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