drivers/infiniband/ulp/iser/iser_verbs.c

   1 /*
   2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/module.h>
  35 #include <linux/delay.h>
  36
  37 #include "iscsi_iser.h"
  38
  39 #define ISCSI_ISER_MAX_CONN     8
  40 #define ISER_MAX_RX_CQ_LEN      (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
  41 #define ISER_MAX_TX_CQ_LEN      (ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
  42
  43 static void iser_cq_tasklet_fn(unsigned long data);
  44 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
  45
  46 static void iser_cq_event_callback(struct ib_event *cause, void *context)
  47 {
  48         iser_err("got cq event %d \n", cause->event);
  49 }
  50
  51 static void iser_qp_event_callback(struct ib_event *cause, void *context)
  52 {
  53         iser_err("got qp event %d\n",cause->event);
  54 }
  55
  56 /**
  57  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  58  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  59  * the adapator.
  60  *
  61  * returns 0 on success, -1 on failure
  62  */
  63 static int iser_create_device_ib_res(struct iser_device *device)
  64 {
  65         device->pd = ib_alloc_pd(device->ib_device);
  66         if (IS_ERR(device->pd))
  67                 goto pd_err;
  68
  69         device->rx_cq = ib_create_cq(device->ib_device,
  70                                   iser_cq_callback,
  71                                   iser_cq_event_callback,
  72                                   (void *)device,
  73                                   ISER_MAX_RX_CQ_LEN, 0);
  74         if (IS_ERR(device->rx_cq))
  75                 goto rx_cq_err;
  76
  77         device->tx_cq = ib_create_cq(device->ib_device,
  78                                   NULL, iser_cq_event_callback,
  79                                   (void *)device,
  80                                   ISER_MAX_TX_CQ_LEN, 0);
  81
  82         if (IS_ERR(device->tx_cq))
  83                 goto tx_cq_err;
  84
  85         if (ib_req_notify_cq(device->rx_cq, IB_CQ_NEXT_COMP))
  86                 goto cq_arm_err;
  87
  88         tasklet_init(&device->cq_tasklet,
  89                      iser_cq_tasklet_fn,
  90                      (unsigned long)device);
  91
  92         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
  93                                    IB_ACCESS_REMOTE_WRITE |
  94                                    IB_ACCESS_REMOTE_READ);
  95         if (IS_ERR(device->mr))
  96                 goto dma_mr_err;
  97
  98         return 0;
  99
 100 dma_mr_err:
 101         tasklet_kill(&device->cq_tasklet);
 102 cq_arm_err:
 103         ib_destroy_cq(device->tx_cq);
 104 tx_cq_err:
 105         ib_destroy_cq(device->rx_cq);
 106 rx_cq_err:
 107         ib_dealloc_pd(device->pd);
 108 pd_err:
 109         iser_err("failed to allocate an IB resource\n");
 110         return -1;
 111 }
 112
 113 /**
 114  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 115  * CQ and PD created with the device associated with the adapator.
