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         int i, j;
  74         struct iser_cq_desc *cq_desc;
  75
  76         device->cqs_used = min(ISER_MAX_CQ, device->ib_device->num_comp_vectors);
  77         iser_info("using %d CQs, device %s supports %d vectors\n",
  78                   device->cqs_used, device->ib_device->name,
  79                   device->ib_device->num_comp_vectors);
  80
  81         device->cq_desc = kmalloc(sizeof(struct iser_cq_desc) * device->cqs_used,
  82                                   GFP_KERNEL);
  83         if (device->cq_desc == NULL)
  84                 goto cq_desc_err;
  85         cq_desc = device->cq_desc;
  86
  87         device->pd = ib_alloc_pd(device->ib_device);
  88         if (IS_ERR(device->pd))
  89                 goto pd_err;
  90
  91         for (i = 0; i < device->cqs_used; i++) {
  92                 cq_desc[i].device   = device;
  93                 cq_desc[i].cq_index = i;
  94
  95                 device->rx_cq[i] = ib_create_cq(device->ib_device,
  96                                           iser_cq_callback,
  97                                           iser_cq_event_callback,
  98                                           (void *)&cq_desc[i],
  99                                           ISER_MAX_RX_CQ_LEN, i);
 100                 if (IS_ERR(device->rx_cq[i]))
 101                         goto cq_err;
 102
 103                 device->tx_cq[i] = ib_create_cq(device->ib_device,
 104                                           NULL, iser_cq_event_callback,
 105                                           (void *)&cq_desc[i],
 106                                           ISER_MAX_TX_CQ_LEN, i);
 107
 108                 if (IS_ERR(device->tx_cq[i]))
 109                         goto cq_err;
 110
 111                 if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
 112                         goto cq_err;
 113
 114                 tasklet_init(&device->cq_tasklet[i],
 115                              iser_cq_tasklet_fn,
 116                         (unsigned long)&cq_desc[i]);
 117         }
 118
 119         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
 120                                    IB_ACCESS_REMOTE_WRITE |
 121                                    IB_ACCESS_REMOTE_READ);
 122         if (IS_ERR(device->mr))
 123                 goto dma_mr_err;
 124
 125         INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
 126                                 iser_event_handler);
 127         if (ib_register_event_handler(&device->event_handler))
 128                 goto handler_err;
 129
 130         return 0;
 131
 132 handler_err:
 133         ib_dereg_mr(device->mr);
 134 dma_mr_err:
 135         for (j = 0; j < device->cqs_used; j++)
 136                 tasklet_kill(&device->cq_tasklet[j]);
 137 cq_err:
 138         for (j = 0; j < i; j++) {
 139                 if (device->tx_cq[j])
 140                         ib_destroy_cq(device->tx_cq[j]);
 141                 if (device->rx_cq[j])
 142                         ib_destroy_cq(device->rx_cq[j]);
 143         }
 144         ib_dealloc_pd(device->pd);
 145 pd_err:
 146         kfree(device->cq_desc);
 147 cq_desc_err:
 148         iser_err("failed to allocate an IB resource\n");
 149         return -1;
 150 }
 151
 152 /**
 153  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 154  * CQ and PD created with the device associated with the adapator.
