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  * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34 #include <linux/kernel.h>
  35 #include <linux/module.h>
  36 #include <linux/slab.h>
  37 #include <linux/delay.h>
  38
  39 #include "iscsi_iser.h"
  40
  41 #define ISCSI_ISER_MAX_CONN     8
  42 #define ISER_MAX_RX_LEN         (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
  43 #define ISER_MAX_TX_LEN         (ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
  44 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_LEN + ISER_MAX_TX_LEN)
  45
  46 static int iser_cq_poll_limit = 512;
  47
  48 static void iser_cq_tasklet_fn(unsigned long data);
  49 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
  50
  51 static void iser_cq_event_callback(struct ib_event *cause, void *context)
  52 {
  53         iser_err("got cq event %d \n", cause->event);
  54 }
  55
  56 static void iser_qp_event_callback(struct ib_event *cause, void *context)
  57 {
  58         iser_err("got qp event %d\n",cause->event);
  59 }
  60
  61 static void iser_event_handler(struct ib_event_handler *handler,
  62                                 struct ib_event *event)
  63 {
  64         iser_err("async event %d on device %s port %d\n", event->event,
  65                 event->device->name, event->element.port_num);
  66 }
  67
  68 /**
  69  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  70  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  71  * the adapator.
  72  *
  73  * returns 0 on success, -1 on failure
  74  */
  75 static int iser_create_device_ib_res(struct iser_device *device)
  76 {
  77         struct ib_device_attr *dev_attr = &device->dev_attr;
  78         int ret, i;
  79
  80         ret = ib_query_device(device->ib_device, dev_attr);
  81         if (ret) {
  82                 pr_warn("Query device failed for %s\n", device->ib_device->name);
  83                 return ret;
  84         }
  85
  86         /* Assign function handles  - based on FMR support */
  87         if (device->ib_device->alloc_fmr && device->ib_device->dealloc_fmr &&
  88             device->ib_device->map_phys_fmr && device->ib_device->unmap_fmr) {
  89                 iser_info("FMR supported, using FMR for registration\n");
  90                 device->iser_alloc_rdma_reg_res = iser_create_fmr_pool;
  91                 device->iser_free_rdma_reg_res = iser_free_fmr_pool;
  92                 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fmr;
  93                 device->iser_unreg_rdma_mem = iser_unreg_mem_fmr;
  94         } else
  95         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
  96                 iser_info("FastReg supported, using FastReg for registration\n");
  97                 device->iser_alloc_rdma_reg_res = iser_create_fastreg_pool;
  98                 device->iser_free_rdma_reg_res = iser_free_fastreg_pool;
  99                 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fastreg;
 100                 device->iser_unreg_rdma_mem = iser_unreg_mem_fastreg;
 101         } else {
 102                 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
 103                 return -1;
 104         }
 105
 106         device->comps_used = min(ISER_MAX_CQ,
 107                                  device->ib_device->num_comp_vectors);
 108         iser_info("using %d CQs, device %s supports %d vectors\n",
 109                   device->comps_used, device->ib_device->name,
 110                   device->ib_device->num_comp_vectors);
 111
 112         device->pd = ib_alloc_pd(device->ib_device);
 113         if (IS_ERR(device->pd))
 114                 goto pd_err;
 115
 116         for (i = 0; i < device->comps_used; i++) {
 117                 struct iser_comp *comp = &device->comps[i];
 118
 119                 comp->device = device;
 120                 comp->cq = ib_create_cq(device->ib_device,
 121                                         iser_cq_callback,
 122                                         iser_cq_event_callback,
 123                                         (void *)comp,
 124                                         ISER_MAX_CQ_LEN, i);
 125                 if (IS_ERR(comp->cq)) {
 126                         comp->cq = NULL;
 127                         goto cq_err;
 128                 }
 129
 130                 if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
 131                         goto cq_err;
 132
 133                 tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
 134                              (unsigned long)comp);
 135         }
 136
 137         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
 138                                    IB_ACCESS_REMOTE_WRITE |
 139                                    IB_ACCESS_REMOTE_READ);
 140         if (IS_ERR(device->mr))
 141                 goto dma_mr_err;
 142
 143         INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
 144                                 iser_event_handler);
 145         if (ib_register_event_handler(&device->event_handler))
 146                 goto handler_err;
 147
 148         return 0;
 149
 150 handler_err:
 151         ib_dereg_mr(device->mr);
 152 dma_mr_err:
 153         for (i = 0; i < device->comps_used; i++)
 154                 tasklet_kill(&device->comps[i].tasklet);
 155 cq_err:
 156         for (i = 0; i < device->comps_used; i++) {
 157                 struct iser_comp *comp = &device->comps[i];
 158
 159                 if (comp->cq)
 160                         ib_destroy_cq(comp->cq);
 161         }
 162         ib_dealloc_pd(device->pd);
 163 pd_err:
 164         iser_err("failed to allocate an IB resource\n");
 165         return -1;
 166 }
 167
 168 /**
 169  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 170  * CQ and PD created with the device associated with the adapator.
