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 alloc_err;
 168                 ret = -ENOMEM;
 169         }
 170
 171         ib_conn->login_dma = ib_dma_map_single(ib_conn->device->ib_device,
 172                                 (void *)ib_conn->login_buf, ISER_RX_LOGIN_SIZE,
 173                                 DMA_FROM_DEVICE);
 174
 175         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 176                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 177                                     GFP_KERNEL);
 178         if (!ib_conn->page_vec) {
 179                 ret = -ENOMEM;
 180                 goto alloc_err;
 181         }
 182         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 183
 184         params.page_shift        = SHIFT_4K;
 185         /* when the first/last SG element are not start/end *
 186          * page aligned, the map whould be of N+1 pages     */
 187         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 188         /* make the pool size twice the max number of SCSI commands *
 189          * the ML is expected to queue, watermark for unmap at 50%  */
 190         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 191         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 192         params.cache             = 0;
 193         params.flush_function    = NULL;
 194         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 195                                     IB_ACCESS_REMOTE_WRITE |
 196                                     IB_ACCESS_REMOTE_READ);
 197
 198         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 199         if (IS_ERR(ib_conn->fmr_pool)) {
 200                 ret = PTR_ERR(ib_conn->fmr_pool);
 201                 goto fmr_pool_err;
 202         }
 203
 204         memset(&init_attr, 0, sizeof init_attr);
 205
 206         init_attr.event_handler = iser_qp_event_callback;
 207         init_attr.qp_context    = (void *)ib_conn;
 208         init_attr.send_cq       = device->tx_cq;
 209         init_attr.recv_cq       = device->rx_cq;
 210         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 211         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 212         init_attr.cap.max_send_sge = 2;
 213         init_attr.cap.max_recv_sge = 1;
 214         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 215         init_attr.qp_type       = IB_QPT_RC;
 216
 217         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 218         if (ret)
 219                 goto qp_err;
 220
 221         ib_conn->qp = ib_conn->cma_id->qp;
 222         iser_err("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 223                  ib_conn, ib_conn->cma_id,
 224                  ib_conn->fmr_pool, ib_conn->cma_id->qp);
 225         return ret;
 226
 227 qp_err:
 228         (void)ib_destroy_fmr_pool(ib_conn->fmr_pool);
 229 fmr_pool_err:
 230         kfree(ib_conn->page_vec);
 231         kfree(ib_conn->login_buf);
 232 alloc_err:
 233         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 234         return ret;
 235 }
 236
 237 /**
 238  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 239  * -1 on failure
 240  */
 241 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
 242 {
 243         BUG_ON(ib_conn == NULL);
 244
 245         iser_err("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 246                  ib_conn, ib_conn->cma_id,
 247                  ib_conn->fmr_pool, ib_conn->qp);
 248
 249         /* qp is created only once both addr & route are resolved */
 250         if (ib_conn->fmr_pool != NULL)
 251                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 252
 253         if (ib_conn->qp != NULL)
 254                 rdma_destroy_qp(ib_conn->cma_id);
 255
 256         if (ib_conn->cma_id != NULL)
 257                 rdma_destroy_id(ib_conn->cma_id);
 258
 259         ib_conn->fmr_pool = NULL;
 260         ib_conn->qp       = NULL;
 261         ib_conn->cma_id   = NULL;
 262         kfree(ib_conn->page_vec);
 263
 264         return 0;
 265 }
 266
 267 /**
 268  * based on the resolved device node GUID see if there already allocated
 269  * device for this device. If there's no such, create one.