 155  */
 156 static void iser_free_device_ib_res(struct iser_device *device)
 157 {
 158         int i;
 159         BUG_ON(device->mr == NULL);
 160
 161         for (i = 0; i < device->cqs_used; i++) {
 162                 tasklet_kill(&device->cq_tasklet[i]);
 163                 (void)ib_destroy_cq(device->tx_cq[i]);
 164                 (void)ib_destroy_cq(device->rx_cq[i]);
 165                 device->tx_cq[i] = NULL;
 166                 device->rx_cq[i] = NULL;
 167         }
 168
 169         (void)ib_unregister_event_handler(&device->event_handler);
 170         (void)ib_dereg_mr(device->mr);
 171         (void)ib_dealloc_pd(device->pd);
 172
 173         kfree(device->cq_desc);
 174
 175         device->mr = NULL;
 176         device->pd = NULL;
 177 }
 178
 179 /**
 180  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
 181  *
 182  * returns 0 on success, -1 on failure
 183  */
 184 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
 185 {
 186         struct iser_device      *device;
 187         struct ib_qp_init_attr  init_attr;
 188         int                     req_err, resp_err, ret = -ENOMEM;
 189         struct ib_fmr_pool_param params;
 190         int index, min_index = 0;
 191
 192         BUG_ON(ib_conn->device == NULL);
 193
 194         device = ib_conn->device;
 195
 196         ib_conn->login_buf = kmalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 197                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 198         if (!ib_conn->login_buf)
 199                 goto out_err;
 200
 201         ib_conn->login_req_buf  = ib_conn->login_buf;
 202         ib_conn->login_resp_buf = ib_conn->login_buf + ISCSI_DEF_MAX_RECV_SEG_LEN;
 203
 204         ib_conn->login_req_dma = ib_dma_map_single(ib_conn->device->ib_device,
 205                                 (void *)ib_conn->login_req_buf,
 206                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
 207
 208         ib_conn->login_resp_dma = ib_dma_map_single(ib_conn->device->ib_device,
 209                                 (void *)ib_conn->login_resp_buf,
 210                                 ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
 211
 212         req_err  = ib_dma_mapping_error(device->ib_device, ib_conn->login_req_dma);
 213         resp_err = ib_dma_mapping_error(device->ib_device, ib_conn->login_resp_dma);
 214
 215         if (req_err || resp_err) {
 216                 if (req_err)
 217                         ib_conn->login_req_dma = 0;
 218                 if (resp_err)
 219                         ib_conn->login_resp_dma = 0;
 220                 goto out_err;
 221         }
 222
 223         ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
 224                                     (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
 225                                     GFP_KERNEL);
 226         if (!ib_conn->page_vec)
 227                 goto out_err;
 228
 229         ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
 230
 231         params.page_shift        = SHIFT_4K;
 232         /* when the first/last SG element are not start/end *
 233          * page aligned, the map whould be of N+1 pages     */
 234         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 235         /* make the pool size twice the max number of SCSI commands *
 236          * the ML is expected to queue, watermark for unmap at 50%  */
 237         params.pool_size         = ISCSI_DEF_XMIT_CMDS_MAX * 2;
 238         params.dirty_watermark   = ISCSI_DEF_XMIT_CMDS_MAX;
 239         params.cache             = 0;
 240         params.flush_function    = NULL;
 241         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 242                                     IB_ACCESS_REMOTE_WRITE |
 243                                     IB_ACCESS_REMOTE_READ);
 244
 245         ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
 246         ret = PTR_ERR(ib_conn->fmr_pool);
 247         if (IS_ERR(ib_conn->fmr_pool) && ret != -ENOSYS) {
 248                 ib_conn->fmr_pool = NULL;
 249                 goto out_err;
 250         } else if (ret == -ENOSYS) {
 251                 ib_conn->fmr_pool = NULL;
 252                 iser_warn("FMRs are not supported, using unaligned mode\n");
 253                 ret = 0;
 254         }
 255
 256         memset(&init_attr, 0, sizeof init_attr);
 257
 258         mutex_lock(&ig.connlist_mutex);
 259         /* select the CQ with the minimal number of usages */
