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         return 0;
 278 }
 279
 280 /**
 281  * based on the resolved device node GUID see if there already allocated
 282  * device for this device. If there's no such, create one.
 283  */
 284 static
 285 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 286 {
 287         struct iser_device *device;
 288
 289         mutex_lock(&ig.device_list_mutex);
 290
 291         list_for_each_entry(device, &ig.device_list, ig_list)
 292                 /* find if there's a match using the node GUID */
 293                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 294                         goto inc_refcnt;
 295
 296         device = kzalloc(sizeof *device, GFP_KERNEL);
 297         if (device == NULL)
 298                 goto out;
 299
 300         /* assign this device to the device */
 301         device->ib_device = cma_id->device;
 302         /* init the device and link it into ig device list */
 303         if (iser_create_device_ib_res(device)) {
 304                 kfree(device);
 305                 device = NULL;
 306                 goto out;
 307         }
 308         list_add(&device->ig_list, &ig.device_list);
 309
 310 inc_refcnt:
 311         device->refcount++;
 312 out:
 313         mutex_unlock(&ig.device_list_mutex);
 314         return device;
 315 }
 316
 317 /* if there's no demand for this device, release it */
 318 static void iser_device_try_release(struct iser_device *device)
 319 {
 320         mutex_lock(&ig.device_list_mutex);
 321         device->refcount--;
 322         iser_err("device %p refcount %d\n",device,device->refcount);
 323         if (!device->refcount) {
 324                 iser_free_device_ib_res(device);
 325                 list_del(&device->ig_list);
 326                 kfree(device);
 327         }
 328         mutex_unlock(&ig.device_list_mutex);
 329 }
 330
 331 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 332                                      enum iser_ib_conn_state comp,
 333                                      enum iser_ib_conn_state exch)
 334 {
 335         int ret;
 336
 337         spin_lock_bh(&ib_conn->lock);
 338         if ((ret = (ib_conn->state == comp)))
 339                 ib_conn->state = exch;
 340         spin_unlock_bh(&ib_conn->lock);
 341         return ret;
 342 }
 343
 344 /**
 345  * Frees all conn objects and deallocs conn descriptor
 346  */
 347 static void iser_conn_release(struct iser_conn *ib_conn, int can_destroy_id)
 348 {
 349         struct iser_device  *device = ib_conn->device;
 350
 351         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 352
 353         mutex_lock(&ig.connlist_mutex);
 354         list_del(&ib_conn->conn_list);
 355         mutex_unlock(&ig.connlist_mutex);
 356         iser_free_rx_descriptors(ib_conn);
 357         iser_free_ib_conn_res(ib_conn, can_destroy_id);
 358         ib_conn->device = NULL;
 359         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 360         if (device != NULL)
 361                 iser_device_try_release(device);
 362         iscsi_destroy_endpoint(ib_conn->ep);
 363 }
 364
 365 void iser_conn_get(struct iser_conn *ib_conn)
 366 {
 367         atomic_inc(&ib_conn->refcount);
 368 }
 369
 370 int iser_conn_put(struct iser_conn *ib_conn, int can_destroy_id)
 371 {
 372         if (atomic_dec_and_test(&ib_conn->refcount)) {
 373                 iser_conn_release(ib_conn, can_destroy_id);
 374                 return 1;
 375         }
 376         return 0;
 377 }
 378
 379 /**
 380  * triggers start of the disconnect procedures and wait for them to be done
 381  */
 382 void iser_conn_terminate(struct iser_conn *ib_conn)
 383 {
 384         int err = 0;
 385
 386         /* change the ib conn state only if the conn is UP, however always call
 387          * rdma_disconnect since this is the only way to cause the CMA to change
 388          * the QP state to ERROR
 389          */
 390
 391         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 392         err = rdma_disconnect(ib_conn->cma_id);
 393         if (err)
 394                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 395                          ib_conn,err);
 396
 397         wait_event_interruptible(ib_conn->wait,
 398                                  ib_conn->state == ISER_CONN_DOWN);
 399
 400         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 401 }
 402
 403 static int iser_connect_error(struct rdma_cm_id *cma_id)
 404 {
 405         struct iser_conn *ib_conn;
 406         ib_conn = (struct iser_conn *)cma_id->context;
 407
 408         ib_conn->state = ISER_CONN_DOWN;