 171  */
 172 static void iser_free_device_ib_res(struct iser_device *device)
 173 {
 174         int i;
 175         BUG_ON(device->mr == NULL);
 176
 177         for (i = 0; i < device->comps_used; i++) {
 178                 struct iser_comp *comp = &device->comps[i];
 179
 180                 tasklet_kill(&comp->tasklet);
 181                 ib_destroy_cq(comp->cq);
 182                 comp->cq = NULL;
 183         }
 184
 185         (void)ib_unregister_event_handler(&device->event_handler);
 186         (void)ib_dereg_mr(device->mr);
 187         (void)ib_dealloc_pd(device->pd);
 188
 189         device->mr = NULL;
 190         device->pd = NULL;
 191 }
 192
 193 /**
 194  * iser_create_fmr_pool - Creates FMR pool and page_vector
 195  *
 196  * returns 0 on success, or errno code on failure
 197  */
 198 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max)
 199 {
 200         struct iser_device *device = ib_conn->device;
 201         struct ib_fmr_pool_param params;
 202         int ret = -ENOMEM;
 203
 204         ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
 205                                         (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
 206                                         GFP_KERNEL);
 207         if (!ib_conn->fmr.page_vec)
 208                 return ret;
 209
 210         ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
 211
 212         params.page_shift        = SHIFT_4K;
 213         /* when the first/last SG element are not start/end *
 214          * page aligned, the map whould be of N+1 pages     */
 215         params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
 216         /* make the pool size twice the max number of SCSI commands *
 217          * the ML is expected to queue, watermark for unmap at 50%  */
 218         params.pool_size         = cmds_max * 2;
 219         params.dirty_watermark   = cmds_max;
 220         params.cache             = 0;
 221         params.flush_function    = NULL;
 222         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 223                                     IB_ACCESS_REMOTE_WRITE |
 224                                     IB_ACCESS_REMOTE_READ);
 225
 226         ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, &params);
 227         if (!IS_ERR(ib_conn->fmr.pool))
 228                 return 0;
 229
 230         /* no FMR => no need for page_vec */
 231         kfree(ib_conn->fmr.page_vec);
 232         ib_conn->fmr.page_vec = NULL;
 233
 234         ret = PTR_ERR(ib_conn->fmr.pool);
 235         ib_conn->fmr.pool = NULL;
 236         if (ret != -ENOSYS) {
 237                 iser_err("FMR allocation failed, err %d\n", ret);
 238                 return ret;
 239         } else {
 240                 iser_warn("FMRs are not supported, using unaligned mode\n");
 241                 return 0;
 242         }
 243 }
 244
 245 /**
 246  * iser_free_fmr_pool - releases the FMR pool and page vec
 247  */
 248 void iser_free_fmr_pool(struct ib_conn *ib_conn)
 249 {
 250         iser_info("freeing conn %p fmr pool %p\n",
 251                   ib_conn, ib_conn->fmr.pool);
 252
 253         if (ib_conn->fmr.pool != NULL)
 254                 ib_destroy_fmr_pool(ib_conn->fmr.pool);
 255
 256         ib_conn->fmr.pool = NULL;
 257
 258         kfree(ib_conn->fmr.page_vec);
 259         ib_conn->fmr.page_vec = NULL;
 260 }
 261
 262 static int
 263 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
 264                          bool pi_enable, struct fast_reg_descriptor *desc)
 265 {
 266         int ret;
 267
 268         desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
 269                                                       ISCSI_ISER_SG_TABLESIZE + 1);
 270         if (IS_ERR(desc->data_frpl)) {
 271                 ret = PTR_ERR(desc->data_frpl);
 272                 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
 273                          ret);
 274                 return PTR_ERR(desc->data_frpl);
 275         }
 276
 277         desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
 278         if (IS_ERR(desc->data_mr)) {
 279                 ret = PTR_ERR(desc->data_mr);
 280                 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
 281                 goto fast_reg_mr_failure;
 282         }
 283         desc->reg_indicators |= ISER_DATA_KEY_VALID;
 284
 285         if (pi_enable) {
 286                 struct ib_mr_init_attr mr_init_attr = {0};
 287                 struct iser_pi_context *pi_ctx = NULL;
 288
 289                 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 290                 if (!desc->pi_ctx) {
 291                         iser_err("Failed to allocate pi context\n");
 292                         ret = -ENOMEM;
 293                         goto pi_ctx_alloc_failure;
 294                 }
 295                 pi_ctx = desc->pi_ctx;
 296
 297                 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
 298                                                     ISCSI_ISER_SG_TABLESIZE);
 299                 if (IS_ERR(pi_ctx->prot_frpl)) {
 300                         ret = PTR_ERR(pi_ctx->prot_frpl);
 301                         iser_err("Failed to allocate prot frpl ret=%d\n",
 302                                  ret);
 303                         goto prot_frpl_failure;
 304                 }
 305
 306                 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
 307                                                 ISCSI_ISER_SG_TABLESIZE + 1);
 308                 if (IS_ERR(pi_ctx->prot_mr)) {
 309                         ret = PTR_ERR(pi_ctx->prot_mr);
 310                         iser_err("Failed to allocate prot frmr ret=%d\n",
 311                                  ret);
 312                         goto prot_mr_failure;
 313                 }
 314                 desc->reg_indicators |= ISER_PROT_KEY_VALID;
 315
 316                 mr_init_attr.max_reg_descriptors = 2;
 317                 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
 318                 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
 319                 if (IS_ERR(pi_ctx->sig_mr)) {
 320                         ret = PTR_ERR(pi_ctx->sig_mr);
 321                         iser_err("Failed to allocate signature enabled mr err=%d\n",
 322                                  ret);
 323                         goto sig_mr_failure;
 324                 }
 325                 desc->reg_indicators |= ISER_SIG_KEY_VALID;
 326         }
 327         desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
 328
 329         iser_dbg("Create fr_desc %p page_list %p\n",
 330                  desc, desc->data_frpl->page_list);
 331
 332         return 0;
 333 sig_mr_failure:
 334         ib_dereg_mr(desc->pi_ctx->prot_mr);
 335 prot_mr_failure:
 336         ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
 337 prot_frpl_failure:
 338         kfree(desc->pi_ctx);
 339 pi_ctx_alloc_failure:
 340         ib_dereg_mr(desc->data_mr);
 341 fast_reg_mr_failure:
 342         ib_free_fast_reg_page_list(desc->data_frpl);
 343
 344         return ret;
 345 }
 346
 347 /**
 348  * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
 349  * for fast registration work requests.