 270  */
 271 static
 272 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 273 {
 274         struct iser_device *device;
 275
 276         mutex_lock(&ig.device_list_mutex);
 277
 278         list_for_each_entry(device, &ig.device_list, ig_list)
 279                 /* find if there's a match using the node GUID */
 280                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 281                         goto inc_refcnt;
 282
 283         device = kzalloc(sizeof *device, GFP_KERNEL);
 284         if (device == NULL)
 285                 goto out;
 286
 287         /* assign this device to the device */
 288         device->ib_device = cma_id->device;
 289         /* init the device and link it into ig device list */
 290         if (iser_create_device_ib_res(device)) {
 291                 kfree(device);
 292                 device = NULL;
 293                 goto out;
 294         }
 295         list_add(&device->ig_list, &ig.device_list);
 296
 297 inc_refcnt:
 298         device->refcount++;
 299 out:
 300         mutex_unlock(&ig.device_list_mutex);
 301         return device;
 302 }
 303
 304 /* if there's no demand for this device, release it */
 305 static void iser_device_try_release(struct iser_device *device)
 306 {
 307         mutex_lock(&ig.device_list_mutex);
 308         device->refcount--;
 309         iser_err("device %p refcount %d\n",device,device->refcount);
 310         if (!device->refcount) {
 311                 iser_free_device_ib_res(device);
 312                 list_del(&device->ig_list);
 313                 kfree(device);
 314         }
 315         mutex_unlock(&ig.device_list_mutex);
 316 }
 317
 318 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 319                                      enum iser_ib_conn_state comp,
 320                                      enum iser_ib_conn_state exch)
 321 {
 322         int ret;
 323
 324         spin_lock_bh(&ib_conn->lock);
 325         if ((ret = (ib_conn->state == comp)))
 326                 ib_conn->state = exch;
 327         spin_unlock_bh(&ib_conn->lock);
 328         return ret;
 329 }
 330
 331 /**
 332  * Frees all conn objects and deallocs conn descriptor
 333  */
 334 static void iser_conn_release(struct iser_conn *ib_conn)
 335 {
 336         struct iser_device  *device = ib_conn->device;
 337
 338         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 339
 340         mutex_lock(&ig.connlist_mutex);
 341         list_del(&ib_conn->conn_list);
 342         mutex_unlock(&ig.connlist_mutex);
 343         iser_free_rx_descriptors(ib_conn);
 344         iser_free_ib_conn_res(ib_conn);
 345         ib_conn->device = NULL;
 346         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 347         if (device != NULL)
 348                 iser_device_try_release(device);
 349         if (ib_conn->iser_conn)
 350                 ib_conn->iser_conn->ib_conn = NULL;
 351         iscsi_destroy_endpoint(ib_conn->ep);
 352 }
 353
 354 void iser_conn_get(struct iser_conn *ib_conn)
 355 {
 356         atomic_inc(&ib_conn->refcount);
 357 }
 358
 359 void iser_conn_put(struct iser_conn *ib_conn)
 360 {
 361         if (atomic_dec_and_test(&ib_conn->refcount))
 362                 iser_conn_release(ib_conn);
 363 }
 364
 365 /**
 366  * triggers start of the disconnect procedures and wait for them to be done
 367  */
 368 void iser_conn_terminate(struct iser_conn *ib_conn)
 369 {
 370         int err = 0;
 371
 372         /* change the ib conn state only if the conn is UP, however always call
 373          * rdma_disconnect since this is the only way to cause the CMA to change
 374          * the QP state to ERROR
 375          */
 376
 377         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 378         err = rdma_disconnect(ib_conn->cma_id);
 379         if (err)
 380                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 381                          ib_conn,err);
 382
 383         wait_event_interruptible(ib_conn->wait,
 384                                  ib_conn->state == ISER_CONN_DOWN);
 385
 386         iser_conn_put(ib_conn);
 387 }
 388
 389 static void iser_connect_error(struct rdma_cm_id *cma_id)
 390 {
 391         struct iser_conn *ib_conn;
 392         ib_conn = (struct iser_conn *)cma_id->context;
 393
 394         ib_conn->state = ISER_CONN_DOWN;
 395         wake_up_interruptible(&ib_conn->wait);
 396 }
 397
 398 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 399 {
 400         struct iser_device *device;
 401         struct iser_conn   *ib_conn;
 402         int    ret;
 403
 404         device = iser_device_find_by_ib_device(cma_id);
 405         if (!device) {
 406                 iser_err("device lookup/creation failed\n");
 407                 iser_connect_error(cma_id);
 408                 return;
 409         }
 410
 411         ib_conn = (struct iser_conn *)cma_id->context;
 412         ib_conn->device = device;
 413
 414         ret = rdma_resolve_route(cma_id, 1000);
 415         if (ret) {