 260         for (index = 0; index < device->cqs_used; index++)
 261                 if (device->cq_active_qps[index] <
 262                     device->cq_active_qps[min_index])
 263                         min_index = index;
 264         device->cq_active_qps[min_index]++;
 265         mutex_unlock(&ig.connlist_mutex);
 266         iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 267
 268         init_attr.event_handler = iser_qp_event_callback;
 269         init_attr.qp_context    = (void *)ib_conn;
 270         init_attr.send_cq       = device->tx_cq[min_index];
 271         init_attr.recv_cq       = device->rx_cq[min_index];
 272         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 273         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 274         init_attr.cap.max_send_sge = 2;
 275         init_attr.cap.max_recv_sge = 1;
 276         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 277         init_attr.qp_type       = IB_QPT_RC;
 278
 279         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 280         if (ret)
 281                 goto out_err;
 282
 283         ib_conn->qp = ib_conn->cma_id->qp;
 284         iser_info("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
 285                   ib_conn, ib_conn->cma_id,
 286                   ib_conn->fmr_pool, ib_conn->cma_id->qp);
 287         return ret;
 288
 289 out_err:
 290         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 291         return ret;
 292 }
 293
 294 /**
 295  * releases the FMR pool, QP and CMA ID objects, returns 0 on success,
 296  * -1 on failure
 297  */
 298 static int iser_free_ib_conn_res(struct iser_conn *ib_conn, int can_destroy_id)
 299 {
 300         int cq_index;
 301         BUG_ON(ib_conn == NULL);
 302
 303         iser_info("freeing conn %p cma_id %p fmr pool %p qp %p\n",
 304                   ib_conn, ib_conn->cma_id,
 305                   ib_conn->fmr_pool, ib_conn->qp);
 306
 307         /* qp is created only once both addr & route are resolved */
 308         if (ib_conn->fmr_pool != NULL)
 309                 ib_destroy_fmr_pool(ib_conn->fmr_pool);
 310
 311         if (ib_conn->qp != NULL) {
 312                 cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
 313                 ib_conn->device->cq_active_qps[cq_index]--;
 314
 315                 rdma_destroy_qp(ib_conn->cma_id);
 316         }
 317         /* if cma handler context, the caller acts s.t the cma destroy the id */
 318         if (ib_conn->cma_id != NULL && can_destroy_id)
 319                 rdma_destroy_id(ib_conn->cma_id);
 320
 321         ib_conn->fmr_pool = NULL;
 322         ib_conn->qp       = NULL;
 323         ib_conn->cma_id   = NULL;
 324         kfree(ib_conn->page_vec);
 325
 326         if (ib_conn->login_buf) {
 327                 if (ib_conn->login_req_dma)
 328                         ib_dma_unmap_single(ib_conn->device->ib_device,
 329                                 ib_conn->login_req_dma,
 330                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
 331                 if (ib_conn->login_resp_dma)
 332                         ib_dma_unmap_single(ib_conn->device->ib_device,
 333                                 ib_conn->login_resp_dma,
 334                                 ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
 335                 kfree(ib_conn->login_buf);
 336         }
 337
 338         return 0;
 339 }
 340
 341 /**
 342  * based on the resolved device node GUID see if there already allocated
 343  * device for this device. If there's no such, create one.
 344  */
 345 static
 346 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 347 {
 348         struct iser_device *device;
 349
 350         mutex_lock(&ig.device_list_mutex);
 351
 352         list_for_each_entry(device, &ig.device_list, ig_list)
 353                 /* find if there's a match using the node GUID */
 354                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 355                         goto inc_refcnt;
 356
 357         device = kzalloc(sizeof *device, GFP_KERNEL);
 358         if (device == NULL)
 359                 goto out;
 360
 361         /* assign this device to the device */
 362         device->ib_device = cma_id->device;
 363         /* init the device and link it into ig device list */
 364         if (iser_create_device_ib_res(device)) {
 365                 kfree(device);