 116  */
 117 static void iser_free_device_ib_res(struct iser_device *device)
 118 {
 119         BUG_ON(device->mr == NULL);
 120
 121         tasklet_kill(&device->cq_tasklet);
 122
 123         (void)ib_dereg_mr(device->mr);
 124         (void)ib_destroy_cq(device->tx_cq);
 125         (void)ib_destroy_cq(device->rx_cq);
 126         (void)ib_dealloc_pd(device->pd);
 127
 128         device->mr = NULL;
 129         device->tx_cq = NULL;
 130         device->rx_cq = NULL;
 131         device->pd = NULL;
 132 }
 133
 134 /**
 135  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 136  *
 137  * returns 0 on success, -1 on failure
 138  */
 139 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 140 {
 141         struct iser_device      *device;
 142         struct ib_qp_init_attr  init_attr;
 143         int                     ret = -ENOMEM;
 144         struct ib_fmr_pool_param params;
 145
 146         BUG_ON(ib_conn->device == NULL);
 147
 148         device = ib_conn->device;
 149
 150         ib_conn->login_buf = kmalloc(ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 151         if (!ib_conn->login_buf) {
 152                 goto alloc_err;
 153                 ret = -ENOMEM;
 154         }
 155
 156         ib_conn->login_dma = ib_dma_map_single(ib_conn->device->ib_device,
 157                                 (void *)ib_conn->login_buf, ISER_RX_LOGIN_SIZE,
 158                                 DMA_FROM_DEVICE);
 159
 160         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 161                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 162                                     GFP_KERNEL);
 163         if (!ib_conn->page_vec) {
 164                 ret = -ENOMEM;
 165                 goto alloc_err;
 166         }
 167         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 168
 169         params.page_shift        = SHIFT_4K;
 170         /* when the first/last SG element are not start/end *
 171          * page aligned, the map whould be of N+1 pages     */
 172         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 173         /* make the pool size twice the max number of SCSI commands *
 174          * the ML is expected to queue, watermark for unmap at 50%  */
 175         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 176         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 177         params.cache             = 0;
 178         params.flush_function    = NULL;
 179         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 180                                     IB_ACCESS_REMOTE_WRITE |
 181                                     IB_ACCESS_REMOTE_READ);
 182
 183         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 184         if (IS_ERR(ib_conn->fmr_pool)) {
 185                 ret = PTR_ERR(ib_conn->fmr_pool);
 186                 goto fmr_pool_err;
 187         }
 188
 189         memset(&init_attr, 0, sizeof init_attr);
 190
 191         init_attr.event_handler = iser_qp_event_callback;
 192         init_attr.qp_context    = (void *)ib_conn;
 193         init_attr.send_cq       = device->tx_cq;
 194         init_attr.recv_cq       = device->rx_cq;
 195         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 196         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 197         init_attr.cap.max_send_sge = 2;
 198         init_attr.cap.max_recv_sge = 1;
 199         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 200         init_attr.qp_type       = IB_QPT_RC;
 201
 202         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 203         if (ret)
 204                 goto qp_err;
 205
 206         ib_conn->qp = ib_conn->cma_id->qp;
 207         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 208                  ib_conn, ib_conn->cma_id,
 209                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 210         return ret;
 211
 212 qp_err:
 213         (void)ib_destroy_fmr_pool(ib_conn->fmr_pool);
 214 fmr_pool_err:
 215         kfree(ib_conn->page_vec);
 216         kfree(ib_conn->login_buf);
 217 alloc_err:
 218         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 219         return ret;
 220 }
 221
 222 /**
 223  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 224  * -1 on failure
 225  */
 226 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
 227 {
 228         BUG_ON(ib_conn == NULL);
 229
 230         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 231                  ib_conn, ib_conn->cma_id,
 232                  ib_conn->fmr_pool, ib_conn->qp);
 233
 234         /* qp is created only once both addr & route are resolved */
 235         if (ib_conn->fmr_pool != NULL)
 236                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 237
 238         if (ib_conn->qp != NULL)
 239                 rdma_destroy_qp(ib_conn->cma_id);
 240
 241         if (ib_conn->cma_id != NULL)
 242                 rdma_destroy_id(ib_conn->cma_id);
 243
 244         ib_conn->fmr_pool = NULL;
 245         ib_conn->qp       = NULL;
 246         ib_conn->cma_id   = NULL;
 247         kfree(ib_conn->page_vec);
 248
 249         return 0;
 250 }
 251
 252 /**
 253  * based on the resolved device node GUID see if there already allocated
 254  * device for this device. If there's no such, create one.
