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