 366                 device = NULL;
 367                 goto out;
 368         }
 369         list_add(&device->ig_list, &ig.device_list);
 370
 371 inc_refcnt:
 372         device->refcount++;
 373 out:
 374         mutex_unlock(&ig.device_list_mutex);
 375         return device;
 376 }
 377
 378 /* if there's no demand for this device, release it */
 379 static void iser_device_try_release(struct iser_device *device)
 380 {
 381         mutex_lock(&ig.device_list_mutex);
 382         device->refcount--;
 383         iser_info("device %p refcount %d\n", device, device->refcount);
 384         if (!device->refcount) {
 385                 iser_free_device_ib_res(device);
 386                 list_del(&device->ig_list);
 387                 kfree(device);
 388         }
 389         mutex_unlock(&ig.device_list_mutex);
 390 }
 391
 392 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
 393                                      enum iser_ib_conn_state comp,
 394                                      enum iser_ib_conn_state exch)
 395 {
 396         int ret;
 397
 398         spin_lock_bh(&ib_conn->lock);
 399         if ((ret = (ib_conn->state == comp)))
 400                 ib_conn->state = exch;
 401         spin_unlock_bh(&ib_conn->lock);
 402         return ret;
 403 }
 404
 405 /**
 406  * Frees all conn objects and deallocs conn descriptor
 407  */
 408 static void iser_conn_release(struct iser_conn *ib_conn, int can_destroy_id)
 409 {
 410         struct iser_device  *device = ib_conn->device;
 411
 412         BUG_ON(ib_conn->state != ISER_CONN_DOWN);
 413
 414         mutex_lock(&ig.connlist_mutex);
 415         list_del(&ib_conn->conn_list);
 416         mutex_unlock(&ig.connlist_mutex);
 417         iser_free_rx_descriptors(ib_conn);
 418         iser_free_ib_conn_res(ib_conn, can_destroy_id);
 419         ib_conn->device = NULL;
 420         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
 421         if (device != NULL)
 422                 iser_device_try_release(device);
 423         iscsi_destroy_endpoint(ib_conn->ep);
 424 }
 425
 426 void iser_conn_get(struct iser_conn *ib_conn)
 427 {
 428         atomic_inc(&ib_conn->refcount);
 429 }
 430
 431 int iser_conn_put(struct iser_conn *ib_conn, int can_destroy_id)
 432 {
 433         if (atomic_dec_and_test(&ib_conn->refcount)) {
 434                 iser_conn_release(ib_conn, can_destroy_id);
 435                 return 1;
 436         }
 437         return 0;
 438 }
 439
 440 /**
 441  * triggers start of the disconnect procedures and wait for them to be done
 442  */
 443 void iser_conn_terminate(struct iser_conn *ib_conn)
 444 {
 445         int err = 0;
 446
 447         /* change the ib conn state only if the conn is UP, however always call
 448          * rdma_disconnect since this is the only way to cause the CMA to change
 449          * the QP state to ERROR
 450          */
 451
 452         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
 453         err = rdma_disconnect(ib_conn->cma_id);
 454         if (err)
 455                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 456                          ib_conn,err);
 457
 458         wait_event_interruptible(ib_conn->wait,
 459                                  ib_conn->state == ISER_CONN_DOWN);
 460
 461         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 462 }
 463
 464 static int iser_connect_error(struct rdma_cm_id *cma_id)
 465 {
 466         struct iser_conn *ib_conn;
 467         ib_conn = (struct iser_conn *)cma_id->context;
 468
 469         ib_conn->state = ISER_CONN_DOWN;
 470         wake_up_interruptible(&ib_conn->wait);
 471         return iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 472 }
 473
 474 static int iser_addr_handler(struct rdma_cm_id *cma_id)
 475 {
 476         struct iser_device *device;
 477         struct iser_conn   *ib_conn;
 478         int    ret;
 479
 480         device = iser_device_find_by_ib_device(cma_id);
 481         if (!device) {
 482                 iser_err("device lookup/creation failed\n");
 483                 return iser_connect_error(cma_id);
 484         }
 485
 486         ib_conn = (struct iser_conn *)cma_id->context;
 487         ib_conn->device = device;
 488
 489         ret = rdma_resolve_route(cma_id, 1000);
 490         if (ret) {
 491                 iser_err("resolve route failed: %d\n", ret);
 492                 return iser_connect_error(cma_id);
 493         }
 494
 495         return 0;