 255  */
 256 static
 257 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 258 {
 259         struct iser_device *device;
 260
 261         mutex_lock(&ig.device_list_mutex);
 262
 263         list_for_each_entry(device, &ig.device_list, ig_list)
 264                 /* find if there's a match using the node GUID */
 265                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 266                         goto inc_refcnt;
 267
 268         device = kzalloc(sizeof *device, GFP_KERNEL);
 269         if (device == NULL)
 270                 goto out;
 271
 272         /* assign this device to the device */
 273         device->ib_device = cma_id->device;
 274         /* init the device and link it into ig device list */
 275         if (iser_create_device_ib_res(device)) {
 276                 kfree(device);
 277                 device = NULL;
 278                 goto out;
 279         }
 280         list_add(&device->ig_list, &ig.device_list);
 281
 282 inc_refcnt:
 283         device->refcount++;
 284 out:
 285         mutex_unlock(&ig.device_list_mutex);
 286         return device;
 287 }
 288
 289 /* if there's no demand for this device, release it */
 290 static void iser_device_try_release(struct iser_device *device)
 291 {
 292         mutex_lock(&ig.device_list_mutex);
 293         device->refcount--;
 294         iser_err("device %p refcount %d\n",device,device->refcount);
 295         if (!device->refcount) {
 296                 iser_free_device_ib_res(device);
 297                 list_del(&device->ig_list);
 298                 kfree(device);
 299         }
 300         mutex_unlock(&ig.device_list_mutex);
 301 }
 302
 303 int iser_conn_state_comp(struct iser_conn *ib_conn,
 304                         enum iser_ib_conn_state comp)
 305 {
 306         int ret;
 307
 308         spin_lock_bh(&ib_conn->lock);
 309         ret = (ib_conn->state == comp);
 310         spin_unlock_bh(&ib_conn->lock);
 311         return ret;
 312 }
 313
 314 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 315                                      enum iser_ib_conn_state comp,
 316                                      enum iser_ib_conn_state exch)
 317 {
 318         int ret;
 319
 320         spin_lock_bh(&ib_conn->lock);
 321         if ((ret = (ib_conn->state == comp)))
 322                 ib_conn->state = exch;
 323         spin_unlock_bh(&ib_conn->lock);
 324         return ret;
 325 }
 326
 327 /**
 328  * Frees all conn objects and deallocs conn descriptor
 329  */
 330 static void iser_conn_release(struct iser_conn *ib_conn)
 331 {
 332         struct iser_device  *device = ib_conn->device;
 333
 334         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 335
 336         mutex_lock(&ig.connlist_mutex);
 337         list_del(&ib_conn->conn_list);
 338         mutex_unlock(&ig.connlist_mutex);
 339         iser_free_rx_descriptors(ib_conn);
 340         iser_free_ib_conn_res(ib_conn);
 341         ib_conn->device = NULL;
 342         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 343         if (device != NULL)
 344                 iser_device_try_release(device);
 345         if (ib_conn->iser_conn)
 346                 ib_conn->iser_conn->ib_conn = NULL;
 347         iscsi_destroy_endpoint(ib_conn->ep);
 348 }
 349
 350 void iser_conn_get(struct iser_conn *ib_conn)
 351 {
 352         atomic_inc(&ib_conn->refcount);
 353 }
 354
 355 void iser_conn_put(struct iser_conn *ib_conn)
 356 {
 357         if (atomic_dec_and_test(&ib_conn->refcount))
 358                 iser_conn_release(ib_conn);
 359 }
 360
 361 /**
 362  * triggers start of the disconnect procedures and wait for them to be done
 363  */
 364 void iser_conn_terminate(struct iser_conn *ib_conn)
 365 {
 366         int err = 0;
 367
 368         /* change the ib conn state only if the conn is UP, however always call
 369          * rdma_disconnect since this is the only way to cause the CMA to change
 370          * the QP state to ERROR
 371          */
 372
 373         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 374         err = rdma_disconnect(ib_conn->cma_id);
 375         if (err)
 376                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 377                          ib_conn,err);
 378
 379         wait_event_interruptible(ib_conn->wait,
 380                                  ib_conn->state == ISER_CONN_DOWN);
 381
 382         iser_conn_put(ib_conn);
 383 }
 384
 385 static void iser_connect_error(struct rdma_cm_id *cma_id)