 409         wake_up_interruptible(&ib_conn->wait);
 410         return iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 411 }
 412
 413 static int iser_addr_handler(struct rdma_cm_id *cma_id)
 414 {
 415         struct iser_device *device;
 416         struct iser_conn   *ib_conn;
 417         int    ret;
 418
 419         device = iser_device_find_by_ib_device(cma_id);
 420         if (!device) {
 421                 iser_err("device lookup/creation failed\n");
 422                 return iser_connect_error(cma_id);
 423         }
 424
 425         ib_conn = (struct iser_conn *)cma_id->context;
 426         ib_conn->device = device;
 427
 428         ret = rdma_resolve_route(cma_id, 1000);
 429         if (ret) {
 430                 iser_err("resolve route failed: %d\n", ret);
 431                 return iser_connect_error(cma_id);
 432         }
 433
 434         return 0;
 435 }
 436
 437 static int iser_route_handler(struct rdma_cm_id *cma_id)
 438 {
 439         struct rdma_conn_param conn_param;
 440         int    ret;
 441
 442         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 443         if (ret)
 444                 goto failure;
 445
 446         memset(&conn_param, 0, sizeof conn_param);
 447         conn_param.responder_resources = 4;
 448         conn_param.initiator_depth     = 1;
 449         conn_param.retry_count         = 7;
 450         conn_param.rnr_retry_count     = 6;
 451
 452         ret = rdma_connect(cma_id, &conn_param);
 453         if (ret) {
 454                 iser_err("failure connecting: %d\n", ret);
 455                 goto failure;
 456         }
 457
 458         return 0;
 459 failure:
 460         return iser_connect_error(cma_id);
 461 }
 462
 463 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 464 {
 465         struct iser_conn *ib_conn;
 466
 467         ib_conn = (struct iser_conn *)cma_id->context;
 468         ib_conn->state = ISER_CONN_UP;
 469         wake_up_interruptible(&ib_conn->wait);
 470 }
 471
 472 static int iser_disconnected_handler(struct rdma_cm_id *cma_id)
 473 {
 474         struct iser_conn *ib_conn;
 475         int ret;
 476
 477         ib_conn = (struct iser_conn *)cma_id->context;
 478
 479         /* getting here when the state is UP means that the conn is being *
 480          * terminated asynchronously from the iSCSI layer's perspective.  */
 481         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 482                                       ISER_CONN_TERMINATING))
 483                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 484                                    ISCSI_ERR_CONN_FAILED);
 485
 486         /* Complete the termination process if no posts are pending */
 487         if (ib_conn->post_recv_buf_count == 0 &&
 488             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 489                 ib_conn->state = ISER_CONN_DOWN;
 490                 wake_up_interruptible(&ib_conn->wait);
 491         }
 492
 493         ret = iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 494         return ret;
 495 }
 496
 497 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 498 {
 499         int ret = 0;
 500
 501         iser_err("event %d status %d conn %p id %p\n",
 502                 event->event, event->status, cma_id->context, cma_id);
 503
 504         switch (event->event) {
 505         case RDMA_CM_EVENT_ADDR_RESOLVED:
 506                 ret = iser_addr_handler(cma_id);
 507                 break;
 508         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 509                 ret = iser_route_handler(cma_id);
 510                 break;
 511         case RDMA_CM_EVENT_ESTABLISHED:
 512                 iser_connected_handler(cma_id);
 513                 break;
 514         case RDMA_CM_EVENT_ADDR_ERROR:
 515         case RDMA_CM_EVENT_ROUTE_ERROR:
 516         case RDMA_CM_EVENT_CONNECT_ERROR:
 517         case RDMA_CM_EVENT_UNREACHABLE:
 518         case RDMA_CM_EVENT_REJECTED:
 519                 ret = iser_connect_error(cma_id);
 520                 break;
 521         case RDMA_CM_EVENT_DISCONNECTED:
 522         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 523         case RDMA_CM_EVENT_ADDR_CHANGE:
 524                 ret = iser_disconnected_handler(cma_id);
 525                 break;
 526         default:
 527                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 528                 break;
 529         }
 530         return ret;
 531 }
 532
 533 void iser_conn_init(struct iser_conn *ib_conn)
 534 {
 535         ib_conn->state = ISER_CONN_INIT;
 536         init_waitqueue_head(&ib_conn->wait);
 537         ib_conn->post_recv_buf_count = 0;
