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