kernel/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         if (!iser_always_reg) {
 137                 int access = IB_ACCESS_LOCAL_WRITE |
 138                              IB_ACCESS_REMOTE_WRITE |
 139                              IB_ACCESS_REMOTE_READ;
 140
 141                 device->mr = ib_get_dma_mr(device->pd, access);
 142                 if (IS_ERR(device->mr))
 143                         goto dma_mr_err;
 144         }
 145
 146         INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
 147                                 iser_event_handler);
 148         if (ib_register_event_handler(&device->event_handler))
 149                 goto handler_err;
 150
 151         return 0;
 152
 153 handler_err:
 154         if (device->mr)
 155                 ib_dereg_mr(device->mr);
 156 dma_mr_err:
 157         for (i = 0; i < device->comps_used; i++)
 158                 tasklet_kill(&device->comps[i].tasklet);
 159 cq_err:
 160         for (i = 0; i < device->comps_used; i++) {
 161                 struct iser_comp *comp = &device->comps[i];
 162
 163                 if (comp->cq)
 164                         ib_destroy_cq(comp->cq);
 165         }
 166         ib_dealloc_pd(device->pd);
 167 pd_err:
 168         kfree(device->comps);
 169 comps_err:
 170         iser_err("failed to allocate an IB resource\n");
 171         return -1;
 172 }
 173
 174 /**
 175  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 176  * CQ and PD created with the device associated with the adapator.
 177  */
 178 static void iser_free_device_ib_res(struct iser_device *device)
 179 {
 180         int i;
 181
 182         for (i = 0; i < device->comps_used; i++) {
 183                 struct iser_comp *comp = &device->comps[i];
 184
 185                 tasklet_kill(&comp->tasklet);
 186                 ib_destroy_cq(comp->cq);
 187                 comp->cq = NULL;
 188         }
 189
 190         (void)ib_unregister_event_handler(&device->event_handler);
 191         if (device->mr)
 192                 (void)ib_dereg_mr(device->mr);
 193         ib_dealloc_pd(device->pd);
 194
 195         kfree(device->comps);
 196         device->comps = NULL;
 197
 198         device->mr = NULL;
 199         device->pd = NULL;
 200 }
 201
 202 /**
 203  * iser_alloc_fmr_pool - Creates FMR pool and page_vector
 204  *
 205  * returns 0 on success, or errno code on failure
 206  */
 207 int iser_alloc_fmr_pool(struct ib_conn *ib_conn,
 208                         unsigned cmds_max,
 209                         unsigned int size)
 210 {
 211         struct iser_device *device = ib_conn->device;
 212         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 213         struct iser_page_vec *page_vec;
 214         struct iser_fr_desc *desc;
 215         struct ib_fmr_pool *fmr_pool;
 216         struct ib_fmr_pool_param params;
 217         int ret;
 218
 219         INIT_LIST_HEAD(&fr_pool->list);
 220         spin_lock_init(&fr_pool->lock);
 221
 222         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 223         if (!desc)
 224                 return -ENOMEM;
 225
 226         page_vec = kmalloc(sizeof(*page_vec) + (sizeof(u64) * size),
 227                            GFP_KERNEL);
 228         if (!page_vec) {
 229                 ret = -ENOMEM;
 230                 goto err_frpl;
 231         }
 232
 233         page_vec->pages = (u64 *)(page_vec + 1);
 234
 235         params.page_shift        = SHIFT_4K;
 236         params.max_pages_per_fmr = size;
 237         /* make the pool size twice the max number of SCSI commands *
 238          * the ML is expected to queue, watermark for unmap at 50%  */
 239         params.pool_size         = cmds_max * 2;
 240         params.dirty_watermark   = cmds_max;
 241         params.cache             = 0;
 242         params.flush_function    = NULL;
 243         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 244                                     IB_ACCESS_REMOTE_WRITE |
 245                                     IB_ACCESS_REMOTE_READ);
 246
 247         fmr_pool = ib_create_fmr_pool(device->pd, &params);
 248         if (IS_ERR(fmr_pool)) {
 249                 ret = PTR_ERR(fmr_pool);
 250                 iser_err("FMR allocation failed, err %d\n", ret);
 251                 goto err_fmr;
 252         }
 253
 254         desc->rsc.page_vec = page_vec;
 255         desc->rsc.fmr_pool = fmr_pool;
 256         list_add(&desc->list, &fr_pool->list);
 257
 258         return 0;
 259
 260 err_fmr:
 261         kfree(page_vec);
 262 err_frpl:
 263         kfree(desc);
 264
 265         return ret;
 266 }
 267
 268 /**
 269  * iser_free_fmr_pool - releases the FMR pool and page vec
 270  */
 271 void iser_free_fmr_pool(struct ib_conn *ib_conn)
 272 {
 273         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 274         struct iser_fr_desc *desc;
 275
 276         desc = list_first_entry(&fr_pool->list,
 277                                 struct iser_fr_desc, list);
 278         list_del(&desc->list);
 279
 280         iser_info("freeing conn %p fmr pool %p\n",
 281                   ib_conn, desc->rsc.fmr_pool);
 282
 283         ib_destroy_fmr_pool(desc->rsc.fmr_pool);
 284         kfree(desc->rsc.page_vec);
 285         kfree(desc);
 286 }
 287
 288 static int
 289 iser_alloc_reg_res(struct ib_device *ib_device,
 290                    struct ib_pd *pd,
 291                    struct iser_reg_resources *res,
 292                    unsigned int size)
 293 {
 294         int ret;
 295
 296         res->mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, size);
 297         if (IS_ERR(res->mr)) {
 298                 ret = PTR_ERR(res->mr);
 299                 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
 300                 return ret;
 301         }