 538         atomic_set(&ib_conn->post_send_buf_count, 0);
 539         atomic_set(&ib_conn->refcount, 1); /* ref ib conn allocation */
 540         INIT_LIST_HEAD(&ib_conn->conn_list);
 541         spin_lock_init(&ib_conn->lock);
 542 }
 543
 544  /**
 545  * starts the process of connecting to the target
 546  * sleeps until the connection is established or rejected
 547  */
 548 int iser_connect(struct iser_conn   *ib_conn,
 549                  struct sockaddr_in *src_addr,
 550                  struct sockaddr_in *dst_addr,
 551                  int                 non_blocking)
 552 {
 553         struct sockaddr *src, *dst;
 554         int err = 0;
 555
 556         sprintf(ib_conn->name, "%pI4:%d",
 557                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 558
 559         /* the device is known only --after-- address resolution */
 560         ib_conn->device = NULL;
 561
 562         iser_err("connecting to: %pI4, port 0x%x\n",
 563                  &dst_addr->sin_addr, dst_addr->sin_port);
 564
 565         ib_conn->state = ISER_CONN_PENDING;
 566
 567         iser_conn_get(ib_conn); /* ref ib conn's cma id */
 568         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 569                                              (void *)ib_conn,
 570                                              RDMA_PS_TCP, IB_QPT_RC);
 571         if (IS_ERR(ib_conn->cma_id)) {
 572                 err = PTR_ERR(ib_conn->cma_id);
 573                 iser_err("rdma_create_id failed: %d\n", err);
 574                 goto id_failure;
 575         }
 576
 577         src = (struct sockaddr *)src_addr;
 578         dst = (struct sockaddr *)dst_addr;
 579         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 580         if (err) {
 581                 iser_err("rdma_resolve_addr failed: %d\n", err);
 582                 goto addr_failure;
 583         }
 584
 585         if (!non_blocking) {
 586                 wait_event_interruptible(ib_conn->wait,
 587                                          (ib_conn->state != ISER_CONN_PENDING));
 588
 589                 if (ib_conn->state != ISER_CONN_UP) {
 590                         err =  -EIO;
 591                         goto connect_failure;
 592                 }
 593         }
 594
 595         mutex_lock(&ig.connlist_mutex);
 596         list_add(&ib_conn->conn_list, &ig.connlist);
 597         mutex_unlock(&ig.connlist_mutex);
 598         return 0;
 599
 600 id_failure:
 601         ib_conn->cma_id = NULL;
 602 addr_failure:
 603         ib_conn->state = ISER_CONN_DOWN;
 604 connect_failure:
 605         iser_conn_release(ib_conn, 1);
 606         return err;
 607 }
 608
 609 /**
 610  * iser_reg_page_vec - Register physical memory
 611  *
 612  * returns: 0 on success, errno code on failure
 613  */
 614 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 615                       struct iser_page_vec *page_vec,
 616                       struct iser_mem_reg  *mem_reg)
 617 {
 618         struct ib_pool_fmr *mem;
 619         u64                io_addr;
 620         u64                *page_list;
 621         int                status;
 622
 623         page_list = page_vec->pages;
 624         io_addr   = page_list[0];
 625
 626         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 627                                     page_list,
 628                                     page_vec->length,
 629                                     io_addr);
 630
 631         if (IS_ERR(mem)) {
 632                 status = (int)PTR_ERR(mem);
 633                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 634                 return status;
 635         }
 636
 637         mem_reg->lkey  = mem->fmr->lkey;
 638         mem_reg->rkey  = mem->fmr->rkey;
 639         mem_reg->len   = page_vec->length * SIZE_4K;
 640         mem_reg->va    = io_addr;
 641         mem_reg->is_fmr = 1;
 642         mem_reg->mem_h = (void *)mem;
 643
 644         mem_reg->va   += page_vec->offset;
 645         mem_reg->len   = page_vec->data_size;
 646
 647         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 648                  "entry[0]: (0x%08lx,%ld)] -> "
 649                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 650                  page_vec, page_vec->length,
 651                  (unsigned long)page_vec->pages[0],
 652                  (unsigned long)page_vec->data_size,
 653                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 654                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 655         return 0;
 656 }
 657
 658 /**
 659  * Unregister (previosuly registered) memory.