 350  * returns 0 on success, or errno code on failure
 351  */
 352 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
 353 {
 354         struct iser_device *device = ib_conn->device;
 355         struct fast_reg_descriptor *desc;
 356         int i, ret;
 357
 358         INIT_LIST_HEAD(&ib_conn->fastreg.pool);
 359         ib_conn->fastreg.pool_size = 0;
 360         for (i = 0; i < cmds_max; i++) {
 361                 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 362                 if (!desc) {
 363                         iser_err("Failed to allocate a new fast_reg descriptor\n");
 364                         ret = -ENOMEM;
 365                         goto err;
 366                 }
 367
 368                 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
 369                                                ib_conn->pi_support, desc);
 370                 if (ret) {
 371                         iser_err("Failed to create fastreg descriptor err=%d\n",
 372                                  ret);
 373                         kfree(desc);
 374                         goto err;
 375                 }
 376
 377                 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
 378                 ib_conn->fastreg.pool_size++;
 379         }
 380
 381         return 0;
 382
 383 err:
 384         iser_free_fastreg_pool(ib_conn);
 385         return ret;
 386 }
 387
 388 /**
 389  * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
 390  */
 391 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
 392 {
 393         struct fast_reg_descriptor *desc, *tmp;
 394         int i = 0;
 395
 396         if (list_empty(&ib_conn->fastreg.pool))
 397                 return;
 398
 399         iser_info("freeing conn %p fr pool\n", ib_conn);
 400
 401         list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
 402                 list_del(&desc->list);
 403                 ib_free_fast_reg_page_list(desc->data_frpl);
 404                 ib_dereg_mr(desc->data_mr);
 405                 if (desc->pi_ctx) {
 406                         ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
 407                         ib_dereg_mr(desc->pi_ctx->prot_mr);
 408                         ib_destroy_mr(desc->pi_ctx->sig_mr);
 409                         kfree(desc->pi_ctx);
 410                 }
 411                 kfree(desc);
 412                 ++i;
 413         }
 414
 415         if (i < ib_conn->fastreg.pool_size)
 416                 iser_warn("pool still has %d regions registered\n",
 417                           ib_conn->fastreg.pool_size - i);
 418 }
 419
 420 /**
 421  * iser_create_ib_conn_res - Queue-Pair (QP)
 422  *
 423  * returns 0 on success, -1 on failure
 424  */
 425 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
 426 {
 427         struct iser_device      *device;
 428         struct ib_qp_init_attr  init_attr;
 429         int                     ret = -ENOMEM;
 430         int index, min_index = 0;
 431
 432         BUG_ON(ib_conn->device == NULL);
 433
 434         device = ib_conn->device;
 435
 436         memset(&init_attr, 0, sizeof init_attr);
 437
 438         mutex_lock(&ig.connlist_mutex);
 439         /* select the CQ with the minimal number of usages */
 440         for (index = 0; index < device->comps_used; index++) {
 441                 if (device->comps[index].active_qps <
 442                     device->comps[min_index].active_qps)
 443                         min_index = index;
 444         }
 445         ib_conn->comp = &device->comps[min_index];
 446         ib_conn->comp->active_qps++;
 447         mutex_unlock(&ig.connlist_mutex);
 448         iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 449
 450         init_attr.event_handler = iser_qp_event_callback;
 451         init_attr.qp_context    = (void *)ib_conn;
 452         init_attr.send_cq       = ib_conn->comp->cq;
 453         init_attr.recv_cq       = ib_conn->comp->cq;
 454         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 455         init_attr.cap.max_send_sge = 2;
 456         init_attr.cap.max_recv_sge = 1;
 457         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 458         init_attr.qp_type       = IB_QPT_RC;
 459         if (ib_conn->pi_support) {
 460                 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS;
 461                 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 462         } else {
 463                 init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
 464         }
 465
 466         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 467         if (ret)
 468                 goto out_err;
 469
 470         ib_conn->qp = ib_conn->cma_id->qp;
 471         iser_info("setting conn %p cma_id %p qp %p\n",
 472                   ib_conn, ib_conn->cma_id,
 473                   ib_conn->cma_id->qp);
 474         return ret;
 475
 476 out_err:
 477         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 478         return ret;
 479 }
 480
 481 /**
 482  * based on the resolved device node GUID see if there already allocated
 483  * device for this device. If there's no such, create one.
