SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 /**
  58  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  59  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  60  * the adapator.
  61  *
  62  * returns 0 on success, -1 on failure
  63  */
  64 static int iser_create_device_ib_res(struct iser_device *device)
  65 {
  66         device->pd = ib_alloc_pd(device->ib_device);
  67         if (IS_ERR(device->pd))
  68                 goto pd_err;
  69
  70         device->rx_cq = ib_create_cq(device->ib_device,
  71                                   iser_cq_callback,
  72                                   iser_cq_event_callback,
  73                                   (void *)device,
  74                                   ISER_MAX_RX_CQ_LEN, 0);
  75         if (IS_ERR(device->rx_cq))
  76                 goto rx_cq_err;
  77
  78         device->tx_cq = ib_create_cq(device->ib_device,
  79                                   NULL, iser_cq_event_callback,
  80                                   (void *)device,
  81                                   ISER_MAX_TX_CQ_LEN, 0);
  82
  83         if (IS_ERR(device->tx_cq))
  84                 goto tx_cq_err;
  85
  86         if (ib_req_notify_cq(device->rx_cq, IB_CQ_NEXT_COMP))
  87                 goto cq_arm_err;
  88
  89         tasklet_init(&device->cq_tasklet,
  90                      iser_cq_tasklet_fn,
  91                      (unsigned long)device);
  92
  93         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
  94                                    IB_ACCESS_REMOTE_WRITE |
  95                                    IB_ACCESS_REMOTE_READ);
  96         if (IS_ERR(device->mr))
  97                 goto dma_mr_err;
  98
  99         return 0;
 100
 101 dma_mr_err:
 102         tasklet_kill(&device->cq_tasklet);
 103 cq_arm_err:
 104         ib_destroy_cq(device->tx_cq);
 105 tx_cq_err:
 106         ib_destroy_cq(device->rx_cq);
 107 rx_cq_err:
 108         ib_dealloc_pd(device->pd);
 109 pd_err:
 110         iser_err("failed to allocate an IB resource\n");
 111         return -1;
 112 }
 113
 114 /**
 115  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 116  * CQ and PD created with the device associated with the adapator.
 117  */
 118 static void iser_free_device_ib_res(struct iser_device *device)
 119 {
 120         BUG_ON(device->mr == NULL);
 121
 122         tasklet_kill(&device->cq_tasklet);
 123
 124         (void)ib_dereg_mr(device->mr);
 125         (void)ib_destroy_cq(device->tx_cq);
 126         (void)ib_destroy_cq(device->rx_cq);
 127         (void)ib_dealloc_pd(device->pd);
 128
 129         device->mr = NULL;
 130         device->tx_cq = NULL;
 131         device->rx_cq = NULL;
 132         device->pd = NULL;
 133 }
 134
 135 /**
 136  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 137  *
 138  * returns 0 on success, -1 on failure
 139  */
 140 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 141 {
 142         struct iser_device      *device;
 143         struct ib_qp_init_attr  init_attr;
 144         int                     ret = -ENOMEM;
 145         struct ib_fmr_pool_param params;
 146
 147         BUG_ON(ib_conn->device == NULL);
 148
 149         device = ib_conn->device;
 150
 151         ib_conn->login_buf = kmalloc(ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 152         if (!ib_conn->login_buf) {
 153                 goto alloc_err;
 154                 ret = -ENOMEM;
 155         }
 156
 157         ib_conn->login_dma = ib_dma_map_single(ib_conn->device->ib_device,
 158                                 (void *)ib_conn->login_buf, ISER_RX_LOGIN_SIZE,
 159                                 DMA_FROM_DEVICE);
 160
 161         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 162                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 163                                     GFP_KERNEL);
 164         if (!ib_conn->page_vec) {
 165                 ret = -ENOMEM;
 166                 goto alloc_err;
 167         }
 168         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 169
 170         params.page_shift        = SHIFT_4K;
 171         /* when the first/last SG element are not start/end *
 172          * page aligned, the map whould be of N+1 pages     */
 173         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 174         /* make the pool size twice the max number of SCSI commands *
 175          * the ML is expected to queue, watermark for unmap at 50%  */
 176         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 177         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 178         params.cache             = 0;
 179         params.flush_function    = NULL;
 180         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 181                                     IB_ACCESS_REMOTE_WRITE |
 182                                     IB_ACCESS_REMOTE_READ);