 302         res->mr_valid = 1;
 303
 304         return 0;
 305 }
 306
 307 static void
 308 iser_free_reg_res(struct iser_reg_resources *rsc)
 309 {
 310         ib_dereg_mr(rsc->mr);
 311 }
 312
 313 static int
 314 iser_alloc_pi_ctx(struct ib_device *ib_device,
 315                   struct ib_pd *pd,
 316                   struct iser_fr_desc *desc,
 317                   unsigned int size)
 318 {
 319         struct iser_pi_context *pi_ctx = NULL;
 320         int ret;
 321
 322         desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 323         if (!desc->pi_ctx)
 324                 return -ENOMEM;
 325
 326         pi_ctx = desc->pi_ctx;
 327
 328         ret = iser_alloc_reg_res(ib_device, pd, &pi_ctx->rsc, size);
 329         if (ret) {
 330                 iser_err("failed to allocate reg_resources\n");
 331                 goto alloc_reg_res_err;
 332         }
 333
 334         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 335         if (IS_ERR(pi_ctx->sig_mr)) {
 336                 ret = PTR_ERR(pi_ctx->sig_mr);
 337                 goto sig_mr_failure;
 338         }
 339         pi_ctx->sig_mr_valid = 1;
 340         desc->pi_ctx->sig_protected = 0;
 341
 342         return 0;
 343
 344 sig_mr_failure:
 345         iser_free_reg_res(&pi_ctx->rsc);
 346 alloc_reg_res_err:
 347         kfree(desc->pi_ctx);
 348
 349         return ret;
 350 }
 351
 352 static void
 353 iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
 354 {
 355         iser_free_reg_res(&pi_ctx->rsc);
 356         ib_dereg_mr(pi_ctx->sig_mr);
 357         kfree(pi_ctx);
 358 }
 359
 360 static struct iser_fr_desc *
 361 iser_create_fastreg_desc(struct ib_device *ib_device,
 362                          struct ib_pd *pd,
 363                          bool pi_enable,
 364                          unsigned int size)
 365 {
 366         struct iser_fr_desc *desc;
 367         int ret;
 368
 369         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 370         if (!desc)
 371                 return ERR_PTR(-ENOMEM);
 372
 373         ret = iser_alloc_reg_res(ib_device, pd, &desc->rsc, size);
 374         if (ret)
 375                 goto reg_res_alloc_failure;
 376
 377         if (pi_enable) {
 378                 ret = iser_alloc_pi_ctx(ib_device, pd, desc, size);
 379                 if (ret)
 380                         goto pi_ctx_alloc_failure;
 381         }
 382
 383         return desc;
 384
 385 pi_ctx_alloc_failure:
 386         iser_free_reg_res(&desc->rsc);
 387 reg_res_alloc_failure:
 388         kfree(desc);
 389
 390         return ERR_PTR(ret);
 391 }
 392
 393 /**
 394  * iser_alloc_fastreg_pool - Creates pool of fast_reg descriptors
 395  * for fast registration work requests.
 396  * returns 0 on success, or errno code on failure
 397  */
 398 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
 399                             unsigned cmds_max,
 400                             unsigned int size)
 401 {
 402         struct iser_device *device = ib_conn->device;
 403         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 404         struct iser_fr_desc *desc;
 405         int i, ret;
 406
 407         INIT_LIST_HEAD(&fr_pool->list);
 408         spin_lock_init(&fr_pool->lock);
 409         fr_pool->size = 0;
 410         for (i = 0; i < cmds_max; i++) {
 411                 desc = iser_create_fastreg_desc(device->ib_device, device->pd,
 412                                                 ib_conn->pi_support, size);
 413                 if (IS_ERR(desc)) {
 414                         ret = PTR_ERR(desc);
 415                         goto err;
 416                 }
 417
 418                 list_add_tail(&desc->list, &fr_pool->list);
 419                 fr_pool->size++;
 420         }
 421
 422         return 0;
 423
 424 err:
 425         iser_free_fastreg_pool(ib_conn);
 426         return ret;
 427 }
 428
 429 /**
 430  * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
 431  */
 432 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
 433 {
 434         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 435         struct iser_fr_desc *desc, *tmp;
 436         int i = 0;
 437
 438         if (list_empty(&fr_pool->list))
 439                 return;
 440
 441         iser_info("freeing conn %p fr pool\n", ib_conn);
 442
 443         list_for_each_entry_safe(desc, tmp, &fr_pool->list, list) {
 444                 list_del(&desc->list);
 445                 iser_free_reg_res(&desc->rsc);
 446                 if (desc->pi_ctx)
 447                         iser_free_pi_ctx(desc->pi_ctx);
 448                 kfree(desc);
 449                 ++i;
 450         }
 451
 452         if (i < fr_pool->size)
 453                 iser_warn("pool still has %d regions registered\n",
 454                           fr_pool->size - i);
 455 }
 456
 457 /**
 458  * iser_create_ib_conn_res - Queue-Pair (QP)
 459  *
 460  * returns 0 on success, -1 on failure
 461  */
 462 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
 463 {
 464         struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
 465                                                    ib_conn);
 466         struct iser_device      *device;
 467         struct ib_device_attr *dev_attr;
 468         struct ib_qp_init_attr  init_attr;
 469         int                     ret = -ENOMEM;
 470         int index, min_index = 0;
 471
 472         BUG_ON(ib_conn->device == NULL);
 473
 474         device = ib_conn->device;
 475         dev_attr = &device->dev_attr;
 476
 477         memset(&init_attr, 0, sizeof init_attr);
 478
 479         mutex_lock(&ig.connlist_mutex);
 480         /* select the CQ with the minimal number of usages */
 481         for (index = 0; index < device->comps_used; index++) {