 416                 iser_err("resolve route failed: %d\n", ret);
 417                 iser_connect_error(cma_id);
 418         }
 419 }
 420
 421 static void iser_route_handler(struct rdma_cm_id *cma_id)
 422 {
 423         struct rdma_conn_param conn_param;
 424         int    ret;
 425
 426         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 427         if (ret)
 428                 goto failure;
 429
 430         memset(&conn_param, 0, sizeof conn_param);
 431         conn_param.responder_resources = 4;
 432         conn_param.initiator_depth     = 1;
 433         conn_param.retry_count         = 7;
 434         conn_param.rnr_retry_count     = 6;
 435
 436         ret = rdma_connect(cma_id, &conn_param);
 437         if (ret) {
 438                 iser_err("failure connecting: %d\n", ret);
 439                 goto failure;
 440         }
 441
 442         return;
 443 failure:
 444         iser_connect_error(cma_id);
 445 }
 446
 447 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 448 {
 449         struct iser_conn *ib_conn;
 450
 451         ib_conn = (struct iser_conn *)cma_id->context;
 452         ib_conn->state = ISER_CONN_UP;
 453         wake_up_interruptible(&ib_conn->wait);
 454 }
 455
 456 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 457 {
 458         struct iser_conn *ib_conn;
 459
 460         ib_conn = (struct iser_conn *)cma_id->context;
 461         ib_conn->disc_evt_flag = 1;
 462
 463         /* getting here when the state is UP means that the conn is being *
 464          * terminated asynchronously from the iSCSI layer's perspective.  */
 465         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 466                                       ISER_CONN_TERMINATING))
 467                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 468                                    ISCSI_ERR_CONN_FAILED);
 469
 470         /* Complete the termination process if no posts are pending */
 471         if (ib_conn->post_recv_buf_count == 0 &&
 472             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 473                 ib_conn->state = ISER_CONN_DOWN;
 474                 wake_up_interruptible(&ib_conn->wait);
 475         }
 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 conn %p id %p\n",event->event,cma_id->context,cma_id);
 483
 484         switch (event->event) {
 485         case RDMA_CM_EVENT_ADDR_RESOLVED:
 486                 iser_addr_handler(cma_id);
 487                 break;
 488         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 489                 iser_route_handler(cma_id);
 490                 break;
 491         case RDMA_CM_EVENT_ESTABLISHED:
 492                 iser_connected_handler(cma_id);
 493                 break;
 494         case RDMA_CM_EVENT_ADDR_ERROR:
 495         case RDMA_CM_EVENT_ROUTE_ERROR:
 496         case RDMA_CM_EVENT_CONNECT_ERROR:
 497         case RDMA_CM_EVENT_UNREACHABLE:
 498         case RDMA_CM_EVENT_REJECTED:
 499                 iser_err("event: %d, error: %d\n", event->event, event->status);
 500                 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                 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);
 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         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 549                                              (void *)ib_conn,
 550                                              RDMA_PS_TCP);
 551         if (IS_ERR(ib_conn->cma_id)) {
 552                 err = PTR_ERR(ib_conn->cma_id);
 553                 iser_err("rdma_create_id failed: %d\n", err);
 554                 goto id_failure;
 555         }
 556
 557         src = (struct sockaddr *)src_addr;
 558         dst = (struct sockaddr *)dst_addr;
 559         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 560         if (err) {
 561                 iser_err("rdma_resolve_addr failed: %d\n", err);
 562                 goto addr_failure;
 563         }
 564
 565         if (!non_blocking) {
 566                 wait_event_interruptible(ib_conn->wait,
 567                                          (ib_conn->state != ISER_CONN_PENDING));
 568
 569                 if (ib_conn->state != ISER_CONN_UP) {
 570                         err =  -EIO;
 571                         goto connect_failure;
 572                 }
 573         }
 574
 575         mutex_lock(&ig.connlist_mutex);
 576         list_add(&ib_conn->conn_list, &ig.connlist);
 577         mutex_unlock(&ig.connlist_mutex);
 578         return 0;
 579
 580 id_failure:
 581         ib_conn->cma_id = NULL;
 582 addr_failure:
 583         ib_conn->state = ISER_CONN_DOWN;
 584 connect_failure:
 585         iser_conn_release(ib_conn);
 586         return err;
 587 }
 588
 589 /**
 590  * iser_reg_page_vec - Register physical memory
 591  *
 592  * returns: 0 on success, errno code on failure
 593  */
 594 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 595                       struct iser_page_vec *page_vec,