 183
 184         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 185         if (IS_ERR(ib_conn->fmr_pool)) {
 186                 ret = PTR_ERR(ib_conn->fmr_pool);
 187                 goto fmr_pool_err;
 188         }
 189
 190         memset(&init_attr, 0, sizeof init_attr);
 191
 192         init_attr.event_handler = iser_qp_event_callback;
 193         init_attr.qp_context    = (void *)ib_conn;
 194         init_attr.send_cq       = device->tx_cq;
 195         init_attr.recv_cq       = device->rx_cq;
 196         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 197         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 198         init_attr.cap.max_send_sge = 2;
 199         init_attr.cap.max_recv_sge = 1;
 200         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 201         init_attr.qp_type       = IB_QPT_RC;
 202
 203         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 204         if (ret)
 205                 goto qp_err;
 206
 207         ib_conn->qp = ib_conn->cma_id->qp;
 208         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 209                  ib_conn, ib_conn->cma_id,
 210                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 211         return ret;
 212
 213 qp_err:
 214         (void)ib_destroy_fmr_pool(ib_conn->fmr_pool);
 215 fmr_pool_err:
 216         kfree(ib_conn->page_vec);
 217         kfree(ib_conn->login_buf);
 218 alloc_err:
 219         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 220         return ret;
 221 }
 222
 223 /**
 224  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 225  * -1 on failure
 226  */
 227 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
 228 {
 229         BUG_ON(ib_conn == NULL);
 230
 231         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 232                  ib_conn, ib_conn->cma_id,
 233                  ib_conn->fmr_pool, ib_conn->qp);
 234
 235         /* qp is created only once both addr & route are resolved */
 236         if (ib_conn->fmr_pool != NULL)
 237                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 238
 239         if (ib_conn->qp != NULL)
 240                 rdma_destroy_qp(ib_conn->cma_id);
 241
 242         if (ib_conn->cma_id != NULL)
 243                 rdma_destroy_id(ib_conn->cma_id);
 244
 245         ib_conn->fmr_pool = NULL;
 246         ib_conn->qp       = NULL;
 247         ib_conn->cma_id   = NULL;
 248         kfree(ib_conn->page_vec);
 249
 250         return 0;
 251 }
 252
 253 /**
 254  * based on the resolved device node GUID see if there already allocated
 255  * device for this device. If there's no such, create one.
 256  */
 257 static
 258 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 259 {
 260         struct iser_device *device;
 261
 262         mutex_lock(&ig.device_list_mutex);
 263
 264         list_for_each_entry(device, &ig.device_list, ig_list)
 265                 /* find if there's a match using the node GUID */
 266                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 267                         goto inc_refcnt;
 268
 269         device = kzalloc(sizeof *device, GFP_KERNEL);
 270         if (device == NULL)
 271                 goto out;
 272
 273         /* assign this device to the device */
 274         device->ib_device = cma_id->device;
 275         /* init the device and link it into ig device list */
 276         if (iser_create_device_ib_res(device)) {
 277                 kfree(device);
 278                 device = NULL;
 279                 goto out;
 280         }
 281         list_add(&device->ig_list, &ig.device_list);
 282
 283 inc_refcnt:
 284         device->refcount++;
 285 out:
 286         mutex_unlock(&ig.device_list_mutex);
 287         return device;
 288 }
 289
 290 /* if there's no demand for this device, release it */
 291 static void iser_device_try_release(struct iser_device *device)
 292 {
 293         mutex_lock(&ig.device_list_mutex);
 294         device->refcount--;
 295         iser_err("device %p refcount %d\n",device,device->refcount);
 296         if (!device->refcount) {
 297                 iser_free_device_ib_res(device);
 298                 list_del(&device->ig_list);
 299                 kfree(device);
 300         }
 301         mutex_unlock(&ig.device_list_mutex);
 302 }
 303
 304 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 305                                      enum iser_ib_conn_state comp,
 306                                      enum iser_ib_conn_state exch)
 307 {
 308         int ret;
 309
 310         spin_lock_bh(&ib_conn->lock);
 311         if ((ret = (ib_conn->state == comp)))
 312                 ib_conn->state = exch;
 313         spin_unlock_bh(&ib_conn->lock);
 314         return ret;
 315 }
 316
 317 /**
 318  * Frees all conn objects and deallocs conn descriptor
 319  */
