps[min_index]++;
 264         mutex_unlock(&ig.connlist_mutex);
 265         iser_err("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 266
 267         init_attr.event_handler = iser_qp_event_callback;
 268         init_attr.qp_context    = (void *)ib_conn;
 269         init_attr.send_cq       = device->tx_cq[min_index];
 270         init_attr.recv_cq       = device->rx_cq[min_index];
 271         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 272         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 273         init_attr.cap.max_send_sge = 2;
 274         init_attr.cap.max_recv_sge = 1;
 275         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 276         init_attr.qp_type       = IB_QPT_RC;
 277
 278         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 279         if (ret)
 280                 goto out_err;
 281
 282         ib_conn->qp = ib_conn->cma_id->qp;
 283         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 284                  ib_conn, ib_conn->cma_id,
 285                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 286         return ret;
 287
 288 out_err:
 289         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 290         return ret;
 291 }
 292
 293 /**
 294  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 295  * -1 on failure
 296  */
 297 static int iser_free_ib_conn_res(struct iser_conn *ib_conn, int can_destroy_id)
 298 {
 299         int cq_index;
 300         BUG_ON(ib_conn == NULL);
 301
 302         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 303                  ib_conn, ib_conn->cma_id,
 304                  ib_conn->fmr_pool, ib_conn->qp);
 305
 306         /* qp is created only once both addr & route are resolved */
 307         if (ib_conn->fmr_pool != NULL)
 308                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 309
 310         if (ib_conn->qp != NULL) {
 311                 cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
 312                 ib_conn->device->cq_active_qps[cq_index]--;
 313
 314                 rdma_destroy_qp(ib_conn->cma_id);
 315         }
 316         /* if cma handler context, the caller acts s.t the cma destroy the id */
 317         if (ib_conn->cma_id != NULL && can_destroy_id)
 318                 rdma_destroy_id(ib_conn->cma_id);
 319
 320         ib_conn->fmr_pool = NULL;
 321         ib_conn->qp       = NULL;
 322         ib_conn->cma_id   = NULL;
 323         kfree(ib_conn->page_vec);
 324
 325         if (ib_conn->login_buf) {
 326                 if (ib_conn->login_req_dma)
 327                         ib_dma_unmap_single(ib_conn->device->ib_device,
 328                                 ib_conn->login_req_dma,
 329                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
 330                 if (ib_conn->login_resp_dma)
 331                         ib_dma_unmap_single(ib_conn->device->ib_device,
 332                                 ib_conn->login_resp_dma,
 333                                 ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
 334                 kfree(ib_conn->login_buf);
 335         }
 336
 337         return 0;
 338 }
 339
 340 /**
 341  * based on the resolved device node GUID see if there already allocated
 342  * device for this device. If there's no such, create one.
 343  */
 344 static
 345 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 346 {
 347         struct iser_device *device;
 348
 349         mutex_lock(&ig.device_list_mutex);
 350
 351         list_for_each_entry(device, &ig.device_list, ig_list)
 352                 /* find if there's a match using the node GUID */
 353                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 354                         goto inc_refcnt;
 355
 356         device = kzalloc(sizeof *device, GFP_KERNEL);
 357         if (device == NULL)
 358                 goto out;
 359
 360         /* assign this device to the device */
 361         device->ib_device = cma_id->device;
 362         /* init the device and link it into ig device list */
 363         if (iser_create_device_ib_res(device)) {
 364                 kfree(device);
 365                 device = NULL;
 366                 goto out;
 367         }
 368         list_add(&device->ig_list, &ig.device_list);
 369
 370 inc_refcnt:
 371         device->refcount++;
 372 out:
 373         mutex_unlock(&ig.device_list_mutex);
 374         return device;
 375 }
 376
 377 /* if there's no demand for this device, release it */
 378 static void iser_device_try_release(struct iser_device *device)
 379 {
 380         mutex_lock(&ig.device_list_mutex);
 381         device->refcount--;
 382         iser_err("device %p refcount %d\n",device,device->refcount);
 383         if (!device->refcount) {
 384                 iser_free_device_ib_res(device);
 385                 list_del(&device->ig_list);
 386                 kfree(device);
 387         }
 388         mutex_unlock(&ig.device_list_mutex);
 389 }
 390
 391 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 392                                      enum iser_ib_conn_state comp,