 596                       struct iser_mem_reg  *mem_reg)
 597 {
 598         struct ib_pool_fmr *mem;
 599         u64                io_addr;
 600         u64                *page_list;
 601         int                status;
 602
 603         page_list = page_vec->pages;
 604         io_addr   = page_list[0];
 605
 606         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 607                                     page_list,
 608                                     page_vec->length,
 609                                     io_addr);
 610
 611         if (IS_ERR(mem)) {
 612                 status = (int)PTR_ERR(mem);
 613                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 614                 return status;
 615         }
 616
 617         mem_reg->lkey  = mem->fmr->lkey;
 618         mem_reg->rkey  = mem->fmr->rkey;
 619         mem_reg->len   = page_vec->length * SIZE_4K;
 620         mem_reg->va    = io_addr;
 621         mem_reg->is_fmr = 1;
 622         mem_reg->mem_h = (void *)mem;
 623
 624         mem_reg->va   += page_vec->offset;
 625         mem_reg->len   = page_vec->data_size;
 626
 627         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 628                  "entry[0]: (0x%08lx,%ld)] -> "
 629                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 630                  page_vec, page_vec->length,
 631                  (unsigned long)page_vec->pages[0],
 632                  (unsigned long)page_vec->data_size,
 633                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 634                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 635         return 0;
 636 }
 637
 638 /**
 639  * Unregister (previosuly registered) memory.
 640  */
 641 void iser_unreg_mem(struct iser_mem_reg *reg)
 642 {
 643         int ret;
 644
 645         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 646
 647         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 648         if (ret)
 649                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 650
 651         reg->mem_h = NULL;
 652 }
 653
 654 int iser_post_recvl(struct iser_conn *ib_conn)
 655 {
 656         struct ib_recv_wr rx_wr, *rx_wr_failed;
 657         struct ib_sge     sge;
 658         int ib_ret;
 659
 660         sge.addr   = ib_conn->login_dma;
 661         sge.length = ISER_RX_LOGIN_SIZE;
 662         sge.lkey   = ib_conn->device->mr->lkey;
 663
 664         rx_wr.wr_id   = (unsigned long)ib_conn->login_buf;
 665         rx_wr.sg_list = &sge;
 666         rx_wr.num_sge = 1;
 667         rx_wr.next    = NULL;
 668
 669         ib_conn->post_recv_buf_count++;
 670         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 671         if (ib_ret) {
 672                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 673                 ib_conn->post_recv_buf_count--;
 674         }
 675         return ib_ret;
 676 }
 677
 678 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 679 {
 680         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 681         int i, ib_ret;
 682         unsigned int my_rx_head = ib_conn->rx_desc_head;
 683         struct iser_rx_desc *rx_desc;
 684
 685         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 686                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 687                 rx_wr->wr_id    = (unsigned long)rx_desc;
 688                 rx_wr->sg_list  = &rx_desc->rx_sg;
 689                 rx_wr->num_sge  = 1;
 690                 rx_wr->next     = rx_wr + 1;
 691                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 692         }
 693
 694         rx_wr--;
 695         rx_wr->next = NULL; /* mark end of work requests list */
 696
 697         ib_conn->post_recv_buf_count += count;
 698         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 699         if (ib_ret) {
 700                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 701                 ib_conn->post_recv_buf_count -= count;
 702         } else
 703                 ib_conn->rx_desc_head = my_rx_head;
 704         return ib_ret;
 705 }
 706
 707
 708 /**
 709  * iser_start_send - Initiate a Send DTO operation
 710  *
 711  * returns 0 on success, -1 on failure
 712  */
 713 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 714 {
 715         int               ib_ret;
 716         struct ib_send_wr send_wr, *send_wr_failed;
 717
 718         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 719                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 720
 721         send_wr.next       = NULL;
 722         send_wr.wr_id      = (unsigned long)tx_desc;
 723         send_wr.sg_list    = tx_desc->tx_sg;
 724         send_wr.num_sge    = tx_desc->num_sge;
 725         send_wr.opcode     = IB_WR_SEND;
 726         send_wr.send_flags = IB_SEND_SIGNALED;
 727
 728         atomic_inc(&ib_conn->post_send_buf_count);
 729
 730         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 731         if (ib_ret) {