 320 static void iser_conn_release(struct iser_conn *ib_conn)
 321 {
 322         struct iser_device  *device = ib_conn->device;
 323
 324         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 325
 326         mutex_lock(&ig.connlist_mutex);
 327         list_del(&ib_conn->conn_list);
 328         mutex_unlock(&ig.connlist_mutex);
 329         iser_free_rx_descriptors(ib_conn);
 330         iser_free_ib_conn_res(ib_conn);
 331         ib_conn->device = NULL;
 332         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 333         if (device != NULL)
 334                 iser_device_try_release(device);
 335         if (ib_conn->iser_conn)
 336                 ib_conn->iser_conn->ib_conn = NULL;
 337         iscsi_destroy_endpoint(ib_conn->ep);
 338 }
 339
 340 void iser_conn_get(struct iser_conn *ib_conn)
 341 {
 342         atomic_inc(&ib_conn->refcount);
 343 }
 344
 345 void iser_conn_put(struct iser_conn *ib_conn)
 346 {
 347         if (atomic_dec_and_test(&ib_conn->refcount))
 348                 iser_conn_release(ib_conn);
 349 }
 350
 351 /**
 352  * triggers start of the disconnect procedures and wait for them to be done
 353  */
 354 void iser_conn_terminate(struct iser_conn *ib_conn)
 355 {
 356         int err = 0;
 357
 358         /* change the ib conn state only if the conn is UP, however always call
 359          * rdma_disconnect since this is the only way to cause the CMA to change
 360          * the QP state to ERROR
 361          */
 362
 363         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 364         err = rdma_disconnect(ib_conn->cma_id);
 365         if (err)
 366                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 367                          ib_conn,err);
 368
 369         wait_event_interruptible(ib_conn->wait,
 370                                  ib_conn->state == ISER_CONN_DOWN);
 371
 372         iser_conn_put(ib_conn);
 373 }
 374
 375 static void iser_connect_error(struct rdma_cm_id *cma_id)
 376 {
 377         struct iser_conn *ib_conn;
 378         ib_conn = (struct iser_conn *)cma_id->context;
 379
 380         ib_conn->state = ISER_CONN_DOWN;
 381         wake_up_interruptible(&ib_conn->wait);
 382 }
 383
 384 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 385 {
 386         struct iser_device *device;
 387         struct iser_conn   *ib_conn;
 388         int    ret;
 389
 390         device = iser_device_find_by_ib_device(cma_id);
 391         if (!device) {
 392                 iser_err("device lookup/creation failed\n");
 393                 iser_connect_error(cma_id);
 394                 return;
 395         }
 396
 397         ib_conn = (struct iser_conn *)cma_id->context;
 398         ib_conn->device = device;
 399
 400         ret = rdma_resolve_route(cma_id, 1000);
 401         if (ret) {
 402                 iser_err("resolve route failed: %d\n", ret);
 403                 iser_connect_error(cma_id);
 404         }
 405 }
 406
 407 static void iser_route_handler(struct rdma_cm_id *cma_id)
 408 {
 409         struct rdma_conn_param conn_param;
 410         int    ret;
 411
 412         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 413         if (ret)
 414                 goto failure;
 415
 416         memset(&conn_param, 0, sizeof conn_param);
 417         conn_param.responder_resources = 4;
 418         conn_param.initiator_depth     = 1;
 419         conn_param.retry_count         = 7;
 420         conn_param.rnr_retry_count     = 6;
 421
 422         ret = rdma_connect(cma_id, &conn_param);
 423         if (ret) {
 424                 iser_err("failure connecting: %d\n", ret);
 425                 goto failure;
 426         }
 427
 428         return;
 429 failure:
 430         iser_connect_error(cma_id);
 431 }
 432
 433 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 434 {
 435         struct iser_conn *ib_conn;
 436
 437         ib_conn = (struct iser_conn *)cma_id->context;
 438         ib_conn->state = ISER_CONN_UP;
 439         wake_up_interruptible(&ib_conn->wait);
 440 }
 441
 442 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 443 {
 444         struct iser_conn *ib_conn;
 445
 446         ib_conn = (struct iser_conn *)cma_id->context;
 447         ib_conn->disc_evt_flag = 1;
 448
 449         /* getting here when the state is UP means that the conn is being *
 450          * terminated asynchronously from the iSCSI layer's perspective.  */
 451         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 452                                       ISER_CONN_TERMINATING))
 453                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 454                                    ISCSI_ERR_CONN_FAILED);
 455
 456         /* Complete the termination process if no posts are pending */
 457         if (ib_conn->post_recv_buf_count == 0 &&