 386 {
 387         struct iser_conn *ib_conn;
 388         ib_conn = (struct iser_conn *)cma_id->context;
 389
 390         ib_conn->state = ISER_CONN_DOWN;
 391         wake_up_interruptible(&ib_conn->wait);
 392 }
 393
 394 static void 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                 iser_connect_error(cma_id);
 404                 return;
 405         }
 406
 407         ib_conn = (struct iser_conn *)cma_id->context;
 408         ib_conn->device = device;
 409
 410         ret = rdma_resolve_route(cma_id, 1000);
 411         if (ret) {
 412                 iser_err("resolve route failed: %d\n", ret);
 413                 iser_connect_error(cma_id);
 414         }
 415 }
 416
 417 static void iser_route_handler(struct rdma_cm_id *cma_id)
 418 {
 419         struct rdma_conn_param conn_param;
 420         int    ret;
 421
 422         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 423         if (ret)
 424                 goto failure;
 425
 426         memset(&conn_param, 0, sizeof conn_param);
 427         conn_param.responder_resources = 4;
 428         conn_param.initiator_depth     = 1;
 429         conn_param.retry_count         = 7;
 430         conn_param.rnr_retry_count     = 6;
 431
 432         ret = rdma_connect(cma_id, &conn_param);
 433         if (ret) {
 434                 iser_err("failure connecting: %d\n", ret);
 435                 goto failure;
 436         }
 437
 438         return;
 439 failure:
 440         iser_connect_error(cma_id);
 441 }
 442
 443 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 444 {
 445         struct iser_conn *ib_conn;
 446
 447         ib_conn = (struct iser_conn *)cma_id->context;
 448         ib_conn->state = ISER_CONN_UP;
 449         wake_up_interruptible(&ib_conn->wait);
 450 }
 451
 452 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 453 {
 454         struct iser_conn *ib_conn;
 455
 456         ib_conn = (struct iser_conn *)cma_id->context;
 457         ib_conn->disc_evt_flag = 1;
 458
 459         /* getting here when the state is UP means that the conn is being *
 460          * terminated asynchronously from the iSCSI layer's perspective.  */
 461         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 462                                       ISER_CONN_TERMINATING))
 463                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 464                                    ISCSI_ERR_CONN_FAILED);
 465
 466         /* Complete the termination process if no posts are pending */
 467         if (ib_conn->post_recv_buf_count == 0 &&
 468             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 469                 ib_conn->state = ISER_CONN_DOWN;
 470                 wake_up_interruptible(&ib_conn->wait);
 471         }
 472 }
 473
 474 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 475 {
 476         int ret = 0;
 477
 478         iser_err("event %d conn %p id %p\n",event->event,cma_id->context,cma_id);
 479
 480         switch (event->event) {
 481         case RDMA_CM_EVENT_ADDR_RESOLVED:
 482                 iser_addr_handler(cma_id);
 483                 break;
 484         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 485                 iser_route_handler(cma_id);
 486                 break;
 487         case RDMA_CM_EVENT_ESTABLISHED:
 488                 iser_connected_handler(cma_id);
 489                 break;
 490         case RDMA_CM_EVENT_ADDR_ERROR:
 491         case RDMA_CM_EVENT_ROUTE_ERROR:
 492         case RDMA_CM_EVENT_CONNECT_ERROR:
 493         case RDMA_CM_EVENT_UNREACHABLE:
 494         case RDMA_CM_EVENT_REJECTED:
 495                 iser_err("event: %d, error: %d\n", event->event, event->status);
 496                 iser_connect_error(cma_id);
 497                 break;
 498         case RDMA_CM_EVENT_DISCONNECTED:
 499         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 500         case RDMA_CM_EVENT_ADDR_CHANGE:
 501                 iser_disconnected_handler(cma_id);
 502                 break;
 503         default:
 504                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 505                 break;
 506         }
 507         return ret;
 508 }
 509
 510 void iser_conn_init(struct iser_conn *ib_conn)
 511 {
 512         ib_conn->state = ISER_CONN_INIT;
 513         init_waitqueue_head(&ib_conn->wait);
 514         ib_conn->post_recv_buf_count = 0;
 515         atomic_set(&ib_conn->post_send_buf_count, 0);