 484  */
 485 static
 486 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 487 {
 488         struct iser_device *device;
 489
 490         mutex_lock(&ig.device_list_mutex);
 491
 492         list_for_each_entry(device, &ig.device_list, ig_list)
 493                 /* find if there's a match using the node GUID */
 494                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 495                         goto inc_refcnt;
 496
 497         device = kzalloc(sizeof *device, GFP_KERNEL);
 498         if (device == NULL)
 499                 goto out;
 500
 501         /* assign this device to the device */
 502         device->ib_device = cma_id->device;
 503         /* init the device and link it into ig device list */
 504         if (iser_create_device_ib_res(device)) {
 505                 kfree(device);
 506                 device = NULL;
 507                 goto out;
 508         }
 509         list_add(&device->ig_list, &ig.device_list);
 510
 511 inc_refcnt:
 512         device->refcount++;
 513 out:
 514         mutex_unlock(&ig.device_list_mutex);
 515         return device;
 516 }
 517
 518 /* if there's no demand for this device, release it */
 519 static void iser_device_try_release(struct iser_device *device)
 520 {
 521         mutex_lock(&ig.device_list_mutex);
 522         device->refcount--;
 523         iser_info("device %p refcount %d\n", device, device->refcount);
 524         if (!device->refcount) {
 525                 iser_free_device_ib_res(device);
 526                 list_del(&device->ig_list);
 527                 kfree(device);
 528         }
 529         mutex_unlock(&ig.device_list_mutex);
 530 }
 531
 532 /**
 533  * Called with state mutex held
 534  **/
 535 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
 536                                      enum iser_conn_state comp,
 537                                      enum iser_conn_state exch)
 538 {
 539         int ret;
 540
 541         ret = (iser_conn->state == comp);
 542         if (ret)
 543                 iser_conn->state = exch;
 544
 545         return ret;
 546 }
 547
 548 void iser_release_work(struct work_struct *work)
 549 {
 550         struct iser_conn *iser_conn;
 551
 552         iser_conn = container_of(work, struct iser_conn, release_work);
 553
 554         /* Wait for conn_stop to complete */
 555         wait_for_completion(&iser_conn->stop_completion);
 556         /* Wait for IB resouces cleanup to complete */
 557         wait_for_completion(&iser_conn->ib_completion);
 558
 559         mutex_lock(&iser_conn->state_mutex);
 560         iser_conn->state = ISER_CONN_DOWN;
 561         mutex_unlock(&iser_conn->state_mutex);
 562
 563         iser_conn_release(iser_conn);
 564 }
 565
 566 /**
 567  * iser_free_ib_conn_res - release IB related resources
 568  * @iser_conn: iser connection struct
 569  * @destroy_device: indicator if we need to try to release
 570  *     the iser device (only iscsi shutdown and DEVICE_REMOVAL
 571  *     will use this.
 572  *
 573  * This routine is called with the iser state mutex held
 574  * so the cm_id removal is out of here. It is Safe to
 575  * be invoked multiple times.
 576  */
 577 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
 578                                   bool destroy_device)
 579 {
 580         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 581         struct iser_device *device = ib_conn->device;
 582
 583         iser_info("freeing conn %p cma_id %p qp %p\n",
 584                   iser_conn, ib_conn->cma_id, ib_conn->qp);
 585
 586         iser_free_rx_descriptors(iser_conn);
 587
 588         if (ib_conn->qp != NULL) {
 589                 ib_conn->comp->active_qps--;
 590                 rdma_destroy_qp(ib_conn->cma_id);
 591                 ib_conn->qp = NULL;
 592         }
 593
 594         if (destroy_device && device != NULL) {
 595                 iser_device_try_release(device);
 596                 ib_conn->device = NULL;
 597         }
 598 }
 599
 600 /**
 601  * Frees all conn objects and deallocs conn descriptor
 602  */
 603 void iser_conn_release(struct iser_conn *iser_conn)
 604 {
 605         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 606
 607         mutex_lock(&ig.connlist_mutex);
 608         list_del(&iser_conn->conn_list);
 609         mutex_unlock(&ig.connlist_mutex);
 610
 611         mutex_lock(&iser_conn->state_mutex);
 612         if (iser_conn->state != ISER_CONN_DOWN)
 613                 iser_warn("iser conn %p state %d, expected state down.\n",
 614                           iser_conn, iser_conn->state);
 615         /*
 616          * In case we never got to bind stage, we still need to
 617          * release IB resources (which is safe to call more than once).
 618          */
 619         iser_free_ib_conn_res(iser_conn, true);
 620         mutex_unlock(&iser_conn->state_mutex);
 621
 622         if (ib_conn->cma_id != NULL) {
 623                 rdma_destroy_id(ib_conn->cma_id);
 624                 ib_conn->cma_id = NULL;
 625         }
 626
 627         kfree(iser_conn);
 628 }
 629
 630 /**
 631  * triggers start of the disconnect procedures and wait for them to be done
 632  * Called with state mutex held
 633  */
 634 int iser_conn_terminate(struct iser_conn *iser_conn)
 635 {
 636         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 637         int err = 0;
 638
 639         /* terminate the iser conn only if the conn state is UP */
 640         if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
 641                                        ISER_CONN_TERMINATING))
 642                 return 0;
 643
 644         iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
 645
 646         /* suspend queuing of new iscsi commands */
 647         if (iser_conn->iscsi_conn)
 648                 iscsi_suspend_queue(iser_conn->iscsi_conn);
 649
 650         /*
 651          * In case we didn't already clean up the cma_id (peer initiated
 652          * a disconnection), we need to Cause the CMA to change the QP
 653          * state to ERROR.