 458             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 459                 ib_conn->state = ISER_CONN_DOWN;
 460                 wake_up_interruptible(&ib_conn->wait);
 461         }
 462 }
 463
 464 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 465 {
 466         int ret = 0;
 467
 468         iser_err("event %d conn %p id %p\n",event->event,cma_id->context,cma_id);
 469
 470         switch (event->event) {
 471         case RDMA_CM_EVENT_ADDR_RESOLVED:
 472                 iser_addr_handler(cma_id);
 473                 break;
 474         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 475                 iser_route_handler(cma_id);
 476                 break;
 477         case RDMA_CM_EVENT_ESTABLISHED:
 478                 iser_connected_handler(cma_id);
 479                 break;
 480         case RDMA_CM_EVENT_ADDR_ERROR:
 481         case RDMA_CM_EVENT_ROUTE_ERROR:
 482         case RDMA_CM_EVENT_CONNECT_ERROR:
 483         case RDMA_CM_EVENT_UNREACHABLE:
 484         case RDMA_CM_EVENT_REJECTED:
 485                 iser_err("event: %d, error: %d\n", event->event, event->status);
 486                 iser_connect_error(cma_id);
 487                 break;
 488         case RDMA_CM_EVENT_DISCONNECTED:
 489         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 490         case RDMA_CM_EVENT_ADDR_CHANGE:
 491                 iser_disconnected_handler(cma_id);
 492                 break;
 493         default:
 494                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 495                 break;
 496         }
 497         return ret;
 498 }
 499
 500 void iser_conn_init(struct iser_conn *ib_conn)
 501 {
 502         ib_conn->state = ISER_CONN_INIT;
 503         init_waitqueue_head(&ib_conn->wait);
 504         ib_conn->post_recv_buf_count = 0;
 505         atomic_set(&ib_conn->post_send_buf_count, 0);
 506         atomic_set(&ib_conn->refcount, 1);
 507         INIT_LIST_HEAD(&ib_conn->conn_list);
 508         spin_lock_init(&ib_conn->lock);
 509 }
 510
 511  /**
 512  * starts the process of connecting to the target
 513  * sleeps until the connection is established or rejected
 514  */
 515 int iser_connect(struct iser_conn   *ib_conn,
 516                  struct sockaddr_in *src_addr,
 517                  struct sockaddr_in *dst_addr,
 518                  int                 non_blocking)
 519 {
 520         struct sockaddr *src, *dst;
 521         int err = 0;
 522
 523         sprintf(ib_conn->name, "%pI4:%d",
 524                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 525
 526         /* the device is known only --after-- address resolution */
 527         ib_conn->device = NULL;
 528
 529         iser_err("connecting to: %pI4, port 0x%x\n",
 530                  &dst_addr->sin_addr, dst_addr->sin_port);
 531
 532         ib_conn->state = ISER_CONN_PENDING;
 533
 534         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 535                                              (void *)ib_conn,
 536                                              RDMA_PS_TCP);
 537         if (IS_ERR(ib_conn->cma_id)) {
 538                 err = PTR_ERR(ib_conn->cma_id);
 539                 iser_err("rdma_create_id failed: %d\n", err);
 540                 goto id_failure;
 541         }
 542
 543         src = (struct sockaddr *)src_addr;
 544         dst = (struct sockaddr *)dst_addr;
 545         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 546         if (err) {
 547                 iser_err("rdma_resolve_addr failed: %d\n", err);
 548                 goto addr_failure;
 549         }
 550
 551         if (!non_blocking) {
 552                 wait_event_interruptible(ib_conn->wait,
 553                                          (ib_conn->state != ISER_CONN_PENDING));
 554
 555                 if (ib_conn->state != ISER_CONN_UP) {
 556                         err =  -EIO;
 557                         goto connect_failure;
 558                 }
 559         }
 560
 561         mutex_lock(&ig.connlist_mutex);
 562         list_add(&ib_conn->conn_list, &ig.connlist);
 563         mutex_unlock(&ig.connlist_mutex);
 564         return 0;
 565
 566 id_failure:
 567         ib_conn->cma_id = NULL;
 568 addr_failure:
 569         ib_conn->state = ISER_CONN_DOWN;
 570 connect_failure:
 571         iser_conn_release(ib_conn);
 572         return err;
 573 }
 574
 575 /**
 576  * iser_reg_page_vec - Register physical memory
 577  *
 578  * returns: 0 on success, errno code on failure
 579  */
 580 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 581                       struct iser_page_vec *page_vec,
 582                       struct iser_mem_reg  *mem_reg)
 583 {
 584         struct ib_pool_fmr *mem;
 585         u64                io_addr;
 586         u64                *page_list;
 587         int                status;
 588
 589         page_list = page_vec->pages;