 660  */
 661 void iser_unreg_mem(struct iser_mem_reg *reg)
 662 {
 663         int ret;
 664
 665         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 666
 667         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 668         if (ret)
 669                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 670
 671         reg->mem_h = NULL;
 672 }
 673
 674 int iser_post_recvl(struct iser_conn *ib_conn)
 675 {
 676         struct ib_recv_wr rx_wr, *rx_wr_failed;
 677         struct ib_sge     sge;
 678         int ib_ret;
 679
 680         sge.addr   = ib_conn->login_resp_dma;
 681         sge.length = ISER_RX_LOGIN_SIZE;
 682         sge.lkey   = ib_conn->device->mr->lkey;
 683
 684         rx_wr.wr_id   = (unsigned long)ib_conn->login_resp_buf;
 685         rx_wr.sg_list = &sge;
 686         rx_wr.num_sge = 1;
 687         rx_wr.next    = NULL;
 688
 689         ib_conn->post_recv_buf_count++;
 690         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 691         if (ib_ret) {
 692                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 693                 ib_conn->post_recv_buf_count--;
 694         }
 695         return ib_ret;
 696 }
 697
 698 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 699 {
 700         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 701         int i, ib_ret;
 702         unsigned int my_rx_head = ib_conn->rx_desc_head;
 703         struct iser_rx_desc *rx_desc;
 704
 705         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 706                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 707                 rx_wr->wr_id    = (unsigned long)rx_desc;
 708                 rx_wr->sg_list  = &rx_desc->rx_sg;
 709                 rx_wr->num_sge  = 1;
 710                 rx_wr->next     = rx_wr + 1;
 711                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 712         }
 713
 714         rx_wr--;
 715         rx_wr->next = NULL; /* mark end of work requests list */
 716
 717         ib_conn->post_recv_buf_count += count;
 718         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 719         if (ib_ret) {
 720                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 721                 ib_conn->post_recv_buf_count -= count;
 722         } else
 723                 ib_conn->rx_desc_head = my_rx_head;
 724         return ib_ret;
 725 }
 726
 727
 728 /**
 729  * iser_start_send - Initiate a Send DTO operation
 730  *
 731  * returns 0 on success, -1 on failure
 732  */
 733 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 734 {
 735         int               ib_ret;
 736         struct ib_send_wr send_wr, *send_wr_failed;
 737
 738         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 739                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 740
 741         send_wr.next       = NULL;
 742         send_wr.wr_id      = (unsigned long)tx_desc;
 743         send_wr.sg_list    = tx_desc->tx_sg;
 744         send_wr.num_sge    = tx_desc->num_sge;
 745         send_wr.opcode     = IB_WR_SEND;
 746         send_wr.send_flags = IB_SEND_SIGNALED;
 747
 748         atomic_inc(&ib_conn->post_send_buf_count);
 749
 750         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 751         if (ib_ret) {
 752                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 753                 atomic_dec(&ib_conn->post_send_buf_count);
 754         }
 755         return ib_ret;
 756 }
 757
 758 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 759                                 struct iser_conn *ib_conn)
 760 {
 761         if (desc && desc->type == ISCSI_TX_DATAOUT)
 762                 kmem_cache_free(ig.desc_cache, desc);
 763
 764         if (ib_conn->post_recv_buf_count == 0 &&
 765             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 766                 /* getting here when the state is UP means that the conn is *
 767                  * being terminated asynchronously from the iSCSI layer's   *
 768                  * perspective.                                             */