 516         atomic_set(&ib_conn->refcount, 1);
 517         INIT_LIST_HEAD(&ib_conn->conn_list);
 518         spin_lock_init(&ib_conn->lock);
 519 }
 520
 521  /**
 522  * starts the process of connecting to the target
 523  * sleeps until the connection is established or rejected
 524  */
 525 int iser_connect(struct iser_conn   *ib_conn,
 526                  struct sockaddr_in *src_addr,
 527                  struct sockaddr_in *dst_addr,
 528                  int                 non_blocking)
 529 {
 530         struct sockaddr *src, *dst;
 531         int err = 0;
 532
 533         sprintf(ib_conn->name, "%pI4:%d",
 534                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 535
 536         /* the device is known only --after-- address resolution */
 537         ib_conn->device = NULL;
 538
 539         iser_err("connecting to: %pI4, port 0x%x\n",
 540                  &dst_addr->sin_addr, dst_addr->sin_port);
 541
 542         ib_conn->state = ISER_CONN_PENDING;
 543
 544         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 545                                              (void *)ib_conn,
 546                                              RDMA_PS_TCP);
 547         if (IS_ERR(ib_conn->cma_id)) {
 548                 err = PTR_ERR(ib_conn->cma_id);
 549                 iser_err("rdma_create_id failed: %d\n", err);
 550                 goto id_failure;
 551         }
 552
 553         src = (struct sockaddr *)src_addr;
 554         dst = (struct sockaddr *)dst_addr;
 555         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 556         if (err) {
 557                 iser_err("rdma_resolve_addr failed: %d\n", err);
 558                 goto addr_failure;
 559         }
 560
 561         if (!non_blocking) {
 562                 wait_event_interruptible(ib_conn->wait,
 563                                          (ib_conn->state != ISER_CONN_PENDING));
 564
 565                 if (ib_conn->state != ISER_CONN_UP) {
 566                         err =  -EIO;
 567                         goto connect_failure;
 568                 }
 569         }
 570
 571         mutex_lock(&ig.connlist_mutex);
 572         list_add(&ib_conn->conn_list, &ig.connlist);
 573         mutex_unlock(&ig.connlist_mutex);
 574         return 0;
 575
 576 id_failure:
 577         ib_conn->cma_id = NULL;
 578 addr_failure:
 579         ib_conn->state = ISER_CONN_DOWN;
 580 connect_failure:
 581         iser_conn_release(ib_conn);
 582         return err;
 583 }
 584
 585 /**
 586  * iser_reg_page_vec - Register physical memory
 587  *
 588  * returns: 0 on success, errno code on failure
 589  */
 590 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 591                       struct iser_page_vec *page_vec,
 592                       struct iser_mem_reg  *mem_reg)
 593 {
 594         struct ib_pool_fmr *mem;
 595         u64                io_addr;
 596         u64                *page_list;
 597         int                status;
 598
 599         page_list = page_vec->pages;
 600         io_addr   = page_list[0];
 601
 602         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 603                                     page_list,
 604                                     page_vec->length,
 605                                     io_addr);
 606
 607         if (IS_ERR(mem)) {
 608                 status = (int)PTR_ERR(mem);
 609                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 610                 return status;
 611         }
 612
 613         mem_reg->lkey  = mem->fmr->lkey;
 614         mem_reg->rkey  = mem->fmr->rkey;
 615         mem_reg->len   = page_vec->length * SIZE_4K;
 616         mem_reg->va    = io_addr;
 617         mem_reg->is_fmr = 1;
 618         mem_reg->mem_h = (void *)mem;
 619
 620         mem_reg->va   += page_vec->offset;
 621         mem_reg->len   = page_vec->data_size;
 622
 623         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 624                  "entry[0]: (0x%08lx,%ld)] -> "
 625                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 626                  page_vec, page_vec->length,
 627                  (unsigned long)page_vec->pages[0],
 628                  (unsigned long)page_vec->data_size,
 629                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 630                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 631         return 0;
 632 }
 633
 634 /**
 635  * Unregister (previosuly registered) memory.