 482                 if (device->comps[index].active_qps <
 483                     device->comps[min_index].active_qps)
 484                         min_index = index;
 485         }
 486         ib_conn->comp = &device->comps[min_index];
 487         ib_conn->comp->active_qps++;
 488         mutex_unlock(&ig.connlist_mutex);
 489         iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 490
 491         init_attr.event_handler = iser_qp_event_callback;
 492         init_attr.qp_context    = (void *)ib_conn;
 493         init_attr.send_cq       = ib_conn->comp->cq;
 494         init_attr.recv_cq       = ib_conn->comp->cq;
 495         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 496         init_attr.cap.max_send_sge = 2;
 497         init_attr.cap.max_recv_sge = 1;
 498         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 499         init_attr.qp_type       = IB_QPT_RC;
 500         if (ib_conn->pi_support) {
 501                 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
 502                 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 503                 iser_conn->max_cmds =
 504                         ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
 505         } else {
 506                 if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
 507                         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS + 1;
 508                         iser_conn->max_cmds =
 509                                 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
 510                 } else {
 511                         init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
 512                         iser_conn->max_cmds =
 513                                 ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
 514                         iser_dbg("device %s supports max_send_wr %d\n",
 515                                  device->ib_device->name, dev_attr->max_qp_wr);
 516                 }
 517         }
 518
 519         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 520         if (ret)
 521                 goto out_err;
 522
 523         ib_conn->qp = ib_conn->cma_id->qp;
 524         iser_info("setting conn %p cma_id %p qp %p\n",
 525                   ib_conn, ib_conn->cma_id,
 526                   ib_conn->cma_id->qp);
 527         return ret;
 528
 529 out_err:
 530         mutex_lock(&ig.connlist_mutex);
 531         ib_conn->comp->active_qps--;
 532         mutex_unlock(&ig.connlist_mutex);
 533         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 534
 535         return ret;
 536 }
 537
 538 /**
 539  * based on the resolved device node GUID see if there already allocated
 540  * device for this device. If there's no such, create one.
 541  */
 542 static
 543 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 544 {
 545         struct iser_device *device;
 546
 547         mutex_lock(&ig.device_list_mutex);
 548
 549         list_for_each_entry(device, &ig.device_list, ig_list)
 550                 /* find if there's a match using the node GUID */
 551                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 552                         goto inc_refcnt;
 553
 554         device = kzalloc(sizeof *device, GFP_KERNEL);
 555         if (device == NULL)
 556                 goto out;
 557
 558         /* assign this device to the device */
 559         device->ib_device = cma_id->device;
 560         /* init the device and link it into ig device list */
 561         if (iser_create_device_ib_res(device)) {
 562                 kfree(device);
 563                 device = NULL;
 564                 goto out;
 565         }
 566         list_add(&device->ig_list, &ig.device_list);
 567
 568 inc_refcnt:
 569         device->refcount++;
 570 out:
 571         mutex_unlock(&ig.device_list_mutex);
 572         return device;
 573 }
 574
 575 /* if there's no demand for this device, release it */
 576 static void iser_device_try_release(struct iser_device *device)
 577 {
 578         mutex_lock(&ig.device_list_mutex);
 579         device->refcount--;
 580         iser_info("device %p refcount %d\n", device, device->refcount);
 581         if (!device->refcount) {
 582                 iser_free_device_ib_res(device);
 583                 list_del(&device->ig_list);
 584                 kfree(device);
 585         }
 586         mutex_unlock(&ig.device_list_mutex);
 587 }
 588
 589 /**
 590  * Called with state mutex held
 591  **/
 592 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
 593                                      enum iser_conn_state comp,
 594                                      enum iser_conn_state exch)
 595 {
 596         int ret;
 597
 598         ret = (iser_conn->state == comp);
 599         if (ret)
 600                 iser_conn->state = exch;
 601
 602         return ret;
 603 }
 604
 605 void iser_release_work(struct work_struct *work)
 606 {
 607         struct iser_conn *iser_conn;
 608
 609         iser_conn = container_of(work, struct iser_conn, release_work);
 610
 611         /* Wait for conn_stop to complete */
 612         wait_for_completion(&iser_conn->stop_completion);
 613         /* Wait for IB resouces cleanup to complete */
 614         wait_for_completion(&iser_conn->ib_completion);
 615
 616         mutex_lock(&iser_conn->state_mutex);
 617         iser_conn->state = ISER_CONN_DOWN;
 618         mutex_unlock(&iser_conn->state_mutex);
 619
 620         iser_conn_release(iser_conn);
 621 }
 622
 623 /**
 624  * iser_free_ib_conn_res - release IB related resources
 625  * @iser_conn: iser connection struct
 626  * @destroy: indicator if we need to try to release the
 627  *     iser device and memory regoins pool (only iscsi
 628  *     shutdown and DEVICE_REMOVAL will use this).