 496 }
 497
 498 static int iser_route_handler(struct rdma_cm_id *cma_id)
 499 {
 500         struct rdma_conn_param conn_param;
 501         int    ret;
 502         struct iser_cm_hdr req_hdr;
 503
 504         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
 505         if (ret)
 506                 goto failure;
 507
 508         memset(&conn_param, 0, sizeof conn_param);
 509         conn_param.responder_resources = 4;
 510         conn_param.initiator_depth     = 1;
 511         conn_param.retry_count         = 7;
 512         conn_param.rnr_retry_count     = 6;
 513
 514         memset(&req_hdr, 0, sizeof(req_hdr));
 515         req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
 516                         ISER_SEND_W_INV_NOT_SUPPORTED);
 517         conn_param.private_data         = (void *)&req_hdr;
 518         conn_param.private_data_len     = sizeof(struct iser_cm_hdr);
 519
 520         ret = rdma_connect(cma_id, &conn_param);
 521         if (ret) {
 522                 iser_err("failure connecting: %d\n", ret);
 523                 goto failure;
 524         }
 525
 526         return 0;
 527 failure:
 528         return iser_connect_error(cma_id);
 529 }
 530
 531 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 532 {
 533         struct iser_conn *ib_conn;
 534
 535         ib_conn = (struct iser_conn *)cma_id->context;
 536         ib_conn->state = ISER_CONN_UP;
 537         wake_up_interruptible(&ib_conn->wait);
 538 }
 539
 540 static int iser_disconnected_handler(struct rdma_cm_id *cma_id)
 541 {
 542         struct iser_conn *ib_conn;
 543         int ret;
 544
 545         ib_conn = (struct iser_conn *)cma_id->context;
 546
 547         /* getting here when the state is UP means that the conn is being *
 548          * terminated asynchronously from the iSCSI layer's perspective.  */
 549         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 550                                       ISER_CONN_TERMINATING))
 551                 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 552                                    ISCSI_ERR_CONN_FAILED);
 553
 554         /* Complete the termination process if no posts are pending */
 555         if (ib_conn->post_recv_buf_count == 0 &&
 556             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
 557                 ib_conn->state = ISER_CONN_DOWN;
 558                 wake_up_interruptible(&ib_conn->wait);
 559         }
 560
 561         ret = iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
 562         return ret;
 563 }
 564
 565 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 566 {
 567         int ret = 0;
 568
 569         iser_info("event %d status %d conn %p id %p\n",
 570                   event->event, event->status, cma_id->context, cma_id);
 571
 572         switch (event->event) {
 573         case RDMA_CM_EVENT_ADDR_RESOLVED:
 574                 ret = iser_addr_handler(cma_id);
 575                 break;
 576         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 577                 ret = iser_route_handler(cma_id);
 578                 break;
 579         case RDMA_CM_EVENT_ESTABLISHED:
 580                 iser_connected_handler(cma_id);
 581                 break;
 582         case RDMA_CM_EVENT_ADDR_ERROR:
 583         case RDMA_CM_EVENT_ROUTE_ERROR:
 584         case RDMA_CM_EVENT_CONNECT_ERROR:
 585         case RDMA_CM_EVENT_UNREACHABLE:
 586         case RDMA_CM_EVENT_REJECTED:
 587                 ret = iser_connect_error(cma_id);
 588                 break;
 589         case RDMA_CM_EVENT_DISCONNECTED:
 590         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 591         case RDMA_CM_EVENT_ADDR_CHANGE:
 592                 ret = iser_disconnected_handler(cma_id);
 593                 break;
 594         default:
 595                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 596                 break;
 597         }
 598         return ret;
 599 }
 600
 601 void iser_conn_init(struct iser_conn *ib_conn)
 602 {
 603         ib_conn->state = ISER_CONN_INIT;
 604         init_waitqueue_head(&ib_conn->wait);
 605         ib_conn->post_recv_buf_count = 0;
 606         atomic_set(&ib_conn->post_send_buf_count, 0);
 607         atomic_set(&ib_conn->refcount, 1); /* ref ib conn allocation */
 608         INIT_LIST_HEAD(&ib_conn->conn_list);
 609         spin_lock_init(&ib_conn->lock);
 610 }
 611
 612  /**
 613  * starts the process of connecting to the target