 393                                      enum iser_ib_conn_state exch)
 394 {
 395         int ret;
 396
 397         spin_lock_bh(&ib_conn->lock);
 398         if ((ret = (ib_conn->state == comp)))
 399                 ib_conn->state = exch;
 400         spin_unlock_bh(&ib_conn->lock);
 401         return ret;
 402 }
 403
 404 /**
 405  * Frees all conn objects and deallocs conn descriptor
 406  */
 407 static void iser_conn_release(struct iser_conn *ib_conn, int can_destroy_id)
 408 {
 409         struct iser_device  *device = ib_conn->device;
 410
 411         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 412
 413         mutex_lock(&ig.connlist_mutex);
 414         list_del(&ib_conn->conn_list);
 415         mutex_unlock(&ig.connlist_mutex);
 416         iser_free_rx_descriptors(ib_conn);
 417         iser_free_ib_conn_res(ib_conn, can_destroy_id);
 418         ib_conn->device = NULL;
 419         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 420         if (device != NULL)
 421                 iser_device_try_release(device);
 422         iscsi_destroy_endpoint(ib_conn->ep);
 423 }
 424
 425 void iser_conn_get(struct iser_conn *ib_conn)
 426 {
 427         atomic_inc(&ib_conn->refcount);
 428 }
 429
 430 int iser_conn_put(struct iser_conn *ib_conn, int can_destroy_id)
 431 {
 432         if (atomic_dec_and_test(&ib_conn->refcount)) {
 433                 iser_conn_release(ib_conn, can_destroy_id);
 434                 return 1;
 435         }
 436         return 0;
 437 }
 438
 439 /**
 440  * triggers start of the disconnect procedures and wait for them to be done
 441  */
 442 void iser_conn_terminate(struct iser_conn *ib_conn)
 443 {
 444         int err = 0;
 445
 446         /* change the ib conn state only if the conn is UP, however always call
 447          * rdma_disconnect since this is the only way to cause the CMA to change
 448          * the QP state to ERROR
 449          */
 450
 451         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 452         err = rdma_disconnect(ib_conn->cma_id);
 453         if (err)
 454                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 455                          ib_conn,err);
 456
 457         wait_event_interruptible(ib_conn->wait,
 458                                  ib_conn->state == ISER_CONN_DOWN);
 459
 460         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 461 }
 462
 463 static int iser_connect_error(struct rdma_cm_id *cma_id)
 464 {
 465         struct iser_conn *ib_conn;
 466         ib_conn = (struct iser_conn *)cma_id->context;
 467
 468         ib_conn->state = ISER_CONN_DOWN;
 469         wake_up_interruptible(&ib_conn->wait);
 470         return iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 471 }
 472
 473 static int iser_addr_handler(struct rdma_cm_id *cma_id)
 474 {
 475         struct iser_device *device;
 476         struct iser_conn   *ib_conn;
 477         int    ret;
 478
 479         device = iser_device_find_by_ib_device(cma_id);
 480         if (!device) {
 481                 iser_err("device lookup/creation failed\n");
 482                 return iser_connect_error(cma_id);
 483         }
 484
 485         ib_conn = (struct iser_conn *)cma_id->context;
 486         ib_conn->device = device;
 487
 488         ret = rdma_resolve_route(cma_id, 1000);
 489         if (ret) {
 490                 iser_err("resolve route failed: %d\n", ret);
 491                 return iser_connect_error(cma_id);
 492         }
 493
 494         return 0;
 495 }
 496
 497 static int iser_route_handler(struct rdma_cm_id *cma_id)
 498 {
 499         struct rdma_conn_param conn_param;
 500         int    ret;
 501
 502         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 503         if (ret)
 504                 goto failure;
 505
 506         memset(&conn_param, 0, sizeof conn_param);
 507         conn_param.responder_resources = 4;
 508         conn_param.initiator_depth     = 1;
 509         conn_param.retry_count         = 7;
 510         conn_param.rnr_retry_count     = 6;
 511
 512         ret = rdma_connect(cma_id, &conn_param);
 513         if (ret) {
 514                 iser_err("failure connecting: %d\n", ret);
 515                 goto failure;
 516         }
 517
 518         return 0;
 519 failure:
 520         return iser_connect_error(cma_id);
 521 }
 522
 523 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 524 {
 525         struct iser_conn *ib_conn;
 526
 527         ib_conn = (struct iser_conn *)cma_id->context;
 528         ib_conn->state = ISER_CONN_UP;
 529         wake_up_interruptible(&ib_conn->wait);
 530 }
 531
 532 static int iser_disconnected_handler(struct rdma_cm_id *cma_id)
 533 {
 534         struct iser_conn *ib_conn;
 535         int ret;
 536
 537         ib_conn = (struct iser_conn *)cma_id->context;
 538
 539         /* getting here when the state is UP means that the conn is being *
 540          * terminated asynchronously from the iSCSI layer's perspective.  */
 541         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 542                                       ISER_CONN_TERMINATING))