 590         io_addr   = page_list[0];
 591
 592         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 593                                     page_list,
 594                                     page_vec->length,
 595                                     io_addr);
 596
 597         if (IS_ERR(mem)) {
 598                 status = (int)PTR_ERR(mem);
 599                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 600                 return status;
 601         }
 602
 603         mem_reg->lkey  = mem->fmr->lkey;
 604         mem_reg->rkey  = mem->fmr->rkey;
 605         mem_reg->len   = page_vec->length * SIZE_4K;
 606         mem_reg->va    = io_addr;
 607         mem_reg->is_fmr = 1;
 608         mem_reg->mem_h = (void *)mem;
 609
 610         mem_reg->va   += page_vec->offset;
 611         mem_reg->len   = page_vec->data_size;
 612
 613         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 614                  "entry[0]: (0x%08lx,%ld)] -> "
 615                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 616                  page_vec, page_vec->length,
 617                  (unsigned long)page_vec->pages[0],
 618                  (unsigned long)page_vec->data_size,
 619                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 620                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 621         return 0;
 622 }
 623
 624 /**
 625  * Unregister (previosuly registered) memory.
 626  */
 627 void iser_unreg_mem(struct iser_mem_reg *reg)
 628 {
 629         int ret;
 630
 631         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 632
 633         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 634         if (ret)
 635                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 636
 637         reg->mem_h = NULL;
 638 }
 639
 640 int iser_post_recvl(struct iser_conn *ib_conn)
 641 {
 642         struct ib_recv_wr rx_wr, *rx_wr_failed;
 643         struct ib_sge     sge;
 644         int ib_ret;
 645
 646         sge.addr   = ib_conn->login_dma;
 647         sge.length = ISER_RX_LOGIN_SIZE;
 648         sge.lkey   = ib_conn->device->mr->lkey;
 649
 650         rx_wr.wr_id   = (unsigned long)ib_conn->login_buf;
 651         rx_wr.sg_list = &sge;
 652         rx_wr.num_sge = 1;
 653         rx_wr.next    = NULL;
 654
 655         ib_conn->post_recv_buf_count++;
 656         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 657         if (ib_ret) {
 658                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 659                 ib_conn->post_recv_buf_count--;
 660         }
 661         return ib_ret;
 662 }
 663
 664 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 665 {
 666         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 667         int i, ib_ret;
 668         unsigned int my_rx_head = ib_conn->rx_desc_head;
 669         struct iser_rx_desc *rx_desc;
 670
 671         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 672                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 673                 rx_wr->wr_id    = (unsigned long)rx_desc;
 674                 rx_wr->sg_list  = &rx_desc->rx_sg;
 675                 rx_wr->num_sge  = 1;
 676                 rx_wr->next     = rx_wr + 1;
 677                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 678         }
 679
 680         rx_wr--;
 681         rx_wr->next = NULL; /* mark end of work requests list */
 682
 683         ib_conn->post_recv_buf_count += count;
 684         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 685         if (ib_ret) {
 686                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 687                 ib_conn->post_recv_buf_count -= count;
 688         } else
 689                 ib_conn->rx_desc_head = my_rx_head;
 690         return ib_ret;
 691 }
 692
 693
 694 /**
 695  * iser_start_send - Initiate a Send DTO operation
 696  *
 697  * returns 0 on success, -1 on failure
 698  */
 699 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 700 {
 701         int               ib_ret;
 702         struct ib_send_wr send_wr, *send_wr_failed;
 703
 704         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 705                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 706
 707         send_wr.next       = NULL;
 708         send_wr.wr_id      = (unsigned long)tx_desc;
 709         send_wr.sg_list    = tx_desc->tx_sg;
 710         send_wr.num_sge    = tx_desc->num_sge;
 711         send_wr.opcode     = IB_WR_SEND;
 712         send_wr.send_flags = IB_SEND_SIGNALED;
 713
 714         atomic_inc(&ib_conn->post_send_buf_count);
 715
 716         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 717         if (ib_ret) {
 718                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 719                 atomic_dec(&ib_conn->post_send_buf_count);
 720         }