 614  * sleeps until the connection is established or rejected
 615  */
 616 int iser_connect(struct iser_conn   *ib_conn,
 617                  struct sockaddr_in *src_addr,
 618                  struct sockaddr_in *dst_addr,
 619                  int                 non_blocking)
 620 {
 621         struct sockaddr *src, *dst;
 622         int err = 0;
 623
 624         sprintf(ib_conn->name, "%pI4:%d",
 625                 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
 626
 627         /* the device is known only --after-- address resolution */
 628         ib_conn->device = NULL;
 629
 630         iser_info("connecting to: %pI4, port 0x%x\n",
 631                   &dst_addr->sin_addr, dst_addr->sin_port);
 632
 633         ib_conn->state = ISER_CONN_PENDING;
 634
 635         iser_conn_get(ib_conn); /* ref ib conn's cma id */
 636         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 637                                              (void *)ib_conn,
 638                                              RDMA_PS_TCP, IB_QPT_RC);
 639         if (IS_ERR(ib_conn->cma_id)) {
 640                 err = PTR_ERR(ib_conn->cma_id);
 641                 iser_err("rdma_create_id failed: %d\n", err);
 642                 goto id_failure;
 643         }
 644
 645         src = (struct sockaddr *)src_addr;
 646         dst = (struct sockaddr *)dst_addr;
 647         err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
 648         if (err) {
 649                 iser_err("rdma_resolve_addr failed: %d\n", err);
 650                 goto addr_failure;
 651         }
 652
 653         if (!non_blocking) {
 654                 wait_event_interruptible(ib_conn->wait,
 655                                          (ib_conn->state != ISER_CONN_PENDING));
 656
 657                 if (ib_conn->state != ISER_CONN_UP) {
 658                         err =  -EIO;
 659                         goto connect_failure;
 660                 }
 661         }
 662
 663         mutex_lock(&ig.connlist_mutex);
 664         list_add(&ib_conn->conn_list, &ig.connlist);
 665         mutex_unlock(&ig.connlist_mutex);
 666         return 0;
 667
 668 id_failure:
 669         ib_conn->cma_id = NULL;
 670 addr_failure:
 671         ib_conn->state = ISER_CONN_DOWN;
 672         iser_conn_put(ib_conn, 1); /* deref ib conn's cma id */
 673 connect_failure:
 674         iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
 675         return err;
 676 }
 677
 678 /**
 679  * iser_reg_page_vec - Register physical memory
 680  *
 681  * returns: 0 on success, errno code on failure
 682  */
 683 int iser_reg_page_vec(struct iser_conn     *ib_conn,
 684                       struct iser_page_vec *page_vec,
 685                       struct iser_mem_reg  *mem_reg)
 686 {
 687         struct ib_pool_fmr *mem;
 688         u64                io_addr;
 689         u64                *page_list;
 690         int                status;
 691
 692         page_list = page_vec->pages;
 693         io_addr   = page_list[0];
 694
 695         mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
 696                                     page_list,
 697                                     page_vec->length,
 698                                     io_addr);
 699
 700         if (IS_ERR(mem)) {
 701                 status = (int)PTR_ERR(mem);
 702                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 703                 return status;
 704         }
 705
 706         mem_reg->lkey  = mem->fmr->lkey;
 707         mem_reg->rkey  = mem->fmr->rkey;
 708         mem_reg->len   = page_vec->length * SIZE_4K;
 709         mem_reg->va    = io_addr;
 710         mem_reg->is_fmr = 1;
 711         mem_reg->mem_h = (void *)mem;
 712
 713         mem_reg->va   += page_vec->offset;
 714         mem_reg->len   = page_vec->data_size;
 715
 716         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 717                  "entry[0]: (0x%08lx,%ld)] -> "
 718                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 719                  page_vec, page_vec->length,
 720                  (unsigned long)page_vec->pages[0],
 721                  (unsigned long)page_vec->data_size,
 722                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 723                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 724         return 0;
 725 }
 726
 727 /**
 728  * Unregister (previosuly registered) memory.