 654          */
 655         if (ib_conn->cma_id) {
 656                 err = rdma_disconnect(ib_conn->cma_id);
 657                 if (err)
 658                         iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 659                                  iser_conn, err);
 660
 661                 wait_for_completion(&ib_conn->flush_comp);
 662         }
 663
 664         return 1;
 665 }
 666
 667 /**
 668  * Called with state mutex held
 669  **/
 670 static void iser_connect_error(struct rdma_cm_id *cma_id)
 671 {
 672         struct iser_conn *iser_conn;
 673
 674         iser_conn = (struct iser_conn *)cma_id->context;
 675         iser_conn->state = ISER_CONN_DOWN;
 676 }
 677
 678 /**
 679  * Called with state mutex held
 680  **/
 681 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 682 {
 683         struct iser_device *device;
 684         struct iser_conn   *iser_conn;
 685         struct ib_conn   *ib_conn;
 686         int    ret;
 687
 688         iser_conn = (struct iser_conn *)cma_id->context;
 689         if (iser_conn->state != ISER_CONN_PENDING)
 690                 /* bailout */
 691                 return;
 692
 693         ib_conn = &iser_conn->ib_conn;
 694         device = iser_device_find_by_ib_device(cma_id);
 695         if (!device) {
 696                 iser_err("device lookup/creation failed\n");
 697                 iser_connect_error(cma_id);
 698                 return;
 699         }
 700
 701         ib_conn->device = device;
 702
 703         /* connection T10-PI support */
 704         if (iser_pi_enable) {
 705                 if (!(device->dev_attr.device_cap_flags &
 706                       IB_DEVICE_SIGNATURE_HANDOVER)) {
 707                         iser_warn("T10-PI requested but not supported on %s, "
 708                                   "continue without T10-PI\n",
 709                                   ib_conn->device->ib_device->name);
 710                         ib_conn->pi_support = false;
 711                 } else {
 712                         ib_conn->pi_support = true;
 713                 }
 714         }
 715
 716         ret = rdma_resolve_route(cma_id, 1000);
 717         if (ret) {
 718                 iser_err("resolve route failed: %d\n", ret);
 719                 iser_connect_error(cma_id);
 720                 return;
 721         }
 722 }
 723
 724 /**
 725  * Called with state mutex held
 726  **/
 727 static void iser_route_handler(struct rdma_cm_id *cma_id)
 728 {
 729         struct rdma_conn_param conn_param;
 730         int    ret;
 731         struct iser_cm_hdr req_hdr;
 732         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 733         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 734         struct iser_device *device = ib_conn->device;
 735
 736         if (iser_conn->state != ISER_CONN_PENDING)
 737                 /* bailout */
 738                 return;
 739
 740         ret = iser_create_ib_conn_res(ib_conn);
 741         if (ret)
 742                 goto failure;
 743
 744         memset(&conn_param, 0, sizeof conn_param);
 745         conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
 746         conn_param.initiator_depth     = 1;
 747         conn_param.retry_count         = 7;
 748         conn_param.rnr_retry_count     = 6;
 749
 750         memset(&req_hdr, 0, sizeof(req_hdr));
 751         req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
 752                         ISER_SEND_W_INV_NOT_SUPPORTED);
 753         conn_param.private_data         = (void *)&req_hdr;
 754         conn_param.private_data_len     = sizeof(struct iser_cm_hdr);
 755
 756         ret = rdma_connect(cma_id, &conn_param);
 757         if (ret) {
 758                 iser_err("failure connecting: %d\n", ret);
 759                 goto failure;
 760         }
 761
 762         return;
 763 failure:
 764         iser_connect_error(cma_id);
 765 }
 766
 767 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 768 {
 769         struct iser_conn *iser_conn;
 770         struct ib_qp_attr attr;
 771         struct ib_qp_init_attr init_attr;
 772
 773         iser_conn = (struct iser_conn *)cma_id->context;
 774         if (iser_conn->state != ISER_CONN_PENDING)
 775                 /* bailout */
 776                 return;
 777
 778         (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
 779         iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
 780
 781         iser_conn->state = ISER_CONN_UP;
 782         complete(&iser_conn->up_completion);
 783 }
 784
 785 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 786 {
 787         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 788
 789         if (iser_conn_terminate(iser_conn)) {
 790                 if (iser_conn->iscsi_conn)
 791                         iscsi_conn_failure(iser_conn->iscsi_conn,
 792                                            ISCSI_ERR_CONN_FAILED);
 793                 else
 794                         iser_err("iscsi_iser connection isn't bound\n");
 795         }
 796 }
 797
 798 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
 799                                  bool destroy_device)
 800 {
 801         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 802
 803         /*
 804          * We are not guaranteed that we visited disconnected_handler
 805          * by now, call it here to be safe that we handle CM drep
 806          * and flush errors.
 807          */
 808         iser_disconnected_handler(cma_id);
 809         iser_free_ib_conn_res(iser_conn, destroy_device);
 810         complete(&iser_conn->ib_completion);
 811 };
 812
 813 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 814 {
 815         struct iser_conn *iser_conn;
 816         int ret = 0;
 817
 818         iser_conn = (struct iser_conn *)cma_id->context;
 819         iser_info("event %d status %d conn %p id %p\n",
 820                   event->event, event->status, cma_id->context, cma_id);
 821
 822         mutex_lock(&iser_conn->state_mutex);
 823         switch (event->event) {
 824         case RDMA_CM_EVENT_ADDR_RESOLVED:
 825                 iser_addr_handler(cma_id);
 826                 break;
 827         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 828                 iser_route_handler(cma_id);
 829                 break;
 830         case RDMA_CM_EVENT_ESTABLISHED:
 831                 iser_connected_handler(cma_id);
 832                 break;
 833         case RDMA_CM_EVENT_ADDR_ERROR:
 834         case RDMA_CM_EVENT_ROUTE_ERROR:
 835         case RDMA_CM_EVENT_CONNECT_ERROR:
 836         case RDMA_CM_EVENT_UNREACHABLE:
 837         case RDMA_CM_EVENT_REJECTED:
 838                 iser_connect_error(cma_id);
 839                 break;
 840         case RDMA_CM_EVENT_DISCONNECTED:
 841         case RDMA_CM_EVENT_ADDR_CHANGE:
 842                 iser_disconnected_handler(cma_id);
 843                 break;
 844         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 845                 /*
 846                  * we *must* destroy the device as we cannot rely
 847                  * on iscsid to be around to initiate error handling.