 543                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 544                                    ISCSI_ERR_CONN_FAILED);
 545
 546         /* Complete the termination process if no posts are pending */
 547         if (ib_conn->post_recv_buf_count == 0 &&
 548             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 549                 ib_conn->state = ISER_CONN_DOWN;
 550                 wake_up_interruptible(&ib_conn->wait);
 551         }
 552
 553         ret = iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 554         return ret;
 555 }
 556
 557 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 558 {
 559         int ret = 0;
 560
 561         iser_err("event %d status %d conn %p id %p\n",
 562                 event->event, event->status, cma_id->context, cma_id);
 563
 564         switch (event->event) {
 565         case RDMA_CM_EVENT_ADDR_RESOLVED:
 566                 ret = iser_addr_handler(cma_id);
 567                 break;
 568         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 569                 ret = iser_route_handler(cma_id);
 570                 break;
 571         case RDMA_CM_EVENT_ESTABLISHED:
 572                 iser_connected_handler(cma_id);
 573                 break;
 574         case RDMA_CM_EVENT_ADDR_ERROR:
 575         case RDMA_CM_EVENT_ROUTE_ERROR:
 576         case RDMA_CM_EVENT_CONNECT_ERROR:
 577         case RDMA_CM_EVENT_UNREACHABLE:
 578         case RDMA_CM_EVENT_REJECTED:
 579                 ret = iser_connect_error(cma_id);
 580                 break;
 581         case RDMA_CM_EVENT_DISCONNECTED:
 582         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 583         case RDMA_CM_EVENT_ADDR_CHANGE:
 584                 ret = iser_disconnected_handler(cma_id);
 585                 break;
 586         default:
 587                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 588                 break;
 589         }
 590         return ret;
 591 }
 592
 593 void iser_conn_init(struct iser_conn *ib_conn)
 594 {
 595         ib_conn->state = ISER_CONN_INIT;
 596         init_waitqueue_head(&ib_conn->wait);
 597         ib_conn->post_recv_buf_count = 0;
 598         atomic_set(&ib_conn->post_send_buf_count, 0);
 599         atomic_set(&ib_conn->refcount, 1); /* ref ib conn allocation */
 600         INIT_LIST_HEAD(&ib_conn->conn_list);
 601         spin_lock_init(&ib_conn->lock);
 602 }
 603
 604  /**
 605  * starts the process of connecting to the target
 606  * sleeps until the connection is established or rejected
 607  */
 608 int iser_connect(struct iser_conn   *ib_conn,
 609                  struct sockaddr_in *src_addr,
 610                  struct sockaddr_in *dst_addr,
 611                  int                 non_blocking)
 612 {
 613         struct sockaddr *src, *dst;
 614         int err = 0;
 615
 616         sprintf(ib_conn->name, "%pI4:%d",
 617                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 618
 619         /* the device is known only --after-- address resolution */
 620         ib_conn->device = NULL;
 621
 622         iser_err("connecting to: %pI4, port 0x%x\n",
 623                  &dst_addr->sin_addr, dst_addr->sin_port);
 624
 625         ib_conn->state = ISER_CONN_PENDING;
 626
 627         iser_conn_get(ib_conn); /* ref ib conn's cma id */
 628         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 629                                              (void *)ib_conn,
 630                                              RDMA_PS_TCP, IB_QPT_RC);
 631         if (IS_ERR(ib_conn->cma_id)) {
 632                 err = PTR_ERR(ib_conn->cma_id);
 633                 iser_err("rdma_create_id failed: %d\n", err);
 634                 goto id_failure;
 635         }
 636
 637         src = (struct sockaddr *)src_addr;
 638         dst = (struct sockaddr *)dst_addr;
 639         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 640         if (err) {
 641                 iser_err("rdma_resolve_addr failed: %d\n", err);
 642                 goto addr_failure;
 643         }
 644
 645         if (!non_blocking) {
 646                 wait_event_interruptible(ib_conn->wait,
 647                                          (ib_conn->state != ISER_CONN_PENDING));
 648
 649                 if (ib_conn->state != ISER_CONN_UP) {
 650                         err =  -EIO;
 651                         goto connect_failure;
 652                 }
 653         }
 654
 655         mutex_lock(&ig.connlist_mutex);
 656         list_add(&ib_conn->conn_list, &ig.connlist);
 657         mutex_unlock(&ig.connlist_mutex);
 658         return 0;
 659
 660 id_failure:
 661         ib_conn->cma_id = NULL;
 662 addr_failure:
 663         ib_conn->state = ISER_CONN_DOWN;
 664         iser_conn_put(ib_conn, 1); /* deref ib conn's cma id */
 665 connect_failure:
 666         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 667         return err;
 668 }
 669
 670 /**
 671  * iser_reg_page_vec - Register physical memory
 672  *
 673  * returns: 0 on success, errno code on failure
 674  */