 729  */
 730 void iser_unreg_mem(struct iser_mem_reg *reg)
 731 {
 732         int ret;
 733
 734         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
 735
 736         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
 737         if (ret)
 738                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
 739
 740         reg->mem_h = NULL;
 741 }
 742
 743 int iser_post_recvl(struct iser_conn *ib_conn)
 744 {
 745         struct ib_recv_wr rx_wr, *rx_wr_failed;
 746         struct ib_sge     sge;
 747         int ib_ret;
 748
 749         sge.addr   = ib_conn->login_resp_dma;
 750         sge.length = ISER_RX_LOGIN_SIZE;
 751         sge.lkey   = ib_conn->device->mr->lkey;
 752
 753         rx_wr.wr_id   = (unsigned long)ib_conn->login_resp_buf;
 754         rx_wr.sg_list = &sge;
 755         rx_wr.num_sge = 1;
 756         rx_wr.next    = NULL;
 757
 758         ib_conn->post_recv_buf_count++;
 759         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
 760         if (ib_ret) {
 761                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 762                 ib_conn->post_recv_buf_count--;
 763         }
 764         return ib_ret;
 765 }
 766
 767 int iser_post_recvm(struct iser_conn *ib_conn, int count)
 768 {
 769         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 770         int i, ib_ret;
 771         unsigned int my_rx_head = ib_conn->rx_desc_head;
 772         struct iser_rx_desc *rx_desc;
 773
 774         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 775                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
 776                 rx_wr->wr_id    = (unsigned long)rx_desc;
 777                 rx_wr->sg_list  = &rx_desc->rx_sg;
 778                 rx_wr->num_sge  = 1;
 779                 rx_wr->next     = rx_wr + 1;
 780                 my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
 781         }
 782
 783         rx_wr--;
 784         rx_wr->next = NULL; /* mark end of work requests list */
 785
 786         ib_conn->post_recv_buf_count += count;
 787         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
 788         if (ib_ret) {
 789                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 790                 ib_conn->post_recv_buf_count -= count;
 791         } else
 792                 ib_conn->rx_desc_head = my_rx_head;
 793         return ib_ret;
 794 }
 795
 796
 797 /**
 798  * iser_start_send - Initiate a Send DTO operation
 799  *
 800  * returns 0 on success, -1 on failure
 801  */
 802 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
 803 {
 804         int               ib_ret;
 805         struct ib_send_wr send_wr, *send_wr_failed;
 806
 807         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
 808                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
 809
 810         send_wr.next       = NULL;
 811         send_wr.wr_id      = (unsigned long)tx_desc;
 812         send_wr.sg_list    = tx_desc->tx_sg;
 813         send_wr.num_sge    = tx_desc->num_sge;
 814         send_wr.opcode     = IB_WR_SEND;
 815         send_wr.send_flags = IB_SEND_SIGNALED;
 816
 817         atomic_inc(&ib_conn->post_send_buf_count);
 818
 819         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
 820         if (ib_ret) {
 821                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
 822                 atomic_dec(&ib_conn->post_send_buf_count);
 823         }
 824         return ib_ret;
 825 }
 826
 827 static void iser_handle_comp_error(struct iser_tx_desc *desc,
 828                                 struct iser_conn *ib_conn)
 829 {
 830         if (desc && desc->type == ISCSI_TX_DATAOUT)
 831                 kmem_cache_free(ig.desc_cache, desc);
 832
 833         if (ib_conn->post_recv_buf_count == 0 &&
 834             atomic_read(&ib_conn->post_send_buf_count) == 0) {
 835                 /* getting here when the state is UP means that the conn is *
 836                  * being terminated asynchronously from the iSCSI layer's   *
 837                  * perspective.                                             */
 838                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
 839                     ISER_CONN_TERMINATING))