 636  */
 637 void iser_unreg_mem(struct iser_mem_reg *reg)
 638 {
 639         int ret;
 640
 641         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 642
 643         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 644         if (ret)
 645                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 646
 647         reg->mem_h = NULL;
 648 }
 649
 650 int iser_post_recvl(struct iser_conn *ib_conn)
 651 {
 652         struct ib_recv_wr rx_wr, *rx_wr_failed;
 653         struct ib_sge     sge;
 654         int ib_ret;
 655
 656         sge.addr   = ib_conn->login_dma;
 657         sge.length = ISER_RX_LOGIN_SIZE;
 658         sge.lkey   = ib_conn->device->mr->lkey;
 659
 660         rx_wr.wr_id   = (unsigned long)ib_conn->login_buf;
 661         rx_wr.sg_list = &sge;
 662         rx_wr.num_sge = 1;
 663         rx_wr.next    = NULL;
 664
 665         ib_conn->post_recv_buf_count++;
 666         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 667         if (ib_ret) {
 668                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 669                 ib_conn->post_recv_buf_count--;
 670         }
 671         return ib_ret;
 672 }
 673
 674 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 675 {
 676         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 677         int i, ib_ret;
 678         unsigned int my_rx_head = ib_conn->rx_desc_head;
 679         struct iser_rx_desc *rx_desc;
 680
 681         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 682                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 683                 rx_wr->wr_id    = (unsigned long)rx_desc;
 684                 rx_wr->sg_list  = &rx_desc->rx_sg;
 685                 rx_wr->num_sge  = 1;
 686                 rx_wr->next     = rx_wr + 1;
 687                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 688         }
 689
 690         rx_wr--;
 691         rx_wr->next = NULL; /* mark end of work requests list */
 692
 693         ib_conn->post_recv_buf_count += count;
 694         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 695         if (ib_ret) {
 696                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 697                 ib_conn->post_recv_buf_count -= count;
 698         } else
 699                 ib_conn->rx_desc_head = my_rx_head;
 700         return ib_ret;
 701 }
 702
 703
 704 /**
 705  * iser_start_send - Initiate a Send DTO operation
 706  *
 707  * returns 0 on success, -1 on failure
 708  */
 709 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 710 {
 711         int               ib_ret;
 712         struct ib_send_wr send_wr, *send_wr_failed;
 713
 714         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 715                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 716
 717         send_wr.next       = NULL;
 718         send_wr.wr_id      = (unsigned long)tx_desc;
 719         send_wr.sg_list    = tx_desc->tx_sg;
 720         send_wr.num_sge    = tx_desc->num_sge;
 721         send_wr.opcode     = IB_WR_SEND;
 722         send_wr.send_flags = IB_SEND_SIGNALED;
 723
 724         atomic_inc(&ib_conn->post_send_buf_count);
 725
 726         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 727         if (ib_ret) {
 728                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 729                 atomic_dec(&ib_conn->post_send_buf_count);
 730         }
 731         return ib_ret;
 732 }
 733
 734 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 735                                 struct iser_conn *ib_conn)
 736 {
 737         if (desc && desc->type == ISCSI_TX_DATAOUT)
 738                 kmem_cache_free(ig.desc_cache, desc);
 739
 740         if (ib_conn->post_recv_buf_count == 0 &&
 741             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 742                 /* getting here when the state is UP means that the conn is *
 743                  * being terminated asynchronously from the iSCSI layer's   *
 744                  * perspective.                                             */