 629  *
 630  * This routine is called with the iser state mutex held
 631  * so the cm_id removal is out of here. It is Safe to
 632  * be invoked multiple times.
 633  */
 634 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
 635                                   bool destroy)
 636 {
 637         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 638         struct iser_device *device = ib_conn->device;
 639
 640         iser_info("freeing conn %p cma_id %p qp %p\n",
 641                   iser_conn, ib_conn->cma_id, ib_conn->qp);
 642
 643         if (ib_conn->qp != NULL) {
 644                 ib_conn->comp->active_qps--;
 645                 rdma_destroy_qp(ib_conn->cma_id);
 646                 ib_conn->qp = NULL;
 647         }
 648
 649         if (destroy) {
 650                 if (iser_conn->rx_descs)
 651                         iser_free_rx_descriptors(iser_conn);
 652
 653                 if (device != NULL) {
 654                         iser_device_try_release(device);
 655                         ib_conn->device = NULL;
 656                 }
 657         }
 658 }
 659
 660 /**
 661  * Frees all conn objects and deallocs conn descriptor
 662  */
 663 void iser_conn_release(struct iser_conn *iser_conn)
 664 {
 665         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 666
 667         mutex_lock(&ig.connlist_mutex);
 668         list_del(&iser_conn->conn_list);
 669         mutex_unlock(&ig.connlist_mutex);
 670
 671         mutex_lock(&iser_conn->state_mutex);
 672         /* In case we endup here without ep_disconnect being invoked. */
 673         if (iser_conn->state != ISER_CONN_DOWN) {
 674                 iser_warn("iser conn %p state %d, expected state down.\n",
 675                           iser_conn, iser_conn->state);
 676                 iscsi_destroy_endpoint(iser_conn->ep);
 677                 iser_conn->state = ISER_CONN_DOWN;
 678         }
 679         /*
 680          * In case we never got to bind stage, we still need to
 681          * release IB resources (which is safe to call more than once).
 682          */
 683         iser_free_ib_conn_res(iser_conn, true);
 684         mutex_unlock(&iser_conn->state_mutex);
 685
 686         if (ib_conn->cma_id != NULL) {
 687                 rdma_destroy_id(ib_conn->cma_id);
 688                 ib_conn->cma_id = NULL;
 689         }
 690
 691         kfree(iser_conn);
 692 }
 693
 694 /**
 695  * triggers start of the disconnect procedures and wait for them to be done
 696  * Called with state mutex held
 697  */
 698 int iser_conn_terminate(struct iser_conn *iser_conn)
 699 {
 700         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 701         struct ib_send_wr *bad_wr;
 702         int err = 0;
 703
 704         /* terminate the iser conn only if the conn state is UP */
 705         if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
 706                                        ISER_CONN_TERMINATING))
 707                 return 0;
 708
 709         iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
 710
 711         /* suspend queuing of new iscsi commands */
 712         if (iser_conn->iscsi_conn)
 713                 iscsi_suspend_queue(iser_conn->iscsi_conn);
 714
 715         /*
 716          * In case we didn't already clean up the cma_id (peer initiated
 717          * a disconnection), we need to Cause the CMA to change the QP
 718          * state to ERROR.