 675 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 676                       struct iser_page_vec *page_vec,
 677                       struct iser_mem_reg  *mem_reg)
 678 {
 679         struct ib_pool_fmr *mem;
 680         u64                io_addr;
 681         u64                *page_list;
 682         int                status;
 683
 684         page_list = page_vec->pages;
 685         io_addr   = page_list[0];
 686
 687         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 688                                     page_list,
 689                                     page_vec->length,
 690                                     io_addr);
 691
 692         if (IS_ERR(mem)) {
 693                 status = (int)PTR_ERR(mem);
 694                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 695                 return status;
 696         }
 697
 698         mem_reg->lkey  = mem->fmr->lkey;
 699         mem_reg->rkey  = mem->fmr->rkey;
 700         mem_reg->len   = page_vec->length * SIZE_4K;
 701         mem_reg->va    = io_addr;
 702         mem_reg->is_fmr = 1;
 703         mem_reg->mem_h = (void *)mem;
 704
 705         mem_reg->va   += page_vec->offset;
 706         mem_reg->len   = page_vec->data_size;
 707
 708         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 709                  "entry[0]: (0x%08lx,%ld)] -> "
 710                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 711                  page_vec, page_vec->length,
 712                  (unsigned long)page_vec->pages[0],
 713                  (unsigned long)page_vec->data_size,
 714                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 715                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 716         return 0;
 717 }
 718
 719 /**
 720  * Unregister (previosuly registered) memory.
 721  */
 722 void iser_unreg_mem(struct iser_mem_reg *reg)
 723 {
 724         int ret;
 725
 726         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 727
 728         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 729         if (ret)
 730                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 731
 732         reg->mem_h = NULL;
 733 }
 734
 735 int iser_post_recvl(struct iser_conn *ib_conn)
 736 {
 737         struct ib_recv_wr rx_wr, *rx_wr_failed;
 738         struct ib_sge     sge;
 739         int ib_ret;
 740
 741         sge.addr   = ib_conn->login_resp_dma;
 742         sge.length = ISER_RX_LOGIN_SIZE;
 743         sge.lkey   = ib_conn->device->mr->lkey;
 744
 745         rx_wr.wr_id   = (unsigned long)ib_conn->login_resp_buf;
 746         rx_wr.sg_list = &sge;
 747         rx_wr.num_sge = 1;
 748         rx_wr.next    = NULL;
 749
 750         ib_conn->post_recv_buf_count++;
 751         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 752         if (ib_ret) {
 753                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 754                 ib_conn->post_recv_buf_count--;
 755         }
 756         return ib_ret;
 757 }
 758
 759 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 760 {
 761         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 762         int i, ib_ret;
 763         unsigned int my_rx_head = ib_conn->rx_desc_head;
 764         struct iser_rx_desc *rx_desc;
 765
 766         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 767                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 768                 rx_wr->wr_id    = (unsigned long)rx_desc;
 769                 rx_wr->sg_list  = &rx_desc->rx_sg;
 770                 rx_wr->num_sge  = 1;
 771                 rx_wr->next     = rx_wr + 1;
 772                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 773         }
 774
 775         rx_wr--;
 776         rx_wr->next = NULL; /* mark end of work requests list */
 777
 778         ib_conn->post_recv_buf_count += count;
 779         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 780         if (ib_ret) {
 781                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 782                 ib_conn->post_recv_buf_count -= count;
 783         } else
 784                 ib_conn->rx_desc_head = my_rx_head;
 785         return ib_ret;
 786 }
 787
 788
 789 /**
 790  * iser_start_send - Initiate a Send DTO operation
 791  *
 792  * returns 0 on success, -1 on failure
 793  */
 794 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 795 {
 796         int               ib_ret;
 797         struct ib_send_wr send_wr, *send_wr_failed;
 798
 799         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 800                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 801
 802         send_wr.next       = NULL;
 803         send_wr.wr_id      = (unsigned long)tx_desc;
 804         send_wr.sg_list    = tx_desc->tx_sg;
 805         send_wr.num_sge    = tx_desc->num_sge;
 806         send_wr.opcode     = IB_WR_SEND;
 807         send_wr.send_flags = IB_SEND_SIGNALED;
 808
 809         atomic_inc(&ib_conn->post_send_buf_count);
 810
 811         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 812         if (ib_ret) {