 732                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 733                 atomic_dec(&ib_conn->post_send_buf_count);
 734         }
 735         return ib_ret;
 736 }
 737
 738 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 739                                 struct iser_conn *ib_conn)
 740 {
 741         if (desc && desc->type == ISCSI_TX_DATAOUT)
 742                 kmem_cache_free(ig.desc_cache, desc);
 743
 744         if (ib_conn->post_recv_buf_count == 0 &&
 745             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 746                 /* getting here when the state is UP means that the conn is *
 747                  * being terminated asynchronously from the iSCSI layer's   *
 748                  * perspective.                                             */
 749                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 750                     ISER_CONN_TERMINATING))
 751                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 752                                            ISCSI_ERR_CONN_FAILED);
 753
 754                 /* complete the termination process if disconnect event was delivered *
 755                  * note there are no more non completed posts to the QP               */
 756                 if (ib_conn->disc_evt_flag) {
 757                         ib_conn->state = ISER_CONN_DOWN;
 758                         wake_up_interruptible(&ib_conn->wait);
 759                 }
 760         }
 761 }
 762
 763 static int iser_drain_tx_cq(struct iser_device  *device)
 764 {
 765         struct ib_cq  *cq = device->tx_cq;
 766         struct ib_wc  wc;
 767         struct iser_tx_desc *tx_desc;
 768         struct iser_conn *ib_conn;
 769         int completed_tx = 0;
 770
 771         while (ib_poll_cq(cq, 1, &wc) == 1) {
 772                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 773                 ib_conn = wc.qp->qp_context;
 774                 if (wc.status == IB_WC_SUCCESS) {
 775                         if (wc.opcode == IB_WC_SEND)
 776                                 iser_snd_completion(tx_desc, ib_conn);
 777                         else
 778                                 iser_err("expected opcode %d got %d\n",
 779                                         IB_WC_SEND, wc.opcode);
 780                 } else {
 781                         iser_err("tx id %llx status %d vend_err %x\n",
 782                                 wc.wr_id, wc.status, wc.vendor_err);
 783                         atomic_dec(&ib_conn->post_send_buf_count);
 784                         iser_handle_comp_error(tx_desc, ib_conn);
 785                 }
 786                 completed_tx++;
 787         }
 788         return completed_tx;
 789 }
 790
 791
 792 static void iser_cq_tasklet_fn(unsigned long data)
 793 {
 794          struct iser_device  *device = (struct iser_device *)data;
 795          struct ib_cq        *cq = device->rx_cq;
 796          struct ib_wc        wc;
 797          struct iser_rx_desc *desc;
 798          unsigned long       xfer_len;
 799         struct iser_conn *ib_conn;
 800         int completed_tx, completed_rx;
 801         completed_tx = completed_rx = 0;
 802
 803         while (ib_poll_cq(cq, 1, &wc) == 1) {
 804                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 805                 BUG_ON(desc == NULL);
 806                 ib_conn = wc.qp->qp_context;
 807                 if (wc.status == IB_WC_SUCCESS) {
 808                         if (wc.opcode == IB_WC_RECV) {
 809                                 xfer_len = (unsigned long)wc.byte_len;
 810                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 811                         } else
 812                                 iser_err("expected opcode %d got %d\n",
 813                                         IB_WC_RECV, wc.opcode);
 814                 } else {
 815                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 816                                 iser_err("rx id %llx status %d vend_err %x\n",
 817                                         wc.wr_id, wc.status, wc.vendor_err);
 818                         ib_conn->post_recv_buf_count--;
 819                         iser_handle_comp_error(NULL, ib_conn);
 820                 }
 821                 completed_rx++;
 822                 if (!(completed_rx & 63))
 823                         completed_tx += iser_drain_tx_cq(device);
 824         }
 825         /* #warning "it is assumed here that arming CQ only once its empty" *
 826          * " would not cause interrupts to be missed"                       */
 827         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 828
 829         completed_tx += iser_drain_tx_cq(device);
 830         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 831 }
 832
 833 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 834 {
 835         struct iser_device  *device = (struct iser_device *)cq_context;
 836
 837         tasklet_schedule(&device->cq_tasklet);
 838 }