 719          */
 720         if (ib_conn->cma_id) {
 721                 err = rdma_disconnect(ib_conn->cma_id);
 722                 if (err)
 723                         iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 724                                  iser_conn, err);
 725
 726                 /* post an indication that all flush errors were consumed */
 727                 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
 728                 if (err) {
 729                         iser_err("conn %p failed to post beacon", ib_conn);
 730                         return 1;
 731                 }
 732
 733                 wait_for_completion(&ib_conn->flush_comp);
 734         }
 735
 736         return 1;
 737 }
 738
 739 /**
 740  * Called with state mutex held
 741  **/
 742 static void iser_connect_error(struct rdma_cm_id *cma_id)
 743 {
 744         struct iser_conn *iser_conn;
 745
 746         iser_conn = (struct iser_conn *)cma_id->context;
 747         iser_conn->state = ISER_CONN_TERMINATING;
 748 }
 749
 750 static void
 751 iser_calc_scsi_params(struct iser_conn *iser_conn,
 752                       unsigned int max_sectors)
 753 {
 754         struct iser_device *device = iser_conn->ib_conn.device;
 755         unsigned short sg_tablesize, sup_sg_tablesize;
 756
 757         sg_tablesize = DIV_ROUND_UP(max_sectors * 512, SIZE_4K);
 758         sup_sg_tablesize = min_t(unsigned, ISCSI_ISER_MAX_SG_TABLESIZE,
 759                                  device->dev_attr.max_fast_reg_page_list_len);
 760
 761         if (sg_tablesize > sup_sg_tablesize) {
 762                 sg_tablesize = sup_sg_tablesize;
 763                 iser_conn->scsi_max_sectors = sg_tablesize * SIZE_4K / 512;
 764         } else {
 765                 iser_conn->scsi_max_sectors = max_sectors;
 766         }
 767
 768         iser_conn->scsi_sg_tablesize = sg_tablesize;
 769
 770         iser_dbg("iser_conn %p, sg_tablesize %u, max_sectors %u\n",
 771                  iser_conn, iser_conn->scsi_sg_tablesize,
 772                  iser_conn->scsi_max_sectors);
 773 }
 774
 775 /**
 776  * Called with state mutex held
 777  **/
 778 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 779 {
 780         struct iser_device *device;
 781         struct iser_conn   *iser_conn;
 782         struct ib_conn   *ib_conn;
 783         int    ret;
 784
 785         iser_conn = (struct iser_conn *)cma_id->context;
 786         if (iser_conn->state != ISER_CONN_PENDING)
 787                 /* bailout */
 788                 return;
 789
 790         ib_conn = &iser_conn->ib_conn;
 791         device = iser_device_find_by_ib_device(cma_id);
 792         if (!device) {
 793                 iser_err("device lookup/creation failed\n");
 794                 iser_connect_error(cma_id);
 795                 return;
 796         }
 797
 798         ib_conn->device = device;
 799
 800         /* connection T10-PI support */
 801         if (iser_pi_enable) {
 802                 if (!(device->dev_attr.device_cap_flags &
 803                       IB_DEVICE_SIGNATURE_HANDOVER)) {
 804                         iser_warn("T10-PI requested but not supported on %s, "
 805                                   "continue without T10-PI\n",
 806                                   ib_conn->device->ib_device->name);
 807                         ib_conn->pi_support = false;
 808                 } else {
 809                         ib_conn->pi_support = true;
 810                 }
 811         }
 812
 813         iser_calc_scsi_params(iser_conn, iser_max_sectors);
 814
 815         ret = rdma_resolve_route(cma_id, 1000);
 816         if (ret) {
 817                 iser_err("resolve route failed: %d\n", ret);
 818                 iser_connect_error(cma_id);
 819                 return;
 820         }
 821 }
 822
 823 /**
 824  * Called with state mutex held
 825  **/
 826 static void iser_route_handler(struct rdma_cm_id *cma_id)
 827 {
 828         struct rdma_conn_param conn_param;
 829         int    ret;
 830         struct iser_cm_hdr req_hdr;
 831         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 832         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 833         struct iser_device *device = ib_conn->device;
 834
 835         if (iser_conn->state != ISER_CONN_PENDING)
 836                 /* bailout */
 837                 return;
 838
 839         ret = iser_create_ib_conn_res(ib_conn);
 840         if (ret)
 841                 goto failure;
 842
 843         memset(&conn_param, 0, sizeof conn_param);
 844         conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
 845         conn_param.initiator_depth     = 1;
 846         conn_param.retry_count         = 7;
 847         conn_param.rnr_retry_count     = 6;
 848
 849         memset(&req_hdr, 0, sizeof(req_hdr));
 850         req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
 851                         ISER_SEND_W_INV_NOT_SUPPORTED);
 852         conn_param.private_data         = (void *)&req_hdr;
 853         conn_param.private_data_len     = sizeof(struct iser_cm_hdr);
 854
 855         ret = rdma_connect(cma_id, &conn_param);
 856         if (ret) {
 857                 iser_err("failure connecting: %d\n", ret);
 858                 goto failure;
 859         }
 860
 861         return;
 862 failure:
 863         iser_connect_error(cma_id);
 864 }
 865
 866 static void iser_connected_handler(struct rdma_cm_id *cma_id)
 867 {
 868         struct iser_conn *iser_conn;
 869         struct ib_qp_attr attr;
 870         struct ib_qp_init_attr init_attr;
 871
 872         iser_conn = (struct iser_conn *)cma_id->context;
 873         if (iser_conn->state != ISER_CONN_PENDING)
 874                 /* bailout */
 875                 return;
 876
 877         (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
 878         iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
 879
 880         iser_conn->state = ISER_CONN_UP;
 881         complete(&iser_conn->up_completion);
 882 }
 883
 884 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 885 {
 886         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 887
 888         if (iser_conn_terminate(iser_conn)) {
 889                 if (iser_conn->iscsi_conn)
 890                         iscsi_conn_failure(iser_conn->iscsi_conn,
 891                                            ISCSI_ERR_CONN_FAILED);
 892                 else
 893                         iser_err("iscsi_iser connection isn't bound\n");
 894         }
 895 }
 896
 897 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
 898                                  bool destroy)
 899 {
 900         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 901
 902         /*
 903          * We are not guaranteed that we visited disconnected_handler
 904          * by now, call it here to be safe that we handle CM drep
 905          * and flush errors.