 813                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 814                 atomic_dec(&ib_conn->post_send_buf_count);
 815         }
 816         return ib_ret;
 817 }
 818
 819 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 820                                 struct iser_conn *ib_conn)
 821 {
 822         if (desc && desc->type == ISCSI_TX_DATAOUT)
 823                 kmem_cache_free(ig.desc_cache, desc);
 824
 825         if (ib_conn->post_recv_buf_count == 0 &&
 826             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 827                 /* getting here when the state is UP means that the conn is *
 828                  * being terminated asynchronously from the iSCSI layer's   *
 829                  * perspective.                                             */
 830                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 831                     ISER_CONN_TERMINATING))
 832                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 833                                            ISCSI_ERR_CONN_FAILED);
 834
 835                 /* no more non completed posts to the QP, complete the
 836                  * termination process w.o worrying on disconnect event */
 837                 ib_conn->state = ISER_CONN_DOWN;
 838                 wake_up_interruptible(&ib_conn->wait);
 839         }
 840 }
 841
 842 static int iser_drain_tx_cq(struct iser_device  *device, int cq_index)
 843 {
 844         struct ib_cq  *cq = device->tx_cq[cq_index];
 845         struct ib_wc  wc;
 846         struct iser_tx_desc *tx_desc;
 847         struct iser_conn *ib_conn;
 848         int completed_tx = 0;
 849
 850         while (ib_poll_cq(cq, 1, &wc) == 1) {
 851                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 852                 ib_conn = wc.qp->qp_context;
 853                 if (wc.status == IB_WC_SUCCESS) {
 854                         if (wc.opcode == IB_WC_SEND)
 855                                 iser_snd_completion(tx_desc, ib_conn);
 856                         else
 857                                 iser_err("expected opcode %d got %d\n",
 858                                         IB_WC_SEND, wc.opcode);
 859                 } else {
 860                         iser_err("tx id %llx status %d vend_err %x\n",
 861                                 wc.wr_id, wc.status, wc.vendor_err);
 862                         atomic_dec(&ib_conn->post_send_buf_count);
 863                         iser_handle_comp_error(tx_desc, ib_conn);
 864                 }
 865                 completed_tx++;
 866         }
 867         return completed_tx;
 868 }
 869
 870
 871 static void iser_cq_tasklet_fn(unsigned long data)
 872 {
 873         struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
 874         struct iser_device  *device = cq_desc->device;
 875         int cq_index = cq_desc->cq_index;
 876         struct ib_cq         *cq = device->rx_cq[cq_index];
 877          struct ib_wc        wc;
 878          struct iser_rx_desc *desc;
 879          unsigned long       xfer_len;
 880         struct iser_conn *ib_conn;
 881         int completed_tx, completed_rx;
 882         completed_tx = completed_rx = 0;
 883
 884         while (ib_poll_cq(cq, 1, &wc) == 1) {
 885                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 886                 BUG_ON(desc == NULL);
 887                 ib_conn = wc.qp->qp_context;
 888                 if (wc.status == IB_WC_SUCCESS) {
 889                         if (wc.opcode == IB_WC_RECV) {
 890                                 xfer_len = (unsigned long)wc.byte_len;
 891                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 892                         } else
 893                                 iser_err("expected opcode %d got %d\n",
 894                                         IB_WC_RECV, wc.opcode);
 895                 } else {
 896                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 897                                 iser_err("rx id %llx status %d vend_err %x\n",
 898                                         wc.wr_id, wc.status, wc.vendor_err);
 899                         ib_conn->post_recv_buf_count--;
 900                         iser_handle_comp_error(NULL, ib_conn);
 901                 }
 902                 completed_rx++;
 903                 if (!(completed_rx & 63))
 904                         completed_tx += iser_drain_tx_cq(device, cq_index);
 905         }
 906         /* #warning "it is assumed here that arming CQ only once its empty" *
 907          * " would not cause interrupts to be missed"                       */
 908         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 909
 910         completed_tx += iser_drain_tx_cq(device, cq_index);
 911         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 912 }
 913
 914 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 915 {
 916         struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
 917         struct iser_device  *device = cq_desc->device;
 918         int cq_index = cq_desc->cq_index;
 919
 920         tasklet_schedule(&device->cq_tasklet[cq_index]);
 921 }