 840                         iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
 841                                            ISCSI_ERR_CONN_FAILED);
 842
 843                 /* no more non completed posts to the QP, complete the
 844                  * termination process w.o worrying on disconnect event */
 845                 ib_conn->state = ISER_CONN_DOWN;
 846                 wake_up_interruptible(&ib_conn->wait);
 847         }
 848 }
 849
 850 static int iser_drain_tx_cq(struct iser_device  *device, int cq_index)
 851 {
 852         struct ib_cq  *cq = device->tx_cq[cq_index];
 853         struct ib_wc  wc;
 854         struct iser_tx_desc *tx_desc;
 855         struct iser_conn *ib_conn;
 856         int completed_tx = 0;
 857
 858         while (ib_poll_cq(cq, 1, &wc) == 1) {
 859                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
 860                 ib_conn = wc.qp->qp_context;
 861                 if (wc.status == IB_WC_SUCCESS) {
 862                         if (wc.opcode == IB_WC_SEND)
 863                                 iser_snd_completion(tx_desc, ib_conn);
 864                         else
 865                                 iser_err("expected opcode %d got %d\n",
 866                                         IB_WC_SEND, wc.opcode);
 867                 } else {
 868                         iser_err("tx id %llx status %d vend_err %x\n",
 869                                 wc.wr_id, wc.status, wc.vendor_err);
 870                         atomic_dec(&ib_conn->post_send_buf_count);
 871                         iser_handle_comp_error(tx_desc, ib_conn);
 872                 }
 873                 completed_tx++;
 874         }
 875         return completed_tx;
 876 }
 877
 878
 879 static void iser_cq_tasklet_fn(unsigned long data)
 880 {
 881         struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
 882         struct iser_device  *device = cq_desc->device;
 883         int cq_index = cq_desc->cq_index;
 884         struct ib_cq         *cq = device->rx_cq[cq_index];
 885          struct ib_wc        wc;
 886          struct iser_rx_desc *desc;
 887          unsigned long       xfer_len;
 888         struct iser_conn *ib_conn;
 889         int completed_tx, completed_rx;
 890         completed_tx = completed_rx = 0;
 891
 892         while (ib_poll_cq(cq, 1, &wc) == 1) {
 893                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
 894                 BUG_ON(desc == NULL);
 895                 ib_conn = wc.qp->qp_context;
 896                 if (wc.status == IB_WC_SUCCESS) {
 897                         if (wc.opcode == IB_WC_RECV) {
 898                                 xfer_len = (unsigned long)wc.byte_len;
 899                                 iser_rcv_completion(desc, xfer_len, ib_conn);
 900                         } else
 901                                 iser_err("expected opcode %d got %d\n",
 902                                         IB_WC_RECV, wc.opcode);
 903                 } else {
 904                         if (wc.status != IB_WC_WR_FLUSH_ERR)
 905                                 iser_err("rx id %llx status %d vend_err %x\n",
 906                                         wc.wr_id, wc.status, wc.vendor_err);
 907                         ib_conn->post_recv_buf_count--;
 908                         iser_handle_comp_error(NULL, ib_conn);
 909                 }
 910                 completed_rx++;
 911                 if (!(completed_rx & 63))
 912                         completed_tx += iser_drain_tx_cq(device, cq_index);
 913         }
 914         /* #warning "it is assumed here that arming CQ only once its empty" *
 915          * " would not cause interrupts to be missed"                       */
 916         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 917
 918         completed_tx += iser_drain_tx_cq(device, cq_index);
 919         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
 920 }
 921
 922 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
 923 {
 924         struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
 925         struct iser_device  *device = cq_desc->device;
 926         int cq_index = cq_desc->cq_index;
 927
 928         tasklet_schedule(&device->cq_tasklet[cq_index]);
 929 }