 906          */
 907         iser_disconnected_handler(cma_id);
 908         iser_free_ib_conn_res(iser_conn, destroy);
 909         complete(&iser_conn->ib_completion);
 910 };
 911
 912 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 913 {
 914         struct iser_conn *iser_conn;
 915         int ret = 0;
 916
 917         iser_conn = (struct iser_conn *)cma_id->context;
 918         iser_info("%s (%d): status %d conn %p id %p\n",
 919                   rdma_event_msg(event->event), event->event,
 920                   event->status, cma_id->context, cma_id);
 921
 922         mutex_lock(&iser_conn->state_mutex);
 923         switch (event->event) {
 924         case RDMA_CM_EVENT_ADDR_RESOLVED:
 925                 iser_addr_handler(cma_id);
 926                 break;
 927         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 928                 iser_route_handler(cma_id);
 929                 break;
 930         case RDMA_CM_EVENT_ESTABLISHED:
 931                 iser_connected_handler(cma_id);
 932                 break;
 933         case RDMA_CM_EVENT_ADDR_ERROR:
 934         case RDMA_CM_EVENT_ROUTE_ERROR:
 935         case RDMA_CM_EVENT_CONNECT_ERROR:
 936         case RDMA_CM_EVENT_UNREACHABLE:
 937         case RDMA_CM_EVENT_REJECTED:
 938                 iser_connect_error(cma_id);
 939                 break;
 940         case RDMA_CM_EVENT_DISCONNECTED:
 941         case RDMA_CM_EVENT_ADDR_CHANGE:
 942         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 943                 iser_cleanup_handler(cma_id, false);
 944                 break;
 945         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 946                 /*
 947                  * we *must* destroy the device as we cannot rely
 948                  * on iscsid to be around to initiate error handling.
 949                  * also if we are not in state DOWN implicitly destroy
 950                  * the cma_id.
 951                  */
 952                 iser_cleanup_handler(cma_id, true);
 953                 if (iser_conn->state != ISER_CONN_DOWN) {
 954                         iser_conn->ib_conn.cma_id = NULL;
 955                         ret = 1;
 956                 }
 957                 break;
 958         default:
 959                 iser_err("Unexpected RDMA CM event: %s (%d)\n",
 960                          rdma_event_msg(event->event), event->event);
 961                 break;
 962         }
 963         mutex_unlock(&iser_conn->state_mutex);
 964
 965         return ret;
 966 }
 967
 968 void iser_conn_init(struct iser_conn *iser_conn)
 969 {
 970         iser_conn->state = ISER_CONN_INIT;
 971         iser_conn->ib_conn.post_recv_buf_count = 0;
 972         init_completion(&iser_conn->ib_conn.flush_comp);
 973         init_completion(&iser_conn->stop_completion);
 974         init_completion(&iser_conn->ib_completion);
 975         init_completion(&iser_conn->up_completion);
 976         INIT_LIST_HEAD(&iser_conn->conn_list);
 977         mutex_init(&iser_conn->state_mutex);
 978 }
 979
 980  /**
 981  * starts the process of connecting to the target
 982  * sleeps until the connection is established or rejected
 983  */
 984 int iser_connect(struct iser_conn   *iser_conn,
 985                  struct sockaddr    *src_addr,
 986                  struct sockaddr    *dst_addr,
 987                  int                 non_blocking)
 988 {
 989         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 990         int err = 0;
 991
 992         mutex_lock(&iser_conn->state_mutex);
 993
 994         sprintf(iser_conn->name, "%pISp", dst_addr);
 995
 996         iser_info("connecting to: %s\n", iser_conn->name);
 997
 998         /* the device is known only --after-- address resolution */
 999         ib_conn->device = NULL;
1000
1001         iser_conn->state = ISER_CONN_PENDING;
1002
1003         ib_conn->beacon.wr_id = ISER_BEACON_WRID;
1004         ib_conn->beacon.opcode = IB_WR_SEND;
1005
1006         ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
1007                                          (void *)iser_conn,
1008                                          RDMA_PS_TCP, IB_QPT_RC);
1009         if (IS_ERR(ib_conn->cma_id)) {
1010                 err = PTR_ERR(ib_conn->cma_id);
1011                 iser_err("rdma_create_id failed: %d\n", err);
1012                 goto id_failure;
1013         }
1014
1015         err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
1016         if (err) {
1017                 iser_err("rdma_resolve_addr failed: %d\n", err);
1018                 goto addr_failure;
1019         }
1020
1021         if (!non_blocking) {
1022                 wait_for_completion_interruptible(&iser_conn->up_completion);
1023
1024                 if (iser_conn->state != ISER_CONN_UP) {
1025                         err =  -EIO;
1026                         goto connect_failure;
1027                 }
1028         }