 848                  * also implicitly destroy the cma_id.
 849                  */
 850                 iser_cleanup_handler(cma_id, true);
 851                 iser_conn->ib_conn.cma_id = NULL;
 852                 ret = 1;
 853                 break;
 854         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 855                 iser_cleanup_handler(cma_id, false);
 856                 break;
 857         default:
 858                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
 859                 break;
 860         }
 861         mutex_unlock(&iser_conn->state_mutex);
 862
 863         return ret;
 864 }
 865
 866 void iser_conn_init(struct iser_conn *iser_conn)
 867 {
 868         iser_conn->state = ISER_CONN_INIT;
 869         iser_conn->ib_conn.post_recv_buf_count = 0;
 870         atomic_set(&iser_conn->ib_conn.post_send_buf_count, 0);
 871         init_completion(&iser_conn->ib_conn.flush_comp);
 872         init_completion(&iser_conn->stop_completion);
 873         init_completion(&iser_conn->ib_completion);
 874         init_completion(&iser_conn->up_completion);
 875         INIT_LIST_HEAD(&iser_conn->conn_list);
 876         spin_lock_init(&iser_conn->ib_conn.lock);
 877         mutex_init(&iser_conn->state_mutex);
 878 }
 879
 880  /**
 881  * starts the process of connecting to the target
 882  * sleeps until the connection is established or rejected
 883  */
 884 int iser_connect(struct iser_conn   *iser_conn,
 885                  struct sockaddr    *src_addr,
 886                  struct sockaddr    *dst_addr,
 887                  int                 non_blocking)
 888 {
 889         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 890         int err = 0;
 891
 892         mutex_lock(&iser_conn->state_mutex);
 893
 894         sprintf(iser_conn->name, "%pISp", dst_addr);
 895
 896         iser_info("connecting to: %s\n", iser_conn->name);
 897
 898         /* the device is known only --after-- address resolution */
 899         ib_conn->device = NULL;
 900
 901         iser_conn->state = ISER_CONN_PENDING;
 902
 903         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
 904                                          (void *)iser_conn,
 905                                          RDMA_PS_TCP, IB_QPT_RC);
 906         if (IS_ERR(ib_conn->cma_id)) {
 907                 err = PTR_ERR(ib_conn->cma_id);
 908                 iser_err("rdma_create_id failed: %d\n", err);
 909                 goto id_failure;
 910         }
 911
 912         err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
 913         if (err) {
 914                 iser_err("rdma_resolve_addr failed: %d\n", err);
 915                 goto addr_failure;
 916         }
 917
 918         if (!non_blocking) {
 919                 wait_for_completion_interruptible(&iser_conn->up_completion);
 920
 921                 if (iser_conn->state != ISER_CONN_UP) {
 922                         err =  -EIO;
 923                         goto connect_failure;
 924                 }
 925         }
 926         mutex_unlock(&iser_conn->state_mutex);
 927
 928         mutex_lock(&ig.connlist_mutex);
 929         list_add(&iser_conn->conn_list, &ig.connlist);
 930         mutex_unlock(&ig.connlist_mutex);
 931         return 0;
 932
 933 id_failure:
 934         ib_conn->cma_id = NULL;
 935 addr_failure:
 936         iser_conn->state = ISER_CONN_DOWN;
 937 connect_failure:
 938         mutex_unlock(&iser_conn->state_mutex);
 939         iser_conn_release(iser_conn);
 940         return err;
 941 }
 942
 943 /**
 944  * iser_reg_page_vec - Register physical memory
 945  *
 946  * returns: 0 on success, errno code on failure
 947  */
 948 int iser_reg_page_vec(struct ib_conn *ib_conn,
 949                       struct iser_page_vec *page_vec,
 950                       struct iser_mem_reg  *mem_reg)
 951 {
 952         struct ib_pool_fmr *mem;
 953         u64                io_addr;
 954         u64                *page_list;
 955         int                status;
 956
 957         page_list = page_vec->pages;
 958         io_addr   = page_list[0];
 959
 960         mem  = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
 961                                     page_list,
 962                                     page_vec->length,
 963                                     io_addr);
 964
 965         if (IS_ERR(mem)) {
 966                 status = (int)PTR_ERR(mem);
 967                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
 968                 return status;
 969         }
 970
 971         mem_reg->lkey  = mem->fmr->lkey;
 972         mem_reg->rkey  = mem->fmr->rkey;
 973         mem_reg->len   = page_vec->length * SIZE_4K;
 974         mem_reg->va    = io_addr;
 975         mem_reg->is_mr = 1;
 976         mem_reg->mem_h = (void *)mem;
 977
 978         mem_reg->va   += page_vec->offset;
 979         mem_reg->len   = page_vec->data_size;
 980
 981         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
 982                  "entry[0]: (0x%08lx,%ld)] -> "
 983                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
 984                  page_vec, page_vec->length,
 985                  (unsigned long)page_vec->pages[0],
 986                  (unsigned long)page_vec->data_size,
 987                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
 988                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
 989         return 0;
 990 }
 991
 992 /**
 993  * Unregister (previosuly registered using FMR) memory.