 769                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 770                     ISER_CONN_TERMINATING))
 771                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 772                                            ISCSI_ERR_CONN_FAILED);
 773
 774                 /* no more non completed posts to the QP, complete the
 775                  * termination process w.o worrying on disconnect event */
 776                 ib_conn->state = ISER_CONN_DOWN;
 777                 wake_up_interruptible(&ib_conn->wait);
 778         }
 779 }
 780
 781 static int iser_drain_tx_cq(struct iser_device  *device)
 782 {
 783         struct ib_cq  *cq = device->tx_cq;
 784         struct ib_wc  wc;
 785         struct iser_tx_desc *tx_desc;
 786         struct iser_conn *ib_conn;
 787         int completed_tx = 0;
 788
 789         while (ib_poll_cq(cq, 1, &wc) == 1) {
 790                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 791                 ib_conn = wc.qp->qp_context;
 792                 if (wc.status == IB_WC_SUCCESS) {
 793                         if (wc.opcode == IB_WC_SEND)
 794                                 iser_snd_completion(tx_desc, ib_conn);
 795                         else
 796                                 iser_err("expected opcode %d got %d\n",
 797                                         IB_WC_SEND, wc.opcode);
 798                 } else {
 799                         iser_err("tx id %llx status %d vend_err %x\n",
 800                                 wc.wr_id, wc.status, wc.vendor_err);
 801                         atomic_dec(&ib_conn->post_send_buf_count);
 802                         iser_handle_comp_error(tx_desc, ib_conn);
 803                 }
 804                 completed_tx++;
 805         }
 806         return completed_tx;
 807 }
 808
 809
 810 static void iser_cq_tasklet_fn(unsigned long data)
 811 {
 812          struct iser_device  *device = (struct iser_device *)data;
 813          struct ib_cq        *cq = device->rx_cq;
 814          struct ib_wc        wc;
 815          struct iser_rx_desc *desc;
 816          unsigned long       xfer_len;
 817         struct iser_conn *ib_conn;
 818         int completed_tx, completed_rx;
 819         completed_tx = completed_rx = 0;
 820
 821         while (ib_poll_cq(cq, 1, &wc) == 1) {
 822                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 823                 BUG_ON(desc == NULL);
 824                 ib_conn = wc.qp->qp_context;
 825                 if (wc.status == IB_WC_SUCCESS) {
 826                         if (wc.opcode == IB_WC_RECV) {
 827                                 xfer_len = (unsigned long)wc.byte_len;
 828                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 829                         } else
 830                                 iser_err("expected opcode %d got %d\n",
 831                                         IB_WC_RECV, wc.opcode);
 832                 } else {
 833                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 834                                 iser_err("rx id %llx status %d vend_err %x\n",
 835                                         wc.wr_id, wc.status, wc.vendor_err);
 836                         ib_conn->post_recv_buf_count--;
 837                         iser_handle_comp_error(NULL, ib_conn);
 838                 }
 839                 completed_rx++;
 840                 if (!(completed_rx & 63))
 841                         completed_tx += iser_drain_tx_cq(device);
 842         }
 843         /* #warning "it is assumed here that arming CQ only once its empty" *
 844          * " would not cause interrupts to be missed"                       */
 845         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 846
 847         completed_tx += iser_drain_tx_cq(device);
 848         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 849 }
 850
 851 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 852 {
 853         struct iser_device  *device = (struct iser_device *)cq_context;
 854
 855         tasklet_schedule(&device->cq_tasklet);
 856 }