1029         mutex_unlock(&iser_conn->state_mutex);
1030
1031         mutex_lock(&ig.connlist_mutex);
1032         list_add(&iser_conn->conn_list, &ig.connlist);
1033         mutex_unlock(&ig.connlist_mutex);
1034         return 0;
1035
1036 id_failure:
1037         ib_conn->cma_id = NULL;
1038 addr_failure:
1039         iser_conn->state = ISER_CONN_DOWN;
1040 connect_failure:
1041         mutex_unlock(&iser_conn->state_mutex);
1042         iser_conn_release(iser_conn);
1043         return err;
1044 }
1045
1046 int iser_post_recvl(struct iser_conn *iser_conn)
1047 {
1048         struct ib_recv_wr rx_wr, *rx_wr_failed;
1049         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1050         struct ib_sge     sge;
1051         int ib_ret;
1052
1053         sge.addr   = iser_conn->login_resp_dma;
1054         sge.length = ISER_RX_LOGIN_SIZE;
1055         sge.lkey   = ib_conn->device->pd->local_dma_lkey;
1056
1057         rx_wr.wr_id   = (uintptr_t)iser_conn->login_resp_buf;
1058         rx_wr.sg_list = &sge;
1059         rx_wr.num_sge = 1;
1060         rx_wr.next    = NULL;
1061
1062         ib_conn->post_recv_buf_count++;
1063         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1064         if (ib_ret) {
1065                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1066                 ib_conn->post_recv_buf_count--;
1067         }
1068         return ib_ret;
1069 }
1070
1071 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1072 {
1073         struct ib_recv_wr *rx_wr, *rx_wr_failed;
1074         int i, ib_ret;
1075         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1076         unsigned int my_rx_head = iser_conn->rx_desc_head;
1077         struct iser_rx_desc *rx_desc;
1078
1079         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1080                 rx_desc         = &iser_conn->rx_descs[my_rx_head];
1081                 rx_wr->wr_id    = (uintptr_t)rx_desc;
1082                 rx_wr->sg_list  = &rx_desc->rx_sg;
1083                 rx_wr->num_sge  = 1;
1084                 rx_wr->next     = rx_wr + 1;
1085                 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1086         }
1087
1088         rx_wr--;
1089         rx_wr->next = NULL; /* mark end of work requests list */
1090
1091         ib_conn->post_recv_buf_count += count;
1092         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1093         if (ib_ret) {
1094                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1095                 ib_conn->post_recv_buf_count -= count;
1096         } else
1097                 iser_conn->rx_desc_head = my_rx_head;
1098         return ib_ret;
1099 }
1100
1101
1102 /**
1103  * iser_start_send - Initiate a Send DTO operation
1104  *
1105  * returns 0 on success, -1 on failure
1106  */
1107 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1108                    bool signal)
1109 {
1110         struct ib_send_wr *bad_wr, *wr = iser_tx_next_wr(tx_desc);
1111         int ib_ret;
1112
1113         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1114                                       tx_desc->dma_addr, ISER_HEADERS_LEN,
1115                                       DMA_TO_DEVICE);
1116
1117         wr->next = NULL;
1118         wr->wr_id = (uintptr_t)tx_desc;
1119         wr->sg_list = tx_desc->tx_sg;
1120         wr->num_sge = tx_desc->num_sge;
1121         wr->opcode = IB_WR_SEND;
1122         wr->send_flags = signal ? IB_SEND_SIGNALED : 0;
1123
1124         ib_ret = ib_post_send(ib_conn->qp, &tx_desc->wrs[0].send, &bad_wr);
1125         if (ib_ret)
1126                 iser_err("ib_post_send failed, ret:%d opcode:%d\n",
1127                          ib_ret, bad_wr->opcode);
1128
1129         return ib_ret;
1130 }
1131
1132 /**
1133  * is_iser_tx_desc - Indicate if the completion wr_id
1134  *     is a TX descriptor or not.
1135  * @iser_conn: iser connection
1136  * @wr_id: completion WR identifier
1137  *
1138  * Since we cannot rely on wc opcode in FLUSH errors
1139  * we must work around it by checking if the wr_id address
1140  * falls in the iser connection rx_descs buffer. If so
1141  * it is an RX descriptor, otherwize it is a TX.
1142  */
1143 static inline bool
1144 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1145 {
1146         void *start = iser_conn->rx_descs;
1147         int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1148
1149         if (wr_id >= start && wr_id < start + len)
1150                 return false;
1151
1152         return true;
1153 }
1154
1155 /**
1156  * iser_handle_comp_error() - Handle error completion
1157  * @ib_conn:   connection RDMA resources
1158  * @wc:        work completion
1159  *
1160  * Notes: We may handle a FLUSH error completion and in this case
1161  *        we only cleanup in case TX type was DATAOUT. For non-FLUSH
1162  *        error completion we should also notify iscsi layer that
1163  *        connection is failed (in case we passed bind stage).