 994  * If memory is non-FMR does nothing.
 995  */
 996 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
 997                         enum iser_data_dir cmd_dir)
 998 {
 999         struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1000         int ret;
1001
1002         if (!reg->is_mr)
1003                 return;
1004
1005         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
1006
1007         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
1008         if (ret)
1009                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
1010
1011         reg->mem_h = NULL;
1012 }
1013
1014 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
1015                             enum iser_data_dir cmd_dir)
1016 {
1017         struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1018         struct iser_conn *iser_conn = iser_task->iser_conn;
1019         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1020         struct fast_reg_descriptor *desc = reg->mem_h;
1021
1022         if (!reg->is_mr)
1023                 return;
1024
1025         reg->mem_h = NULL;
1026         reg->is_mr = 0;
1027         spin_lock_bh(&ib_conn->lock);
1028         list_add_tail(&desc->list, &ib_conn->fastreg.pool);
1029         spin_unlock_bh(&ib_conn->lock);
1030 }
1031
1032 int iser_post_recvl(struct iser_conn *iser_conn)
1033 {
1034         struct ib_recv_wr rx_wr, *rx_wr_failed;
1035         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1036         struct ib_sge     sge;
1037         int ib_ret;
1038
1039         sge.addr   = iser_conn->login_resp_dma;
1040         sge.length = ISER_RX_LOGIN_SIZE;
1041         sge.lkey   = ib_conn->device->mr->lkey;
1042
1043         rx_wr.wr_id   = (unsigned long)iser_conn->login_resp_buf;
1044         rx_wr.sg_list = &sge;
1045         rx_wr.num_sge = 1;
1046         rx_wr.next    = NULL;
1047
1048         ib_conn->post_recv_buf_count++;
1049         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1050         if (ib_ret) {
1051                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1052                 ib_conn->post_recv_buf_count--;
1053         }
1054         return ib_ret;
1055 }
1056
1057 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1058 {
1059         struct ib_recv_wr *rx_wr, *rx_wr_failed;
1060         int i, ib_ret;
1061         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1062         unsigned int my_rx_head = iser_conn->rx_desc_head;
1063         struct iser_rx_desc *rx_desc;
1064
1065         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1066                 rx_desc         = &iser_conn->rx_descs[my_rx_head];
1067                 rx_wr->wr_id    = (unsigned long)rx_desc;
1068                 rx_wr->sg_list  = &rx_desc->rx_sg;
1069                 rx_wr->num_sge  = 1;
1070                 rx_wr->next     = rx_wr + 1;
1071                 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1072         }
1073
1074         rx_wr--;
1075         rx_wr->next = NULL; /* mark end of work requests list */
1076
1077         ib_conn->post_recv_buf_count += count;
1078         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1079         if (ib_ret) {
1080                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1081                 ib_conn->post_recv_buf_count -= count;
1082         } else
1083                 iser_conn->rx_desc_head = my_rx_head;
1084         return ib_ret;
1085 }
1086
1087
1088 /**
1089  * iser_start_send - Initiate a Send DTO operation
1090  *
1091  * returns 0 on success, -1 on failure
1092  */
1093 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc)
1094 {
1095         int               ib_ret;
1096         struct ib_send_wr send_wr, *send_wr_failed;
1097
1098         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1099                                       tx_desc->dma_addr, ISER_HEADERS_LEN,
1100                                       DMA_TO_DEVICE);
1101
1102         send_wr.next       = NULL;
1103         send_wr.wr_id      = (unsigned long)tx_desc;
1104         send_wr.sg_list    = tx_desc->tx_sg;
1105         send_wr.num_sge    = tx_desc->num_sge;
1106         send_wr.opcode     = IB_WR_SEND;
1107         send_wr.send_flags = IB_SEND_SIGNALED;
1108
1109         atomic_inc(&ib_conn->post_send_buf_count);
1110
1111         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1112         if (ib_ret) {
1113                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1114                 atomic_dec(&ib_conn->post_send_buf_count);
1115         }
1116         return ib_ret;
1117 }
1118
1119 /**
1120  * is_iser_tx_desc - Indicate if the completion wr_id
1121  *     is a TX descriptor or not.
1122  * @iser_conn: iser connection
1123  * @wr_id: completion WR identifier
1124  *
1125  * Since we cannot rely on wc opcode in FLUSH errors
1126  * we must work around it by checking if the wr_id address
1127  * falls in the iser connection rx_descs buffer. If so
1128  * it is an RX descriptor, otherwize it is a TX.
1129  */
1130 static inline bool
1131 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1132 {
1133         void *start = iser_conn->rx_descs;
1134         int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1135
1136         if (wr_id >= start && wr_id < start + len)
1137                 return false;
1138
1139         return true;
1140 }
1141
1142 /**
1143  * iser_handle_comp_error() - Handle error completion
1144  * @ib_conn:   connection RDMA resources
1145  * @wc:        work completion
1146  *
1147  * Notes: We may handle a FLUSH error completion and in this case
1148  *        we only cleanup in case TX type was DATAOUT. For non-FLUSH
1149  *        error completion we should also notify iscsi layer that
1150  *        connection is failed (in case we passed bind stage).