 745                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 746                     ISER_CONN_TERMINATING))
 747                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 748                                            ISCSI_ERR_CONN_FAILED);
 749
 750                 /* complete the termination process if disconnect event was delivered *
 751                  * note there are no more non completed posts to the QP               */
 752                 if (ib_conn->disc_evt_flag) {
 753                         ib_conn->state = ISER_CONN_DOWN;
 754                         wake_up_interruptible(&ib_conn->wait);
 755                 }
 756         }
 757 }
 758
 759 static int iser_drain_tx_cq(struct iser_device  *device)
 760 {
 761         struct ib_cq  *cq = device->tx_cq;
 762         struct ib_wc  wc;
 763         struct iser_tx_desc *tx_desc;
 764         struct iser_conn *ib_conn;
 765         int completed_tx = 0;
 766
 767         while (ib_poll_cq(cq, 1, &wc) == 1) {
 768                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 769                 ib_conn = wc.qp->qp_context;
 770                 if (wc.status == IB_WC_SUCCESS) {
 771                         if (wc.opcode == IB_WC_SEND)
 772                                 iser_snd_completion(tx_desc, ib_conn);
 773                         else
 774                                 iser_err("expected opcode %d got %d\n",
 775                                         IB_WC_SEND, wc.opcode);
 776                 } else {
 777                         iser_err("tx id %llx status %d vend_err %x\n",
 778                                 wc.wr_id, wc.status, wc.vendor_err);
 779                         atomic_dec(&ib_conn->post_send_buf_count);
 780                         iser_handle_comp_error(tx_desc, ib_conn);
 781                 }
 782                 completed_tx++;
 783         }
 784         return completed_tx;
 785 }
 786
 787
 788 static void iser_cq_tasklet_fn(unsigned long data)
 789 {
 790          struct iser_device  *device = (struct iser_device *)data;
 791          struct ib_cq        *cq = device->rx_cq;
 792          struct ib_wc        wc;
 793          struct iser_rx_desc *desc;
 794          unsigned long       xfer_len;
 795         struct iser_conn *ib_conn;
 796         int completed_tx, completed_rx;
 797         completed_tx = completed_rx = 0;
 798
 799         while (ib_poll_cq(cq, 1, &wc) == 1) {
 800                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 801                 BUG_ON(desc == NULL);
 802                 ib_conn = wc.qp->qp_context;
 803                 if (wc.status == IB_WC_SUCCESS) {
 804                         if (wc.opcode == IB_WC_RECV) {
 805                                 xfer_len = (unsigned long)wc.byte_len;
 806                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 807                         } else
 808                                 iser_err("expected opcode %d got %d\n",
 809                                         IB_WC_RECV, wc.opcode);
 810                 } else {
 811                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 812                                 iser_err("rx id %llx status %d vend_err %x\n",
 813                                         wc.wr_id, wc.status, wc.vendor_err);
 814                         ib_conn->post_recv_buf_count--;
 815                         iser_handle_comp_error(NULL, ib_conn);
 816                 }
 817                 completed_rx++;
 818                 if (!(completed_rx & 63))
 819                         completed_tx += iser_drain_tx_cq(device);
 820         }
 821         /* #warning "it is assumed here that arming CQ only once its empty" *
 822          * " would not cause interrupts to be missed"                       */
 823         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 824
 825         completed_tx += iser_drain_tx_cq(device);
 826         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 827 }
 828
 829 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 830 {
 831         struct iser_device  *device = (struct iser_device *)cq_context;
 832
 833         tasklet_schedule(&device->cq_tasklet);
 834 }