1164  */
1165 static void
1166 iser_handle_comp_error(struct ib_conn *ib_conn,
1167                        struct ib_wc *wc)
1168 {
1169         void *wr_id = (void *)(uintptr_t)wc->wr_id;
1170         struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1171                                                    ib_conn);
1172
1173         if (wc->status != IB_WC_WR_FLUSH_ERR)
1174                 if (iser_conn->iscsi_conn)
1175                         iscsi_conn_failure(iser_conn->iscsi_conn,
1176                                            ISCSI_ERR_CONN_FAILED);
1177
1178         if (wc->wr_id == ISER_FASTREG_LI_WRID)
1179                 return;
1180
1181         if (is_iser_tx_desc(iser_conn, wr_id)) {
1182                 struct iser_tx_desc *desc = wr_id;
1183
1184                 if (desc->type == ISCSI_TX_DATAOUT)
1185                         kmem_cache_free(ig.desc_cache, desc);
1186         } else {
1187                 ib_conn->post_recv_buf_count--;
1188         }
1189 }
1190
1191 /**
1192  * iser_handle_wc - handle a single work completion
1193  * @wc: work completion
1194  *
1195  * Soft-IRQ context, work completion can be either
1196  * SEND or RECV, and can turn out successful or
1197  * with error (or flush error).
1198  */
1199 static void iser_handle_wc(struct ib_wc *wc)
1200 {
1201         struct ib_conn *ib_conn;
1202         struct iser_tx_desc *tx_desc;
1203         struct iser_rx_desc *rx_desc;
1204
1205         ib_conn = wc->qp->qp_context;
1206         if (likely(wc->status == IB_WC_SUCCESS)) {
1207                 if (wc->opcode == IB_WC_RECV) {
1208                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1209                         iser_rcv_completion(rx_desc, wc->byte_len,
1210                                             ib_conn);
1211                 } else
1212                 if (wc->opcode == IB_WC_SEND) {
1213                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1214                         iser_snd_completion(tx_desc, ib_conn);
1215                 } else {
1216                         iser_err("Unknown wc opcode %d\n", wc->opcode);
1217                 }
1218         } else {
1219                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1220                         iser_err("%s (%d): wr id %llx vend_err %x\n",
1221                                  ib_wc_status_msg(wc->status), wc->status,
1222                                  wc->wr_id, wc->vendor_err);
1223                 else
1224                         iser_dbg("%s (%d): wr id %llx\n",
1225                                  ib_wc_status_msg(wc->status), wc->status,
1226                                  wc->wr_id);
1227
1228                 if (wc->wr_id == ISER_BEACON_WRID)
1229                         /* all flush errors were consumed */
1230                         complete(&ib_conn->flush_comp);
1231                 else
1232                         iser_handle_comp_error(ib_conn, wc);
1233         }
1234 }
1235
1236 /**
1237  * iser_cq_tasklet_fn - iSER completion polling loop
1238  * @data: iSER completion context
1239  *
1240  * Soft-IRQ context, polling connection CQ until
1241  * either CQ was empty or we exausted polling budget
1242  */
1243 static void iser_cq_tasklet_fn(unsigned long data)
1244 {
1245         struct iser_comp *comp = (struct iser_comp *)data;
1246         struct ib_cq *cq = comp->cq;
1247         struct ib_wc *const wcs = comp->wcs;
1248         int i, n, completed = 0;
1249
1250         while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1251                 for (i = 0; i < n; i++)
1252                         iser_handle_wc(&wcs[i]);
1253
1254                 completed += n;
1255                 if (completed >= iser_cq_poll_limit)
1256                         break;
1257         }
1258
1259         /*
1260          * It is assumed here that arming CQ only once its empty
1261          * would not cause interrupts to be missed.
1262          */
1263         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1264
1265         iser_dbg("got %d completions\n", completed);
1266 }
1267
1268 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1269 {
1270         struct iser_comp *comp = cq_context;
1271
1272         tasklet_schedule(&comp->tasklet);
1273 }
1274
1275 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1276                              enum iser_data_dir cmd_dir, sector_t *sector)
1277 {
1278         struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1279         struct iser_fr_desc *desc = reg->mem_h;
1280         unsigned long sector_size = iser_task->sc->device->sector_size;
1281         struct ib_mr_status mr_status;
1282         int ret;
1283
1284         if (desc && desc->pi_ctx->sig_protected) {
1285                 desc->pi_ctx->sig_protected = 0;
1286                 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1287                                          IB_MR_CHECK_SIG_STATUS, &mr_status);
1288                 if (ret) {
1289                         pr_err("ib_check_mr_status failed, ret %d\n", ret);
1290                         goto err;
1291                 }
1292
1293                 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1294                         sector_t sector_off = mr_status.sig_err.sig_err_offset;
1295
1296                         sector_div(sector_off, sector_size + 8);
1297                         *sector = scsi_get_lba(iser_task->sc) + sector_off;
1298
1299                         pr_err("PI error found type %d at sector %llx "
1300                                "expected %x vs actual %x\n",
1301                                mr_status.sig_err.err_type,
1302                                (unsigned long long)*sector,
1303                                mr_status.sig_err.expected,
1304                                mr_status.sig_err.actual);
1305
1306                         switch (mr_status.sig_err.err_type) {
1307                         case IB_SIG_BAD_GUARD:
1308                                 return 0x1;
1309                         case IB_SIG_BAD_REFTAG:
1310                                 return 0x3;
1311                         case IB_SIG_BAD_APPTAG:
1312                                 return 0x2;
1313                         }
1314                 }
1315         }
1316
1317         return 0;
1318 err:
1319         /* Not alot we can do here, return ambiguous guard error */
1320         return 0x1;
1321 }