1151  */
1152 static void
1153 iser_handle_comp_error(struct ib_conn *ib_conn,
1154                        struct ib_wc *wc)
1155 {
1156         struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1157                                                    ib_conn);
1158
1159         if (wc->status != IB_WC_WR_FLUSH_ERR)
1160                 if (iser_conn->iscsi_conn)
1161                         iscsi_conn_failure(iser_conn->iscsi_conn,
1162                                            ISCSI_ERR_CONN_FAILED);
1163
1164         if (is_iser_tx_desc(iser_conn, (void *)wc->wr_id)) {
1165                 struct iser_tx_desc *desc = (struct iser_tx_desc *)wc->wr_id;
1166
1167                 atomic_dec(&ib_conn->post_send_buf_count);
1168                 if (desc->type == ISCSI_TX_DATAOUT)
1169                         kmem_cache_free(ig.desc_cache, desc);
1170         } else {
1171                 ib_conn->post_recv_buf_count--;
1172         }
1173 }
1174
1175 /**
1176  * iser_handle_wc - handle a single work completion
1177  * @wc: work completion
1178  *
1179  * Soft-IRQ context, work completion can be either
1180  * SEND or RECV, and can turn out successful or
1181  * with error (or flush error).
1182  */
1183 static void iser_handle_wc(struct ib_wc *wc)
1184 {
1185         struct ib_conn *ib_conn;
1186         struct iser_tx_desc *tx_desc;
1187         struct iser_rx_desc *rx_desc;
1188
1189         ib_conn = wc->qp->qp_context;
1190         if (wc->status == IB_WC_SUCCESS) {
1191                 if (wc->opcode == IB_WC_RECV) {
1192                         rx_desc = (struct iser_rx_desc *)wc->wr_id;
1193                         iser_rcv_completion(rx_desc, wc->byte_len,
1194                                             ib_conn);
1195                 } else
1196                 if (wc->opcode == IB_WC_SEND) {
1197                         tx_desc = (struct iser_tx_desc *)wc->wr_id;
1198                         iser_snd_completion(tx_desc, ib_conn);
1199                         atomic_dec(&ib_conn->post_send_buf_count);
1200                 } else {
1201                         iser_err("Unknown wc opcode %d\n", wc->opcode);
1202                 }
1203         } else {
1204                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1205                         iser_err("wr id %llx status %d vend_err %x\n",
1206                                  wc->wr_id, wc->status, wc->vendor_err);
1207                 else
1208                         iser_dbg("flush error: wr id %llx\n", wc->wr_id);
1209
1210                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
1211                         iser_handle_comp_error(ib_conn, wc);
1212
1213                 /* complete in case all flush errors were consumed */
1214                 if (ib_conn->post_recv_buf_count == 0 &&
1215                     atomic_read(&ib_conn->post_send_buf_count) == 0)
1216                         complete(&ib_conn->flush_comp);
1217         }
1218 }
1219
1220 /**
1221  * iser_cq_tasklet_fn - iSER completion polling loop
1222  * @data: iSER completion context
1223  *
1224  * Soft-IRQ context, polling connection CQ until
1225  * either CQ was empty or we exausted polling budget
1226  */
1227 static void iser_cq_tasklet_fn(unsigned long data)
1228 {
1229         struct iser_comp *comp = (struct iser_comp *)data;
1230         struct ib_cq *cq = comp->cq;
1231         struct ib_wc wc;
1232         int completed = 0;
1233
1234         while (ib_poll_cq(cq, 1, &wc) == 1) {
1235                 iser_handle_wc(&wc);
1236
1237                 if (++completed >= iser_cq_poll_limit)
1238                         break;
1239         }
1240
1241         /*
1242          * It is assumed here that arming CQ only once its empty
1243          * would not cause interrupts to be missed.
1244          */
1245         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1246
1247         iser_dbg("got %d completions\n", completed);
1248 }
1249
1250 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1251 {
1252         struct iser_comp *comp = cq_context;
1253
1254         tasklet_schedule(&comp->tasklet);
1255 }
1256
1257 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1258                              enum iser_data_dir cmd_dir, sector_t *sector)
1259 {
1260         struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1261         struct fast_reg_descriptor *desc = reg->mem_h;
1262         unsigned long sector_size = iser_task->sc->device->sector_size;
1263         struct ib_mr_status mr_status;
1264         int ret;
1265
1266         if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1267                 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1268                 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1269                                          IB_MR_CHECK_SIG_STATUS, &mr_status);
1270                 if (ret) {
1271                         pr_err("ib_check_mr_status failed, ret %d\n", ret);
1272                         goto err;
1273                 }
1274
1275                 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1276                         sector_t sector_off = mr_status.sig_err.sig_err_offset;
1277
1278                         do_div(sector_off, sector_size + 8);
1279                         *sector = scsi_get_lba(iser_task->sc) + sector_off;
1280
1281                         pr_err("PI error found type %d at sector %llx "
1282                                "expected %x vs actual %x\n",
1283                                mr_status.sig_err.err_type,
1284                                (unsigned long long)*sector,
1285                                mr_status.sig_err.expected,
1286                                mr_status.sig_err.actual);
1287
1288                         switch (mr_status.sig_err.err_type) {
1289                         case IB_SIG_BAD_GUARD:
1290                                 return 0x1;
1291                         case IB_SIG_BAD_REFTAG:
1292                                 return 0x3;
1293                         case IB_SIG_BAD_APPTAG:
1294                                 return 0x2;
1295                         }
1296                 }
1297         }
1298
1299         return 0;
1300 err:
1301         /* Not alot we can do here, return ambiguous guard error */
1302         return 0x1;
1303 }