 721         return ib_ret;
 722 }
 723
 724 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 725                                 struct iser_conn *ib_conn)
 726 {
 727         if (desc && desc->type == ISCSI_TX_DATAOUT)
 728                 kmem_cache_free(ig.desc_cache, desc);
 729
 730         if (ib_conn->post_recv_buf_count == 0 &&
 731             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 732                 /* getting here when the state is UP means that the conn is *
 733                  * being terminated asynchronously from the iSCSI layer's   *
 734                  * perspective.                                             */
 735                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 736                     ISER_CONN_TERMINATING))
 737                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 738                                            ISCSI_ERR_CONN_FAILED);
 739
 740                 /* complete the termination process if disconnect event was delivered *
 741                  * note there are no more non completed posts to the QP               */
 742                 if (ib_conn->disc_evt_flag) {
 743                         ib_conn->state = ISER_CONN_DOWN;
 744                         wake_up_interruptible(&ib_conn->wait);
 745                 }
 746         }
 747 }
 748
 749 static int iser_drain_tx_cq(struct iser_device  *device)
 750 {
 751         struct ib_cq  *cq = device->tx_cq;
 752         struct ib_wc  wc;
 753         struct iser_tx_desc *tx_desc;
 754         struct iser_conn *ib_conn;
 755         int completed_tx = 0;
 756
 757         while (ib_poll_cq(cq, 1, &wc) == 1) {
 758                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 759                 ib_conn = wc.qp->qp_context;
 760                 if (wc.status == IB_WC_SUCCESS) {
 761                         if (wc.opcode == IB_WC_SEND)
 762                                 iser_snd_completion(tx_desc, ib_conn);
 763                         else
 764                                 iser_err("expected opcode %d got %d\n",
 765                                         IB_WC_SEND, wc.opcode);
 766                 } else {
 767                         iser_err("tx id %llx status %d vend_err %x\n",
 768                                 wc.wr_id, wc.status, wc.vendor_err);
 769                         atomic_dec(&ib_conn->post_send_buf_count);
 770                         iser_handle_comp_error(tx_desc, ib_conn);
 771                 }
 772                 completed_tx++;
 773         }
 774         return completed_tx;
 775 }
 776
 777
 778 static void iser_cq_tasklet_fn(unsigned long data)
 779 {
 780          struct iser_device  *device = (struct iser_device *)data;
 781          struct ib_cq        *cq = device->rx_cq;
 782          struct ib_wc        wc;
 783          struct iser_rx_desc *desc;
 784          unsigned long       xfer_len;
 785         struct iser_conn *ib_conn;
 786         int completed_tx, completed_rx;
 787         completed_tx = completed_rx = 0;
 788
 789         while (ib_poll_cq(cq, 1, &wc) == 1) {
 790                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 791                 BUG_ON(desc == NULL);
 792                 ib_conn = wc.qp->qp_context;
 793                 if (wc.status == IB_WC_SUCCESS) {
 794                         if (wc.opcode == IB_WC_RECV) {
 795                                 xfer_len = (unsigned long)wc.byte_len;
 796                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 797                         } else
 798                                 iser_err("expected opcode %d got %d\n",
 799                                         IB_WC_RECV, wc.opcode);
 800                 } else {
 801                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 802                                 iser_err("rx id %llx status %d vend_err %x\n",
 803                                         wc.wr_id, wc.status, wc.vendor_err);
 804                         ib_conn->post_recv_buf_count--;
 805                         iser_handle_comp_error(NULL, ib_conn);
 806                 }
 807                 completed_rx++;
 808                 if (!(completed_rx & 63))
 809                         completed_tx += iser_drain_tx_cq(device);
 810         }
 811         /* #warning "it is assumed here that arming CQ only once its empty" *
 812          * " would not cause interrupts to be missed"                       */
 813         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 814
 815         completed_tx += iser_drain_tx_cq(device);
 816         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 817 }
 818
 819 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 820 {
 821         struct iser_device  *device = (struct iser_device *)cq_context;
 822
 823         tasklet_schedule(&device->cq_tasklet);
 824 }